简体中文 | English
本文档给出了PaddleVideo系列模型在各平台预测耗时benchmark。
我们从Kinetics-400数据集中,随机选择提供100条用于benchmark时间测试,测试数据可以点击下载。
解压后文件目录:
time-test
├── data # 测试视频文件
└── file.list # 文件列表
视频属性如下:
mean video time: 9.67s
mean video width: 373
mean video height: 256
mean fps: 25
硬件环境:
- CPU: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
- GPU: Tesla V100 16G
软件环境:
- Python 3.7
- PaddlePaddle 2.3.1
- CUDA 10.2
- CUDNN 8.1.1
- 各python库版本参考requirement.txt
各模型性能数据按预测总时间排序,结果如下:
模型名称 | 骨干网络 | 配置文件 | 精度% | 预处理时间ms | 模型推理时间ms | 预测总时间ms |
---|---|---|---|---|---|---|
PP-TSM | MobileNetV2 | pptsm_mv2_k400_videos_uniform.yaml | 68.09 | 51.5 | 3.31 | 54.81 |
PP-TSM | MobileNetV3 | pptsm_mv3_k400_frames_uniform.yaml | 69.84 | 51 | 4.34 | 55.34 |
PP-TSMv2 | PP-LCNet_v2.8f | pptsm_lcnet_k400_8frames_uniform.yaml | 72.45 | 55.31 | 4.37 | 59.68 |
TSM | R50 | tsm_k400_frames.yaml | 71.06 | 52.02 | 9.87 | 61.89 |
PP-TSM | R50 | pptsm_k400_frames_uniform.yaml | 75.11 | 51.84 | 11.26 | 63.1 |
PP-TSM | R101 | pptsm_k400_frames_dense_r101.yaml | 76.35 | 52.1 | 17.91 | 70.01 |
PP-TSMv2 | PP-LCNet_v2.16f | pptsm_lcnet_k400_16frames_uniform.yaml | 74.38 | 69.4 | 7.55 | 76.95 |
SlowFast | 4*16 | slowfast.yaml | 74.35 | 99.27 | 27.4 | 126.67 |
*VideoSwin | B | videoswin_k400_videos.yaml | 82.4 | 95.65 | 117.22 | 212.88 |
MoViNet | A0 | movinet_k400_frame.yaml | 66.62 | 150.36 | 47.24 | 197.60 |
*PP-TimeSformer | base | pptimesformer_k400_videos.yaml | 78.87 | 299.48 | 133.41 | 432.90 |
*TimeSformer | base | timesformer_k400_videos.yaml | 77.29 | 301.54 | 136.12 | 437.67 |
TSN | R50 | tsn_k400_frames.yaml | 69.81 | 794.30 | 168.70 | 963.00 |
PP-TSN | R50 | pptsn_k400_frames.yaml | 75.06 | 837.75 | 175.12 | 1012.87 |
- 注:带
*
表示该模型未使用tensorRT进行预测加速。
- TSN预测时采用TenCrop,比TSM采用的CenterCrop更加耗时。TSN如果使用CenterCrop,则速度稍优于TSM,但精度会低3.5个点。
各模型性能数据按预测总时间排序,结果如下:
模型名称 | 骨干网络 | 配置文件 | 精度% | 预处理时间ms | 模型推理时间ms | 预测总时间ms |
---|---|---|---|---|---|---|
PP-TSM | MobileNetV2 | pptsm_mv2_k400_videos_uniform.yaml | 68.09 | 52.62 | 137.03 | 189.65 |
PP-TSM | MobileNetV3 | pptsm_mv3_k400_frames_uniform.yaml | 69.84 | 53.44 | 139.13 | 192.58 |
PP-TSMv2 | PP-LCNet_v2.8f | pptsm_lcnet_k400_8frames_uniform.yaml | 72.45 | 53.37 | 189.62 | 242.99 |
PP-TSMv2 | PP-LCNet_v2.16f | pptsm_lcnet_k400_16frames_uniform.yaml | 74.38 | 68.07 | 388.64 | 456.71 |
SlowFast | 4*16 | slowfast.yaml | 74.35 | 110.04 | 1201.36 | 1311.41 |
TSM | R50 | tsm_k400_frames.yaml | 71.06 | 52.47 | 1302.49 | 1354.96 |
PP-TSM | R50 | pptsm_k400_frames_uniform.yaml | 75.11 | 52.26 | 1354.21 | 1406.48 |
*MoViNet | A0 | movinet_k400_frame.yaml | 66.62 | 148.30 | 1290.46 | 1438.76 |
PP-TSM | R101 | pptsm_k400_frames_dense_r101.yaml | 76.35 | 52.50 | 2236.94 | 2289.45 |
PP-TimeSformer | base | pptimesformer_k400_videos.yaml | 78.87 | 294.89 | 13426.53 | 13721.43 |
TimeSformer | base | timesformer_k400_videos.yaml | 77.29 | 297.33 | 14034.77 | 14332.11 |
