Seq2seq: Teacher-Forced Decoder Inputs
Beginner
The problem, from first principles
During training the decoder is fed the correct previous word, not its own guess. This is teacher forcing. The inputs are the targets shifted right by one position, with a begin-of-sentence token at the front.
From theory to code
Implement decoder_inputs(targets, bos), which returns [bos] + targets[:-1].
Constraints
- The last target token is never an input.
Hints
Hint 1
Concatenate the start token with all targets except the final one.
Keywords
research-paperstransformersllmseq2seq
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.