I’m not very sure that what is the parameter “tgt” means.
The document says it is the sequence of decoder. However, I have a question that since I have the sequence to decoder, thus I have the result of decoder, why do I have to use transformer for inferencing?In another word, in inferencing process, we have no corresponded tgt to input.
For example, I have a seq2seq task that translate “I love you” from English to Franch “Je t’aime”, what is the parameter of “tgt”?
Looking forward answering. Thanks!
tgt is the target input, in the training you need tgt input as golden truth to do teacher forcing learning, but in the inference, you don’t need tgt, you only need to input the encoded_vector from encoder and the first token in the tgt, usually it’s BOS.