A Pytorch implementation of Neural Network Compression (pruning, quantization, encoding/decoding)
Most work of this repo is better done in distiller. However, they have not implement channel pruning and coding yet. With coding in this repo, you can save the model with actually much smaller memory size.
Neural Network Pruning reduces the number of nonzero parameters and thus computation amount (FLOPs).
Deep Compression uses vanilla pruning method. It prunes the parameters with the least importance.
-
Elementwise Pruning: prune those with the smallest magnitude
-
Kernelwise Pruning: prune 2D kernels with the smallest L1(default)/L2 norm
-
Filterwise Pruning: prune 3D filters with the smallest L1(default)/L2 norm
# vanilla pruner usage
from modules.prune import VanillaPruner
rule = [
('0.weight', 'element', [0.3, 0.5], 'abs'),
('1.weight', 'kernel', [0.4, 0.6], 'default')
('2.weight', 'filter', [0.5, 0.7], 'l2norm')
]
pruner = VanillaPruner(rule=rule)
"""
:param rule: str, path to the rule file, each line formats
'param_name granularity sparsity_stage_0, sparstiy_stage_1, ...'
list of tuple, [(param_name(str), granularity(str),
sparsity(float) or [sparsity_stage_0(float), sparstiy_stage_1,],
fn_importance(optional, str or function))]
'granularity': str, choose from ['element', 'kernel', 'filter']
'fn_importance': str, choose from ['abs', 'l1norm', 'l2norm', 'default']
"""
stage = 0
for epoch in range(0, 90):
if epoch == 0:
pruner.prune(model=model, stage=stage, update_masks=True)
best_prec1 = validate(val_loader, model, criterion, epoch)
# in train function
for i, (input, target) in enumerate(train_loader):
output = model(input)
loss = criterion(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
pruner.prune(model=model, stage=stage, update_masks=False)Channel Pruning is another set of neural network pruning methods. It reduces the number of output channels in every convolution or fully-connected layers. Therefore, it can directly speed up the inference.
Channel Pruning takes 2 steps:
- Channel Selection: select channels with least impact to prune
- Parameter Reconstruction: reconstruct the parameter values to optimize the output feature of the next layer to the pruned one
These two steps are conducted layer by layer.
# channel pruning usage
def prune_channel(sparsity, module, next_module, fn_next_input_feature, input_feature,
method='greedy', cpu=True):
"""
channel pruning core function
:param sparsity: float, pruning sparsity
:param module: torch.nn.module, module of the layer being pruned
:param next_module: torch.nn.module, module of the next layer to the one being pruned
:param fn_next_input_feature: function, function to calculate the input feature map for next_module
:param input_feature: torch.(cuda.)Tensor, input feature map of the layer being pruned
:param method: str
'greedy': select one contributed to the smallest next feature after another
'lasso': pruned channels by lasso regression
'random': randomly select
:param cpu: bool, whether done in cpu for larger reconstruction batch size
:return:
void
"""Detailed example shows in here.
Neural Network Quantization is to represent the parameters with fewer bits.
There are several ways to quantize neural network parameters:
-
Fixed-point Quantization: the most common way, uses (i+f)-bits to represent the number, where i-bits for integer and f-bits for fraction.
-
Uniform/Linear Quantization: quantization centroids lies uniformly in the range of parameter values, i.e., the quantization step equals
$(max - min) / k$ , where k is the quantization levels -
K-Means Quantization: quantization centroids calculated by K-Means clustering
# vanilla quantizer usage
from modules.quantize import Quantizer
rule = [
('0.weight', 'k-means', 4, 'k-means++'),
('1.weight', 'fixed_point', 6, 1),
]
quantizer = Quantizer(rule=rule, fix_zeros=True)
"""
:param rule: str, path to the rule file, each line formats
'param_name method bit_length initial_guess_or_bit_length_of_integer'
list of tuple,
[(param_name(str), method(str), bit_length(int),
initial_guess(str)_or_bit_length_of_integer(int))]
:param fix_zeros: whether to fix zeros when quantizing
"""
for epoch in range(0, 90):
# in the train loop
# in train function
for i, (input, target) in enumerate(train_loader):
output = model(input)
loss = criterion(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
quantizer.quantize(model=model, update_labels=True, re_quantize=False)
"""
:param update_labels: bool, whether to re-allocate the param elements
to the latest centroids when using k-means
:param re_quantize: bool, whether to re-quantize the param when using k-means
"""Coding is the last step to compress the neural network in Deep Compression:
-
Fixed-point Coding: it actually is not a coding method, just in case if we want to actually save the model in fixed-point style.
-
Vanilla (Linear) Coding: it uses
$log_2 (N)$ -bits to represent N float number in the codebook, i.e., there are only N possible values in a parameter matrix -
Huffman Coding: it uses huffman coding to represent N float number in the codebook
# coding codec usage (encode)
import torch
from modules.coding import Codec
rule = [
('0.weight', 'huffman', 0, 0, 4),
('1.weight', 'fixed_point', 6, 1, 4)
]
codec = Codec(rule=rule)
"""
:param rule: str, path to the rule file, each line formats
'param_name coding_method bit_length_fixed_point bit_length_fixed_point_of_integer_part
bit_length_of_zero_run_length'
list of tuple,
[(param_name(str), coding_method(str), bit_length_fixed_point(int),
bit_length_fixed_point_of_integer_part(int), bit_length_of_zero_run_length(int))]
"""
encoded_model = codec.encode(model=model)
torch.save({'state_dict': encoded_model.state_dict()}, 'encode.pth.tar', pickle_protocol=4)# coding codec usage (decode)
import torch
from modules.coding import Codec
checkpoint = torch.load('encode.pth.tar')
model = Codec.decode(model=model, state_dict=checkpoint['state_dict']) # initial model is created before
torch.save({'state_dict': model.state_dict()}, 'decode.pth.tar')@article{han2015deep,
title={Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding},
author={Han, Song and Mao, Huizi and Dally, William J},
journal={arXiv preprint arXiv:1510.00149},
year={2015}
}
@inproceedings{han2015learning,
title={Learning both weights and connections for efficient neural network},
author={Han, Song and Pool, Jeff and Tran, John and Dally, William},
booktitle={Advances in neural information processing systems},
pages={1135--1143},
year={2015}
}
@article{luo2017thinet,
title={Thinet: A filter level pruning method for deep neural network compression},
author={Luo, Jian-Hao and Wu, Jianxin and Lin, Weiyao},
journal={arXiv preprint arXiv:1707.06342},
year={2017}
}
@inproceedings{he2017channel,
title={Channel pruning for accelerating very deep neural networks},
author={He, Yihui and Zhang, Xiangyu and Sun, Jian},
booktitle={International Conference on Computer Vision (ICCV)},
volume={2},
number={6},
year={2017}
}