Soroush Omranpour
Soroush Omranpour
Hi, as I have commented in the code log_probs dimensions are (batch_size, num_classes, output_len) and the "y" tensor's dimensions are also (batch_size, target_len). output_len and target_len indicate the time steps...
Hi @hieuhv94 , Can you print the whole shape of x and log_probs?
So, the "log_probs" which are your model outputs, have lengths lower than your "y" which is the ground truth. This problem is due to the value of the strides used...
Hi, well I didn’t have enough GPUs resources to fully train my model so no I don’t have reliable results on LibriSpeech. But I tried to apply the main papers...