Add Post-Training Quantization and export function in dygraph mode

[wanghaoshuang]

PR types

New features

PR changes

APIs

Describe

1. Add Post-Training Quantization
   1.1 Abstract some functions from QAT to Quantization class
   1.2 Add Post-Training Quantization by extending Quantization class
   1.3 Add observers for PTQ
   1.4 Add unittest for PTQ
2. Add exporting function for QAT and PTQ

TODO

动态图量化功能重构子任务：

• Configure: Add configure of quantization for dynamic graph #48000
• QAT: Add API for quantization-aware training in dygraph mode #49398
• PTQ & Export：Add Post-Training Quantization and export function in dygraph mode #50107
• Stub: Add stub class for quantization #50510
• More quanters & observers:

PTQ示例

原始模型：

LeNetDygraph(
  (features): Sequential(
    (0): Conv2D(1, 6, kernel_size=[3, 3], padding=1, data_format=NCHW)
    (1): ReLU()
    (2): MaxPool2D(kernel_size=2, stride=2, padding=0)
    (3): Conv2D(6, 16, kernel_size=[5, 5], data_format=NCHW)
    (4): ReLU()
    (5): MaxPool2D(kernel_size=2, stride=2, padding=0)
  )
  (fc): Sequential(
    (0): Linear(in_features=400, out_features=120, dtype=float32)
    (1): Linear(in_features=120, out_features=84, dtype=float32)
    (2): Linear(in_features=84, out_features=10, dtype=float32)
  )
)

转换为PTQ模型：

LeNetDygraph(
  (features): Sequential(
    (0): QuantedConv2D(
      (weight_quanter): AbsmaxObserverLayer()
      (activation_quanter): AbsmaxObserverLayer()
    )
    (1): ObserveWrapper(
      (_observer): AbsmaxObserverLayer()
      (_observed): ReLU()
    )
    (2): ObserveWrapper(
      (_observer): AbsmaxObserverLayer()
      (_observed): MaxPool2D(kernel_size=2, stride=2, padding=0)
    )
    (3): QuantedConv2D(
      (weight_quanter): AbsmaxObserverLayer()
      (activation_quanter): AbsmaxObserverLayer()
    )
    (4): ObserveWrapper(
      (_observer): AbsmaxObserverLayer()
      (_observed): ReLU()
    )
    (5): ObserveWrapper(
      (_observer): AbsmaxObserverLayer()
      (_observed): MaxPool2D(kernel_size=2, stride=2, padding=0)
    )
  )
  (fc): Sequential(
    (0): QuantedLinear(
      (weight_quanter): AbsmaxObserverLayer()
      (activation_quanter): AbsmaxObserverLayer()
    )
    (1): QuantedLinear(
      (weight_quanter): AbsmaxObserverLayer()
      (activation_quanter): AbsmaxObserverLayer()
    )
    (2): QuantedLinear(
      (weight_quanter): AbsmaxObserverLayer()
      (activation_quanter): AbsmaxObserverLayer()
    )
  )
)

模型导出示例

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[ceci3]

LGTM

[chenwhql]

LGTM for setup.py.in

[zhangbo9674]

LGTM for setup

[jeff41404]

LGTM for APIs

[jzhang533]

LGTM

[zhangbo9674]

Approve for setup