A implementation of SlowFast Networks for Video Recognition with tf.keras This code use half-period cosine schedule of learning rate decaying and warm-up strategy, but I don't kown if it's the same as the paper.

tensorflow >= 1.12
pillow>=5.1.0

##Get code

git clone https://github.com/xuzheyuan624/slowfast-keras.git
cd slowfast-keras

You can use UCF101 or other datasets, which should be orgnized as :

• UCF101
  • ApplyEyeMakeUp
  • ApplyLipstick
  • Archery
  • ......
• ucfTrainTestlist
  • classInd.txt
  • trainlist01.txt
  • testlist01.txt
  • ......

convert video to jpgs with:

python utils/ucf_hmdb51_frames.py UCF101 UCF101_jpg

calculate the video's frames

python utils/ucf_hmdb51_frames.py UCF101_jpg

You must change the root_path, video_path, name_path or others in train.sh for your own. See details in opts.py
For example: root_pathis path to slowfast-keras
video_pathispath to UCF101_jpg
name_pathis path to classInd.txt
......
Then you can train with:

bash train.sh

[1] SlowFastNetworks
[2] 3D-ResNets-Pytorch
[3] SGDR