TSN | R50 | tsn_k400_frames.yaml | 69.81 | 860.41 | 18359.26 | 19219.68 |
PP-TSN | R50 | pptsn_k400_frames.yaml | 75.06 | 835.86 | 19778.60 | 20614.46 |
*VideoSwin | B | videoswin_k400_videos.yaml | 82.4 | 76.21 | 32983.49 | 33059.70 |
- 注: 带
*
表示该模型未使用mkldnn进行预测加速。
在进行测试之前,需要安装requirements.txt相关依赖,并且还需安装AutoLog
用于记录计算时间,使用如下命令安装:
python3.7 -m pip install --upgrade pip
pip3.7 install --upgrade -r requirements.txt
python3.7 -m pip install git+https://github.com/LDOUBLEV/AutoLog
以PP-TSM模型为例,请先参考PP-TSM文档导出推理模型,之后使用如下命令进行速度测试:
python3.7 tools/predict.py --input_file time-test/file.list \
--time_test_file=True \
--config configs/recognition/pptsm/pptsm_k400_frames_uniform.yaml \
--model_file inference/ppTSM/ppTSM.pdmodel \
--params_file inference/ppTSM/ppTSM.pdiparams \
--use_gpu=False \
--use_tensorrt=False \
--enable_mkldnn=True \
--enable_benchmark=True \
--disable_glog True
- 各参数含义如下:
input_file: 指定测试文件/文件列表, 示例使用1.1小节提供的测试数据
time_test_file: 是否进行时间测试,请设为True
config: 指定模型配置文件
model_file: 指定推理文件pdmodel路径
params_file: 指定推理文件pdiparams路径
use_gpu: 是否使用GPU预测, False则使用CPU预测
use_tensorrt: 是否开启TensorRT预测
enable_mkldnn: 开启benchmark时间测试,默认设为True
disable_glog: 是否关闭推理时的日志,请设为True
- 测试时,GPU推理使用FP32+TensorRT配置下,CPU使用mkldnn加速。运行100次,去除前3次的warmup时间,得到推理平均时间。
使用以下批量测试脚本,可以方便的将性能结果进行复现:
-
- 下载预训练模型:
mkdir ckpt
cd ckpt
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.1/PPTSM/ppTSM_k400_uniform_distill.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.2/ppTSM_k400_uniform_distill_r101.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.3/ppTSM_mv2_k400.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.3/ppTSM_mv3_k400.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.3/PPTSMv2_k400_16f_dml.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.2/ppTSN_k400_8.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.2/ppTimeSformer_k400_8f_distill.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.1/TSM/TSM_k400.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.2/TSN_k400.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.2/TimeSformer_k400.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo/SlowFast/SlowFast.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.3/MoViNetA0_k400.pdparams
wget https://videotag.bj.bcebos.com/PaddleVideo-release2.2/VideoSwin_k400.pdparams
-
- 准备各模型配置参数列表文件
model.list
- 准备各模型配置参数列表文件
PP-TSM_R50 configs/recognition/pptsm/pptsm_k400_frames_uniform.yaml ckpt/ppTSM_k400_uniform_distill.pdparams ppTSM
PP-TSM_R101 configs/recognition/pptsm/pptsm_k400_frames_dense_r101.yaml ckpt/ppTSM_k400_uniform_distill_r101.pdparams ppTSM
PP-TSM_MobileNetV2 configs/recognition/pptsm/pptsm_mv2_k400_videos_uniform.yaml ckpt/ppTSM_mv2_k400.pdparams ppTSM
PP-TSM_MobileNetV3 configs/recognition/pptsm/pptsm_mv3_k400_frames_uniform.yaml ckpt/ppTSM_mv3_k400.pdparams ppTSM
