Stack/nets_train.py at master · Delisseu/Stack · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
import gc
import json
import pickle
import sys

from neuralnet.Layers_Features import BatchNorm, Pooling, XavierUniform
from neuralnet.Optimizers import Adam, InverseSqrtSchedulerMod1
from neuralnet.core_gpu import *

var_1 = [
    {"input_dim": 300, 'neurons': 128, "learn_params": {"lr": 0.0001}}, {"layer": Relu},
    {'neurons': 64, "learn_params": {"lr": 0.0001}}, {"layer": Relu},
    {'neurons': 1, "learn_params": {"lr": 0.0001}, "init_dict": {"init_cls": XavierUniform}}]

var_2 = [
    {"input_dim": (75, 90, 1), "out_channels": 4, "layer": Conv2D, "learn_params": {"lr": 0.0001}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0001}}, {"layer": Relu}, {"layer": Pooling},

    {"out_channels": 8, "layer": Conv2D, "learn_params": {"lr": 0.0001}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0001}}, {"layer": Relu}, {"layer": Pooling},

    {'neurons': 512, "learn_params": {"lr": 0.0001}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0001}}, {"layer": Relu},
    {'neurons': 64, "learn_params": {"lr": 0.0001}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0001}}, {"layer": Relu},
    {'neurons': 8, "learn_params": {"lr": 0.0001}}, {"layer": Relu},
    {'neurons': 1, "learn_params": {"lr": 0.0001}, "init_dict": {"init_cls": XavierUniform}}]

var_3 = [
    {"input_dim": (100, 100, 1), "out_channels": 8, "layer": Conv2D, "learn_params": {"lr": 0.0001}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0001}}, {"layer": Relu}, {"layer": Pooling},

    {"out_channels": 16, "layer": Conv2D, "learn_params": {"lr": 0.0001}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0001}}, {"layer": Relu},

    {"out_channels": 1, "layer": Conv2D, "learn_params": {"lr": 0.0005}, "kernel_size": (1, 1), "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0005}}, {"layer": Relu},

    {'neurons': 512, "learn_params": {"lr": 0.0005}, "bias": False},
    {"layer": BatchNorm, "learn_params": {"lr": 0.0005}}, {"layer": Relu},
    {'neurons': 64, "bias": False}, {"layer": BatchNorm, "learn_params": {"lr": 0.0005}}, {"layer": Relu},
    {'neurons': 8, "learn_params": {"lr": 0.0005}}, {"layer": Relu},
    {'neurons': 1, "learn_params": {"lr": 0.0001}, "init_dict": {"init_cls": XavierUniform}}]  # 0.000643


def get_metrics(y_pred, y_true):
    y_true = np.array(y_true)
    y_pred = np.array(y_pred)

    tp = np.sum((y_pred == 1) & (y_true == 1))
    tn = np.sum((y_pred == 0) & (y_true == 0))
    fp = np.sum((y_pred == 1) & (y_true == 0))
    fn = np.sum((y_pred == 0) & (y_true == 1))

    precision = tp / (tp + fp) if (tp + fp) else 0
    recall = tp / (tp + fn) if (tp + fn) else 0
    f1 = 2 * precision * recall / (precision + recall) if (precision + recall) else 0
    accuracy = (tp + tn) / (tp + tn + fp + fn)

    return {
        'precision': precision,
        'recall': recall,
        'f1': f1,
        'accuracy': accuracy
    }


def train_and_save(model, loader, name, test_loader=None, epochs=50, early_stop=0, folder="Models",
                   min_delta=0.001):
    var = model.train(loader, epochs=epochs, early_stop=early_stop, test_loader=test_loader, min_delta=min_delta)
    if var is None:
        var = model.export()
    with open(f"{folder}/{name}.pkl", "wb") as f:
        pickle.dump(var, f)

    return var


def best_threshold_Logits(model, X, y, metric="f1"):
    best_val = {metric: float("-inf")}
    best_thr = None
    result = Sigmoid().forward(model.predict(AsyncCupyDataLoader(X, shuffle=False, batch_size=512), numpy=False)).get()
    for threshold in range(1, 100):
        threshold /= 100
        val = get_metrics(result > threshold, y)

        if val[metric] >= best_val[metric]:
            best_val = val
            best_thr = threshold

    return best_thr, best_val


def append_sigmoid(name):
    with open(f"Models/{name}.pkl", "rb") as f:
        var = pickle.load(f)

    var.append({"layer": Sigmoid})
    with open(f"Models/{name}.pkl", "wb") as f:
        pickle.dump(var, f)


if __name__ == "__main__":
    if "final" in sys.argv:

        X_base = np.load(rf"Data\X_nodup.npy").astype(np.float32) / np.asarray(255, dtype=np.float32)
        y_base = np.load(rf"Data\y_nodup.npy")
        final_agent = NeuralNetwork(var_3, BCELogits(), Adam(scheduler=InverseSqrtSchedulerMod1(warmup_steps=3500)))
        train_and_save(final_agent, AsyncCupyDataLoader(X_base, y_base, batch_size=128), "final_agent",
                       epochs=20)
        append_sigmoid("final_agent")
        thr_final_agent, final_agent_metrics = best_threshold_Logits(final_agent, X_base, y_base)

        print(f"{thr_final_agent=}, {final_agent_metrics=}")

        with open(rf"Models\nets_thresholds.json", "r") as f:
            thresholds = json.load(f)

        with open(rf"Models\nets_thresholds.json", "w") as f:
            thresholds |= {"final_agent": thr_final_agent}
            json.dump(thresholds, f)

        del final_agent
        gc.collect()  # очищает память CPU, а CuPy освобождает привязанную GPU-память
        cp._default_memory_pool.free_all_blocks()
    else:
        X_base = np.load(rf"Data\pixels_all_aug.npy").astype(np.float32) / np.asarray(255, dtype=np.float32)
        y_base = np.load(rf"Data\target_all_aug_cls.npy")

        X_game_over = np.load(r"Data\game_over_data_aug.npy").astype(np.float32) / np.asarray(255, dtype=np.float32)
        y_game_over = np.load(r"Data\game_over_target_aug.npy")

        X_success = np.load(r"Data\success_collect.npy").astype(np.float32) / np.asarray(255, dtype=np.float32)
        y_success = np.load(r"Data\success_collect_target.npy")

        online_agent = NeuralNetwork(var_3, BCELogits(), Adam())
        game_over_net = NeuralNetwork(var_1, BCELogits(), Adam())
        success_net = NeuralNetwork(var_2, BCELogits(), Adam())

        train_and_save(online_agent, AsyncCupyDataLoader(X_base, y_base, batch_size=128), "online_agent", epochs=1)

        train_and_save(game_over_net, AsyncCupyDataLoader(X_game_over, y_game_over, batch_size=64), "game_over_net",
                       epochs=15)

        train_and_save(success_net, AsyncCupyDataLoader(X_success, y_success, batch_size=128), "success_net",
                       epochs=5)

        append_sigmoid("online_agent")
        append_sigmoid("game_over_net")
        append_sigmoid("success_net")

        thr_online_agent, online_agent_metrics = best_threshold_Logits(online_agent, X_base,
                                                                       (y_base > 0.5).astype(np.int32))
        thr_game_over_net, game_over_net_metrics = best_threshold_Logits(game_over_net, X_game_over, y_game_over)
        thr_success_net, success_net_metrics = best_threshold_Logits(success_net, X_success, y_success)

        print(f"{thr_online_agent=}, {online_agent_metrics=}\n"
              f"{thr_game_over_net=}, {game_over_net_metrics=}\n"
              f"{thr_success_net=}, {success_net_metrics=}")

        with open(rf"Models\nets_thresholds.json", "w") as f:
            json.dump(
                {"online_agent": thr_online_agent,
                 "game_over_net": thr_game_over_net,
                 "success_net": thr_success_net}, f)

        del game_over_net, success_net, online_agent
        gc.collect()  # очищает память CPU, а CuPy освобождает привязанную GPU-память
        cp._default_memory_pool.free_all_blocks()