search

poojahira / image-captioning-bottom-up-top-down

PyTorch implementation of Image captioning with Bottom-up, Top-down Attention
164 stars 39 forks source link

question about discriminatiive supervision #8

Closed homelifes closed 5 years ago

homelifes commented 5 years ago

Hi and Thanks for your code, I would like to ask. You are obtaining the max of the tensor predictions1 = torch.zeros(batch_size, max(decode_lengths), vocab_size) along the timesteps and not along the words (you are not predicting the max word from the softmax output as in the predictionstensor). You are doing: scores_d = scores_d.max(1)[0]. So you are getting the max across timesteps for each vocabulary word. Is that supposed to be correct? THanks!

poojahira commented 5 years ago

Hi, that's the way I understood discriminative supervision at the time. You could implement it differently as well.