Awesome

ModelPruningLibrary (Updated 3/3/2021)

Plan for the Next Version

We plan to further complete ModelPruningLibrary with the following:

1. c++ implementation conv2d with groups > 1 and depthwise conv2d, as well as missing models in torchvision.models.
2. more optimizers as in torch.optim.
3. well-known pruning algorithms such as SNIP [1].
4. we also plan to implement tools for federated learning (e.g. well-known datasets for FL).

Suggestions/comments are welcome!

Description

This is a PyTorch-based library that implements

1. model pruning: various magnitude-based pruning algorithms (by percentage, random pruning, etc.);
2. conv2d module with sparse kernels as well as fully-connected module implementations;
3. SGD optimizer designed for our sparse modules;
4. two types of save-load functionalities for sparse tensors, determined automatically according to tensor's density (fraction of non-zero entries). If density < 1/32, we save value-index pairs, and otherwise, we use bitmap to save sparse tensors.

It is originally from the following paper:

• Jiang, Y., Wang, S., Valls, V., Ko, B. J., Lee, W. H., Leung, K. K. & Tassiulas, L. (2019). Model pruning enables efficient federated learning on edge devices. arXiv preprint arXiv:1909.12326.

When using this code for scientific publications, please kindly cite the above paper.

The library consists of the following components:

• setup.py: installs the c++ extension and mpl (model pruning library) module
• extension: the extension.cpp c++ file extends the current PyTorch implementation with sparse kernels (the installed module is called sparse_conv2d). However, please note that we only extend PyTorch's slow, cpu version of conv2d forward/backward with no groups and dilation = 1 (see PyTorch's c++ code here). In other words, we do not use acceleration packages such as MKL (which are not available on Raspberry Pis on which our paper experimented). Do not compare the speed of our implementation with the acceleration packages.
• autograd: the AddmmFunction and SparseConv2dFunction functions provide the forward and backward functions to our customized modules.
• models: this is similar to torchvision's implementations (link). Note that we do not implement mnasnet, mobilenet and shufflenetv2 since they have groups > 1 in the models. We also implement popular models such as models in leaf.
• nn: conv2d.py and linear.py implement the prunable modules and their to_sparse functionalities.
• optim: implements a compatible version of SGD optimizer.

Our code has been validated on Ubuntu 20.04. Contact me if you encounter any issues!

Examples

Setup Library:

sudo python3 setup.py install

Importing and Using Model

from mpl.models import conv2

model = conv2()
print(model)

output:

Conv2(
  (features): Sequential(
    (0): DenseConv2d(1, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (1): ReLU(inplace=True)
    (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (3): DenseConv2d(32, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
    (4): ReLU(inplace=True)
    (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (classifier): Sequential(
    (0): DenseLinear(in_features=3136, out_features=2048, bias=True)
    (1): ReLU(inplace=True)
    (2): DenseLinear(in_features=2048, out_features=62, bias=True)
  )
)

Model Pruning:

import mpl.models

model = mpl.models.conv2()
print("Before pruning:")
model.calc_num_prunable_params(display=True)

print("After pruning:")
model.prune_by_pct([0.1, 0, None, 0.9])
model.calc_num_prunable_params(display=True)

output:

Before pruning:
Layer name: features.0. remaining/all: 832/832 = 1.0
Layer name: features.3. remaining/all: 51264/51264 = 1.0
Layer name: classifier.0. remaining/all: 6424576/6424576 = 1.0
Layer name: classifier.2. remaining/all: 127038/127038 = 1.0
Total: remaining/all: 6603710/6603710 = 1.0
After pruning:
Layer name: features.0. remaining/all: 752/832 = 0.9038461538461539
Layer name: features.3. remaining/all: 51264/51264 = 1.0
Layer name: classifier.0. remaining/all: 6424576/6424576 = 1.0
Layer name: classifier.2. remaining/all: 12760/127038 = 0.10044238731718069
Total: remaining/all: 6489352/6603710 = 0.9826827646883343

Dense to Sparse Conversion:

from mpl.models import conv2

model = conv2()
print(model.to_sparse())

output:

Conv2(
  (features): Sequential(
    (0): SparseConv2d(1, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), bias=True)
    (1): ReLU(inplace=True)
    (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
    (3): SparseConv2d(32, 64, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), bias=True)
    (4): ReLU(inplace=True)
    (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (classifier): Sequential(
    (0): SparseLinear(in_features=3136, out_features=2048, bias=True)
    (1): ReLU(inplace=True)
    (2): SparseLinear(in_features=2048, out_features=62, bias=True)
  )
)

Note that DenseConv2d and DenseLinear layers are converted to SparseConv2d and SparseLinear layers, respectively.

SGD Training with a Sparse Model:

from mpl.models import conv2
from mpl.optim import SGD
import torch

inp = torch.rand(size=(10, 1, 28, 28))
model = conv2().to_sparse()
optimizer = SGD(model.parameters(), lr=0.01)
optimizer.zero_grad()
model(inp).sum().backward()
optimizer.step()

Save/Load a Tensor:

from mpl.utils.save_load import save, load
import torch

torch.manual_seed(0)
x = torch.randn(size=(1000, 1000))
mask = torch.rand_like(x) <= 0.5
x = (x * mask).to_sparse()
save(x, "sparse_x.pt")

x_loaded = load("sparse_x.pt")

Using our implementation, the size of sparse_x.pt file is 2.1 MB, while the default torch.save results in a file size of 10 MB (4.8x).

References

<a id="1">[1]</a> Lee, Namhoon, Thalaiyasingam Ajanthan, and Philip HS Torr. "Snip: Single-shot network pruning based on connection sensitivity." arXiv preprint arXiv:1810.02340 (2018).