From 5f64f6f93251b1a57e6640842800ccde56d5423d Mon Sep 17 00:00:00 2001 From: Carousel126 Date: Mon, 9 Feb 2026 11:17:00 +0800 Subject: [PATCH 1/3] feat: use static capture for squeezenet1_1 json --- .../vision/squeezenet1_1/__init__.py | 0 .../vision/squeezenet1_1/graph_hash.txt | 1 + .../vision/squeezenet1_1/input_meta.py | 8 + .../vision/squeezenet1_1/model.py | 13 ++ .../vision/squeezenet1_1/squeezenet1_1.json | 221 ++++++++++++++++++ .../vision/squeezenet1_1/weight_meta.py | 3 + graph_net/tests/test_squeezenet1_1_extract.py | 33 +++ my_extractor.py | 33 +++ my_samples/squeezenet1_1/__init__.py | 0 my_samples/squeezenet1_1/graph_hash.txt | 1 + my_samples/squeezenet1_1/input_meta.py | 8 + my_samples/squeezenet1_1/model.py | 13 ++ my_samples/squeezenet1_1/squeezenet1_1.json | 221 ++++++++++++++++++ my_samples/squeezenet1_1/weight_meta.py | 3 + 14 files changed, 558 insertions(+) create mode 100644 graph_net/samples/paddle_samples/vision/squeezenet1_1/__init__.py create mode 100644 graph_net/samples/paddle_samples/vision/squeezenet1_1/graph_hash.txt create mode 100644 graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py create mode 100644 graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py create mode 100644 graph_net/samples/paddle_samples/vision/squeezenet1_1/squeezenet1_1.json create mode 100644 graph_net/samples/paddle_samples/vision/squeezenet1_1/weight_meta.py create mode 100644 graph_net/tests/test_squeezenet1_1_extract.py create mode 100644 my_extractor.py create mode 100644 my_samples/squeezenet1_1/__init__.py create mode 100644 my_samples/squeezenet1_1/graph_hash.txt create mode 100644 my_samples/squeezenet1_1/input_meta.py create mode 100644 my_samples/squeezenet1_1/model.py create mode 100644 my_samples/squeezenet1_1/squeezenet1_1.json create mode 100644 my_samples/squeezenet1_1/weight_meta.py diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/__init__.py b/graph_net/samples/paddle_samples/vision/squeezenet1_1/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/graph_hash.txt b/graph_net/samples/paddle_samples/vision/squeezenet1_1/graph_hash.txt new file mode 100644 index 000000000..8ebaa39af --- /dev/null +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/graph_hash.txt @@ -0,0 +1 @@ +6de9a1959b0f3ccd2b9e70f1b42a4f295af1bea89e4779c1e0fe5753bc609be7 \ No newline at end of file diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py b/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py new file mode 100644 index 000000000..6453218cf --- /dev/null +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py @@ -0,0 +1,8 @@ +import paddle + +input_meta = { + "inputs": { + "shape": [1, 3, 224, 224], + "dtype": "float32" + } +} diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py b/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py new file mode 100644 index 000000000..a149a8966 --- /dev/null +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py @@ -0,0 +1,13 @@ +import paddle +from paddle.vision.models import squeezenet1_1 + +class GraphModule(paddle.nn.Layer): + def __init__(self): + super(GraphModule, self).__init__() + self.model = squeezenet1_1(pretrained=False) + + def forward(self, inputs=None): + # 防御性修复：如果 inputs 为空，生成一个符合规格的 dummy tensor + if inputs is None: + inputs = paddle.randn([1, 3, 224, 224]) + return self.model(inputs) diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/squeezenet1_1.json b/graph_net/samples/paddle_samples/vision/squeezenet1_1/squeezenet1_1.json new file mode 100644 index 000000000..8cdf0fc0b --- /dev/null +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/squeezenet1_1.json @@ -0,0 +1,221 @@ +{ + (%0) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_25.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<1000xf32> + (%1) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_25.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<1000x512x1x1xf32> + (%2) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_24.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%3) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_24.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x3x3xf32> + (%4) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_23.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%5) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_23.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x1x1xf32> + (%6) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_22.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%7) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_22.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x512x1x1xf32> + (%8) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_21.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%9) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_21.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x3x3xf32> + (%10) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_20.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%11) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_20.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x1x1xf32> + (%12) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_19.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%13) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_19.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x384x1x1xf32> + (%14) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_18.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%15) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_18.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x3x3xf32> + (%16) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_17.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%17) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_17.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x1x1xf32> + (%18) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_16.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48xf32> + (%19) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_16.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48x384x1x1xf32> + (%20) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_15.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%21) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_15.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x3x3xf32> + (%22) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_14.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%23) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_14.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x1x1xf32> + (%24) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_13.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48xf32> + (%25) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_13.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48x256x1x1xf32> + (%26) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_12.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%27) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_12.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x3x3xf32> + (%28) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_11.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%29) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_11.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x1x1xf32> + (%30) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_10.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32xf32> + (%31) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_10.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32x256x1x1xf32> + (%32) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_9.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%33) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_9.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x3x3xf32> + (%34) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_8.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%35) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_8.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x1x1xf32> + (%36) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_7.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32xf32> + (%37) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_7.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32x128x1x1xf32> + (%38) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_6.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%39) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_6.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x3x3xf32> + (%40) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_5.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%41) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_5.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x1x1xf32> + (%42) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_4.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16xf32> + (%43) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_4.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16x128x1x1xf32> + (%44) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_3.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%45) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_3.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x3x3xf32> + (%46) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_2.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%47) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_2.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x1x1xf32> + (%48) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_1.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16xf32> + (%49) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_1.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16x64x1x1xf32> + (%50) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_0.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%51) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_0.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x3x3x3xf32> + (%52) = "pd_op.data" () {dtype:float32,name:"inputs",place:Place(undefined:0),shape:[1,3,224,224],stop_gradient:[true]} : () -> tensor<1x3x224x224xf32> + (%53) = "pd_op.conv2d" (%52, %51) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<1x3x224x224xf32>, tensor<64x3x3x3xf32>) -> tensor<1x64x112x112xf32> + (%54) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%55) = "pd_op.reshape" (%50, %54) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%56) = "pd_op.add" (%53, %55) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32> + (%57) = "pd_op.relu" (%56) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x64x112x112xf32>) -> tensor<1x64x112x112xf32> + (%58) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D/",value:[3,3]} : () -> tensor<2xi64> + (%59) = "pd_op.pool2d" (%57, %58) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D/"} : (tensor<1x64x112x112xf32>, tensor<2xi64>) -> tensor<1x64x55x55xf32> + (%60) = "pd_op.conv2d" (%59, %49) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<16x64x1x1xf32>) -> tensor<1x16x55x55xf32> + (%61) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%62) = "pd_op.reshape" (%48, %61) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<16xf32>, tensor<4xi64>) -> tensor<1x16x1x1xf32> + (%63) = "pd_op.add" (%60, %62) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<1x16x1x1xf32>) -> tensor<1x16x55x55xf32> + (%64) = "pd_op.relu" (%63) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/"} : (tensor<1x16x55x55xf32>) -> tensor<1x16x55x55xf32> + (%65) = "pd_op.conv2d" (%64, %47) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x1x1xf32>) -> tensor<1x64x55x55xf32> + (%66) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%67) = "pd_op.reshape" (%46, %66) