search

Paranioar / SGRAF

[AAAI2021] The code of “Similarity Reasoning and Filtration for Image-Text Matching”
212 stars 36 forks source link

bug #8

Closed KevinLight831 closed 3 years ago

KevinLight831 commented 3 years ago

data.py line44: image = torch.Tensor(self.images[img_id]) --> image = torch.Tensor(self.images[int(img_id)])

KevinLight831 commented 3 years ago

model.py line122:cap_emb = (cap_emb[:, :, :cap_emb.size(2)/2] + cap_emb[:, :, cap_emb.size(2)/2:])/2 -->cap_emb = (cap_emb[:,:,:cap_emb.size(2)//2] + cap_emb[:,:,cap_emb.size(2)//2:])/2

Paranioar commented 3 years ago

model.py line122:cap_emb = (cap_emb[:, :, :cap_emb.size(2)/2] + cap_emb[:, :, cap_emb.size(2)/2:])/2 -->cap_emb = (cap_emb[:,:,:cap_emb.size(2)//2] + cap_emb[:,:,cap_emb.size(2)//2:])/2

image

KevinLight831 commented 3 years ago

sorry, i was too rude and careless.

Paranioar commented 3 years ago

sorry, i was too rude and careless.

Never mind. Actually, most recent SOTAs are based on python2.7 and achieve lower performance with the same codes on python3. It is really meaningful to find out the difference, such as '/' or '//'，'open(path,'rb')' or 'open(path,'r')', ‘iteritems()’, etc. I’m looking forward to your reproduction with python3. where you may debug and adjust some details.

Paranioar commented 3 years ago

pytorch_python版本.pdf Thanks for the exploration and discussion from KevinLight831. We will get an insight into this significant issue and create a new branch for python3.6. Other issues are also instructive, VSRN, VSE++, etc.

zt-1026 commented 2 years ago

sorry, i was too rude and careless.

Hello, what's the results of your final experiment with Python 3 and whether it has been reduced，Looking forward to your reply.