PP-TSMv2_PP-LCNet_v2 configs/recognition/pptsm/v2/pptsm_lcnet_k400_16frames_uniform_dml_distillation.yaml ckpt/PPTSMv2_k400_16f_dml.pdparams ppTSMv2
PP-TSN_R50 configs/recognition/pptsn/pptsn_k400_frames.yaml ckpt/ppTSN_k400_8.pdparams ppTSN
PP-TimeSformer_base configs/recognition/pptimesformer/pptimesformer_k400_videos.yaml ckpt/ppTimeSformer_k400_8f_distill.pdparams ppTimeSformer
TSM_R50 configs/recognition/tsm/tsm_k400_frames.yaml ckpt/TSM_k400.pdparams TSM
TSN_R50 configs/recognition/tsn/tsn_k400_frames.yaml ckpt/TSN_k400.pdparams TSN
TimeSformer_base configs/recognition/timesformer/timesformer_k400_videos.yaml ckpt/TimeSformer_k400.pdparams TimeSformer
SlowFast_416 configs/recognition/slowfast/slowfast.yaml ckpt/SlowFast.pdparams SlowFast
MoViNet_A0 configs/recognition/movinet/movinet_k400_frame.yaml ckpt/MoViNetA0_k400.pdparams MoViNet
VideoSwin_B configs/recognition/videoswin/videoswin_k400_videos.yaml ckpt/VideoSwin_k400.pdparams VideoSwin
-
- 批量导出模型,执行时传入model.list文件
file=$1
while read line
do
arr=($line)
ModelName=${arr[0]}
ConfigFile=${arr[1]}
ParamsPath=${arr[2]}
echo $ModelName
python3.7 tools/export_model.py -c $ConfigFile \
-p $ParamsPath \
-o inference/$ModelName
done <$file
-
- 测试时间,执行时传入model.list文件
file=$1
while read line
do
arr=($line)
ModelName=${arr[0]}
ConfigFile=${arr[1]}
ParamsPath=${arr[2]}
Model=${arr[3]}
python3.7 tools/predict.py --input_file ../../time-test/file.list \
--time_test_file=True \
--config $ConfigFile \
--model_file inference/$ModelName/$Model.pdmodel \
--params_file inference/$ModelName/$Model.pdiparams \
--use_gpu=False \
--use_tensorrt=False \
--enable_mkldnn=False \
--enable_benchmark=True \
--disable_glog True
echo =====$ModelName END====
done <$file
硬件环境:
- 8 NVIDIA Tesla V100 (16G) GPUs
- Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
软件环境:
- Python 3.7
- PaddlePaddle2.0
- CUDA 10.1
- CUDNN 7.6.3
- NCCL 2.1.15
- GCC 8.2.0
本仓库提供经典和热门时序动作分割模型的性能和精度对比
Model | Metrics | Value | Flops(M) | Params(M) | test time(ms) bs=1 | test time(ms) bs=2 | inference time(ms) bs=1 | inference time(ms) bs=2 |
---|---|---|---|---|---|---|---|---|
MS-TCN | [email protected] | 38.8% | 791.360 | 0.8 | 170 | - | 10.68 | - |
ASRF | [email protected] | 55.7% | 1,283.328 | 1.3 | 190 | - | 16.34 | - |
- 模型名称:填写模型的具体名字,比如PP-TSM
- Metrics:填写模型测试时所用的指标,使用的数据集为breakfast
- Value:填写Metrics指标对应的数值,一般保留小数点后两位
- Flops(M):模型一次前向运算所需的浮点运算量,可以调用PaddleVideo/tools/summary.py脚本计算(不同模型可能需要稍作修改),保留小数点后一位,使用数据输入形状为(1, 2048, 1000)的张量测得
- Params(M):模型参数量,和Flops一起会被脚本计算出来,保留小数点后一位
- test time(ms) bs=1:python脚本开batchsize=1测试时,一个样本所需的耗时,保留小数点后两位。测试使用的数据集为breakfast。
- test time(ms) bs=2:python脚本开batchsize=2测试时,一个样本所需的耗时,保留小数点后两位。时序动作分割模型一般是全卷积网络,所以训练、测试和推理的batch_size都是1。测试使用的数据集为breakfast。
- inference time(ms) bs=1:推理模型用GPU(默认V100)开batchsize=1测试时,一个样本所需的耗时,保留小数点后两位。推理使用的数据集为breakfast。
- inference time(ms) bs=2:推理模型用GPU(默认V100)开batchsize=1测试时,一个样本所需的耗时,保留小数点后两位。时序动作分割模型一般是全卷积网络,所以训练、测试和推理的batch_size都是1。推理使用的数据集为breakfast。