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
Here taking SomethingV1 dataset for an example:
cat 20bn-something-something-v1-?? | tar zx
common.py
, modify ROOT_DIR
, DATA_PATH
and EXPS_PATH
to setup the dataset path and logs path (where to save checkpoints and logs)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.
EXPS_PATH
sh zz_eval.sh
and you will see the expected/real results (4 GPUS are needed to guarantee the exact results)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.