ramakanth-pasunuru / video_captioning_rl Goto Github PK

Hi. May I ask if you forget to upload the file named automatic_evaluation ? Thx.

Hello, thank you very much for your contribution. I tried to reproduce your code on the MSVD data set, but I failed to try for a long time. Therefore, I would like to ask you what code did you use when extracting features, and how did you merge the two features into 4086 dimensions in the end?

Excuse me, how do you extract the ResNeXt-101 motion features?

Hi, I tried to use InceptionV4 as CNN encoder to train the Baseline-XE model, but I cannot get the results like your paper. When finishing trarining, I got three .pth, i.e., epoch1, epoch2, and epoch3. I didn't konw which is more better, so I tested them respectively. I found epoch3.pth is the best among three. But the results are still lower than your paper.
I got the results as follows:
2020-04-22 18:00:49,045:INFO::Results:
2020-04-22 18:00:49,046:INFO::BLEU-4:0.3745473407866488
2020-04-22 18:00:49,046:INFO::METEOR:0.26116567752846037
2020-04-22 18:00:49,046:INFO::ROUGE_L:0.5878605132246427
2020-04-22 18:00:49,046:INFO::CIDEr:0.4295072766573666
2020-04-22 18:00:49,046:INFO::AVG:0.4132702020492796,
which are much lower than your paper, B4 38.6, M 27.7, R-L 59.5, and C 44.6.

I trained the model with python3.6.10, pytorch0.3.1, and TitanXP gpu.

Thank you for your excellent work. Can you provide the code to train and test on MSVD?

excellent work. However, if I want to use the in a new dataset, how can I extract the resnet feature? Can you give me some hints? Thank you very much.

Hello, Sir:

     When I employ the pretrained baseline-XE model using 'python main.py --mode test --load_path "path_to_model_ending_with_*.pth" --beam_size 5 ', there are losts of [UNK] and the result is too low.

    However, the results employed by pretrained model like 'CIDEr-RL' and 'CIDEnt-RL' is good.

   Are there any difference of command between baseline-XE and CIDEr-RL? Or is there something wrong with the baseline-XE model? 

can you check your results, thank you

Thank you for your fantastic work, I wander whether you can provide the data processing code.

I cannot download the ResNet-152 frame-level features + ResNeXt-101 motion features for MSR-VTT videos data from the Google website. Can you share it with me in any other way?

Hello, I cloned your code and found a missing entailment classifier model during runtime. I tried to find a suitable entailment classifier model on the Internet but failed to find a suitable model. So I hope you can share your entailment classifier model to me, thank you very much.

There is an error in your code.

video_captioning_rl/models/seq2seq_atten.py

frame_packed = nn.utils.rnn.pack_padded_sequence(frames, flengths, batch_first=True)

ValueError: some of the strides of a given numpy array are negative. This is currently not supported, b
ut will be added in future releases.