chenxinpeng / arnet Goto Github PK

It's a good work for image caption. But I got the problem when I try to run the code.
I can not find the variable 'drop_prob_lm' in the opts.py and '.sh' files. So, I do not know the value of it.
My environment is Pytorch 0.3.1 with python 2.7.
Please help. Thanks a lot!

Hi, I have read your masterpiece recently. Because I am doing adaptive attention with my model, I want to mix adaptive attention model with ARNet together. After the first stage of training, because the adaptive attention model uses hidden state with attention to form a new hidden, where should I put ARNet. Only on LSTM or I need to count the attention together. If you have any suggestions, please teach me. Thank you.

It's a good work for image caption. But I got the problem when I try to run the code.
I can not find the variable 't_SNE_model_path' in the opts.py and '.sh' files. So, I do not know the value of it.
Please help. Thanks a lot!

when I run the image_caption_ende_rcst_lstm.py ，there are some errors。
KeyError: 'missing keys in state_dict: "{'rcst_lstm.h2h.bias', 'rcst_lstm.h2h.weight', 'h_2_pre_h.weight', 'rcst_lstm.i2h.bias', 'h_2_pre_h.bias', 'rcst_lstm.i2h.weight'}"'

When I try to run this project, after a few steps, I meet this error at line:
output = torch.sum(output) / batch_size in the 'image_caption_ende_xe.py'
the output is Variable of 'torch.cuda.FloatTensor' and the type of 'batch_size' is 'float'. But I do not know where is the problem.
My environment: Pytorch-0.3.1 with python2.7.
Please Help! Thanks a lot!

Hi! Thank you for your sharing your code and a comprehensive tutorial to use it. However, after I excuted the command "./bash_image_caption_soft_att_xe.sh", some weird things happened as follows:
"
idx: 13024 epoch: 0 lr:0.00050000 loss: 43.086 time: 0.239
idx: 13040 epoch: 0 lr:0.00050000 loss: 41.745 time: 0.287

/home/xxx/extend/ARNet/image_captioning/utils_model.py(162)forward()
161
--> 162 output = - input.gather(1, target) * mask
163 output = torch.sum(output) / batch_size

ipdb> list
157
158 target_cpu = target.data.cpu().numpy()
159 if 10516 in target_cpu:
160 ipdb.set_trace()
161
--> 162 output = - input.gather(1, target) * mask
163 output = torch.sum(output) / batch_size
164
165 return output
166
167
"
What's wrong with it ? Thank you for your answering.

how to get the hidden states of the sentences？

Sorry for asking about im2p question here. Because I didn't find way to establish an issue in your im2p project.
I'm confused how to evaluate the paragraph generated by im2p.
Should I regard the whole paragraph (multi-sentences) as a large sentence and regard the ground truth as a sentence, either? Then put them into bleu, cider (and so on) to evaluate?
Or should I change the code of bleu.py and cider.py to evaluate the paragraphs by one sentence (generated) matching one sentence (ground truth)?
Hope you can help me with this! Thank you!