I am using the nn.CTCLoss(zero_infinity=True) loss function on CRNN model. The output from training the model after a few epochs on the validation set is shown below.
The batch size used is 8 and max_length for each sequence is 16. All the different outputs can be better viewed in this colab notebook: Link to colab notebook
The first line contains input_lengths and target_lengths. The second line contains shape of log_probs and targets. From the third line onwards, first string is the originally predicted output, second string is the processed predicted output and third string is the ground truth.
tensor([16, 16, 16, 16, 16, 16, 16, 16], dtype=torch.int32) tensor([5, 4, 1, 5, 5, 4, 5, 1])
torch.Size([16, 8, 41]) torch.Size([8, 16])
eeeeeeeeeeeee222 e2 22.44___________
eeeeeeeeeeeee222 e2 8.70____________
eeeeeeeeeeeee222 e2 0_______________
eeeeeeeeeeeee222 e2 12.90___________
eeeeeeeeeeeee222 e2 15.80___________
eeeeeeeeeeeee222 e2 2.80____________
eeeeeeeeeeeee222 e2 11.50___________
eeeeeeeeeeeee222 e2 4_______________
The problem can be either the arguments passed to the CTC loss function are wrong or something else entirely.