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%68) = "pd_op.add" (%65, %67) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%69) = "pd_op.relu" (%68) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%70) = "pd_op.conv2d" (%64, %45) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x3x3xf32>) -> tensor<1x64x55x55xf32> + (%71) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%72) = "pd_op.reshape" (%44, %71) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%73) = "pd_op.add" (%70, %72) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%74) = "pd_op.relu" (%73) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%75) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/",value:1} : () -> tensor<1xi32> + (%76) = "builtin.combine" (%69, %74) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/"} : (tensor<1x64x55x55xf32>, tensor<1x64x55x55xf32>) -> vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>] + (%77) = "pd_op.concat" (%76, %75) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/"} : (vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>], tensor<1xi32>) -> tensor<1x128x55x55xf32> + (%78) = "pd_op.conv2d" (%77, %43) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<1x128x55x55xf32>, tensor<16x128x1x1xf32>) -> tensor<1x16x55x55xf32> + (%79) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%80) = "pd_op.reshape" (%42, %79) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<16xf32>, tensor<4xi64>) -> tensor<1x16x1x1xf32> + (%81) = "pd_op.add" (%78, %80) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<1x16x1x1xf32>) -> tensor<1x16x55x55xf32> + (%82) = "pd_op.relu" (%81) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/"} : (tensor<1x16x55x55xf32>) -> tensor<1x16x55x55xf32> + (%83) = "pd_op.conv2d" (%82, %41) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x1x1xf32>) -> tensor<1x64x55x55xf32> + (%84) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%85) = "pd_op.reshape" (%40, %84) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%86) = "pd_op.add" (%83, %85) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%87) = "pd_op.relu" (%86) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%88) = "pd_op.conv2d" (%82, %39) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x3x3xf32>) -> tensor<1x64x55x55xf32> + (%89) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%90) = "pd_op.reshape" (%38, %89) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%91) = "pd_op.add" (%88, %90) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%92) = "pd_op.relu" (%91) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%93) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/",value:1} : () -> tensor<1xi32> + (%94) = "builtin.combine" (%87, %92) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/"} : (tensor<1x64x55x55xf32>, tensor<1x64x55x55xf32>) -> vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>] + (%95) = "pd_op.concat" (%94, %93) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/"} : (vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>], tensor<1xi32>) -> tensor<1x128x55x55xf32> + (%96) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D_1/",value:[3,3]} : () -> tensor<2xi64> + (%97) = "pd_op.pool2d" (%95, %96) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D_1/"} : (tensor<1x128x55x55xf32>, tensor<2xi64>) -> tensor<1x128x27x27xf32> + (%98) = "pd_op.conv2d" (%97, %37) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<32x128x1x1xf32>) -> tensor<1x32x27x27xf32> + (%99) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%100) = "pd_op.reshape" (%36, %99) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<32xf32>, tensor<4xi64>) -> tensor<1x32x1x1xf32> + (%101) = "pd_op.add" (%98, %100) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<1x32x1x1xf32>) -> tensor<1x32x27x27xf32> + (%102) = "pd_op.relu" (%101) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/"} : (tensor<1x32x27x27xf32>) -> tensor<1x32x27x27xf32> + (%103) = "pd_op.conv2d" (%102, %35) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x1x1xf32>) -> tensor<1x128x27x27xf32> + (%104) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%105) = "pd_op.reshape" (%34, %104) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%106) = "pd_op.add" (%103, %105) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%107) = "pd_op.relu" (%106) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%108) = "pd_op.conv2d" (%102, %33) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x3x3xf32>) -> tensor<1x128x27x27xf32> + (%109) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%110) = "pd_op.reshape" (%32, %109) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%111) = "pd_op.add" (%108, %110) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%112) = "pd_op.relu" (%111) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%113) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/",value:1} : () -> tensor<1xi32> + (%114) = "builtin.combine" (%107, %112) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/"} : (tensor<1x128x27x27xf32>, tensor<1x128x27x27xf32>) -> vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>] + (%115) = "pd_op.concat" (%114, %113) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/"} : (vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>], tensor<1xi32>) -> tensor<1x256x27x27xf32> + (%116) = "pd_op.conv2d" (%115, %31) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<1x256x27x27xf32>, tensor<32x256x1x1xf32>) -> tensor<1x32x27x27xf32> + (%117) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%118) = "pd_op.reshape" (%30, %117) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<32xf32>, tensor<4xi64>) -> tensor<1x32x1x1xf32> + (%119) = "pd_op.add" (%116, %118) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<1x32x1x1xf32>) -> tensor<1x32x27x27xf32> + (%120) = "pd_op.relu" (%119) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/"} : (tensor<1x32x27x27xf32>) -> tensor<1x32x27x27xf32> + (%121) = "pd_op.conv2d" (%120, %29) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x1x1xf32>) -> tensor<1x128x27x27xf32> + (%122) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%123) = "pd_op.reshape" (%28, %122) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%124) = "pd_op.add" (%121, %123) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%125) = "pd_op.relu" (%124) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%126) = "pd_op.conv2d" (%120, %27) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x3x3xf32>) -> tensor<1x128x27x27xf32> + (%127) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%128) = "pd_op.reshape" (%26, %127) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%129) = "pd_op.add" (%126, %128) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%130) = "pd_op.relu" (%129) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%131) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/",value:1} : () -> tensor<1xi32> + (%132) = "builtin.combine" (%125, %130) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/"} : (tensor<1x128x27x27xf32>, tensor<1x128x27x27xf32>) -> vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>] + (%133) = "pd_op.concat" (%132, %131) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/"} : (vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>], tensor<1xi32>) -> tensor<1x256x27x27xf32> + (%134) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D_2/",value:[3,3]} : () -> tensor<2xi64> + (%135) = "pd_op.pool2d" (%133, %134) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D_2/"} : (tensor<1x256x27x27xf32>, tensor<2xi64>) -> tensor<1x256x13x13xf32> + (%136) = "pd_op.conv2d" (%135, %25) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<48x256x1x1xf32>) -> tensor<1x48x13x13xf32> + (%137) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%138) = "pd_op.reshape" (%24, %137) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<48xf32>, tensor<4xi64>) -> tensor<1x48x1x1xf32> + (%139) = "pd_op.add" (%136, %138) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<1x48x1x1xf32>) -> tensor<1x48x13x13xf32> + (%140) = "pd_op.relu" (%139) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/"} : (tensor<1x48x13x13xf32>) -> tensor<1x48x13x13xf32> + (%141) = "pd_op.conv2d" (%140, %23) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x1x1xf32>) -> tensor<1x192x13x13xf32> + (%142) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%143) = "pd_op.reshape" (%22, %142) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%144) = "pd_op.add" (%141, %143) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%145) = "pd_op.relu" (%144) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%146) = "pd_op.conv2d" (%140, %21) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x3x3xf32>) -> tensor<1x192x13x13xf32> + (%147) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%148) = "pd_op.reshape" (%20, %147) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%149) = "pd_op.add" (%146, %148) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%150) = "pd_op.relu" (%149) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%151) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/",value:1} : () -> tensor<1xi32> + (%152) = "builtin.combine" (%145, %150) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/"} : (tensor<1x192x13x13xf32>, tensor<1x192x13x13xf32>) -> vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>] + (%153) = "pd_op.concat" (%152, %151) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/"} : (vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>], tensor<1xi32>) -> tensor<1x384x13x13xf32> + (%154) = "pd_op.conv2d" (%153, %19) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<1x384x13x13xf32>, tensor<48x384x1x1xf32>) -> tensor<1x48x13x13xf32> + (%155) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%156) = "pd_op.reshape" (%18, %155) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<48xf32>, tensor<4xi64>) -> tensor<1x48x1x1xf32> + (%157) = "pd_op.add" (%154, %156) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<1x48x1x1xf32>) -> tensor<1x48x13x13xf32> + (%158) = "pd_op.relu" (%157) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/"} : (tensor<1x48x13x13xf32>) -> tensor<1x48x13x13xf32> + (%159) = "pd_op.conv2d" (%158, %17) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x1x1xf32>) -> tensor<1x192x13x13xf32> + (%160) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%161) = "pd_op.reshape" (%16, %160) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%162) = "pd_op.add" (%159, %161) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%163) = "pd_op.relu" (%162) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%164) = "pd_op.conv2d" (%158, %15) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x3x3xf32>) -> tensor<1x192x13x13xf32> + (%165) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%166) = "pd_op.reshape" (%14, %165) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%167) = "pd_op.add" (%164, %166) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%168) = "pd_op.relu" (%167) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%169) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/",value:1} : () -> tensor<1xi32> + (%170) = "builtin.combine" (%163, %168) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/"} : (tensor<1x192x13x13xf32>, tensor<1x192x13x13xf32>) -> vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>] + (%171) = "pd_op.concat" (%170, %169) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/"} : (vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>], tensor<1xi32>) -> tensor<1x384x13x13xf32> + (%172) = "pd_op.conv2d" (%171, %13) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<1x384x13x13xf32>, tensor<64x384x1x1xf32>) -> tensor<1x64x13x13xf32> + (%173) