fix test functions of models; just semantic fix

LabsSolutions/02-pytorch-asr/models.py

@@ -418,7 +418,7 @@ def ex_ctc():
     # The size of our vocabulary (including the blank character)
     vocab_size = 44
     # The class id for the blank token
-    blank_id = 43
+    blank_id = 0
 
     max_spectro_length = 50
     min_transcript_length = 10
@@ -427,7 +427,7 @@ def ex_ctc():
     # Compute a dummy vector of probabilities over the vocabulary (including the blank)
     # log_probs is here batch_first, i.e. (Batch, Tx, vocab_size)
     log_probs = torch.randn(batch_size, max_spectro_length, vocab_size).log_softmax(
-        dim=1
+        dim=2
     )
     spectro_lengths = torch.randint(
         low=max_transcript_length, high=max_spectro_length, size=(batch_size,)
@@ -437,15 +437,15 @@ def ex_ctc():
     # targets is here (batch_size, Ty)
     targets = torch.randint(
         low=0,
-        high=vocab_size + 1,  # include the blank character
+        high=vocab_size,  # include the blank character
         size=(batch_size, max_transcript_length),
     )
     target_lengths = torch.randint(
         low=min_transcript_length, high=max_transcript_length, size=(batch_size,)
     )
     loss = torch.nn.CTCLoss(blank=blank_id)
 
-    # The log_probs must be given as (Tx, Batch, vocab_size)
+    # The log_probs must be given as (Tx, Batch, vocab_size), hence we transpose it
     vloss = loss(log_probs.transpose(0, 1), targets, spectro_lengths, target_lengths)
 
     print(f"Our dummy loss equals {vloss}")
@@ -485,7 +485,6 @@ def ex_pack():
 
 
 if __name__ == "__main__":
-    # ex_ctc()
-    # ex_pack()
-    test_model()
+    ex_ctc()
+    ex_pack()
 # SOL@