简体中文 | English

SlowFast

---

Contents

• Introduction
• Data
• Train
• Test
• Inference
• Reference

Introduction

SlowFast involves (i) a Slow pathway, operating at low frame rate, to capture spatial semantics, and (ii) a Fast path-way, operating at high frame rate, to capture motion at fine temporal resolution. The Fast pathway can be made very lightweight by reducing its channel capacity, yet can learn useful temporal information for video recognition.

SlowFast Overview

Data

We use Kinetics-400 to train this model，data preparation please refer to Kinetics-400 dataset.

Train

You can start training by：

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

python -B -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" --log_dir=log_slowfast  main.py --validate -c configs/recognition/slowfast/slowfast.yaml

• Training would be efficent using our code. The training speed is 2x faster than the original implementation. Details can refer to benchmark.

Speed up training

It's time consuming to train SlowFast model. So we implement Multigrid training stragety to speed up training. Training script:

python -B -m paddle.distributed.launch --selected_gpus="0,1,2,3,4,5,6,7" --log_dir=log-slowfast main.py --validate --multigrid -c configs/recognition/slowfast/slowfast_multigrid.yaml

Performance evaluation:

training stragety	time cost of one epoch/min	total training time/min	speed-up
Multigrid	27.25	9758 (6.7 days)	2.89x
Normal	78.76	15438 (10.7days)	base

For more details, please refer to accelerate doc.

Test

You can start testing by：

python -B -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" --log_dir=log_slowfast_test main.py --test -c  configs/recognition/slowfast/slowfast.yaml -w output/SlowFast/SlowFast_epoch_000196.pdparams

• Args -w is used to specifiy the model path，you can download our model in SlowFast.pdparams.

Test accuracy in Kinetics-400:

Configs	Acc1	Acc5	Weights
slowfast.yaml	74.35	91.33	slowfast_4x16.pdparams
slowfast_multigrid.yaml	75.84	92.33	slowfast_8x8.pdparams

• Acc1 may be lower than that released in papaer, as ~5% data of kinetics-400 is missing. Experiments have verified that if training with the same data, we can get the same accuracy.

Inference

export inference model

To get model architecture file SlowFast.pdmodel and parameters file SlowFast.pdiparams, use:

python3.7 tools/export_model.py -c configs/recognition/slowfast/slowfast.yaml \
                                -p data/SlowFast.pdparams \
                                -o inference/SlowFast

• Args usage please refer to Model Inference.

infer

python3.7 tools/predict.py --input_file data/example.avi \
                           --config configs/recognition/slowfast/slowfast.yaml \
                           --model_file inference/SlowFast/SlowFast.pdmodel \
                           --params_file inference/SlowFast/SlowFast.pdiparams \
                           --use_gpu=True \
                           --use_tensorrt=False

example of logs:

Current video file: data/example.avi
        top-1 class: 5
        top-1 score: 1.0

we can get the class name using class id and map file data/k400/Kinetics-400_label_list.txt. The top1 prediction of data/example.avi is archery.

Reference

• SlowFast Networks for Video Recognition, Feichtenhofer C, Fan H, Malik J, et al.