提交 9bcff00c 编写于 作者: B Bai Yifan 提交者: whs

Add distillation tutorial (#20)

上级 fdb09f05
本示例将介绍如何使用PaddleSlim蒸馏接口来对模型进行蒸馏训练。
## 接口介绍
请参考[蒸馏API文档](https://paddlepaddle.github.io/PaddleSlim/api/single_distiller_api/)
## PaddleSlim蒸馏训练流程
一般情况下,模型参数量越多,结构越复杂,其性能越好,但运算量和资源消耗也越大。**知识蒸馏** 就是一种将大模型学习到的有用信息(Dark Knowledge)压缩进更小更快的模型,而获得可以匹敌大模型结果的方法。
在本示例中精度较高的大模型被称为teacher,精度稍逊但速度更快的小模型被称为student。
### 1. 定义student_program
```python
student_program = fluid.Program()
student_startup = fluid.Program()
with fluid.program_guard(student_program, student_startup):
image = fluid.data(
name='image', shape=[None] + [3, 224, 224], dtype='float32')
label = fluid.data(name='label', shape=[None, 1], dtype='int64')
# student model definition
model = MobileNet()
out = model.net(input=image, class_dim=1000)
cost = fluid.layers.cross_entropy(input=out, label=label)
avg_cost = fluid.layers.mean(x=cost)
```
### 2. 定义teacher_program
在定义好`teacher_program`后,可以一并加载训练好的pretrained_model。
`teacher_program`内需要加上`with fluid.unique_name.guard():`,保证teacher的变量命名不被`student_program`影响,从而能够正确地加载预训练参数。
```python
teacher_program = fluid.Program()
teacher_startup = fluid.Program()
with fluid.program_guard(teacher_program, teacher_startup):
with fluid.unique_name.guard():
image = fluid.data(
name='data', shape=[None] + [3, 224, 224], dtype='float32')
# teacher model definition
teacher_model = ResNet()
predict = teacher_model.net(image, class_dim=1000)
exe.run(teacher_startup)
def if_exist(var):
return os.path.exists(
os.path.join("./pretrained", var.name)
fluid.io.load_vars(
exe,
"./pretrained",
main_program=teacher_program,
predicate=if_exist)
```
### 3.选择特征图
定义好`student_program``teacher_program`后,我们需要从中两两对应地挑选出若干个特征图,留待后续为其添加知识蒸馏损失函数。
```python
# get all student variables
student_vars = []
for v in student_program.list_vars():
try:
student_vars.append((v.name, v.shape))
except:
pass
print("="*50+"student_model_vars"+"="*50)
print(student_vars)
# get all teacher variables
teacher_vars = []
for v in teacher_program.list_vars():
try:
teacher_vars.append((v.name, v.shape))
except:
pass
print("="*50+"teacher_model_vars"+"="*50)
print(teacher_vars)
```
### 4. 合并Program(merge)
PaddlePaddle使用Program来描述计算图,为了同时计算student和teacher两个Program,这里需要将其两者合并(merge)为一个Program。
merge过程操作较多,具体细节请参考[merge API文档](https://paddlepaddle.github.io/PaddleSlim/api/single_distiller_api/#merge)
```python
data_name_map = {'data': 'image'}
student_program = merge(teacher_program, student_program, data_name_map, place)
```
### 5.添加蒸馏loss
在添加蒸馏loss的过程中,可能还会引入部分变量(Variable),为了避免命名重复这里可以使用`with fluid.name_scope("distill"):`为新引入的变量加一个命名作用域。
另外需要注意的是,merge过程为`teacher_program`的变量统一加了名称前缀,默认是`"teacher_"`, 这里在添加`l2_loss`时也要为teacher的变量加上这个前缀。
```python
with fluid.program_guard(student_program, student_startup):
with fluid.name_scope("distill"):
distill_loss = l2_loss('teacher_bn5c_branch2b.output.1.tmp_3',
'depthwise_conv2d_11.tmp_0', student_program)
distill_weight = 1
loss = avg_cost + distill_loss * distill_weight
opt = create_optimizer()
opt.minimize(loss)
exe.run(student_startup)
```
至此,我们就得到了用于蒸馏训练的`student_program`,后面就可以使用一个普通program一样对其开始训练和评估。
......@@ -7,11 +7,12 @@ nav:
- 量化训练: tutorials/quant_aware_demo.md
- Embedding量化: tutorials/quant_embedding_demo.md
- SA搜索: tutorials/nas_demo.md
- 知识蒸馏: tutorials/distillation_demo.md
- API:
- 量化: api/quantization_api.md
- 剪枝与敏感度: api/prune_api.md
- 模型分析: api/analysis_api.md
- 蒸馏: api/single_distiller_api.md
- 知识蒸馏: api/single_distiller_api.md
- SA搜索: api/nas_api.md
- 搜索空间: api/search_space.md
- 硬件延时评估表: table_latency.md
......
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