未验证 提交 879d59d2 编写于 作者: Z Zhen Wang 提交者: GitHub

Merge pull request #2044 from wzzju/add_quant_doc

add GoogleNet support add quant low level api usage doc.
......@@ -215,6 +215,10 @@ compress_pass:
### 2.1 量化训练
**用户须知:** 现阶段的量化训练主要针对卷积层(包括二维卷积和Depthwise卷积)以及全连接层进行量化。卷积层和全连接层在PaddlePaddle框架中对应算子包括`conv2d``depthwise_conv2d``mul`等。量化训练会对所有的`conv2d``depthwise_conv2d``mul`进行量化操作,且要求它们的输入中必须包括激活和参数两部分。
#### 2.1.1 基于High-Level API的量化训练
>注意:多个压缩策略组合使用时,量化训练策略必须放在最后。
```
......@@ -279,6 +283,11 @@ strategies:
- **save_in_nodes:** variable名称列表。在保存量化后模型的时候,需要根据save_in_nodes对eval programg 网络进行前向遍历剪枝。默认为eval_feed_list内指定的variable的名称列表。
- **save_out_nodes:** varibale名称列表。在保存量化后模型的时候,需要根据save_out_nodes对eval programg 网络进行回溯剪枝。默认为eval_fetch_list内指定的variable的名称列表。
#### 2.1.2 基于Low-Level API的量化训练
量化训练High-Level API是对Low-Level API的高层次封装,这使得用户仅需编写少量的代码和配置文件即可进行量化训练。然而,封装必然会带来使用灵活性的降低。因此,若用户在进行量化训练时需要更多的灵活性,可参考 [量化训练Low-Level API使用示例](../quant_low_level_api/README.md)
### 2.2 卷积核剪切
该策略通过减少指定卷积层中卷积核的数量,达到缩减模型大小和计算复杂度的目的。根据选取剪切比例的策略的不同,又细分为以下两个方式:
......
from .mobilenet import MobileNet
from .resnet import ResNet50, ResNet101, ResNet152
from .googlenet import GoogleNet
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
__all__ = ['GoogleNet']
train_parameters = {
"input_size": [3, 224, 224],
"input_mean": [0.485, 0.456, 0.406],
"input_std": [0.229, 0.224, 0.225],
"learning_strategy": {
"name": "piecewise_decay",
"batch_size": 256,
"epochs": [10, 16, 30],
"steps": [0.1, 0.01, 0.001, 0.0001]
}
}
class GoogleNet():
def __init__(self):
self.params = train_parameters
def conv_layer(self,
input,
num_filters,
filter_size,
stride=1,
groups=1,
act=None,
name=None):
channels = input.shape[1]
stdv = (3.0 / (filter_size**2 * channels))**0.5
param_attr = ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + "_weights")
conv = fluid.layers.conv2d(
input=input,
num_filters=num_filters,
filter_size=filter_size,
stride=stride,
padding=(filter_size - 1) // 2,
groups=groups,
act=act,
param_attr=param_attr,
bias_attr=False,
name=name)
return conv
def xavier(self, channels, filter_size, name):
stdv = (3.0 / (filter_size**2 * channels))**0.5
param_attr = ParamAttr(
initializer=fluid.initializer.Uniform(-stdv, stdv),
name=name + "_weights")
return param_attr
def inception(self,
input,
channels,
filter1,
filter3R,
filter3,
filter5R,
filter5,
proj,
name=None):
conv1 = self.conv_layer(
input=input,
num_filters=filter1,
filter_size=1,
stride=1,
act=None,
name="inception_" + name + "_1x1")
conv3r = self.conv_layer(
input=input,
num_filters=filter3R,
filter_size=1,
stride=1,
act=None,
name="inception_" + name + "_3x3_reduce")
conv3 = self.conv_layer(
input=conv3r,
num_filters=filter3,
filter_size=3,
stride=1,
act=None,
name="inception_" + name + "_3x3")
conv5r = self.conv_layer(
input=input,
num_filters=filter5R,
filter_size=1,
stride=1,
act=None,
name="inception_" + name + "_5x5_reduce")
conv5 = self.conv_layer(
input=conv5r,
num_filters=filter5,
filter_size=5,
stride=1,
act=None,
name="inception_" + name + "_5x5")
pool = fluid.layers.pool2d(
input=input,
pool_size=3,
pool_stride=1,
pool_padding=1,
pool_type='max')
convprj = fluid.layers.conv2d(
input=pool,
filter_size=1,
num_filters=proj,
stride=1,
padding=0,
name="inception_" + name + "_3x3_proj",
param_attr=ParamAttr(
name="inception_" + name + "_3x3_proj_weights"),
bias_attr=False)
cat = fluid.layers.concat(input=[conv1, conv3, conv5, convprj], axis=1)
cat = fluid.layers.relu(cat)
return cat
def net(self, input, class_dim=1000):
conv = self.conv_layer(
input=input,
num_filters=64,
filter_size=7,
stride=2,
act=None,
name="conv1")
pool = fluid.layers.pool2d(
input=conv, pool_size=3, pool_type='max', pool_stride=2)
conv = self.conv_layer(
input=pool,
num_filters=64,
filter_size=1,
stride=1,
act=None,
name="conv2_1x1")
conv = self.conv_layer(
input=conv,
num_filters=192,
filter_size=3,
stride=1,
act=None,
name="conv2_3x3")
pool = fluid.layers.pool2d(
input=conv, pool_size=3, pool_type='max', pool_stride=2)
ince3a = self.inception(pool, 192, 64, 96, 128, 16, 32, 32, "ince3a")
ince3b = self.inception(ince3a, 256, 128, 128, 192, 32, 96, 64,
"ince3b")
pool3 = fluid.layers.pool2d(
input=ince3b, pool_size=3, pool_type='max', pool_stride=2)
ince4a = self.inception(pool3, 480, 192, 96, 208, 16, 48, 64, "ince4a")
ince4b = self.inception(ince4a, 512, 160, 112, 224, 24, 64, 64,
"ince4b")
ince4c = self.inception(ince4b, 512, 128, 128, 256, 24, 64, 64,
"ince4c")
ince4d = self.inception(ince4c, 512, 112, 144, 288, 32, 64, 64,
"ince4d")
ince4e = self.inception(ince4d, 528, 256, 160, 320, 32, 128, 128,
"ince4e")
pool4 = fluid.layers.pool2d(
input=ince4e, pool_size=3, pool_type='max', pool_stride=2)
ince5a = self.inception(pool4, 832, 256, 160, 320, 32, 128, 128,
"ince5a")
ince5b = self.inception(ince5a, 832, 384, 192, 384, 48, 128, 128,
"ince5b")
pool5 = fluid.layers.pool2d(
input=ince5b, pool_size=7, pool_type='avg', pool_stride=7)
dropout = fluid.layers.dropout(x=pool5, dropout_prob=0.4)
out = fluid.layers.fc(input=dropout,
size=class_dim,
act='softmax',
param_attr=self.xavier(1024, 1, "out"),
name="out",
bias_attr=ParamAttr(name="out_offset"))
pool_o1 = fluid.layers.pool2d(
input=ince4a, pool_size=5, pool_type='avg', pool_stride=3)
conv_o1 = self.conv_layer(
input=pool_o1,
num_filters=128,
filter_size=1,
stride=1,
act=None,
name="conv_o1")
fc_o1 = fluid.layers.fc(input=conv_o1,
size=1024,
act='relu',
param_attr=self.xavier(2048, 1, "fc_o1"),
name="fc_o1",
bias_attr=ParamAttr(name="fc_o1_offset"))
dropout_o1 = fluid.layers.dropout(x=fc_o1, dropout_prob=0.7)
out1 = fluid.layers.fc(input=dropout_o1,
size=class_dim,
act='softmax',
param_attr=self.xavier(1024, 1, "out1"),
name="out1",
bias_attr=ParamAttr(name="out1_offset"))
pool_o2 = fluid.layers.pool2d(
