torch-simplify 1.0.1

Simplification of pruned models for accelerated inference

Simplify

Simplification of pruned models for accelerated inference.

Installation

Simplify can be installed using pip:

pip3 install torch-simplify

or if you want to run the latest version of the code, you can install from git:

git clone https://github.com/EIDOSlab/simplify
pip3 install -r requirements.txt

Example usage

model.eval()

for name, module in model.named_modules():
    if isinstance(module, nn.Conv2d):
        prune.random_structured(module, 'weight', amount=0.8, dim=0)
        prune.remove(module, 'weight')

zeros = torch.zeros(1, 3, 224, 224)
simplify.simplify(model, zeros)

Benchmarks

Update timestamp 09/06/2021 19:12:17

Random structured pruning amount = 50.0%

Architecture	Pruned time	Simplified time
alexnet	0.2897s ± 0.0503	0.1185s ± 0.0026
vgg11	3.1443s ± 0.1589	1.2348s ± 0.0394
vgg11_bn	3.8797s ± 0.1467	1.2343s ± 0.0454
vgg13	4.4773s ± 0.0932	1.9440s ± 0.1109
vgg13_bn	6.0232s ± 0.2760	1.8734s ± 0.0811
vgg16	5.5083s ± 0.1880	2.2702s ± 0.0811
vgg16_bn	7.1850s ± 0.2267	2.2364s ± 0.0961
vgg19	6.6639s ± 0.1938	2.5652s ± 0.0746
vgg19_bn	8.3650s ± 0.2687	2.6303s ± 0.1530
resnet18	1.1779s ± 0.0498	0.6446s ± 0.0100
resnet34	2.0665s ± 0.1810	1.0954s ± 0.0935
resnet50	4.2819s ± 0.1770	2.5654s ± 0.0893
resnet101	6.6143s ± 0.1600	3.7886s ± 0.0852
resnet152	9.4440s ± 0.2140	5.2877s ± 0.0616
squeezenet1_0	1.0496s ± 0.0329	0.5848s ± 0.0065
squeezenet1_1	0.5824s ± 0.0244	0.3341s ± 0.0037
inception_v3	2.0929s ± 0.0678	0.7564s ± 0.0099
googlenet	1.6310s ± 0.0754	0.6293s ± 0.0103
shufflenet_v2_x0_5	0.4433s ± 0.0133	0.4466s ± 0.0216
shufflenet_v2_x1_0	0.5693s ± 0.0112	0.5300s ± 0.0144
shufflenet_v2_x1_5	0.7649s ± 0.0202	0.7265s ± 0.0208
shufflenet_v2_x2_0	1.2785s ± 0.0286	1.0461s ± 0.0139
mobilenet_v2	2.8491s ± 0.0573	2.5284s ± 0.0488
mobilenet_v3_small	0.7676s ± 0.0215	0.7478s ± 0.0285
mobilenet_v3_large	1.9852s ± 0.0462	1.7503s ± 0.0499
resnext50_32x4d	5.1172s ± 0.1612	3.6673s ± 0.0595
resnext101_32x8d	13.1092s ± 0.2916	8.8046s ± 0.1130
wide_resnet50_2	6.8694s ± 0.2837	3.3443s ± 0.0764
wide_resnet101_2	11.7889s ± 0.2877	4.8303s ± 0.0874
mnasnet0_5	1.4428s ± 0.0438	1.3573s ± 0.0561
mnasnet0_75	2.2419s ± 0.0547	2.1083s ± 0.0418
mnasnet1_0	2.7525s ± 0.0431	2.4746s ± 0.0260
mnasnet1_3	3.6940s ± 0.0754	3.2427s ± 0.0963

Status of torchvision.models

:heavy_check_mark:: all good

:x:: gives different results

:cursing_face:: an exception occurred

:man_shrugging:: test skipped due to failing of the previous one

Update timestamp 09/06/2021 17:00:55

Architecture	BatchNorm Folding	Bias Propagation	Simplification
alexnet	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg11	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg11_bn	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg13	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg13_bn	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg16	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg16_bn	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg19	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
vgg19_bn	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnet18	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnet34	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnet50	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnet101	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnet152	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
squeezenet1_0	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
squeezenet1_1	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
inception_v3	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
googlenet	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
shufflenet_v2_x0_5	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
shufflenet_v2_x1_0	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
shufflenet_v2_x1_5	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
shufflenet_v2_x2_0	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mobilenet_v2	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mobilenet_v3_small	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mobilenet_v3_large	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnext50_32x4d	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
resnext101_32x8d	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
wide_resnet50_2	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
wide_resnet101_2	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mnasnet0_5	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mnasnet0_75	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mnasnet1_0	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark:
mnasnet1_3	:heavy_check_mark:	:heavy_check_mark:	:heavy_check_mark: