General video classification framework implemented by Pytorch for all video classification task.

(Remember first to extract all frames of your videos and put the frames in the same video data dir.)

This directory will store all models you trained.

Please put all your training or test data in this directory and follow the original directory structure.

All log record file will be stored in this directory.

You can put all the network model you design in this directory. I already provide four classic networks: ALSTM, CNNLSTM,LRCN, RNN.

A simple test dataset created by me to validate the test function. Users can keep this directory and make their own change to use their test data, or just delete this directory.

This directory contains the original videos, their frames and the extracted feature.

You could use following command to install all dependencies:

pip -r requirements.txt

PS: for the pytorch version, early version may still be available.

I implement Dataset class in dataset.py. All videos and their frames should put in /data/test and /data/train directory. The child directory is each classes, such as /data/test/ClassA et. al.

Users can change the config setting in conf.py as they need, such as IMAGE_SIZE, EPOCH et al.

After you put your video and their frames well in /data, users should first run extract_features.py to extract the feature of your video data:

>> python extract_features.py -h
usage: extract_features.py [-h] -model MODEL -seq_length SEQ_LENGTH
                           -image_size IMAGE_SIZE -sample_frames SAMPLE_FRAMES
                           -data_dir DATA_DIR [-gpu]

optional arguments:
  -h, --help            show this help message and exit
  -model MODEL          extractor model type
  -seq_length SEQ_LENGTH
                        sample frames length
  -image_size IMAGE_SIZE
                        crop image size
  -sample_frames SAMPLE_FRAMES
                        sampled frames
  -data_dir DATA_DIR    whole image folder dir
  -gpu                  use gpu or not

Once feature extraction done, there will be a /sequences directory in /data which contains all the features and the meta files.

Users can run train.py to start training:

>> python train.py -h
usage: train.py [-h] -model MODEL -seq_dir SEQ_DIR -seq_length SEQ_LENGTH
                -cnn_type CNN_TYPE [-gpu]

optional arguments:
  -h, --help            show this help message and exit
  -model MODEL          model type
  -seq_dir SEQ_DIR      features dir
  -seq_length SEQ_LENGTH
                        sequences length
  -cnn_type CNN_TYPE    features extractor cnn type
  -gpu                  use gpu or not

Users can run test.py to evaluate your model:

>> python test.py -h
usage: test.py [-h] -model MODEL -weights WEIGHTS [-gpu] -data_path DATA_PATH
               -image_size IMAGE_SIZE -cnn_type CNN_TYPE -seq_length
               SEQ_LENGTH

optional arguments:
  -h, --help            show this help message and exit
  -model MODEL          model type
  -weights WEIGHTS      the weights file you want to test
  -gpu                  use gpu or not
  -data_path DATA_PATH  test data path
  -image_size IMAGE_SIZE
                        input image size
  -cnn_type CNN_TYPE    lstm feature extractor type
  -seq_length SEQ_LENGTH
                        sequences length

utils.py: some utils function used in train.py and test.py. Users can modify this file for their convinence.

If this repo do you a favor, a star is my pleasure :)

And if you find any problem, please contact me or open an issue.