input=ince4d, pool_size=5, pool_type='avg', pool_stride=3)
conv_o2 = self.conv_layer(
input=pool_o2,
num_filters=128,
filter_size=1,
stride=1,
act=None,
name="conv_o2")
fc_o2 = fluid.layers.fc(input=conv_o2,
size=1024,
act='relu',
param_attr=self.xavier(2048, 1, "fc_o2"),
name="fc_o2",
bias_attr=ParamAttr(name="fc_o2_offset"))
dropout_o2 = fluid.layers.dropout(x=fc_o2, dropout_prob=0.7)
out2 = fluid.layers.fc(input=dropout_o2,
size=class_dim,
act='softmax',
param_attr=self.xavier(1024, 1, "out2"),
name="out2",
bias_attr=ParamAttr(name="out2_offset"))
# last fc layer is "out"
return out, out1, out2
......@@ -14,7 +14,7 @@ train_parameters = {
"learning_strategy": {
"name": "piecewise_decay",
"batch_size": 256,
"epochs": [30, 60, 90],
"epochs": [10, 16, 30],
"steps": [0.1, 0.01, 0.001, 0.0001]
}
}
......
......@@ -15,7 +15,7 @@ train_parameters = {
"learning_strategy": {
"name": "piecewise_decay",
"batch_size": 256,
"epochs": [30, 60, 90],
"epochs": [10, 16, 30],
"steps": [0.1, 0.01, 0.001, 0.0001]
}
}
......
<div align="center">
<h3>
<a href="../docs/tutorial.md">
算法原理介绍
</a>
<span> | </span>
<a href="../docs/usage.md">
使用文档
</a>
<span> | </span>
<a href="../docs/demo.md">
示例文档
</a>
<span> | </span>
<a href="../docs/model_zoo.md">
Model Zoo
</a>
</h3>
</div>
---
# 量化训练Low-Level API使用示例
## 目录
- [量化训练Low-Level APIs介绍](#1-量化训练low-level-apis介绍)
- [基于Low-Level API的量化训练](#2-基于low-level-api的量化训练)
## 1. 量化训练Low-Level APIs介绍
量化训练Low-Level APIs主要涉及到PaddlePaddle框架中的四个IrPass,即`QuantizationTransformPass``QuantizationFreezePass``ConvertToInt8Pass`以及`TransformForMobilePass`。这四个IrPass的具体功能描述如下:
* `QuantizationTransformPass`: QuantizationTransformPass主要负责在IrGraph的`conv2d``depthwise_conv2d``mul`等算子的各个输入前插入连续的量化op和反量化op,并改变相应反向算子的某些输入,示例如图1:
<p align="center">
<img src="../docs/images/usage/TransformPass.png" height=400 width=520 hspace='10'/> <br />
<strong>图1:应用QuantizationTransformPass后的结果</strong>
</p>
* `QuantizationFreezePass`:QuantizationFreezePass主要用于改变IrGraph中量化op和反量化op的顺序,即将类似图1中的量化op和反量化op顺序改变为图2中的布局。除此之外,QuantizationFreezePass还会将`conv2d``depthwise_conv2d``mul`等算子的权重离线量化为int8_t范围内的值(但数据类型仍为float32),以减少预测过程中对权重的量化操作,示例如图2:
<p align="center">
<img src="../docs/images/usage/FreezePass.png" height=400 width=420 hspace='10'/> <br />
<strong>图2:应用QuantizationFreezePass后的结果</strong>
</p>
* `ConvertToInt8Pass`:ConvertToInt8Pass必须在QuantizationFreezePass之后执行,其主要目的是将执行完QuantizationFreezePass后输出的权重类型由`FP32`更改为`INT8`。换言之,用户可以选择将量化后的权重保存为float32类型(不执行ConvertToInt8Pass)或者int8_t类型(执行ConvertToInt8Pass),示例如图3:
<p align="center">
<img src="../docs/images/usage/ConvertToInt8Pass.png" height=400 width=400 hspace='10'/> <br />
<strong>图3:应用ConvertToInt8Pass后的结果</strong>
</p>
