Skip to content

a question in BERT pytorch scratch notebook #2

@kindlytree

Description

@kindlytree

hi, a question about the loss calculation in BERT notebook:

https://github.com/ChanCheeKean/DataScience/blob/main/13%20-%20NLP/C04%20-%20BERT%20(Pytorch%20Scratch).ipynb

2-1. NLL(negative log likelihood) loss of is_next classification result

next_loss = self.criterion(next_sent_output, data["is_next"])

2-2. NLLLoss of predicting masked token word

transpose to (m, vocab_size, seq_len) vs (m, seq_len)

criterion(mask_lm_output.view(-1, mask_lm_output.size(-1)), data["bert_label"].view(-1))

mask_loss = self.criterion(mask_lm_output.transpose(1, 2), data["bert_label"])

but the loss is defined as self.criterion = torch.nn.NLLLoss(ignore_index=0)
when the gt of the 'is_next' = 0, the gradient will not be backpropagted??

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions