AdaFuse

Dependencies

Red Hat 7.6, Tesla-V100x4

(Here shows how to install the environment using conda)

conda create --name newenv python=3.7.6
conda activate newenv
conda install pytorch=1.2.0 torchvision
pip install thop==0.0.31-2001170342

Preparations

Here taking SomethingV1 dataset for an example:

1. Download all the tar-filed video frames from https://20bn.com/datasets/something-something/v1 and untar them by cat 20bn-something-something-v1-?? | tar zx
2. In common.py, modify ROOT_DIR, DATA_PATH and EXPS_PATH to setup the dataset path and logs path (where to save checkpoints and logs)

Evaluation on SomethingV1

Here we are trying to reproduce the results for AdaFuse-TSN50, AdaFuse-TSM50 and AdaFuse-TSM50Last as shown in Table 3 in the main paper for SomethingV1 dataset.

1. Download the pretrained models from https://drive.google.com/drive/folders/1riSu-E0EzvpLHha5tcbC9OoAAX5i3cBL?usp=sharing and put the folder under your experiment directory EXPS_PATH
2. Run sh zz_eval.sh and you will see the expected/real results (4 GPUS are needed to guarantee the exact results)

Training on SomethingV1

To train AdaFuse-TSM50Last on SomethingV1 dataset, simply run: sh zz_train.sh. It takes less than 14 hours on 4 TeslaV100 GPUs and should get similar accuracy (~ 46.8%) as shown in Table 3 in the main paper.