* `TransformForMobilePass`:经TransformForMobilePass转换后,用户可得到兼容[paddle-mobile](https://github.com/PaddlePaddle/paddle-mobile)移动端预测库的量化模型。paddle-mobile中的量化op和反量化op的名称分别为`quantize``dequantize``quantize`算子和PaddlePaddle框架中的`fake_quantize_abs_max`算子簇的功能类似,`dequantize` 算子和PaddlePaddle框架中的`fake_dequantize_max_abs`算子簇的功能相同。若选择paddle-mobile执行量化训练输出的模型,则需要将`fake_quantize_abs_max`等算子改为`quantize`算子以及将`fake_dequantize_max_abs`等算子改为`dequantize`算子,示例如图4:
<p align="center">
<img src="../docs/images/usage/TransformForMobilePass.png" height=400 width=400 hspace='10'/> <br />
<strong>图4:应用TransformForMobilePass后的结果</strong>
</p>
## 2. 基于Low-Level API的量化训练
本小节以ResNet50和MobileNetV1为例,介绍了PaddlePaddle量化训练Low-Level API的使用方法,具体如下:
1) 执行如下命令clone [Pddle models repo](https://github.com/PaddlePaddle/models)
```bash
git clone https://github.com/PaddlePaddle/models.git
```
2) 准备数据集(包括训练数据集和验证数据集)。以ILSVRC2012数据集为例,数据集应包含如下结构:
```bash
data
└──ILSVRC2012
├── train
├── train_list.txt
├── val
└── val_list.txt
```
3)切换到`models/PaddleSlim/quant_low_level_api`目录下,修改`run_quant.sh`内容,即将**data_dir**设置为第2)步所准备的数据集路径。最后,执行`run_quant.sh`脚本即可进行量化训练。
### 2.1 量化训练Low-Level API使用小结:
* 参照[quant.py](quant.py)文件的内容,总结使用量化训练Low-Level API的方法如下:
```python
#startup_program = fluid.Program()
#train_program = fluid.Program()
#train_cost = build_program(
# main_prog=train_program,
# startup_prog=startup_program,
# is_train=True)
#build_program(
# main_prog=test_program,
# startup_prog=startup_program,
# is_train=False)
#test_program = test_program.clone(for_test=True)
# The above pseudo code is used to build up the model.
# ---------------------------------------------------------------------------------
# The following code are part of Quantization Aware Training logic:
# 0) Convert Programs to IrGraphs.
main_graph = IrGraph(core.Graph(train_program.desc), for_test=False)
test_graph = IrGraph(core.Graph(test_program.desc), for_test=True)
# 1) Make some quantization transforms in the graph before training and testing.
# According to the weight and activation quantization type, the graph will be added
# some fake quantize operators and fake dequantize operators.
transform_pass = QuantizationTransformPass(
scope=fluid.global_scope(), place=place,
activation_quantize_type=activation_quant_type,
weight_quantize_type=weight_quant_type)
transform_pass.apply(main_graph)
transform_pass.apply(test_graph)
# Compile the train_graph for training.
binary = fluid.CompiledProgram(main_graph.graph).with_data_parallel(
loss_name=train_cost.name, build_strategy=build_strategy)
# Convert the transformed test_graph to test program for testing.
test_prog = test_graph.to_program()
# For training
exe.run(binary, fetch_list=train_fetch_list)
# For testing
exe.run(program=test_prog, fetch_list=test_fetch_list)
# 2) Freeze the graph after training by adjusting the quantize
# operators' order for the inference.
freeze_pass = QuantizationFreezePass(
scope=fluid.global_scope(),
place=place,
weight_quantize_type=weight_quant_type)
freeze_pass.apply(test_graph)
