Table Of Contents

4. Computing FLOPS, latency and fps of a model

It is important to have an idea of how to measure a video model’s speed, so that you can choose the model that suits best for your use case. In this tutorial, we provide two simple scripts to help you compute (1) FLOPS, (2) number of parameters, (3) fps and (4) latency. These four numbers will help you evaluate the speed of this model. To be specific, FLOPS means floating point operations per second, and fps means frame per second. In terms of comparison, (1) FLOPS, the lower the better, (2) number of parameters, the lower the better, (3) fps, the higher the better, (4) latency, the lower the better.

In terms of input, we use the setting in each model’s training config. For example, I3D models will use 32 frames with stride 2 in crop size 224, but R2+1D models will use 16 frames with stride 2 in crop size 112. This will make sure that the speed performance here correlates well with the reported accuracy number. We list these four numbers and the models’ accuracy on Kinetics400 dataset in the table below.

Model	FLOPS	# params	fps	Latency	Top-1 Accuracy
resnet18_v1b_kinetics400	1.819	11.382	264.01	0.0038	66.73
resnet34_v1b_kinetics400	3.671	21.49	151.96	0.0066	69.85
resnet50_v1b_kinetics400	4.110	24.328	114.05	0.0088	70.88
resnet101_v1b_kinetics400	7.833	43.320	59.56	0.0167	72.25
resnet152_v1b_kinetics400	11.558	58.963	36.93	0.0271	72.45
i3d_resnet50_v1_kinetics400	33.275	28.863	1719.50	0.0372	74.87
i3d_resnet101_v1_kinetics400	51.864	52.574	1137.74	0.0563	75.10
i3d_nl5_resnet50_v1_kinetics400	47.737	38.069	1403.16	0.0456	75.17
i3d_nl10_resnet50_v1_kinetics400	62.199	42.275	1200.69	0.0533	75.93
i3d_nl5_resnet101_v1_kinetics400	66.326	61.780	999.94	0.0640	75.81
i3d_nl10_resnet101_v1_kinetics400	80.788	70.985	890.33	0.0719	75.93
i3d_slow_resnet50_f8s8_kinetics400	41.919	32.454	1702.60	0.0376	74.41
i3d_slow_resnet50_f16s4_kinetics400	83.838	32.454	1406.00	0.0455	76.36
i3d_slow_resnet50_f32s2_kinetics400	167.675	32.454	860.74	0.0744	77.89
i3d_slow_resnet101_f8s8_kinetics400	85.675	60.359	1114.22	0.0574	76.15
i3d_slow_resnet101_f16s4_kinetics400	171.348	60.359	876.20	0.0730	77.11
i3d_slow_resnet101_f32s2_kinetics400	342.696	60.359	541.16	0.1183	78.57
r2plus1d_v1_resnet18_kinetics400	40.645	31.505	804.31	0.0398	71.72
r2plus1d_v1_resnet34_kinetics400	75.400	61.832	503.17	0.0636	72.63
r2plus1d_v1_resnet50_kinetics400	65.543	53.950	667.06	0.0480	74.92
r2plus1d_v2_resnet152_kinetics400	252.900	118.227	546.19	0.1172	81.34
ircsn_v2_resnet152_f32s2_kinetics400	74.758	29.704	435.77	0.1469	83.18
slowfast_4x16_resnet50_kinetics400	27.820	34.480	1396.45	0.0458	75.25
slowfast_8x8_resnet50_kinetics400	50.583	34.566	1297.24	0.0493	76.66
slowfast_8x8_resnet101_kinetics400	96.794	62.827	889.62	0.0719	76.95
tpn_resnet50_f8s8_kinetics400	50.457	71.800	1350.39	0.0474	77.04
tpn_resnet50_f16s4_kinetics400	99.929	71.800	1128.39	0.0567	77.33
tpn_resnet50_f32s2_kinetics400	198.874	71.800	716.89	0.0893	78.90
tpn_resnet101_f8s8_kinetics400	94.366	99.705	942.61	0.0679	78.10
tpn_resnet101_f16s4_kinetics400	187.594	99.705	754.00	0.0849	79.39
tpn_resnet101_f32s2_kinetics400	374.048	99.705	479.77	0.1334	79.70

Note

Feel free to skip the tutorial because the speed computation scripts are self-complete and ready to launch.

Download Full Python Script: get_flops.py

Download Full Python Script: get_fps.py

You can reproduce the numbers in the above table by

python get_flops.py --config-file CONFIG and python get_fps.py --config-file CONFIG

If you encouter missing dependecy issue of thop, please install the package first.

pip install thop

Previous

3. Extracting video features from pre-trained models

Next

5. DistributedDataParallel (DDP) Framework