/Torch-Pruning

Pruning channels for model acceleration

Primary LanguagePythonMIT LicenseMIT

Torch-Pruning

Pruning channels for model acceleration

Torch-Pruning is a pytorch toolbox for structured neural network pruning. Different from the pruning-by-masking functions in pytorch (unstructured), this toolbox removes entire channels from neural networks for acceleration. Torch-Pruning will automatically detect and handle layer dependencies during pruning. Without too much human effort, it is able to handle various network architectures like DenseNet, ResNet and DeepLab.

Features:

  • Channel pruning for CNNs (e.g. ResNet, DenseNet, Deeplab) and Transformers (e.g. Bert, contributed by @horseee)
  • Graph Tracing and automatic dependency maintaining.
  • Supported modules: Conv, Linear, BatchNorm, LayerNorm, Transposed Conv, PReLU, Embedding and customized modules.
  • Supported operations: split, concatenation, skip connection, flatten, etc.
  • Pruning strategies: Random, L1, L2, etc.

Installation

pip install torch_pruning # v0.2.7

How it works

Torch-Pruning will forward your model with a fake inputs and collect layer information just like torch.jit. A dependency graph is established to describe the computational graph and layer relations. As pruning a certain layer may affect several different layers (see Quick Start), the dependecy will propogate your pruning operation to other layers automatically and provide a PruningPlan. All pruning indices will be correctly transformed if there is torch.split or torch.cat in your models.

Dependency Visualization Example
Conv-Conv AlexNet
Conv-FC (Global Pooling or Flatten) ResNet, VGG
Skip Connection ResNet
Concatenation DenseNet, ASPP
Split torch.chunk

Known Issues:

  • When groups>1, only depthwise conv is supported, i.e. groups=in_channels=out_channels.
  • Customized operations will be treated as element-wise op, e.g. subclass of torch.autograd.Function.

Quickstart

1. A minimal example

import torch
from torchvision.models import resnet18
import torch_pruning as tp

model = resnet18(pretrained=True).eval()

# 1. setup strategy (L1 Norm)
strategy = tp.strategy.L1Strategy() # or tp.strategy.RandomStrategy()

# 2. build layer dependency for resnet18
DG = tp.DependencyGraph()
DG.build_dependency(model, example_inputs=torch.randn(1,3,224,224))

# 3. get a pruning plan from the dependency graph.
pruning_idxs = strategy(model.conv1.weight, amount=0.4) # or manually selected pruning_idxs=[2, 6, 9, ...]
pruning_plan = DG.get_pruning_plan( model.conv1, tp.prune_conv, idxs=pruning_idxs )
print(pruning_plan)

# 4. execute this plan (prune the model)
pruning_plan.exec()

Pruning the resnet.conv1 will affect several layers. Let's inspect the pruning plan (with pruning_idxs=[2, 6, 9]):

-------------
[ <DEP: prune_conv => prune_conv on conv1 (Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False))>, Index=[2, 6, 9], NumPruned=441]
[ <DEP: prune_conv => prune_batchnorm on bn1 (BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))>, Index=[2, 6, 9], NumPruned=6]
[ <DEP: prune_batchnorm => _prune_elementwise_op on _ElementWiseOp()>, Index=[2, 6, 9], NumPruned=0]
[ <DEP: _prune_elementwise_op => _prune_elementwise_op on _ElementWiseOp()>, Index=[2, 6, 9], NumPruned=0]
[ <DEP: _prune_elementwise_op => prune_related_conv on layer1.0.conv1 (Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))>, Index=[2, 6, 9], NumPruned=1728]
[ <DEP: _prune_elementwise_op => _prune_elementwise_op on _ElementWiseOp()>, Index=[2, 6, 9], NumPruned=0]
[ <DEP: _prune_elementwise_op => prune_batchnorm on layer1.0.bn2 (BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))>, Index=[2, 6, 9], NumPruned=6]
[ <DEP: prune_batchnorm => prune_conv on layer1.0.conv2 (Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))>, Index=[2, 6, 9], NumPruned=1728]
[ <DEP: _prune_elementwise_op => _prune_elementwise_op on _ElementWiseOp()>, Index=[2, 6, 9], NumPruned=0]
[ <DEP: _prune_elementwise_op => prune_related_conv on layer1.1.conv1 (Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))>, Index=[2, 6, 9], NumPruned=1728]
[ <DEP: _prune_elementwise_op => _prune_elementwise_op on _ElementWiseOp()>, Index=[2, 6, 9], NumPruned=0]
[ <DEP: _prune_elementwise_op => prune_batchnorm on layer1.1.bn2 (BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True))>, Index=[2, 6, 9], NumPruned=6]
[ <DEP: prune_batchnorm => prune_conv on layer1.1.conv2 (Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False))>, Index=[2, 6, 9], NumPruned=1728]
[ <DEP: _prune_elementwise_op => _prune_elementwise_op on _ElementWiseOp()>, Index=[2, 6, 9], NumPruned=0]
[ <DEP: _prune_elementwise_op => prune_related_conv on layer2.0.conv1 (Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False))>, Index=[2, 6, 9], NumPruned=3456]
[ <DEP: _prune_elementwise_op => prune_related_conv on layer2.0.downsample.0 (Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False))>, Index=[2, 6, 9], NumPruned=384]
11211 parameters will be pruned
-------------

Tip: please remember to save the whole model object (weights+architecture) rather than model weights only:

# save a pruned model
# torch.save(model.state_dict(), 'model.pth') # weights only
torch.save(model, 'model.pth') # obj (arch + weights), recommended.

# load a pruned model
model = torch.load('model.pth') # no load_state_dict

2. Low-level pruning functions

It is equivalent to make a layer-by-layer fixing using the low-level pruning functions.

tp.prune_conv( model.conv1, idxs=[2,6,9] )

# fix the broken dependencies manually
tp.prune_batchnorm( model.bn1, idxs=[2,6,9] )
tp.prune_related_conv( model.layer2[0].conv1, idxs=[2,6,9] )
...

3. Customized Layers

Please refer to examples/customized_layer.py.

4. Rounding channels for device-friendly network pruning

You can round the channels by passing a round_to parameter to strategy. For example, the following script will round the number of channels to 16xN (e.g., 16, 32, 48, 64).

strategy = tp.strategy.L1Strategy()
pruning_idxs = strategy(model.conv1.weight, amount=0.2, round_to=16)

Please refer to VainF#38 for more details.

5. Example: pruning ResNet18 on Cifar10

5.1. Scratch training

cd examples
python prune_resnet18_cifar10.py --mode train # 11.1M, Acc=0.9248

5.2. Pruning and fintuning

python prune_resnet18_cifar10.py --mode prune --round 1 --total_epochs 30 --step_size 20 # 4.5M, Acc=0.9229
python prune_resnet18_cifar10.py --mode prune --round 2 --total_epochs 30 --step_size 20 # 1.9M, Acc=0.9207
python prune_resnet18_cifar10.py --mode prune --round 3 --total_epochs 30 --step_size 20 # 0.8M, Acc=0.9176
python prune_resnet18_cifar10.py --mode prune --round 4 --total_epochs 30 --step_size 20 # 0.4M, Acc=0.9102
python prune_resnet18_cifar10.py --mode prune --round 5 --total_epochs 30 --step_size 20 # 0.2M, Acc=0.9011
...

Layer Dependency

During structured pruning, we need to maintain the channel consistency between different layers.

A Simple Case

More Complicated Cases

the layer dependency becomes much more complicated when the model contains skip connections or concatenations.

Residual Block:

Concatenation:

See paper Pruning Filters for Efficient ConvNets for more details.