C3D 1.1 sports-1m model (conv3d_deepnetA_sport1m_iter_1900000)

[wjb123]

Can you release your C3D 1.1 sports-1m model (conv3d_deepnetA_sport1m_iter_1900000), and I need to use the C3D to retrain my own network.

[InnovArul]

There is a pretrained model available with C3D-v1.1 if that's what you are looking for. Please look into this file for the download link of the model.

https://github.com/facebook/C3D/blob/master/C3D-v1.1/examples/c3d_ucf101_feature_extraction/feature_extraction.sh

[dutran]

But that is a different architecture based on Resnet. For C3D-VGG, I have to convert it into the new format later. Will keep this issue open until I share with you the converted model.

[wjb123]

Hope you release it soon.

[bityangke]

@dutran Hi, dutran, how can I fine-tuning on my data using C3D-ResNet18-Sports1M? Thank you in advance!

[dutran]

@bityangke try the example at https://github.com/facebook/C3D/tree/master/C3D-v1.1/examples/c3d_ucf101_finetuning

[bityangke]

Thank you very much！

[bityangke]

@dutran Hi, bother you again. But I can not find 'c3d_resnet18_ucf101_r2_ft_iter_20000.caffemodel' that you have fine-tuned on UCF-101. Have you shared it online?

[dutran]

@bityangke try this https://www.dropbox.com/s/bf5z2jw1pg07c9n/c3d_resnet18_ucf101_r2_ft_iter_20000.caffemodel?dl=0

[bityangke]

@dutran Thank you very much !

[vivoutlaw]

@dutran : have you been able to convert the original C3D model into the new format for C3DV1.1? Thanks!

[dutran]

@vivoutlaw Opps, I closed it by accident. I haven't done this yet.

[rezvannzri]

@dutran :I'd appreciate you if you could release your C3D 1.1 sports-1m model (conv3d_deepnetA_sport1m_iter_1900000) .

[samiksome92]

@rezvannzri I had converted the v1.0 weights to v1.1 format a while back. You can use them if you want. However, I did not convert the old mean file.

https://github.com/samiksome/C3D-weights

[rezvannzri]

@samiksome : Thank you so much!

[kasparov92]

@dutran @InnovArul is this model (c3d_resnet18_ucf101_r2_ft_iter_20000.caffemodel) trained on UCF101 or trained on sports1m and fine tuned on UCF101

[dutran]

c3d_resnet18_ucf101_r2_ft_iter_20000.caffemodel is finetune on UCF101 split1 (after being trained Sports1M)

[kasparov92]

@dutran I am using the c3d_resnet18_ucf101_r2_ft_iter_20000.caffemodel for feature extraction and extract only the prob. In the output, I expected probability for each of the 101 classes but instead, I find probs for 487 classes, which must be the classes of sports1M. Is something wrong with this model?

[dutran]

you used the wrong prototxt file to extract feature.

[kasparov92]

Yeah, you are right, I used the one in the c3d_ucf101_feature_extraction folder, shall I change the line at the end to make the num of classes 101 or shall I use the prototxt in the c3d_ucf101_finetuning folder called train_resnet18_r2.prototxt

[yzldw333]

@samiksome It works, thank u.

[xiaozeyuan]

@dutran I am using the conv3d_deepnetA_sport1m_iter_1900000 for feature extraction,but here is an error "Check failed: num_axes() <= 4 (5 vs. 4) Cannot use legacy accessors on Blobs with > 4 axes." in Relu layer.How to solve it?

[samiksome92]

@xiaozeyuan I believe you are trying to use old C3D v1.0 weights with C3D v1.1 code.

[xiaozeyuan]

@samiksome I download the weights from "https://github.com/samiksome/C3D-weights", and the prototxt file i used is c3d_sport1m_feature_extractor_frm_v1.1.prototxt

[samiksome92]

@xiaozeyuan are you using this with C3D v1.1 or C3D v1.0?

If you use it with C3D v1.1 it should work

[xiaozeyuan]

@samiksome C3D v1.1

[samiksome92]

@xiaozeyuan I just ran it with C3D v1.1 and it ran fine. You are using both weights and prototxt from "https://github.com/samiksome/C3D-weights" and C3D v1.1 extract_image_features.bin right?

[xiaozeyuan]

@samiksome I do use these files to extract C3D features.The puzzling problem made me confused.

[HAHA-DL]

@dutran Hi, just want to use the C3D to extract some features from the video frames. May I ask has the model 'conv3d_deepnetA_sport1m_iter_1900000' been released or not? Many thanks for your time.

[HAHA-DL]

@samiksome Hi, I found that "you had converted the v1.0 weights to v1.1 format a while back." May I ask what is the difference between the v1.0 weights and v1.1 format? Can I use your converted weights with the v1.0 code?

[samiksome92]

@Darren1988 I believe v1.0 and v1.1 use different versions of caffe which causes the weights to be incompatible. weights saved in v1.0 cannot be used v1.1 and vice-versa. As such the converted weights would only work with v1.1 code.

For v1.0 the weights can be downloaded from https://www.dropbox.com/s/vr8ckp0pxgbldhs/conv3d_deepnetA_sport1m_iter_1900000?dl=0, (this is from https://github.com/facebook/C3D/blob/7d10d27f67a4b496958ad58aff0a271b9746f664/C3D-v1.0/examples/c3d_feature_extraction/extract_C3D_feature.py#L38)

[HAHA-DL]

@samiksome Thanks a lot for your prompt reply and explanation. I think the weights you mentioned is the exact weights I wanted to extract my features.

[Zakia13]

Hello. i'm newbie to Deep learning. I want to use C3D pretrained model as a feature extractor and extract features from each 16 frame long clips. I'm using this code https://github.com/hx173149/C3D-tensorflow. But i'm confused how to use it. Kindly guide me on this. Thanks.

[xuyunlu1030]

I tried both the C3D 1.1 sports-1m pretrained model and UCF101 finetuned model, but the loss does not decrease (the same as training from the scratch). See Issue: https://github.com/facebook/C3D/issues/351 I am confused about the problem. Looking forward your suggestions.

[YoungerGao]

@rezvannzri I had converted the v1.0 weights to v1.1 format a while back. You can use them if you want. However, I did not convert the old mean file.

https://github.com/samiksome/C3D-weights

where can i find the covert tools