Closed SonNguyen2510 closed 1 year ago
Hi, Nguyen~
The 312MB C3D model contains the flattened and full FC layers, which is replaces with global average pooling in 110MB model. (We use the C3D descriptor as 3D Conv feature extractor, so the redundant FC layers are dropped to avoid overfitting.) And it is the right way to load the state_dict with param strict=False.
As for the C3D model always returns non-violence results, is the output normalized or is a proper threshold chosen?
Thank you for the answer. I'm choosing the threshold equal 0.5 and using the demo code to test the model. During the training process, I use the data set that I prepared myself. I have divided the data set into 2 folders, fi and no, and placed them in the VioDB folder. Then I use the code "annotation_kfold.ipynb" to create the json file. Are there any tips when preparing data? Did I prepare the wrong data somewhere?
Could you please provide training and validation logs? I'll try to find whether bug exists.
These are my log files. Thank you for spending your time consider my problem C3D_log_files.tar.gz
It looks normal on train and validation from logs. Maybe, did you use a private dataset for testing?
The data set I used for testing was split from the data set I used for training. In the training code, you use normalize before training the model, however in Demo code you didn't use it. Does it affect the testing process?
Thanks a lot for your help, I think I found my mistake, the problem is normalizing the data before putting it on the network. One more time thank you very much!
You're welcome. ;-)
Hello, maybe I have a same problem when I test by @JimLee1996 's Demo_code, but the results always return no-violence. I'm wondering if I'm making the same mistake as @SonNguyen2510 ?
Hello, maybe I have a same problem when I test by @JimLee1996 's Demo_code, but the results always return no-violence. I'm wondering if I'm making the same mistake as @SonNguyen2510 ?
For my problem, as I mention above, input data must be normalized before feeding to the 3D CNN for violence detection. You can try and see if it solves your problem.
At the moment, I'm using dense-lean to train with my dataset. The results is not good. @SonNguyen2510 , when you trained 3D CNN, which model do you think is better?
I'm using DenseNet Lean too. I have test with C3D but its performance is not good as Dense-lean
At the moment, I'm using dense-lean to train with my dataset. The results is not good. @SonNguyen2510 , when you trained 3D CNN, which model do you think is better?
I'm using DenseNet Lean too. I have test with C3D but its performance is not good as Dense-lean
At the moment, I'm using dense-lean to train with my dataset. The results is not good. @SonNguyen2510 , when you trained 3D CNN, which model do you think is better?
Thank you very much
Hello @SonNguyen2510 , could you give me your checkpoint to compare the pretrained. Thank you so much !!!
Thank you for your great work! I am very interested in your research and I am trying to retrain the C3D model to compare the result with the DenseNet Lean model. When loading the pretrain weight, I have set the strict = False and I get the models with size equal 110 MB while the pretrain weight is 312 MB. Is it normal? And I have test my model with the test set and the result is my model cannot detect any violent behavior. Could you tell me how to retrain the C3D model? Thank you very much!
p/s: with the same training data, the DenseNet Lean model can perform well on test set, however C3D model always returns nonviolence