Andrej
Hi thank you for reporting this. The exp should be guarded, you're right, but despite the warning I don't believe this should be too big of an issue. I will...
oops, i think you're probably right. If you try to meddle with the sizes it would error.
Yep! I took them out and didn't see a huge decrease in performance, so I decided to get rid of it for NeuralTalk for simplicity.
With default parameters? I Thoguht I tuned them so that this doesn't happen, sorry about that. As the message suggests, lowering the learning rate does it. Set learning_rate to be...
@StevenLOL Nice! Looking at the Model Zoo, http://cs.stanford.edu/people/karpathy/neuraltalk/ my LSTM model achieves perplexity of about 15.7 (which is slightly better). I ran it for longer and cross-validated it on our...
doh! Sorry, can you elaborate? I was under the impression that `logprobs` was already normalized due to use of a LogSoftMax layer, so this should already be a correctly-normalized log-perplexity?...
When I create the things I end up sorting, I create them on L218 as ``` table.insert(candidates, {c=ix[{ q,c }], q=q, p=candidate_logprob, r=local_logprob }) ``` so in fact the `.p`...
Wow, that's painful! I didn't realize or encounter this, because preprocessing is done in Python. I'm using torch load for some files (e.g. mscoco test set files), but there it...
I won't take this because 1) it is minor, and 2) it might confuse people who watched the video. upside has to be worth that downside. ty though!