xlliu7 / e2e-tad Goto Github PK

Hi @xlliu7 ,
Thank you so much for your wonderful work.
I have a problem when I use following method to test Pre-trained Models,

python main.py --cfg configs/thumos14_e2e_slowfast_tadtr.yml --eval --resume scripts/test_reference_models_e2e.sh

And I got this error, I don't know why.

magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: could not find MARK

I had some problems converting my model to onnx
Can torchvision.ops.RoIAlign be used instead of RoIAlign（CUDA extensions）

"I used the test set provided by the author, and the mAP value is high, but the confidence score of each data is very low."

Hi @xlliu7 , I wanted to ask, when the image size is changed from 96 to 224 the overall model doesn't learn. Are there any other Hyper-parameters that affect this training collapse?

I am trying to reproduce ActivityNet v1.3 using the provided code and data, but I am running into some issues. Specifically, I am having trouble with a significant performance drop.

I have tried extracting fixed frames from a video, adding the code about the binary class and combining the external classifier (chuk), but so far, mAP is 27.2.

I would greatly appreciate any help or guidance you can provide on this matter. If there are any additional details or information I can provide to help solve the issue, please let me know.

Thank you for your time.

Hi @xlliu7 ,

I would like to use the Honda Driving dataset for the transformer. How would I do that? Any inputs?

Thanks

Hello,
I want to train model by myself,
But I don't know which dataset you use and the train code you use.
Will you release this part?
thanks

Thanks for your nice work. I am trying to try my dataset using your code and encountered a problem said cannot import name 'Align1D' from partially initialized module 'models.ops.roi_align'. I guess it may be a g-tad problem so I tried to install it. However, when I install g-tad, pycharm shows ' \visual studio\VC\Tools\MSVC\14.29.30133\bin\HostX86\x64\link.exe' failed with exit code 1120'. Could you help me?

Hi, thank you so much for your wonderful work.
I am fascinated by the fact that most models can be trained using a single GPU with 12 GB of memory (as claimed in the paper). May I know when the training code will be released? Have been waiting for the update for a while now and I am just curious!
Thank you.

like using a demo.mp4 and output temporal action results