Please Help, issue with WER: 100% During Training

[Onestringlab]

I'm using your code for training on the PHOENIX dataset, but I'm encountering an issue where the Word Error Rate (WER) is consistently 100% during evaluation. Here’s a snippet of my training log:

[ Tue Sep  3 02:31:07 2024 ]    Mean training loss: 9.5473120947.
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 135/135 [02:34<00:00,  1.14s/it]
/share/work/latih15/VAC_CSLR_PHOENIX
preprocess.sh ./work_dir/baseline_res18/output-hypothesis-dev-conv.ctm ./work_dir/baseline_res18/tmp.ctm ./work_dir/baseline_res18/tmp2.ctm
Tue Sep 3 02:33:42 AM UTC 2024
Preprocess Finished.
sh: 1: ./software/sclite: not found
/bin/sh: 1: ./software/sclite: not found
Unexpected error: <class 'IndexError'>
[ Tue Sep  3 02:33:42 2024 ] Epoch 39, dev  100.00%
[ Tue Sep  3 02:33:42 2024 ] Dev WER: 100.00%

I noticed the errors related to sclite not being found, and I'm wondering if this might be causing the issue. I followed the configuration in baseline.yaml, but I'm not sure if I've missed anything. Below is my baseline.yaml configuration:

feeder: dataset.dataloader_video.BaseFeeder
phase: train
dataset: phoenix14
# dataset: phoenix14-si5
num_epoch: 40
work_dir: ./work_dir/baseline_res18/
batch_size: 2
random_seed: 0
test_batch_size: 4 #8
num_worker: 8
device: 0,1
log_interval: 50
eval_interval: 1
save_interval: 5
# python in default
evaluate_tool: sclite
loss_weights:
  SeqCTC: 1.0
  # VAC
  # ConvCTC: 1.0
  # Dist: 10.0
#load_weights: ''

optimizer_args:
  optimizer: Adam
  base_lr: 0.0001
  step: [ 20, 35]
  learning_ratio: 1
  weight_decay: 0.0001
  start_epoch: 0
  nesterov: False

feeder_args:
  mode: 'train'
  datatype: 'video'
  num_gloss: -1
  drop_ratio: 1.0

model: slr_network.SLRModel
decode_mode: beam
model_args:
  num_classes: 1296
  c2d_type: resnet18
  conv_type: 2
  use_bn: 1
  # SMKD
  share_classifier: False
  weight_norm: False

Could you advise on what might be going wrong? Do I need to adjust the sclite path or any other settings? Any help would be greatly appreciated.

Thank you!

[lzzzzzzzd]

This means the system cannot find the sclite executable file, or it lacks permission to run,you can change the evaluate_tool to python in config or make sure sclite is executable.