The training and inference results are inconsistent.

[yujian159]

Hello,

Thank you very much for your article. I followed your tutorial to download the LRS2 dataset and processed it using preprocess_lrs2.py. I attempted to use the command python train.py --conf-dir config/lrs2_RTFSNet_4_layer.yaml for training, keeping everything at default except not loading the pre-trained weights (frcnn_128_512.backbone.pth.tar).

Upon completing the training, the model has converged, and both the training and validation loss are decreasing. The SI-SDR on the validation set is 12.82 (which is probably due to not using pre-trained weights). However, when I use the trained weights for validation, use python test.py --conf-dir config/lrs2_RTFSNet_4_layer.yaml , the results seem a bit strange:
si-snri mean:-6.394 std:19.8
sdi_i mean: 4.434 std:8.85
pesq mean:1.654 std:1.31
stoi mean:0.5188 std:0.405
si-snr mean:-6.395 std:19.4
sdr mean:4.434 std:8.96
testloss 6.3946
The results obtained using train.py and test.py differ too much. Could you please tell me what is causing this? Is such a result normal?