C3M3_Assignment - Decoder definition error

Course Q&A Professional Certificate:PyTorch for Deep Learning PyTorch: Advanced Architectures and Deployment
,
1

I have problem with decoder and task quite big to identify where is the error. Do you have any idea where do dig?

======================================================================
 Decoder - Main Layers
======================================================================

Layer                          Type                           Parameters
----------------------------------------------------------------------
token_emb                      Embedding                       1,280,000
pos_enc                        PositionalEncoding                      0
dropout                        Dropout                                 0
transformer_decoder            TransformerDecoder              2,372,352
output_projection              Linear                          1,285,000
----------------------------------------------------------------------
TOTAL                                                          4,937,352
======================================================================

Expected Output

======================================================================
 Decoder - Main Layers
======================================================================

Layer                          Type                           Parameters
----------------------------------------------------------------------
token_emb                      Embedding                       1,280,000
pos_enc                        PositionalEncoding                      0
dropout                        Dropout                                 0
transformer_decoder            TransformerDecoder              2,372,352
output_projection              Linear                          1,285,000
----------------------------------------------------------------------
TOTAL                                                          4,937,352
======================================================================

unittests.exercise_2(Decoder)

Failed test case: Failed forward pass test.
Expected: None
Got: shape '[1, 1, 8, 8]' is invalid for input of size 16
2

I think it has to do with you definition of self.token_emb. Try to follow the hints as closely as possible.

3

I think I skip Hint 7 but I cannot find any reference in lab template where it must be implemented.

    # Convert into token embeddings

I put new code here but error still exists.

Hint 7: Forward Pass - Embedding with Scaling
What you need: Embed tokens and scale by sqrt(d_model).

4

Hey, even I am struggling with this error. Did you find a solution for this error?

Failed test case: Failed forward pass test.
Expected: None
Got: shape ‘[1, 1, 1, 8]’ is invalid for input of size 16

5

Hi everyone!
Could you DM me your solutions? Several learners are encountering this error, and since our solution passes the unit tests, the issue may be with the tests themselves.

6

Hey Lucas,

I tried reading through the hints again. In forward pass adding positional encoding, I forgot to add to the embedded tensor after getting the positional encodings. The code works now, test passed!

Thanks

1 Like