bestjuly / iic Goto Github PK

Hello, Li Tao, Thanks for your great work.
I recently want to reproduce your work on hmdb51 dataset, but why I self-supervised pretrain model on UCF101 train split1, finetune on HMDB trainlist01.txt, and then evaluate with testlist01.txt, but I got accuracy of 2% ???

Hi could you please provide weights for your SSL method?

Thanks.

Thanks for sharing your great work.

Just wondering which model do you select for the downstream task since you save pertaining models every 40 epochs.

Thank you.

Hi, thanks for the good work.

I'm currently trying to reproduce the action classification results on UCF101. Using the training parameters that you provided, I've trained my own backbone network and the linear classifier. However, I'm only getting a 0.68 accuracy with the RGB+Res+Repeat settup. I've also trained a linear classifier with the backbone network that you provided, I'm also only getting a accuracy around 0.685.

Could you please help me with this problem? Is there anything you think that could go wrong? Can you share the weights of your linear classifier network?

Thanks very much.

I see you didn't freeze the parameters of backbone during the classifier training, which is different from image SSL papers. I wonder is there any reason for this?

I am trying to the training script like this:

python train_ssl.py --dataset=ucf101

{'print_freq': 10, 'tb_freq': 500, 'save_freq': 40, 'batch_size': 16, 'num_workers': 8, 'epochs': 240, 'learning_rate': 0.01, 'lr_decay_epochs': '120,160,200', 'lr_decay_rate': 0.1, 'beta1': 0.5, 'beta2': 0.999, 'weight_decay': 0.0001, 'momentum': 0.9, 'resume': '', 'model': 'r3d', 'softmax': True, 'nce_k': 1024, 'nce_t': 0.07, 'nce_m': 0.5, 'feat_dim': 512, 'dataset': 'ucf101', 'model_path': './ckpt/', 'tb_path': './logs/', 'debug': True, 'modality': 'res', 'intra_neg': True, 'neg': 'repeat', 'seed': 632, 'model_name': 'intraneg_r3d_res_1012', 'model_folder': './ckpt/intraneg_r3d_res_1012', 'tb_folder': './logs/intraneg_r3d_res_1012'}
[Warning] The training modalities are RGB and [res]
Use split1
v_LongJump_g18_c03 896] BT 5.066 (5.538)        DT 4.007 (4.468)        loss 15.277 (20.188))   1_p -0.001 (0.014)      2_p -0.009 (-0.007)


Traceback (most recent call last):
  File "train_ssl.py", line 341, in <module>
    main()
  File "train_ssl.py", line 308, in main
    criterion_1, criterion_2, optimizer, args)
  File "train_ssl.py", line 166, in train
    for idx, (inputs, u_inputs, v_inputs, _, index) in enumerate(train_loader):
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 363, in __next__
    data = self._next_data()
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 989, in _next_data
    return self._process_data(data)
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 1014, in _process_data
    data.reraise()
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/_utils.py", line 395, in reraise
    raise self.exc_type(msg)
RuntimeError: Caught RuntimeError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/utils/data/_utils/worker.py", line 185, in _worker_loop
    data = fetcher.fetch(index)
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/krishna/anaconda3/envs/py3.7/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/tankpool/home/krishna/PycharmProjects/Inter-intra-video-contrastive-learning/datasets/ucf101.py", line 123, in __getitem__
    raise
RuntimeError: No active exception to reraise

It always gives me this runtime error. I followed all the steps in the README. Still unable to figure out. Can you please help me out? Thank you!

该自监督方法对短视频这种特征提取效果怎么样

1. Has anyone reproduced the results with versions above pytorch1.7+?(1.8 1.9.1.10 and so on)
2. For some reason I have to use cuda11, has anyone reproduced better results with this version of cuda11?

Sorry to disturb you, but have you published the code of IIC v2 mentioned in the paper?

Hi, I'm trying to reproduce your results by training on ucf101. I noticed that my loss is not improving at all. I trained with a batch size of 20, and the loss is stuck at around 15.2 for every iteration, and does not decrease. Just wondering what's the configuration that you used to train the model?

when training the ssl model in multi-gpu setting, loss get stuck at around 15, but in single gpu, loss decrease normally. my environment is pytorch 1.5, CUDA10.2, GeForce RTX 2080 Ti

Thanks for your great work.
I finetuned the pretrained model on UCF101 train split1, but evaluation results show about 6.5% accuracy.
I think that is caused by multi gpus and the procedure of loading checkpoints. But, despite of the change, the result was same.
I only change the original code about dataset path, model wrapped with torch.nn.DataParallel().

Hello, Li Tao, Thanks for your great work.
I'am sorry to bother you, but the problem has been bothering me for a long time.
When I run train_ssl.py for pre-train, the --neg is set to repeat. Next, run ft_classify.py to fine-tune.
When I run train_ssl.py for pre-train, the --neg is set to shuffle. Other settings are the same. Next, run ft_classify.py to fine-tune.
Why is there the same accuracy and best_model. I don't understand this very much. Can you help me analyze what's wrong?