简体中文 | English
We aims to speed up TSN model training using DALI in this code. As nvidia DALI not support TSN sampling way, we reimplemented segment sampling in VideoReader.
Test Environment:
Card: Tesla v100
Memory: 4 * 16G
Cuda: 9.0
batch_size of single card: 32
| Training way | batch cost/s | reader cost/s | ips:instance/sec | Speed up |
|---|---|---|---|---|
| DALI | 2.083 | 1.804 | 15.36597 | 1.41x |
| Dataloader: num_workers=4 | 2.943 | 2.649 | 10.87460 | base |
| pytorch实现 | TODO | TODO | TODO | TODO |
docker image:
huangjun12/paddlevideo:tsn_dali_cuda9_0
To build container, you can use:
nvidia-docker run --name tsn-DALI -v /home:/workspace --network=host -it --shm-size 64g -e NVIDIA_DRIVER_CAPABILITIES=compute,utility,video huangjun12/paddlevideo:tsn_dali_cuda9_0 /bin/bash-
Kinetics400 dataset please refer to K400 data
-
UCF101 dataset please refer to UCF101 data
- Please download ResNet50_pretrain.pdparams as pretraind model:
wget https://videotag.bj.bcebos.com/PaddleVideo/PretrainModel/ResNet50_pretrain.pdparamsand add path to MODEL.framework.backbone.pretrained in config file as：
MODEL:
framework: "Recognizer2D"
backbone:
name: "ResNet"
pretrained: your weight pathYou can start training by:
python3.7 -B -m paddle.distributed.launch --gpus="0,1,2,3" --log_dir=log_tsn main.py --train_dali -c configs/recognition/tsn/tsn_dali.yaml -o log_level="INFO"-
Args -c is used to specify config file，default is
configs/recognition/tsn/tsn_dali.yaml。 -
For finetune please download our trained model TSN.pdparamscoming soon，and specify file path with --weights.
-
For the config file usage，please refer to config.
Please refer to TSN Test
Please refer to TSN Inference
- Temporal Segment Networks: Towards Good Practices for Deep Action Recognition, Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, Luc Van Gool