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%174) = "pd_op.reshape" (%12, %173) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%175) = "pd_op.add" (%172, %174) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x13x13xf32> + (%176) = "pd_op.relu" (%175) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/"} : (tensor<1x64x13x13xf32>) -> tensor<1x64x13x13xf32> + (%177) = "pd_op.conv2d" (%176, %11) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x1x1xf32>) -> tensor<1x256x13x13xf32> + (%178) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%179) = "pd_op.reshape" (%10, %178) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%180) = "pd_op.add" (%177, %179) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%181) = "pd_op.relu" (%180) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%182) = "pd_op.conv2d" (%176, %9) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x3x3xf32>) -> tensor<1x256x13x13xf32> + (%183) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%184) = "pd_op.reshape" (%8, %183) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%185) = "pd_op.add" (%182, %184) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%186) = "pd_op.relu" (%185) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%187) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/",value:1} : () -> tensor<1xi32> + (%188) = "builtin.combine" (%181, %186) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/"} : (tensor<1x256x13x13xf32>, tensor<1x256x13x13xf32>) -> vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>] + (%189) = "pd_op.concat" (%188, %187) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/"} : (vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>], tensor<1xi32>) -> tensor<1x512x13x13xf32> + (%190) = "pd_op.conv2d" (%189, %7) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<1x512x13x13xf32>, tensor<64x512x1x1xf32>) -> tensor<1x64x13x13xf32> + (%191) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%192) = "pd_op.reshape" (%6, %191) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%193) = "pd_op.add" (%190, %192) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x13x13xf32> + (%194) = "pd_op.relu" (%193) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/"} : (tensor<1x64x13x13xf32>) -> tensor<1x64x13x13xf32> + (%195) = "pd_op.conv2d" (%194, %5) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x1x1xf32>) -> tensor<1x256x13x13xf32> + (%196) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%197) = "pd_op.reshape" (%4, %196) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%198) = "pd_op.add" (%195, %197) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%199) = "pd_op.relu" (%198) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%200) = "pd_op.conv2d" (%194, %3) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x3x3xf32>) -> tensor<1x256x13x13xf32> + (%201) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%202) = "pd_op.reshape" (%2, %201) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%203) = "pd_op.add" (%200, %202) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%204) = "pd_op.relu" (%203) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%205) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/",value:1} : () -> tensor<1xi32> + (%206) = "builtin.combine" (%199, %204) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/"} : (tensor<1x256x13x13xf32>, tensor<1x256x13x13xf32>) -> vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>] + (%207) = "pd_op.concat" (%206, %205) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/"} : (vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>], tensor<1xi32>) -> tensor<1x512x13x13xf32> + (%208) = "pd_op.full" () {dtype:float32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/Dropout/",value:0.5} : () -> tensor<1xf32> + (%209, %210) = "pd_op.dropout" (%207, <>, %208) {fix_seed:false,is_test:false,mode:"downgrade_in_infer",seed:0,stop_gradient:[false,false],struct_name:"/SqueezeNet/Dropout/"} : (tensor<1x512x13x13xf32>, <>, tensor<1xf32>) -> tensor<1x512x13x13xf32>, tensor<1x512x13x13xu8> + (%211) = "pd_op.conv2d" (%209, %1) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1x512x13x13xf32>, tensor<1000x512x1x1xf32>) -> tensor<1x1000x13x13xf32> + (%212) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/Conv2D_1/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%213) = "pd_op.reshape" (%0, %212) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1000xf32>, tensor<4xi64>) -> tensor<1x1000x1x1xf32> + (%214) = "pd_op.add" (%211, %213) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1x1000x13x13xf32>, tensor<1x1000x1x1xf32>) -> tensor<1x1000x13x13xf32> + (%215) = "pd_op.relu" (%214) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x1000x13x13xf32>) -> tensor<1x1000x13x13xf32> + (%216) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/AdaptiveAvgPool2D/",value:[1,1]} : () -> tensor<2xi64> + (%217) = "pd_op.pool2d" (%215, %216) {adaptive:true,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"avg",stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/AdaptiveAvgPool2D/"} : (tensor<1x1000x13x13xf32>, tensor<2xi64>) -> tensor<1x1000x1x1xf32> + (%218) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/",value:[2,3]} : () -> tensor<2xi64> + (%219) = "pd_op.squeeze" (%217, %218) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x1000x1x1xf32>, tensor<2xi64>) -> tensor<1x1000xf32> +} diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/weight_meta.py b/graph_net/samples/paddle_samples/vision/squeezenet1_1/weight_meta.py new file mode 100644 index 000000000..714737364 --- /dev/null +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/weight_meta.py @@ -0,0 +1,3 @@ +import paddle + +weight_meta = {} diff --git a/graph_net/tests/test_squeezenet1_1_extract.py b/graph_net/tests/test_squeezenet1_1_extract.py new file mode 100644 index 000000000..fc29fc27f --- /dev/null +++ b/graph_net/tests/test_squeezenet1_1_extract.py @@ -0,0 +1,33 @@ +import paddle +import os +from paddle.vision.models import squeezenet1_1 +from graph_net.paddle.extractor import GraphExtractor + +# 1. 环境准备 +os.environ["GRAPH_NET_EXTRACT_WORKSPACE"] = os.path.abspath("./my_samples") +if not os.path.exists("./my_samples"): + os.makedirs("./my_samples") + +# 2. 准备模型 +model = squeezenet1_1(pretrained=False) +model.eval() + +# 3. 定义 InputSpec (关键改动：name='inputs') +input_spec = [paddle.static.InputSpec(shape=[1, 3, 224, 224], dtype='float32', name='inputs')] + +# 4. 手动实例化提取器 +print("正在初始化提取器...") +extractor = GraphExtractor(model, name="squeezenet1_1", dynamic=False, input_spec=input_spec) + +# 5. 执行提取 (关键改动：Key 名改为 'inputs') +print("正在执行提取流程...") +model_dump_path = os.path.join(os.environ["GRAPH_NET_EXTRACT_WORKSPACE"], "squeezenet1_1") +dummy_data = {"inputs": paddle.randn([1, 3, 224, 224])} + +try: + # 绕过 __call__ 直接运行内部 dump 逻辑 + extractor.run_model_with_dump_enabled(model_dump_path, **dummy_data) + print("\n✨✨✨ 奇迹发生了！提取流程成功完成！ ✨✨✨") + print(f"产物已存至: {model_dump_path}") +except Exception as e: + print(f"\n❌ 捕获到错误: {e}") \ No newline at end of file diff --git a/my_extractor.py b/my_extractor.py new file mode 100644 index 000000000..ab69bc13c --- /dev/null +++ b/my_extractor.py @@ -0,0 +1,33 @@ +import paddle +import os +from paddle.vision.models import squeezenet1_1 + +# 1. 环境强制设定 +os.environ["FLAGS_enable_pir_api"] = "1" +paddle.enable_static() # 切换到静态图模式 + +save_dir = "./my_samples/squeezenet1_1" +if not os.path.exists(save_dir): + os.makedirs(save_dir) + +# 2. 定义静态图容器 +main_program = paddle.static.Program() +startup_program = paddle.static.Program() + +with paddle.static.program_guard(main_program, startup_program): + # 3. 在静态图内定义输入 + inputs = paddle.static.data(name='inputs', shape=[1, 3, 224, 224], dtype='float32') + + # 4. 实例化模型并运squeezenet1_1squeezenet1_1 + model = squeezenet1_1(pretrained=False) + out = model(inputs) + +# 5. 此时 main_program 已经包含 PIR 计算图 +save_path = os.path.join(save_dir, "squeezenet1_1.json") +with open(save_path, "w") as f: + # 导出 PIR 的文本序列化 + f.write(str(main_program)) + +print(f"DONE! PIR Graph captured.") +print(f"File: {save_path}") +print(f"Size: {os.path.getsize(save_path)} bytes") diff --git a/my_samples/squeezenet1_1/__init__.py b/my_samples/squeezenet1_1/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/my_samples/squeezenet1_1/graph_hash.txt b/my_samples/squeezenet1_1/graph_hash.txt new file mode 100644 index 000000000..8ebaa39af --- /dev/null +++ b/my_samples/squeezenet1_1/graph_hash.txt @@ -0,0 +1 @@ +6de9a1959b0f3ccd2b9e70f1b42a4f295af1bea89e4779c1e0fe5753bc609be7 \ No newline at end of file diff --git a/my_samples/squeezenet1_1/input_meta.py b/my_samples/squeezenet1_1/input_meta.py new file mode 100644 index 000000000..6453218cf --- /dev/null +++ b/my_samples/squeezenet1_1/input_meta.py @@ -0,0 +1,8 @@ +import paddle + +input_meta = { + "inputs": { + "shape": [1, 3, 224, 224], + "dtype": "float32" + } +} diff --git a/my_samples/squeezenet1_1/model.py b/my_samples/squeezenet1_1/model.py new file mode 100644 index 000000000..a149a8966 --- /dev/null +++ b/my_samples/squeezenet1_1/model.py @@ -0,0 +1,13 @@ +import paddle +from paddle.vision.models import squeezenet1_1 + +class GraphModule(paddle.nn.Layer): + def __init__(self): + super(GraphModule, self).__init__() + self.model = squeezenet1_1(pretrained=False) + + def forward(self, inputs=None): + # 防御性修复：如果 inputs 为空，生成一个符合规格的 dummy tensor + if inputs is None: + inputs = paddle.randn([1, 3, 224, 224]) + return self.model(inputs) diff --git a/my_samples/squeezenet1_1/squeezenet1_1.json b/my_samples/squeezenet1_1/squeezenet1_1.json new file mode 100644 index 000000000..8cdf0fc0b --- /dev/null +++ b/my_samples/squeezenet1_1/squeezenet1_1.json @@ -0,0 +1,221 @@ +{ + (%0) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_25.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<1000xf32> + (%1) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_25.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<1000x512x1x1xf32> + (%2) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_24.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%3) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_24.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x3x3xf32> + (%4) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_23.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%5) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_23.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x1x1xf32> + (%6) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_22.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%7) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_22.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x512x1x1xf32> + (%8) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_21.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%9) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_21.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x3x3xf32> + (%10) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_20.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> + (%11) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_20.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x1x1xf32> + (%12) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_19.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%13) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_19.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x384x1x1xf32> + (%14) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_18.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%15) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_18.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x3x3xf32> + (%16) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_17.