模型剪枝入门

模型剪枝是一种通过减小模型权重规模或中间状态规模来减小模型大小和计算量的技术。 修剪 DNN 模型有三种常见做法：

1. 训练一个模型 -> 对模型进行剪枝 -> 对剪枝后模型进行微调

2. 在模型训练过程中进行剪枝 -> 对剪枝后模型进行微调

3. 对模型进行剪枝 -> 从头训练剪枝后模型

NNI 主要通过在剪枝阶段进行工作来支持上述所有剪枝过程。 通过本教程可以快速了解如何在常见实践中使用 NNI 修剪模型。

准备工作

在本教程中，我们使用一个简单的模型在 MNIST 数据集上进行了预训练。 如果你熟悉在 pytorch 中定义模型和训练模型，可以直接跳到 模型剪枝。

import torch
import torch.nn.functional as F
from torch.optim import SGD

from scripts.compression_mnist_model import TorchModel, trainer, evaluator, device

# define the model
model = TorchModel().to(device)

# show the model structure, note that pruner will wrap the model layer.
print(model)

Out:

TorchModel(
  (conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
  (conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
  (fc1): Linear(in_features=256, out_features=120, bias=True)
  (fc2): Linear(in_features=120, out_features=84, bias=True)
  (fc3): Linear(in_features=84, out_features=10, bias=True)
)

# define the optimizer and criterion for pre-training

optimizer = SGD(model.parameters(), 1e-2)
criterion = F.nll_loss

# pre-train and evaluate the model on MNIST dataset
for epoch in range(3):
    trainer(model, optimizer, criterion)
    evaluator(model)

Out:

Average test loss: 0.5266, Accuracy: 8345/10000 (83%)
Average test loss: 0.2713, Accuracy: 9209/10000 (92%)
Average test loss: 0.1919, Accuracy: 9356/10000 (94%)

模型剪枝

使用 L1NormPruner 对模型进行剪枝并生成掩码。 通常情况下，pruner 需要原始模型和一个 config_list 作为输入参数。 具体关于如何写 config_list 请参考 compression config specification。

以下 config_list 表示 pruner 将修剪类型为 Linear 或 Conv2d 的所有层除了名为 fc3 的层，因为 fc3 被设置为 exclude。 每层的最终稀疏率是 50%。而名为 fc3 的层将不会被修剪。

config_list = [{
    'sparsity_per_layer': 0.5,
    'op_types': ['Linear', 'Conv2d']
}, {
    'exclude': True,
    'op_names': ['fc3']
}]

Pruners usually require model and config_list as input arguments.

from nni.compression.pytorch.pruning import L1NormPruner
pruner = L1NormPruner(model, config_list)

# show the wrapped model structure, `PrunerModuleWrapper` have wrapped the layers that configured in the config_list.
print(model)

Out:

TorchModel(
  (conv1): PrunerModuleWrapper(
    (module): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
  )
  (conv2): PrunerModuleWrapper(
    (module): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
  )
  (fc1): PrunerModuleWrapper(
    (module): Linear(in_features=256, out_features=120, bias=True)
  )
  (fc2): PrunerModuleWrapper(
    (module): Linear(in_features=120, out_features=84, bias=True)
  )
  (fc3): Linear(in_features=84, out_features=10, bias=True)
)

# compress the model and generate the masks
_, masks = pruner.compress()
# show the masks sparsity
for name, mask in masks.items():
    print(name, ' sparsity : ', '{:.2}'.format(mask['weight'].sum() / mask['weight'].numel()))

Out:

conv1  sparsity :  0.5
conv2  sparsity :  0.5
fc1  sparsity :  0.5
fc2  sparsity :  0.5

使用 NNI 的模型加速功能和 pruner 生成好的 masks 对原始模型进行加速，注意 ModelSpeedup 需要 unwrapped 的模型。 模型会在加速之后真正的在规模上变小，并且可能会达到相比于 masks 更大的稀疏率，这是因为 ModelSpeedup 会自动在模型中传播稀疏， 识别由于掩码带来的冗余权重。

# need to unwrap the model, if the model is wrapped before speedup
pruner._unwrap_model()

# speedup the model
from nni.compression.pytorch.speedup import ModelSpeedup

ModelSpeedup(model, torch.rand(3, 1, 28, 28).to(device), masks).speedup_model()

Out:

aten::log_softmax is not Supported! Please report an issue at https://github.com/microsoft/nni. Thanks~
Note: .aten::log_softmax.12 does not have corresponding mask inference object
/home/ningshang/anaconda3/envs/nni-dev/lib/python3.8/site-packages/torch/_tensor.py:1013: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at  aten/src/ATen/core/TensorBody.h:417.)
  return self._grad

模型在加速之后变小了。

print(model)

Out:

TorchModel(
  (conv1): Conv2d(1, 3, kernel_size=(5, 5), stride=(1, 1))
  (conv2): Conv2d(3, 8, kernel_size=(5, 5), stride=(1, 1))
  (fc1): Linear(in_features=128, out_features=60, bias=True)
  (fc2): Linear(in_features=60, out_features=42, bias=True)
  (fc3): Linear(in_features=42, out_features=10, bias=True)
)

微调压缩好的紧凑模型

注意当前的模型已经经过了加速，如果你需要微调模型，请重新生成 optimizer。 这是因为在加速过程中进行了层替换，原来的 optimizer 已经不适用于现在的新模型了。

optimizer = SGD(model.parameters(), 1e-2)
for epoch in range(3):
    trainer(model, optimizer, criterion)