I successfully re-trained the C3D with UCF101 dataset.
How?
Used script to decode the videos
Used the script to generate train/test (fixed the code on CENTOS7 does not have jot command). [non-deterministic split]
Fixes:
made the train code start from previous saved model now is not supported.
Made the train code to have variable num_frames in train and predict code.
generate new crop_mean.npy for the num_frames = 1 as well. I use
new_mean = numpy.mean(numpy.load('crop_mean.py'), axis=0)
Hi,
I successfully re-trained the C3D with UCF101 dataset.
How? Used script to decode the videos Used the script to generate train/test (fixed the code on CENTOS7 does not have jot command). [non-deterministic split]
Fixes:
Simple benchmark: num_frames = 16 Accuracy = 66.82% (Correct_predicted_labels/num_files) num_frames = 1 Accuracy = 64.00%
Is that something that you guys expect too? I am not using sports1M model as a starting model though purely UCF101 training.
When you reduce the num_frames to 1, that is good. Not losing that much.