Skip to content

Latest commit

 

History

History

3D_Network

README.md

3D Network for Video Recognition

Requirements

Please follow the installation instructions in INSTALL.md.

Implementation Details

  1. We uniformly sample 16/32/64 frames for DATA.NUM_FRAMES_L, DATA.NUM_FRAMES_M and DATA.NUM_FRAMES_H during training, and use DATA.NUM_FRAMES to specify the number of frames during inference.
  2. We use 1-clip 1-crop evaluation for 3D network with the resolution of 256x256 following the original implementation.
  3. TRAIN.LAMBDA denotes the coefficient $\lambda$ in the loss function and we set it as 1 without further fine-tuning the hyperparameter.
  4. We train 3D network SlowFast with 4 NVIDIA Tesla V100 (32GB) cards and the model is pretrained on Kinetics400 before training on Something-Something V1.

Training

  1. Specify the directory of datasets with DATA.PATH_PREFIX in exp/slowfast_sthv1_FFN/run.sh.

  2. Specify the directory of output with OUTPUT_DIR in run.sh.

  3. Download the pretrained model on Kinetics400 from the original repo and specify the path with CHECKPOINT_FILE_PATH in configs/sth/SLOWFAST_16x8_R50_FFN.yaml.

  4. Simply run the training scripts in exp as followed:

    bash exp/slowfast_sthv1/run.sh  ## baseline training
bash exp/slowfast_sthv1_FFN/run.sh   ## FFN training

Inference

  1. Specify the directory of datasets with DATA.PATH_PREFIX in exp/slowfast_sthv1_FFN/test.sh.

  2. Please download pretrained models from Google Drive.

  3. Specify the directory of the pretrained model with TEST.CHECKPOINT_FILE_PATH in test.sh.

  4. Run the inference scripts in exp as followed:

    bash exp/slowfast_sthv1/test.sh  ## baseline inference
bash exp/slowfast_sthv1_FFN/test.sh   ## FFN inference