# 3) Convert the weights into int8_t type.
# [This step is optional.]
convert_int8_pass = ConvertToInt8Pass(scope=fluid.global_scope(), place=place)
convert_int8_pass.apply(test_graph)
# 4) Convert the freezed graph for paddle-mobile execution.
# [This step is optional. But, if you execute this step, you must execute the step 3).]
mobile_pass = TransformForMobilePass()
mobile_pass.apply(test_graph)
```
* [run_quant.sh](run_quant.sh)脚本中的命令配置详解:
```bash
--model:指定量化训练的模型,如MobileNet、ResNet50。
--pretrained_fp32_model:指定预训练float32模型参数的位置。
--checkpoint:指定模型断点训练的checkpoint路径。若指定了checkpoint路径,则不应该再指定pretrained_fp32_model路径。
--use_gpu:选择是否使用GPU训练。
--data_dir:指定训练数据集和验证数据集的位置。
--batch_size:设置训练batch size大小。
--total_images:指定训练数据图像的总数。
--class_dim:指定类别总数。
--image_shape:指定图像的尺寸。
--model_save_dir:指定模型保存的路径。
--lr_strategy:学习率衰减策略。
--num_epochs:训练的总epoch数。
--lr:初始学习率,指定预训练模型参数进行fine-tune时一般设置一个较小的初始学习率。
--act_quant_type:激活量化类型,可选abs_max, moving_average_abs_max, range_abs_max。
--wt_quant_type:权重量化类型,可选abs_max, channel_wise_abs_max。
```
> **备注:** 量化训练结束后,用户可在其指定的模型保存路径下看到float、int8和mobile三个目录。下面对三个目录下保存的模型特点进行解释说明:
> - **float目录:** 参数范围为int8范围但参数数据类型为float32的量化模型。
> - **int8目录:** 参数范围为int8范围且参数数据类型为int8的量化模型。
> - **mobile目录:** 参数特点与int8目录相同且兼容[paddle-mobile](https://github.com/PaddlePaddle/paddle-mobile)的量化模型。
>
> **注意:** 目前PaddlePaddle框架在Server端只支持使用float目录下的量化模型做预测。
......@@ -131,7 +131,7 @@ def net_config(image, label, model, args):
avg_cost = avg_cost0 + 0.3 * avg_cost1 + 0.3 * avg_cost2
acc_top1 = fluid.layers.accuracy(input=out0, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out0, label=label, k=5)
out = out2
out = out0
else:
out = model.net(input=image, class_dim=class_dim)
cost = fluid.layers.cross_entropy(input=out, label=label)
......
......@@ -4,6 +4,8 @@
root_url="http://paddle-imagenet-models-name.bj.bcebos.com"
MobileNetV1="MobileNetV1_pretrained.zip"
ResNet50="ResNet50_pretrained.zip"
GoogleNet="GoogleNet_pretrained.tar"
data_dir='Your image dataset path, e.g. ILSVRC2012'
pretrain_dir='../pretrain'
if [ ! -d ${pretrain_dir} ]; then
......@@ -22,6 +24,11 @@ if [ ! -f ${ResNet50} ]; then
unzip ${ResNet50}
fi
if [ ! -f ${GoogleNet} ]; then
wget ${root_url}/${GoogleNet}
tar xf ${GoogleNet}
fi
cd -
......@@ -32,7 +39,7 @@ python quant.py \
--model=MobileNet \
--pretrained_fp32_model=${pretrain_dir}/MobileNetV1_pretrained \
--use_gpu=True \
--data_dir=../data/ILSVRC2012 \
--data_dir=${data_dir} \
--batch_size=256 \
--total_images=1281167 \
--class_dim=1000 \
......@@ -50,7 +57,7 @@ python quant.py \
# --model=ResNet50 \
# --pretrained_fp32_model=${pretrain_dir}/ResNet50_pretrained \
# --use_gpu=True \
# --data_dir=../data/ILSVRC2012 \
# --data_dir=${data_dir} \
# --batch_size=128 \
# --total_images=1281167 \
# --class_dim=1000 \
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册