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
cd simplify
pip3 install -r requirements.txt

Example usage

from torchvision.models import resnet18
from simplify import fuse

model = resnet18()
model.eval()
bn_folding = ...  # List of pairs (conv, bn) to fuse in a single layer
model = fuse(model, bn_folding)

Propagate

The propagate module is used to remove the non-zero bias from zeroed-out neurons in order to be able to remove them.

import torch
from simplify import propagate_bias
from torchvision.models import resnet18

zeros = torch.zeros(1, 3, 224, 224)
model = resnet18()
pinned_out = ...  # List of layers for which the bias should not be propagated
propagate_bias(model, zeros, pinned_out)

Remove

The remove module is used to remove actually remove the zeroed neurons from the model architecture.

import torch
from simplify import remove_zeroed
from torchvision.models import resnet18

zeros = torch.zeros(1, 3, 224, 224)
model = resnet18()
pinned_out = ...  # List of layers in which the output should not change shape
remove_zeroed(model, zeros, pinned_out)

Utilities

We also provide a set of utilities used to define bn_folding and pinned_out for standard PyTorch models.

from torchvision.models import resnet18
from utils import get_bn_folding, get_pinned_out

model = resnet18()
bn_folding = get_bn_folding(model)
pinned_out = get_pinned_out(model)

Tests

Inference time benchmarks

Evaluation mode (fuses BatchNorm)

Update timestamp 08/10/2021 14:26:25

Random structured pruning amount = 50.0%

Architecture	Dense time	Pruned time	Simplified time
alexnet	7.58ms ± 0.29	7.55ms ± 0.28	2.95ms ± 0.02
densenet121	36.41ms ± 4.88	34.31ms ± 3.85	21.87ms ± 1.45
googlenet	15.44ms ± 3.19	13.68ms ± 0.09	10.31ms ± 0.82
inception_v3	25.29ms ± 7.31	21.68ms ± 2.90	13.22ms ± 2.23
mnasnet1_0	17.66ms ± 0.57	13.64ms ± 0.13	11.59ms ± 0.07
mobilenet_v3_large	13.74ms ± 0.67	12.18ms ± 0.46	11.95ms ± 0.21
resnet50	24.39ms ± 4.48	26.19ms ± 5.84	18.21ms ± 1.98
resnext101_32x8d	76.11ms ± 15.79	77.35ms ± 20.04	65.68ms ± 16.41
shufflenet_v2_x2_0	18.07ms ± 2.23	14.32ms ± 0.21	13.06ms ± 0.08
squeezenet1_1	4.50ms ± 0.06	4.39ms ± 0.05	4.09ms ± 0.50
vgg19_bn	40.41ms ± 12.13	38.56ms ± 10.72	12.39ms ± 0.19
wide_resnet101_2	79.40ms ± 25.57	82.86ms ± 22.47	60.16ms ± 10.77

Status of torchvision.models

✔️: all good

❌: gives different results

🤬: an exception occurred

🤷‍♂️: test skipped due to failing of the previous one

Fuse BatchNorm

Update timestamp 06/10/2021 20:26:15

Architecture	BatchNorm Folding	Bias Propagation	Simplification
alexnet	✔️	✔️	✔️
densenet121	✔️	✔️	✔️
googlenet	✔️	✔️	✔️
inception_v3	✔️	✔️	✔️
mnasnet1_0	✔️	✔️	✔️
mobilenet_v3_large	✔️	✔️	✔️
resnet50	✔️	✔️	✔️
resnext101_32x8d	✔️	✔️	✔️
shufflenet_v2_x2_0	✔️	✔️	✔️
squeezenet1_1	✔️	✔️	✔️
vgg19_bn	✔️	✔️	✔️
wide_resnet101_2	✔️	✔️	✔️

Keep BatchNorm

Update timestamp 06/10/2021 20:36:11

Architecture	BatchNorm Folding	Bias Propagation	Simplification
alexnet	✔️	✔️	✔️
densenet121	✔️	✔️	✔️
googlenet	✔️	✔️	✔️
inception_v3	✔️	✔️	✔️
mnasnet1_0	✔️	✔️	✔️
mobilenet_v3_large	✔️	✔️	✔️
resnet50	✔️	✔️	✔️
resnext101_32x8d	✔️	✔️	✔️
shufflenet_v2_x2_0	✔️	✔️	✔️
squeezenet1_1	✔️	✔️	✔️
vgg19_bn	✔️	✔️	✔️
wide_resnet101_2	✔️	✔️	✔️

jeonggunlee/SOFTX-D-21-00143