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%17) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_17.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x1x1xf32> + (%18) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_16.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48xf32> + (%19) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_16.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48x384x1x1xf32> + (%20) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_15.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%21) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_15.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x3x3xf32> + (%22) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_14.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> + (%23) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_14.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x1x1xf32> + (%24) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_13.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48xf32> + (%25) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_13.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48x256x1x1xf32> + (%26) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_12.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%27) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_12.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x3x3xf32> + (%28) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_11.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%29) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_11.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x1x1xf32> + (%30) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_10.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32xf32> + (%31) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_10.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32x256x1x1xf32> + (%32) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_9.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%33) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_9.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x3x3xf32> + (%34) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_8.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> + (%35) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_8.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x1x1xf32> + (%36) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_7.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32xf32> + (%37) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_7.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32x128x1x1xf32> + (%38) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_6.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%39) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_6.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x3x3xf32> + (%40) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_5.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%41) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_5.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x1x1xf32> + (%42) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_4.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16xf32> + (%43) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_4.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16x128x1x1xf32> + (%44) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_3.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%45) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_3.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x3x3xf32> + (%46) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_2.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%47) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_2.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x1x1xf32> + (%48) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_1.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16xf32> + (%49) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_1.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16x64x1x1xf32> + (%50) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_0.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> + (%51) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_0.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x3x3x3xf32> + (%52) = "pd_op.data" () {dtype:float32,name:"inputs",place:Place(undefined:0),shape:[1,3,224,224],stop_gradient:[true]} : () -> tensor<1x3x224x224xf32> + (%53) = "pd_op.conv2d" (%52, %51) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<1x3x224x224xf32>, tensor<64x3x3x3xf32>) -> tensor<1x64x112x112xf32> + (%54) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%55) = "pd_op.reshape" (%50, %54) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%56) = "pd_op.add" (%53, %55) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32> + (%57) = "pd_op.relu" (%56) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x64x112x112xf32>) -> tensor<1x64x112x112xf32> + (%58) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D/",value:[3,3]} : () -> tensor<2xi64> + (%59) = "pd_op.pool2d" (%57, %58) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D/"} : (tensor<1x64x112x112xf32>, tensor<2xi64>) -> tensor<1x64x55x55xf32> + (%60) = "pd_op.conv2d" (%59, %49) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<16x64x1x1xf32>) -> tensor<1x16x55x55xf32> + (%61) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%62) = "pd_op.reshape" (%48, %61) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<16xf32>, tensor<4xi64>) -> tensor<1x16x1x1xf32> + (%63) = "pd_op.add" (%60, %62) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<1x16x1x1xf32>) -> tensor<1x16x55x55xf32> + (%64) = "pd_op.relu" (%63) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/"} : (tensor<1x16x55x55xf32>) -> tensor<1x16x55x55xf32> + (%65) = "pd_op.conv2d" (%64, %47) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x1x1xf32>) -> tensor<1x64x55x55xf32> + (%66) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%67) = "pd_op.reshape" (%46, %66) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%68) = "pd_op.add" (%65, %67) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%69) = "pd_op.relu" (%68) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%70) = "pd_op.conv2d" (%64, %45) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x3x3xf32>) -> tensor<1x64x55x55xf32> + (%71) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%72) = "pd_op.reshape" (%44, %71) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%73) = "pd_op.add" (%70, %72) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%74) = "pd_op.relu" (%73) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%75) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/",value:1} : () -> tensor<1xi32> + (%76) = "builtin.combine" (%69, %74) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/"} : (tensor<1x64x55x55xf32>, tensor<1x64x55x55xf32>) -> vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>] + (%77) = "pd_op.concat" (%76, %75) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/"} : (vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>], tensor<1xi32>) -> tensor<1x128x55x55xf32> + (%78) = "pd_op.conv2d" (%77, %43) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<1x128x55x55xf32>, tensor<16x128x1x1xf32>) -> tensor<1x16x55x55xf32> + (%79) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%80) = "pd_op.reshape" (%42, %79) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<16xf32>, tensor<4xi64>) -> tensor<1x16x1x1xf32> + (%81) = "pd_op.add" (%78, %80) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<1x16x1x1xf32>) -> tensor<1x16x55x55xf32> + (%82) = "pd_op.relu" (%81) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/"} : (tensor<1x16x55x55xf32>) -> tensor<1x16x55x55xf32> + (%83) = "pd_op.conv2d" (%82, %41) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x1x1xf32>) -> tensor<1x64x55x55xf32> + (%84) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%85) = "pd_op.reshape" (%40, %84) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%86) = "pd_op.add" (%83, %85) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%87) = "pd_op.relu" (%86) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%88) = "pd_op.conv2d" (%82, %39) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x3x3xf32>) -> tensor<1x64x55x55xf32> + (%89) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%90) = "pd_op.reshape" (%38, %89) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%91) = "pd_op.add" (%88, %90) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> + (%92) = "pd_op.relu" (%91) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> + (%93) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/",value:1} : () -> tensor<1xi32> + (%94) = "builtin.combine" (%87, %92) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/"} : (tensor<1x64x55x55xf32>, tensor<1x64x55x55xf32>) -> vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>] + (%95) = "pd_op.concat" (%94, %93) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/"} : (vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>], tensor<1xi32>) -> tensor<1x128x55x55xf32> + (%96) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D_1/",value:[3,3]} : () -> tensor<2xi64> + (%97) = "pd_op.pool2d" (%95, %96) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D_1/"} : (tensor<1x128x55x55xf32>, tensor<2xi64>) -> tensor<1x128x27x27xf32> + (%98) = "pd_op.conv2d" (%97, %37) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<32x128x1x1xf32>) -> tensor<1x32x27x27xf32> + (%99) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%100) = "pd_op.reshape" (%36, %99) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<32xf32>, tensor<4xi64>) -> tensor<1x32x1x1xf32> + (%101) = "pd_op.add" (%98, %100) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<1x32x1x1xf32>) -> tensor<1x32x27x27xf32> + (%102) = "pd_op.relu" (%101) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/"} : (tensor<1x32x27x27xf32>) -> tensor<1x32x27x27xf32> + (%103) = "pd_op.conv2d" (%102, %35) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x1x1xf32>) -> tensor<1x128x27x27xf32> + (%104) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%105) = "pd_op.reshape" (%34, %104) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%106) = "pd_op.add" (%103, %105) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%107) = "pd_op.relu" (%106) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%108) = "pd_op.conv2d" (%102, %33) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x3x3xf32>) -> tensor<1x128x27x27xf32> + (%109) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%110) = "pd_op.reshape" (%32, %109) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%111) = "pd_op.add" (%108, %110) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%112) = "pd_op.relu" (%111) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%113) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/",value:1} : () -> tensor<1xi32> + (%114) = "builtin.combine" (%107, %112) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/"} : (tensor<1x128x27x27xf32>, tensor<1x128x27x27xf32>) -> vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>] + (%115) = "pd_op.concat" (%114, %113) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/"} : (vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>], tensor<1xi32>) -> tensor<1x256x27x27xf32> + (%116) = "pd_op.conv2d" (%115, %31) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<1x256x27x27xf32>, tensor<32x256x1x1xf32>) -> tensor<1x32x27x27xf32> + (%117) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%118) = "pd_op.reshape" (%30, %117) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<32xf32>, tensor<4xi64>) -> tensor<1x32x1x1xf32> + (%119) = "pd_op.add" (%116, %118) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<1x32x1x1xf32>) -> tensor<1x32x27x27xf32> + (%120) = "pd_op.relu" (%119) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/"} : (tensor<1x32x27x27xf32>) -> tensor<1x32x27x27xf32> + (%121) = "pd_op.conv2d" (%120, %29) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x1x1xf32>) -> tensor<1x128x27x27xf32> + (%122) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%123) = "pd_op.reshape" (%28, %122) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%124) = "pd_op.add" (%121, %123) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%125) = "pd_op.relu" (%124) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%126) = "pd_op.conv2d" (%120, %27) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x3x3xf32>) -> tensor<1x128x27x27xf32> + (%127) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%128) = "pd_op.reshape" (%26, %127) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> + (%129) = "pd_op.add" (%126, %128) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> + (%130) = "pd_op.relu" (%129) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> + (%131) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/",value:1} : () -> tensor<1xi32> + (%132) = "builtin.combine" (%125, %130) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/"} : (tensor<1x128x27x27xf32>, tensor<1x128x27x27xf32>) -> vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>] + (%133) = "pd_op.concat" (%132, %131) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/"} : (vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>], tensor<1xi32>) -> tensor<1x256x27x27xf32> + (%134) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D_2/",value:[3,3]} : () -> tensor<2xi64> + (%135) = "pd_op.pool2d" (%133, %134) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D_2/"} : (tensor<1x256x27x27xf32>, tensor<2xi64>) -> tensor<1x256x13x13xf32> + (%136) = "pd_op.conv2d" (%135, %25) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<48x256x1x1xf32>) -> tensor<1x48x13x13xf32> + (%137) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%138) = "pd_op.reshape" (%24, %137) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<48xf32>, tensor<4xi64>) -> tensor<1x48x1x1xf32> + (%139) = "pd_op.add" (%136, %138) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<1x48x1x1xf32>) -> tensor<1x48x13x13xf32> + (%140) = "pd_op.relu" (%139) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/"} : (tensor<1x48x13x13xf32>) -> tensor<1x48x13x13xf32> + (%141) = "pd_op.conv2d" (%140, %23) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x1x1xf32>) -> tensor<1x192x13x13xf32> + (%142) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%143) = "pd_op.reshape" (%22, %142) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%144) = "pd_op.add" (%141, %143) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%145) = "pd_op.relu" (%144) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%146) = "pd_op.conv2d" (%140, %21) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x3x3xf32>) -> tensor<1x192x13x13xf32> + (%147) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%148) = "pd_op.reshape" (%20, %147) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%149) = "pd_op.add" (%146, %148) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%150) = "pd_op.relu" (%149) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%151) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/",value:1} : () -> tensor<1xi32> + (%152) = "builtin.combine" (%145, %150) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/"} : (tensor<1x192x13x13xf32>, tensor<1x192x13x13xf32>) -> vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>] + (%153) = "pd_op.concat" (%152, %151) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/"} : (vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>], tensor<1xi32>) -> tensor<1x384x13x13xf32> + (%154) = "pd_op.conv2d" (%153, %19) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<1x384x13x13xf32>, tensor<48x384x1x1xf32>) -> tensor<1x48x13x13xf32> + (%155) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%156) = "pd_op.reshape" (%18, %155) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<48xf32>, tensor<4xi64>) -> tensor<1x48x1x1xf32> + (%157) = "pd_op.add" (%154, %156) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<1x48x1x1xf32>) -> tensor<1x48x13x13xf32> + (%158) = "pd_op.relu" (%157) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/"} : (tensor<1x48x13x13xf32>) -> tensor<1x48x13x13xf32> + (%159) = "pd_op.conv2d" (%158, %17) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x1x1xf32>) -> tensor<1x192x13x13xf32> + (%160) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%161) = "pd_op.reshape" (%16, %160) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%162) = "pd_op.add" (%159, %161) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%163) = "pd_op.relu" (%162) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%164) = "pd_op.conv2d" (%158, %15) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x3x3xf32>) -> tensor<1x192x13x13xf32> + (%165) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%166) = "pd_op.reshape" (%14, %165) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> + (%167) = "pd_op.add" (%164, %166) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> + (%168) = "pd_op.relu" (%167) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> + (%169) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/",value:1} : () -> tensor<1xi32> + (%170) = "builtin.combine" (%163, %168) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/"} : (tensor<1x192x13x13xf32>, tensor<1x192x13x13xf32>) -> vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>] + (%171) = "pd_op.concat" (%170, %169) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/"} : (vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>], tensor<1xi32>) -> tensor<1x384x13x13xf32> + (%172) = "pd_op.conv2d" (%171, %13) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<1x384x13x13xf32>, tensor<64x384x1x1xf32>) -> tensor<1x64x13x13xf32> + (%173) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%174) = "pd_op.reshape" (%12, %173) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%175) = "pd_op.add" (%172, %174) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x13x13xf32> + (%176) = "pd_op.relu" (%175) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/"} : (tensor<1x64x13x13xf32>) -> tensor<1x64x13x13xf32> + (%177) = "pd_op.conv2d" (%176, %11) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x1x1xf32>) -> tensor<1x256x13x13xf32> + (%178) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%179) = "pd_op.reshape" (%10, %178) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%180) = "pd_op.add" (%177, %179) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%181) = "pd_op.relu" (%180) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%182) = "pd_op.conv2d" (%176, %9) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x3x3xf32>) -> tensor<1x256x13x13xf32> + (%183) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%184) = "pd_op.reshape" (%8, %183) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%185) = "pd_op.add" (%182, %184) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%186) = "pd_op.relu" (%185) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%187) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/",value:1} : () -> tensor<1xi32> + (%188) = "builtin.combine" (%181, %186) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/"} : (tensor<1x256x13x13xf32>, tensor<1x256x13x13xf32>) -> vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>] + (%189) = "pd_op.concat" (%188, %187) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/"} : (vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>], tensor<1xi32>) -> tensor<1x512x13x13xf32> + (%190) = "pd_op.conv2d" (%189, %7) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<1x512x13x13xf32>, tensor<64x512x1x1xf32>) -> tensor<1x64x13x13xf32> + (%191) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%192) = "pd_op.reshape" (%6, %191) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> + (%193) = "pd_op.add" (%190, %192) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x13x13xf32> + (%194) = "pd_op.relu" (%193) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/"} : (tensor<1x64x13x13xf32>) -> tensor<1x64x13x13xf32> + (%195) = "pd_op.conv2d" (%194, %5) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x1x1xf32>) -> tensor<1x256x13x13xf32> + (%196) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%197) = "pd_op.reshape" (%4, %196) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%198) = "pd_op.add" (%195, %197) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%199) = "pd_op.relu" (%198) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%200) = "pd_op.conv2d" (%194, %3) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x3x3xf32>) -> tensor<1x256x13x13xf32> + (%201) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%202) = "pd_op.reshape" (%2, %201) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> + (%203) = "pd_op.add" (%200, %202) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> + (%204) = "pd_op.relu" (%203) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> + (%205) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/",value:1} : () -> tensor<1xi32> + (%206) = "builtin.combine" (%199, %204) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/"} : (tensor<1x256x13x13xf32>, tensor<1x256x13x13xf32>) -> vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>] + (%207) = "pd_op.concat" (%206, %205) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/"} : (vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>], tensor<1xi32>) -> tensor<1x512x13x13xf32> + (%208) = "pd_op.full" () {dtype:float32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/Dropout/",value:0.5} : () -> tensor<1xf32> + (%209, %210) = "pd_op.dropout" (%207, <>, %208) {fix_seed:false,is_test:false,mode:"downgrade_in_infer",seed:0,stop_gradient:[false,false],struct_name:"/SqueezeNet/Dropout/"} : (tensor<1x512x13x13xf32>, <>, tensor<1xf32>) -> tensor<1x512x13x13xf32>, tensor<1x512x13x13xu8> + (%211) = "pd_op.conv2d" (%209, %1) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1x512x13x13xf32>, tensor<1000x512x1x1xf32>) -> tensor<1x1000x13x13xf32> + (%212) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/Conv2D_1/",value:[1,-1,1,1]} : () -> tensor<4xi64> + (%213) = "pd_op.reshape" (%0, %212) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1000xf32>, tensor<4xi64>) -> tensor<1x1000x1x1xf32> + (%214) = "pd_op.add" (%211, %213) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1x1000x13x13xf32>, tensor<1x1000x1x1xf32>) -> tensor<1x1000x13x13xf32> + (%215) = "pd_op.relu" (%214) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x1000x13x13xf32>) -> tensor<1x1000x13x13xf32> + (%216) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/AdaptiveAvgPool2D/",value:[1,1]} : () -> tensor<2xi64> + (%217) = "pd_op.pool2d" (%215, %216) {adaptive:true,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"avg",stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/AdaptiveAvgPool2D/"} : (tensor<1x1000x13x13xf32>, tensor<2xi64>) -> tensor<1x1000x1x1xf32> + (%218) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/",value:[2,3]} : () -> tensor<2xi64> + (%219) = "pd_op.squeeze" (%217, %218) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x1000x1x1xf32>, tensor<2xi64>) -> tensor<1x1000xf32> +} diff --git a/my_samples/squeezenet1_1/weight_meta.py b/my_samples/squeezenet1_1/weight_meta.py new file mode 100644 index 000000000..714737364 --- /dev/null +++ b/my_samples/squeezenet1_1/weight_meta.py @@ -0,0 +1,3 @@ +import paddle + +weight_meta = {} From 249ff9e470f891a1709ea80ade20932bf4dc1a83 Mon Sep 17 00:00:00 2001 From: Carousel126 Date: Mon, 9 Feb 2026 11:24:44 +0800 Subject: [PATCH 2/3] chore: cleanup temporary extraction scripts --- my_extractor.py | 33 --- my_samples/squeezenet1_1/__init__.py | 0 my_samples/squeezenet1_1/graph_hash.txt | 1 - my_samples/squeezenet1_1/input_meta.py | 8 - my_samples/squeezenet1_1/model.py | 13 -- my_samples/squeezenet1_1/squeezenet1_1.json | 221 -------------------- my_samples/squeezenet1_1/weight_meta.py | 3 - 7 files changed, 279 deletions(-) delete mode 100644 my_extractor.py delete mode 100644 my_samples/squeezenet1_1/__init__.py delete mode 100644 my_samples/squeezenet1_1/graph_hash.txt delete mode 100644 my_samples/squeezenet1_1/input_meta.py delete mode 100644 my_samples/squeezenet1_1/model.py delete mode 100644 my_samples/squeezenet1_1/squeezenet1_1.json delete mode 100644 my_samples/squeezenet1_1/weight_meta.py diff --git a/my_extractor.py b/my_extractor.py deleted file mode 100644 index ab69bc13c..000000000 --- a/my_extractor.py +++ /dev/null @@ -1,33 +0,0 @@ -import paddle -import os -from paddle.vision.models import squeezenet1_1 - -# 1. 环境强制设定 -os.environ["FLAGS_enable_pir_api"] = "1" -paddle.enable_static() # 切换到静态图模式 - -save_dir = "./my_samples/squeezenet1_1" -if not os.path.exists(save_dir): - os.makedirs(save_dir) - -# 2. 定义静态图容器 -main_program = paddle.static.Program() -startup_program = paddle.static.Program() - -with paddle.static.program_guard(main_program, startup_program): - # 3. 在静态图内定义输入 - inputs = paddle.static.data(name='inputs', shape=[1, 3, 224, 224], dtype='float32') - - # 4. 实例化模型并运squeezenet1_1squeezenet1_1 - model = squeezenet1_1(pretrained=False) - out = model(inputs) - -# 5. 此时 main_program 已经包含 PIR 计算图 -save_path = os.path.join(save_dir, "squeezenet1_1.json") -with open(save_path, "w") as f: - # 导出 PIR 的文本序列化 - f.write(str(main_program)) - -print(f"DONE! PIR Graph captured.") -print(f"File: {save_path}") -print(f"Size: {os.path.getsize(save_path)} bytes") diff --git a/my_samples/squeezenet1_1/__init__.py b/my_samples/squeezenet1_1/__init__.py deleted file mode 100644 index e69de29bb..000000000 diff --git a/my_samples/squeezenet1_1/graph_hash.txt b/my_samples/squeezenet1_1/graph_hash.txt deleted file mode 100644 index 8ebaa39af..000000000 --- a/my_samples/squeezenet1_1/graph_hash.txt +++ /dev/null @@ -1 +0,0 @@ -6de9a1959b0f3ccd2b9e70f1b42a4f295af1bea89e4779c1e0fe5753bc609be7 \ No newline at end of file diff --git a/my_samples/squeezenet1_1/input_meta.py b/my_samples/squeezenet1_1/input_meta.py deleted file mode 100644 index 6453218cf..000000000 --- a/my_samples/squeezenet1_1/input_meta.py +++ /dev/null @@ -1,8 +0,0 @@ -import paddle - -input_meta = { - "inputs": { - "shape": [1, 3, 224, 224], - "dtype": "float32" - } -} diff --git a/my_samples/squeezenet1_1/model.py b/my_samples/squeezenet1_1/model.py deleted file mode 100644 index a149a8966..000000000 --- a/my_samples/squeezenet1_1/model.py +++ /dev/null @@ -1,13 +0,0 @@ -import paddle -from paddle.vision.models import squeezenet1_1 - -class GraphModule(paddle.nn.Layer): - def __init__(self): - super(GraphModule, self).__init__() - self.model = squeezenet1_1(pretrained=False) - - def forward(self, inputs=None): - # 防御性修复：如果 inputs 为空，生成一个符合规格的 dummy tensor - if inputs is None: - inputs = paddle.randn([1, 3, 224, 224]) - return self.model(inputs) diff --git a/my_samples/squeezenet1_1/squeezenet1_1.json b/my_samples/squeezenet1_1/squeezenet1_1.json deleted file mode 100644 index 8cdf0fc0b..000000000 --- a/my_samples/squeezenet1_1/squeezenet1_1.json +++ /dev/null @@ -1,221 +0,0 @@ -{ - (%0) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_25.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<1000xf32> - (%1) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_25.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<1000x512x1x1xf32> - (%2) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_24.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> - (%3) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_24.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x3x3xf32> - (%4) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_23.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> - (%5) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_23.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x1x1xf32> - (%6) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_22.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%7) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_22.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x512x1x1xf32> - (%8) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_21.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> - (%9) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_21.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x3x3xf32> - (%10) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_20.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256xf32> - (%11) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_20.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<256x64x1x1xf32> - (%12) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_19.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%13) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_19.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x384x1x1xf32> - (%14) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_18.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> - (%15) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_18.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x3x3xf32> - (%16) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_17.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> - (%17) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_17.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x1x1xf32> - (%18) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_16.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48xf32> - (%19) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_16.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48x384x1x1xf32> - (%20) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_15.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> - (%21) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_15.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x3x3xf32> - (%22) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_14.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192xf32> - (%23) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_14.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<192x48x1x1xf32> - (%24) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_13.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48xf32> - (%25) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_13.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<48x256x1x1xf32> - (%26) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_12.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> - (%27) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_12.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x3x3xf32> - (%28) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_11.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> - (%29) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_11.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x1x1xf32> - (%30) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_10.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32xf32> - (%31) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_10.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32x256x1x1xf32> - (%32) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_9.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> - (%33) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_9.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x3x3xf32> - (%34) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_8.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128xf32> - (%35) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_8.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<128x32x1x1xf32> - (%36) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_7.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32xf32> - (%37) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_7.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<32x128x1x1xf32> - (%38) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_6.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%39) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_6.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x3x3xf32> - (%40) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_5.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%41) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_5.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x1x1xf32> - (%42) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_4.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16xf32> - (%43) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_4.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16x128x1x1xf32> - (%44) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_3.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%45) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_3.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x3x3xf32> - (%46) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_2.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%47) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_2.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x16x1x1xf32> - (%48) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_1.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16xf32> - (%49) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_1.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<16x64x1x1xf32> - (%50) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_0.b_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64xf32> - (%51) = "builtin.parameter" () {is_distributed:[false],is_parameter:[true],need_clip:[true],parameter_name:"conv2d_0.w_0",persistable:[true],stop_gradient:[false],trainable:[true]} : () -> tensor<64x3x3x3xf32> - (%52) = "pd_op.data" () {dtype:float32,name:"inputs",place:Place(undefined:0),shape:[1,3,224,224],stop_gradient:[true]} : () -> tensor<1x3x224x224xf32> - (%53) = "pd_op.conv2d" (%52, %51) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<1x3x224x224xf32>, tensor<64x3x3x3xf32>) -> tensor<1x64x112x112xf32> - (%54) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%55) = "pd_op.reshape" (%50, %54) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%56) = "pd_op.add" (%53, %55) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D/"} : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32> - (%57) = "pd_op.relu" (%56) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x64x112x112xf32>) -> tensor<1x64x112x112xf32> - (%58) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D/",value:[3,3]} : () -> tensor<2xi64> - (%59) = "pd_op.pool2d" (%57, %58) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D/"} : (tensor<1x64x112x112xf32>, tensor<2xi64>) -> tensor<1x64x55x55xf32> - (%60) = "pd_op.conv2d" (%59, %49) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<16x64x1x1xf32>) -> tensor<1x16x55x55xf32> - (%61) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%62) = "pd_op.reshape" (%48, %61) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<16xf32>, tensor<4xi64>) -> tensor<1x16x1x1xf32> - (%63) = "pd_op.add" (%60, %62) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<1x16x1x1xf32>) -> tensor<1x16x55x55xf32> - (%64) = "pd_op.relu" (%63) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv/"} : (tensor<1x16x55x55xf32>) -> tensor<1x16x55x55xf32> - (%65) = "pd_op.conv2d" (%64, %47) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x1x1xf32>) -> tensor<1x64x55x55xf32> - (%66) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%67) = "pd_op.reshape" (%46, %66) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%68) = "pd_op.add" (%65, %67) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> - (%69) = "pd_op.relu" (%68) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_1/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> - (%70) = "pd_op.conv2d" (%64, %45) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x3x3xf32>) -> tensor<1x64x55x55xf32> - (%71) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%72) = "pd_op.reshape" (%44, %71) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%73) = "pd_op.add" (%70, %72) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> - (%74) = "pd_op.relu" (%73) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/MakeFireConv_2/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> - (%75) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire/",value:1} : () -> tensor<1xi32> - (%76) = "builtin.combine" (%69, %74) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/"} : (tensor<1x64x55x55xf32>, tensor<1x64x55x55xf32>) -> vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>] - (%77) = "pd_op.concat" (%76, %75) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire/"} : (vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>], tensor<1xi32>) -> tensor<1x128x55x55xf32> - (%78) = "pd_op.conv2d" (%77, %43) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<1x128x55x55xf32>, tensor<16x128x1x1xf32>) -> tensor<1x16x55x55xf32> - (%79) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%80) = "pd_op.reshape" (%42, %79) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<16xf32>, tensor<4xi64>) -> tensor<1x16x1x1xf32> - (%81) = "pd_op.add" (%78, %80) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<1x16x1x1xf32>) -> tensor<1x16x55x55xf32> - (%82) = "pd_op.relu" (%81) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv/"} : (tensor<1x16x55x55xf32>) -> tensor<1x16x55x55xf32> - (%83) = "pd_op.conv2d" (%82, %41) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x1x1xf32>) -> tensor<1x64x55x55xf32> - (%84) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%85) = "pd_op.reshape" (%40, %84) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%86) = "pd_op.add" (%83, %85) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> - (%87) = "pd_op.relu" (%86) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_1/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> - (%88) = "pd_op.conv2d" (%82, %39) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<1x16x55x55xf32>, tensor<64x16x3x3xf32>) -> tensor<1x64x55x55xf32> - (%89) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%90) = "pd_op.reshape" (%38, %89) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%91) = "pd_op.add" (%88, %90) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/Conv2D/"} : (tensor<1x64x55x55xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x55x55xf32> - (%92) = "pd_op.relu" (%91) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/MakeFireConv_2/"} : (tensor<1x64x55x55xf32>) -> tensor<1x64x55x55xf32> - (%93) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_1/",value:1} : () -> tensor<1xi32> - (%94) = "builtin.combine" (%87, %92) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/"} : (tensor<1x64x55x55xf32>, tensor<1x64x55x55xf32>) -> vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>] - (%95) = "pd_op.concat" (%94, %93) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_1/"} : (vec[tensor<1x64x55x55xf32>,tensor<1x64x55x55xf32>], tensor<1xi32>) -> tensor<1x128x55x55xf32> - (%96) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D_1/",value:[3,3]} : () -> tensor<2xi64> - (%97) = "pd_op.pool2d" (%95, %96) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D_1/"} : (tensor<1x128x55x55xf32>, tensor<2xi64>) -> tensor<1x128x27x27xf32> - (%98) = "pd_op.conv2d" (%97, %37) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<32x128x1x1xf32>) -> tensor<1x32x27x27xf32> - (%99) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%100) = "pd_op.reshape" (%36, %99) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<32xf32>, tensor<4xi64>) -> tensor<1x32x1x1xf32> - (%101) = "pd_op.add" (%98, %100) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<1x32x1x1xf32>) -> tensor<1x32x27x27xf32> - (%102) = "pd_op.relu" (%101) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv/"} : (tensor<1x32x27x27xf32>) -> tensor<1x32x27x27xf32> - (%103) = "pd_op.conv2d" (%102, %35) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x1x1xf32>) -> tensor<1x128x27x27xf32> - (%104) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%105) = "pd_op.reshape" (%34, %104) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> - (%106) = "pd_op.add" (%103, %105) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> - (%107) = "pd_op.relu" (%106) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_1/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> - (%108) = "pd_op.conv2d" (%102, %33) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x3x3xf32>) -> tensor<1x128x27x27xf32> - (%109) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%110) = "pd_op.reshape" (%32, %109) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> - (%111) = "pd_op.add" (%108, %110) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> - (%112) = "pd_op.relu" (%111) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/MakeFireConv_2/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> - (%113) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_2/",value:1} : () -> tensor<1xi32> - (%114) = "builtin.combine" (%107, %112) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/"} : (tensor<1x128x27x27xf32>, tensor<1x128x27x27xf32>) -> vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>] - (%115) = "pd_op.concat" (%114, %113) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_2/"} : (vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>], tensor<1xi32>) -> tensor<1x256x27x27xf32> - (%116) = "pd_op.conv2d" (%115, %31) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<1x256x27x27xf32>, tensor<32x256x1x1xf32>) -> tensor<1x32x27x27xf32> - (%117) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%118) = "pd_op.reshape" (%30, %117) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<32xf32>, tensor<4xi64>) -> tensor<1x32x1x1xf32> - (%119) = "pd_op.add" (%116, %118) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<1x32x1x1xf32>) -> tensor<1x32x27x27xf32> - (%120) = "pd_op.relu" (%119) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv/"} : (tensor<1x32x27x27xf32>) -> tensor<1x32x27x27xf32> - (%121) = "pd_op.conv2d" (%120, %29) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x1x1xf32>) -> tensor<1x128x27x27xf32> - (%122) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%123) = "pd_op.reshape" (%28, %122) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> - (%124) = "pd_op.add" (%121, %123) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> - (%125) = "pd_op.relu" (%124) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_1/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> - (%126) = "pd_op.conv2d" (%120, %27) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<1x32x27x27xf32>, tensor<128x32x3x3xf32>) -> tensor<1x128x27x27xf32> - (%127) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%128) = "pd_op.reshape" (%26, %127) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<128xf32>, tensor<4xi64>) -> tensor<1x128x1x1xf32> - (%129) = "pd_op.add" (%126, %128) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/Conv2D/"} : (tensor<1x128x27x27xf32>, tensor<1x128x1x1xf32>) -> tensor<1x128x27x27xf32> - (%130) = "pd_op.relu" (%129) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/MakeFireConv_2/"} : (tensor<1x128x27x27xf32>) -> tensor<1x128x27x27xf32> - (%131) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_3/",value:1} : () -> tensor<1xi32> - (%132) = "builtin.combine" (%125, %130) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/"} : (tensor<1x128x27x27xf32>, tensor<1x128x27x27xf32>) -> vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>] - (%133) = "pd_op.concat" (%132, %131) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_3/"} : (vec[tensor<1x128x27x27xf32>,tensor<1x128x27x27xf32>], tensor<1xi32>) -> tensor<1x256x27x27xf32> - (%134) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MaxPool2D_2/",value:[3,3]} : () -> tensor<2xi64> - (%135) = "pd_op.pool2d" (%133, %134) {adaptive:false,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"max",stop_gradient:[false],strides:[2,2],struct_name:"/SqueezeNet/MaxPool2D_2/"} : (tensor<1x256x27x27xf32>, tensor<2xi64>) -> tensor<1x256x13x13xf32> - (%136) = "pd_op.conv2d" (%135, %25) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<48x256x1x1xf32>) -> tensor<1x48x13x13xf32> - (%137) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%138) = "pd_op.reshape" (%24, %137) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<48xf32>, tensor<4xi64>) -> tensor<1x48x1x1xf32> - (%139) = "pd_op.add" (%136, %138) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<1x48x1x1xf32>) -> tensor<1x48x13x13xf32> - (%140) = "pd_op.relu" (%139) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv/"} : (tensor<1x48x13x13xf32>) -> tensor<1x48x13x13xf32> - (%141) = "pd_op.conv2d" (%140, %23) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x1x1xf32>) -> tensor<1x192x13x13xf32> - (%142) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%143) = "pd_op.reshape" (%22, %142) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> - (%144) = "pd_op.add" (%141, %143) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> - (%145) = "pd_op.relu" (%144) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_1/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> - (%146) = "pd_op.conv2d" (%140, %21) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x3x3xf32>) -> tensor<1x192x13x13xf32> - (%147) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%148) = "pd_op.reshape" (%20, %147) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> - (%149) = "pd_op.add" (%146, %148) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> - (%150) = "pd_op.relu" (%149) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/MakeFireConv_2/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> - (%151) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_4/",value:1} : () -> tensor<1xi32> - (%152) = "builtin.combine" (%145, %150) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/"} : (tensor<1x192x13x13xf32>, tensor<1x192x13x13xf32>) -> vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>] - (%153) = "pd_op.concat" (%152, %151) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_4/"} : (vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>], tensor<1xi32>) -> tensor<1x384x13x13xf32> - (%154) = "pd_op.conv2d" (%153, %19) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<1x384x13x13xf32>, tensor<48x384x1x1xf32>) -> tensor<1x48x13x13xf32> - (%155) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%156) = "pd_op.reshape" (%18, %155) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<48xf32>, tensor<4xi64>) -> tensor<1x48x1x1xf32> - (%157) = "pd_op.add" (%154, %156) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<1x48x1x1xf32>) -> tensor<1x48x13x13xf32> - (%158) = "pd_op.relu" (%157) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv/"} : (tensor<1x48x13x13xf32>) -> tensor<1x48x13x13xf32> - (%159) = "pd_op.conv2d" (%158, %17) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x1x1xf32>) -> tensor<1x192x13x13xf32> - (%160) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%161) = "pd_op.reshape" (%16, %160) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> - (%162) = "pd_op.add" (%159, %161) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> - (%163) = "pd_op.relu" (%162) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_1/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> - (%164) = "pd_op.conv2d" (%158, %15) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<1x48x13x13xf32>, tensor<192x48x3x3xf32>) -> tensor<1x192x13x13xf32> - (%165) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%166) = "pd_op.reshape" (%14, %165) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<192xf32>, tensor<4xi64>) -> tensor<1x192x1x1xf32> - (%167) = "pd_op.add" (%164, %166) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/Conv2D/"} : (tensor<1x192x13x13xf32>, tensor<1x192x1x1xf32>) -> tensor<1x192x13x13xf32> - (%168) = "pd_op.relu" (%167) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/MakeFireConv_2/"} : (tensor<1x192x13x13xf32>) -> tensor<1x192x13x13xf32> - (%169) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_5/",value:1} : () -> tensor<1xi32> - (%170) = "builtin.combine" (%163, %168) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/"} : (tensor<1x192x13x13xf32>, tensor<1x192x13x13xf32>) -> vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>] - (%171) = "pd_op.concat" (%170, %169) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_5/"} : (vec[tensor<1x192x13x13xf32>,tensor<1x192x13x13xf32>], tensor<1xi32>) -> tensor<1x384x13x13xf32> - (%172) = "pd_op.conv2d" (%171, %13) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<1x384x13x13xf32>, tensor<64x384x1x1xf32>) -> tensor<1x64x13x13xf32> - (%173) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%174) = "pd_op.reshape" (%12, %173) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%175) = "pd_op.add" (%172, %174) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x13x13xf32> - (%176) = "pd_op.relu" (%175) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv/"} : (tensor<1x64x13x13xf32>) -> tensor<1x64x13x13xf32> - (%177) = "pd_op.conv2d" (%176, %11) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x1x1xf32>) -> tensor<1x256x13x13xf32> - (%178) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%179) = "pd_op.reshape" (%10, %178) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> - (%180) = "pd_op.add" (%177, %179) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> - (%181) = "pd_op.relu" (%180) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_1/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> - (%182) = "pd_op.conv2d" (%176, %9) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x3x3xf32>) -> tensor<1x256x13x13xf32> - (%183) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%184) = "pd_op.reshape" (%8, %183) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> - (%185) = "pd_op.add" (%182, %184) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> - (%186) = "pd_op.relu" (%185) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/MakeFireConv_2/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> - (%187) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_6/",value:1} : () -> tensor<1xi32> - (%188) = "builtin.combine" (%181, %186) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/"} : (tensor<1x256x13x13xf32>, tensor<1x256x13x13xf32>) -> vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>] - (%189) = "pd_op.concat" (%188, %187) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_6/"} : (vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>], tensor<1xi32>) -> tensor<1x512x13x13xf32> - (%190) = "pd_op.conv2d" (%189, %7) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<1x512x13x13xf32>, tensor<64x512x1x1xf32>) -> tensor<1x64x13x13xf32> - (%191) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%192) = "pd_op.reshape" (%6, %191) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<64xf32>, tensor<4xi64>) -> tensor<1x64x1x1xf32> - (%193) = "pd_op.add" (%190, %192) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x13x13xf32> - (%194) = "pd_op.relu" (%193) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv/"} : (tensor<1x64x13x13xf32>) -> tensor<1x64x13x13xf32> - (%195) = "pd_op.conv2d" (%194, %5) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x1x1xf32>) -> tensor<1x256x13x13xf32> - (%196) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%197) = "pd_op.reshape" (%4, %196) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> - (%198) = "pd_op.add" (%195, %197) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> - (%199) = "pd_op.relu" (%198) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_1/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> - (%200) = "pd_op.conv2d" (%194, %3) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[1,1],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<1x64x13x13xf32>, tensor<256x64x3x3xf32>) -> tensor<1x256x13x13xf32> - (%201) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%202) = "pd_op.reshape" (%2, %201) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<256xf32>, tensor<4xi64>) -> tensor<1x256x1x1xf32> - (%203) = "pd_op.add" (%200, %202) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/Conv2D/"} : (tensor<1x256x13x13xf32>, tensor<1x256x1x1xf32>) -> tensor<1x256x13x13xf32> - (%204) = "pd_op.relu" (%203) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/MakeFireConv_2/"} : (tensor<1x256x13x13xf32>) -> tensor<1x256x13x13xf32> - (%205) = "pd_op.full" () {dtype:int32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/MakeFire_7/",value:1} : () -> tensor<1xi32> - (%206) = "builtin.combine" (%199, %204) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/"} : (tensor<1x256x13x13xf32>, tensor<1x256x13x13xf32>) -> vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>] - (%207) = "pd_op.concat" (%206, %205) {stop_gradient:[false],struct_name:"/SqueezeNet/MakeFire_7/"} : (vec[tensor<1x256x13x13xf32>,tensor<1x256x13x13xf32>], tensor<1xi32>) -> tensor<1x512x13x13xf32> - (%208) = "pd_op.full" () {dtype:float32,place:Place(cpu),shape:[1],stop_gradient:[true],struct_name:"/SqueezeNet/Dropout/",value:0.5} : () -> tensor<1xf32> - (%209, %210) = "pd_op.dropout" (%207, <>, %208) {fix_seed:false,is_test:false,mode:"downgrade_in_infer",seed:0,stop_gradient:[false,false],struct_name:"/SqueezeNet/Dropout/"} : (tensor<1x512x13x13xf32>, <>, tensor<1xf32>) -> tensor<1x512x13x13xf32>, tensor<1x512x13x13xu8> - (%211) = "pd_op.conv2d" (%209, %1) {data_format:"NCHW",dilations:[1,1],groups:1,padding_algorithm:"EXPLICIT",paddings:[0,0],stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1x512x13x13xf32>, tensor<1000x512x1x1xf32>) -> tensor<1x1000x13x13xf32> - (%212) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/Conv2D_1/",value:[1,-1,1,1]} : () -> tensor<4xi64> - (%213) = "pd_op.reshape" (%0, %212) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1000xf32>, tensor<4xi64>) -> tensor<1x1000x1x1xf32> - (%214) = "pd_op.add" (%211, %213) {stop_gradient:[false],struct_name:"/SqueezeNet/Conv2D_1/"} : (tensor<1x1000x13x13xf32>, tensor<1x1000x1x1xf32>) -> tensor<1x1000x13x13xf32> - (%215) = "pd_op.relu" (%214) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x1000x13x13xf32>) -> tensor<1x1000x13x13xf32> - (%216) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/AdaptiveAvgPool2D/",value:[1,1]} : () -> tensor<2xi64> - (%217) = "pd_op.pool2d" (%215, %216) {adaptive:true,ceil_mode:false,data_format:"NCHW",exclusive:true,global_pooling:false,padding_algorithm:"EXPLICIT",paddings:[0,0],pooling_type:"avg",stop_gradient:[false],strides:[1,1],struct_name:"/SqueezeNet/AdaptiveAvgPool2D/"} : (tensor<1x1000x13x13xf32>, tensor<2xi64>) -> tensor<1x1000x1x1xf32> - (%218) = "pd_op.full_int_array" () {dtype:int64,place:Place(cpu),stop_gradient:[true],struct_name:"/SqueezeNet/",value:[2,3]} : () -> tensor<2xi64> - (%219) = "pd_op.squeeze" (%217, %218) {stop_gradient:[false],struct_name:"/SqueezeNet/"} : (tensor<1x1000x1x1xf32>, tensor<2xi64>) -> tensor<1x1000xf32> -} diff --git a/my_samples/squeezenet1_1/weight_meta.py b/my_samples/squeezenet1_1/weight_meta.py deleted file mode 100644 index 714737364..000000000 --- a/my_samples/squeezenet1_1/weight_meta.py +++ /dev/null @@ -1,3 +0,0 @@ -import paddle - -weight_meta = {} From 5a07f615753a785a216771106582987c08670ab3 Mon Sep 17 00:00:00 2001 From: Carousel126 Date: Mon, 9 Feb 2026 16:25:43 +0800 Subject: [PATCH 3/3] style: format code with black --- .../vision/squeezenet1_1/input_meta.py | 7 +-- .../vision/squeezenet1_1/model.py | 6 +-- graph_net/tests/test_squeezenet1_1_extract.py | 43 ++++++++----------- 3 files changed, 22 insertions(+), 34 deletions(-) diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py b/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py index 6453218cf..e8938c68a 100644 --- a/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/input_meta.py @@ -1,8 +1,3 @@ import paddle -input_meta = { - "inputs": { - "shape": [1, 3, 224, 224], - "dtype": "float32" - } -} +input_meta = {"inputs": {"shape": [1, 3, 224, 224], "dtype": "float32"}} diff --git a/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py b/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py index a149a8966..e719a3bd4 100644 --- a/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py +++ b/graph_net/samples/paddle_samples/vision/squeezenet1_1/model.py @@ -1,13 +1,11 @@ import paddle from paddle.vision.models import squeezenet1_1 + class GraphModule(paddle.nn.Layer): def __init__(self): super(GraphModule, self).__init__() self.model = squeezenet1_1(pretrained=False) - def forward(self, inputs=None): - # 防御性修复：如果 inputs 为空，生成一个符合规格的 dummy tensor - if inputs is None: - inputs = paddle.randn([1, 3, 224, 224]) + def forward(self, inputs): return self.model(inputs) diff --git a/graph_net/tests/test_squeezenet1_1_extract.py b/graph_net/tests/test_squeezenet1_1_extract.py index fc29fc27f..34aa5af57 100644 --- a/graph_net/tests/test_squeezenet1_1_extract.py +++ b/graph_net/tests/test_squeezenet1_1_extract.py @@ -1,33 +1,28 @@ import paddle import os from paddle.vision.models import squeezenet1_1 -from graph_net.paddle.extractor import GraphExtractor -# 1. 环境准备 -os.environ["GRAPH_NET_EXTRACT_WORKSPACE"] = os.path.abspath("./my_samples") -if not os.path.exists("./my_samples"): - os.makedirs("./my_samples") -# 2. 准备模型 -model = squeezenet1_1(pretrained=False) -model.eval() +def extract_squeezenet(): + # 环境准备 + os.environ["FLAGS_enable_pir_api"] = "1" + paddle.enable_static() -# 3. 定义 InputSpec (关键改动：name='inputs') -input_spec = [paddle.static.InputSpec(shape=[1, 3, 224, 224], dtype='float32', name='inputs')] + # 捕获计算图 + main_program = paddle.static.Program() + with paddle.static.program_guard(main_program): + inputs = paddle.static.data( + name="inputs", shape=[1, 3, 224, 224], dtype="float32" + ) + model = squeezenet1_1(pretrained=False) + out = model(inputs) -# 4. 手动实例化提取器 -print("正在初始化提取器...") -extractor = GraphExtractor(model, name="squeezenet1_1", dynamic=False, input_spec=input_spec) + # 保存结果 + save_path = "squeezenet1_1.json" + with open(save_path, "w") as f: + f.write(str(main_program)) + print(f"提取完成，文件已生成至: {save_path}") -# 5. 执行提取 (关键改动：Key 名改为 'inputs') -print("正在执行提取流程...") -model_dump_path = os.path.join(os.environ["GRAPH_NET_EXTRACT_WORKSPACE"], "squeezenet1_1") -dummy_data = {"inputs": paddle.randn([1, 3, 224, 224])} -try: - # 绕过 __call__ 直接运行内部 dump 逻辑 - extractor.run_model_with_dump_enabled(model_dump_path, **dummy_data) - print("\n✨✨✨ 奇迹发生了！提取流程成功完成！ ✨✨✨") - print(f"产物已存至: {model_dump_path}") -except Exception as e: - print(f"\n❌ 捕获到错误: {e}") \ No newline at end of file +if __name__ == "__main__": + extract_squeezenet()