cmhungsteve / ta3n Goto Github PK

HI,
I'm terribly sorry to disturb you again, but Input data links are expired.
Please update the URL.

Where can I download the UCF-HMDB_full dataset used in the paper?

Hi @cmhungsteve ,

Thanks for sharing the codes. I wonder if the script_train_val file is correct or not. I try to run it but face some errors. (e.g. error: argument --pred_normalize: expected one argument) It seems that this version misses some arguments.

Best.

Nice work!
Thanks for your sharing code!

I am interested in this task,
I want to know some training detail.

• Which GPUs are you using, and how many GPUs are you using?
• Using your setting, I am curious about your UCF -> HMDE and HMDE -> UCF experiments detail

How long do you train on these experiments(hours/days)?

Do we need all of the classes for the source and target domains to match?
Also, when training, are the classes aligned for every mini batch?

Hello sir, I would like to ask the difference between use_attn and use_attn_frame. Is the use_attn for aggregated features/frames and the use_attn_frame for a single frame? Thanks in advance

hi,

for any given iteration, does the network use num_segments number of frames to classify a video?

Hi, I used the pre-extracted features provided and manually edited the paths in the data list files. I am sure my paths are correct but I am getting the following error when running the script_train_val.sh:

main.py: error: argument modality: invalid choice: 'data/classInd_ucf_olympic.txt' (choose from 'RGB', 'Flow', 'RGBDiff', 'RGBDiff2', 'RGBDiffplus') .

Do you have any insight as to where I did it wrong? Thank you in advance.

Hello Sir,

Sorry I forgot to ask this in my previous issue:
When use_target is set to "none", does frame-relation still work?

Thank you

Hi,

Thanks for this awesome codebase!

I would like to train the feature encoders also while reading the data from *.jpgs. Is it possible with this codebase?

I want to visualized data after a training network. could you provide an idea of how can I visualize source and target data after training for comparison of data alignment way? I read your paper you are using T-SNE for visualization but I have no idea how can I use T-SNE after training finished. I prepared T-SNE before training source and target data but after training, I have a problem with visualization.

Hello, thanks for your great job, I want to ask if the code could output the feature processed?and how to visulize the feature processed and unprocessed like the picture in your paper?

Why is formula 7 a minus sign instead of a plus sign? Why not add more weight to the ones that are hard to distinguish.

I am looking forward to your answer. Thank you very much!

Hello Sir,

I have a question about add_loss_DA configuration. I have read the TA3N paper, but the target entropy is not mentioned, may I ask where can I find more information about it? Thank you in advance.

It is better to be provided a download script for kinetics.

Thank you for the wonderful code.

However, when trying to reproduce the results, I am getting errors using the default setting in opts.py. I found that when running it would produce the following error:

Traceback (most recent call last):
  File "main.py", line 836, in <module>
    main()
  File "main.py", line 240, in main
    loss_c, attn_epoch_source, attn_epoch_target = train(num_class, source_loader, target_loader, model, criterion, criterion_domain, optimizer, epoch, train_file, train_short_file, alpha, beta, gamma, mu)
  File "main.py", line 668, in train
    return losses_c.avg, attn_epoch_source.mean(0), attn_epoch_target.mean(0)
RuntimeError: invalid argument 2: invalid dimension 0 at /opt/conda/conda-bld/pytorch_1524584710464/work/aten/src/TH/generic/THTensorMath.c:3897

After checking, I suspect that the save_attention argument was the cause, and I change the default from -1 to 0. Although now it could run, I am getting only 73.89% accuracy. Could you help explain why the save_attention was set to -1? And what is the meaning of it setting to different values? Thank you!

Hi @cmhungsteve

Thanks for the great work. I wanted to execute and obtain the baseline results mentioned in the paper (DANN + [TempPooling/TempRelation]), can you please let me know if you can make the code public?

Thanks

Hi, I want to konw how to get the /models/c3d.pickle? Could this code extract features using C3d networks?

Hi @cmhungsteve,

Thank you for releasing the wonderful code. I am trying to reproduce the UCF-HMDB_{full} results from the ICCV19 paper. i.e. source only, target only, TA^2N, and TA^3N of Table 3 and Table 4 of the paper. But I cannot reproduce the TA^2N and TA^3N results. Can you share the hyperparameter settings of the TA^2N and TA^3N? If you can share the shell script file of those, it would be really helpful.

Thanks!