1.简介
PaddleSlim是一个专注于深度学习模型压缩的工具库，提供剪裁、量化、蒸馏、和模型结构搜索等模型压缩策略，帮助用户快速实现模型的小型化。
2.卷积Filter剪裁
对卷积网络的通道进行一次剪裁。剪裁一个卷积层的通道，是指剪裁该卷积层输出的通道。卷积层的权重形状为 [output_channel, input_channel, kernel_size, kernel_size] ，通过剪裁该权重的第一纬度达到剪裁输出通道数的目的。
实际剪裁时要考虑到每层通道的敏感度，一般剪裁后要在验证集上测试精度得到敏感度，敏感度低的剪裁掉来压缩模型。
2.1paddleSlim的API
在paddleslim中关于动态图的剪裁接口主要有三个：
- L1NormFilterPruner 该剪裁器按
Filters的l1-norm统计值对单个卷积层内的Filters的重要性进行排序，并按指定比例剪裁掉相对不重要的Filters。对Filters的剪裁等价于剪裁卷积层的输出通道数。 - L2NormFilterPruner 该剪裁器按
Filters的l2-norm统计值对单个卷积层内的Filters的重要性进行排序，并按指定比例剪裁掉相对不重要的Filters。对Filters的剪裁等价于剪裁卷积层的输出通道数。 - FPGMFilterPruner 该剪裁器按论文 Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration <https://arxiv.org/abs/1811.00250>_ 中的统计方法对单个卷积层内的
Filters的重要性进行排序，并按指定比例剪裁掉相对不重要的Filters。对Filters的剪裁等价于剪裁卷积层的输出通道数。
2.2示例
导入各种包
import paddle
import paddle.vision.models as models
from paddle.static import InputSpec as Input
from paddle.vision.datasets import Cifar10
import paddle.vision.transforms as T
from paddleslim.dygraph import L1NormFilterPruner网络定义和数据集加载
net = models.mobilenet_v1()
inputs = Input(shape=[None, 3, 32, 32], dtype='float32', name='image')
labels = Input(shape=[None, 1], dtype='int64', name='label')
optmizer = paddle.optimizer.Momentum(learning_rate=0.1, parameters=net.parameters())
model = paddle.Model(net, inputs, labels)
model.prepare(
optimizer=optmizer,
loss=paddle.nn.CrossEntropyLoss(),
metrics=paddle.metric.Accuracy(topk=(1, 5))
)
transforms = T.Compose([
T.Transpose(),
T.Normalize([127.5], [127.5])
])
train_dataset = Cifar10(mode='train', transform=transforms)
test_dataset = Cifar10(mode='train', transform=transforms)训练模型
model.fit(train_dataset, epochs=2, batch_size=128, verbose=1)计算剪裁之前的模型相关信息。使用paddle.flops函数
flops = paddle.flops(net, input_size=[1, 3, 32, 32], print_detail=True)结果会返回模型的详细参数量Params和Flops，此外在Layer Name中的名字就是调用剪裁器pruner要给出剪裁参数名。
+-----------------------+-----------------+-----------------+---------+---------+
| Layer Name | Input Shape | Output Shape | Params | Flops |
+-----------------------+-----------------+-----------------+---------+---------+
| conv2d_0 | [1, 3, 32, 32] | [1, 32, 16, 16] | 864 | 221184 |
| batch_norm2d_0 | [1, 32, 16, 16] | [1, 32, 16, 16] | 128 | 16384 |
| re_lu_0 | [1, 32, 16, 16] | [1, 32, 16, 16] | 0 | 0 |
| conv2d_1 | [1, 32, 16, 16] | [1, 32, 16, 16] | 288 | 73728 |
| batch_norm2d_1 | [1, 32, 16, 16] | [1, 32, 16, 16] | 128 | 16384 |
| re_lu_1 | [1, 32, 16, 16] | [1, 32, 16, 16] | 0 | 0 |
| conv2d_2 | [1, 32, 16, 16] | [1, 64, 16, 16] | 2048 | 524288 |
| batch_norm2d_2 | [1, 64, 16, 16] | [1, 64, 16, 16] | 256 | 32768 |
| re_lu_2 | [1, 64, 16, 16] | [1, 64, 16, 16] | 0 | 0 |
| conv2d_3 | [1, 64, 16, 16] | [1, 64, 8, 8] | 576 | 36864 |
| batch_norm2d_3 | [1, 64, 8, 8] | [1, 64, 8, 8] | 256 | 8192 |
| re_lu_3 | [1, 64, 8, 8] | [1, 64, 8, 8] | 0 | 0 |
| conv2d_4 | [1, 64, 8, 8] | [1, 128, 8, 8] | 8192 | 524288 |
| batch_norm2d_4 | [1, 128, 8, 8] | [1, 128, 8, 8] | 512 | 16384 |
| re_lu_4 | [1, 128, 8, 8] | [1, 128, 8, 8] | 0 | 0 |
| conv2d_5 | [1, 128, 8, 8] | [1, 128, 8, 8] | 1152 | 73728 |
| batch_norm2d_5 | [1, 128, 8, 8] | [1, 128, 8, 8] | 512 | 16384 |
| re_lu_5 | [1, 128, 8, 8] | [1, 128, 8, 8] | 0 | 0 |
| conv2d_6 | [1, 128, 8, 8] | [1, 128, 8, 8] | 16384 | 1048576 |
| batch_norm2d_6 | [1, 128, 8, 8] | [1, 128, 8, 8] | 512 | 16384 |
| re_lu_6 | [1, 128, 8, 8] | [1, 128, 8, 8] | 0 | 0 |
| conv2d_7 | [1, 128, 8, 8] | [1, 128, 4, 4] | 1152 | 18432 |
| batch_norm2d_7 | [1, 128, 4, 4] | [1, 128, 4, 4] | 512 | 4096 |
| re_lu_7 | [1, 128, 4, 4] | [1, 128, 4, 4] | 0 | 0 |
| conv2d_8 | [1, 128, 4, 4] | [1, 256, 4, 4] | 32768 | 524288 |
| batch_norm2d_8 | [1, 256, 4, 4] | [1, 256, 4, 4] | 1024 | 8192 |
| re_lu_8 | [1, 256, 4, 4] | [1, 256, 4, 4] | 0 | 0 |
| conv2d_9 | [1, 256, 4, 4] | [1, 256, 4, 4] | 2304 | 36864 |
| batch_norm2d_9 | [1, 256, 4, 4] | [1, 256, 4, 4] | 1024 | 8192 |
| re_lu_9 | [1, 256, 4, 4] | [1, 256, 4, 4] | 0 | 0 |
| conv2d_10 | [1, 256, 4, 4] | [1, 256, 4, 4] | 65536 | 1048576 |
| batch_norm2d_10 | [1, 256, 4, 4] | [1, 256, 4, 4] | 1024 | 8192 |
| re_lu_10 | [1, 256, 4, 4] | [1, 256, 4, 4] | 0 | 0 |
| conv2d_11 | [1, 256, 4, 4] | [1, 256, 2, 2] | 2304 | 9216 |
| batch_norm2d_11 | [1, 256, 2, 2] | [1, 256, 2, 2] | 1024 | 2048 |
| re_lu_11 | [1, 256, 2, 2] | [1, 256, 2, 2] | 0 | 0 |
| conv2d_12 | [1, 256, 2, 2] | [1, 512, 2, 2] | 131072 | 524288 |
| batch_norm2d_12 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_12 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_13 | [1, 512, 2, 2] | [1, 512, 2, 2] | 4608 | 18432 |
| batch_norm2d_13 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_13 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_14 | [1, 512, 2, 2] | [1, 512, 2, 2] | 262144 | 1048576 |
| batch_norm2d_14 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_14 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_15 | [1, 512, 2, 2] | [1, 512, 2, 2] | 4608 | 18432 |
| batch_norm2d_15 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_15 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_16 | [1, 512, 2, 2] | [1, 512, 2, 2] | 262144 | 1048576 |
| batch_norm2d_16 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_16 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_17 | [1, 512, 2, 2] | [1, 512, 2, 2] | 4608 | 18432 |
| batch_norm2d_17 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_17 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_18 | [1, 512, 2, 2] | [1, 512, 2, 2] | 262144 | 1048576 |
| batch_norm2d_18 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_18 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_19 | [1, 512, 2, 2] | [1, 512, 2, 2] | 4608 | 18432 |
| batch_norm2d_19 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_19 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_20 | [1, 512, 2, 2] | [1, 512, 2, 2] | 262144 | 1048576 |
| batch_norm2d_20 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_20 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_21 | [1, 512, 2, 2] | [1, 512, 2, 2] | 4608 | 18432 |
| batch_norm2d_21 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_21 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_22 | [1, 512, 2, 2] | [1, 512, 2, 2] | 262144 | 1048576 |
| batch_norm2d_22 | [1, 512, 2, 2] | [1, 512, 2, 2] | 2048 | 4096 |
| re_lu_22 | [1, 512, 2, 2] | [1, 512, 2, 2] | 0 | 0 |
| conv2d_23 | [1, 512, 2, 2] | [1, 512, 1, 1] | 4608 | 4608 |
| batch_norm2d_23 | [1, 512, 1, 1] | [1, 512, 1, 1] | 2048 | 1024 |
| re_lu_23 | [1, 512, 1, 1] | [1, 512, 1, 1] | 0 | 0 |
| conv2d_24 | [1, 512, 1, 1] | [1, 1024, 1, 1] | 524288 | 524288 |
| batch_norm2d_24 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 4096 | 2048 |
| re_lu_24 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 0 | 0 |
| conv2d_25 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 9216 | 9216 |
| batch_norm2d_25 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 4096 | 2048 |
| re_lu_25 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 0 | 0 |
| conv2d_26 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 1048576 | 1048576 |
| batch_norm2d_26 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 4096 | 2048 |
| re_lu_26 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 0 | 0 |
| adaptive_avg_pool2d_0 | [1, 1024, 1, 1] | [1, 1024, 1, 1] | 0 | 2048 |
| linear_0 | [1, 1024] | [1, 1000] | 1025000 | 1024000 |
+-----------------------+-----------------+-----------------+---------+---------+
Total Flops: 12817920 Total Params: 4253864评估精度
model.evaluate(test_dataset, batch_size=128, verbose=1)剪裁前的精度为：0.937
{'loss': [1.3092904], 'acc': 0.93706}剪裁 。对网络模型两个不同的网络层按照参数名分别进行比例为50%，60%的裁剪。
pruner = L1NormFilterPruner(net, [1, 3, 32, 32])
pruner.prune_vars({'conv2d_22.w_0':0.5, 'conv2d_20.w_0':0.6}, axis=0)计算剪裁之后的flops
flops = paddle.flops(net, input_size=[1, 3, 32, 32], print_detail=True)这里给出部分结果，剪裁的conv2d_20和conv2d_22比例为0.5，0.6
+-----------------------+-----------------+-----------------+---------+---------+
| Layer Name | Input Shape | Output Shape | Params | Flops |
+-----------------------+-----------------+-----------------+---------+---------+
| conv2d_20 | [1, 512, 2, 2] | [1, 205, 2, 2] | 104960 | 419840 |
| conv2d_22 | [1, 205, 2, 2] | [1, 256, 2, 2] | 52480 | 209920 |
+-----------------------+-----------------+-----------------+---------+---------+
Total Flops: 11067556 Total Params: 3615301剪裁后的精度
model.evaluate(test_dataset, batch_size=128, verbose=1)精度由0.93降为0.76 。对模型进行裁剪会导致模型精度有一定程度下降。
{'loss': [2.2277398], 'acc_top5': 0.76516} 对模型进行微调会有助于模型恢复原有精度。 以下代码对裁剪过后的模型进行评估后执行了一个epoch的微调，再对微调过后的模型重新进行评估：
optimizer = paddle.optimizer.Momentum(
learning_rate=0.1,
parameters=net.parameters())
model.prepare(
optimizer,
paddle.nn.CrossEntropyLoss(),
paddle.metric.Accuracy(topk=(1, 5)))
model.fit(train_dataset, epochs=1, batch_size=128, verbose=1)评估
model.evaluate(test_dataset, batch_size=128, verbose=1)微调后精度恢复到0.94，比剪裁prune前还要高是因为剪裁前模型没有调到最优。
{'loss': [1.2696353], 'acc_top1': 0.54044, 'acc_top5': 0.94396}2.2.1剪裁前后对比
| 剪裁前 | 剪裁后 | |
|---|---|---|
| params | 4253864 | 3615301 |
| flops | 12817920 | 11067556 |
| accuracy | 0.937 | 0.94 |
剪裁前的评估每个样本耗时：1s/step 剪裁后的评估每个样本耗时：905ms/step
