Support user-defined activation/weight quantize and preprocess.

[baiyfbupt]

PR types

New features

PR changes

APIs

Describe

Support user-defined quantification and preprocessing methods, such as PACT.
QAT model without using PACT

-----------------------------------------------------------------------------------------------
           Layer (type)                 Input Shape          Output Shape         Param #
===============================================================================================
     FakeQuantMovingAverage-1         [[1, 1, 28, 28]]      [1, 1, 28, 28]           3
FakeChannelWiseQuantDequantAbsMax-1    [[6, 1, 3, 3]]        [6, 1, 3, 3]            6
         QuantizedConv2D-1            [[1, 1, 28, 28]]      [1, 6, 28, 28]          60
             Pool2D-1                 [[1, 6, 28, 28]]      [1, 6, 14, 14]           0
     FakeQuantMovingAverage-2         [[1, 6, 14, 14]]      [1, 6, 14, 14]           3
FakeChannelWiseQuantDequantAbsMax-2   [[16, 6, 5, 5]]       [16, 6, 5, 5]           16
         QuantizedConv2D-2            [[1, 6, 14, 14]]     [1, 16, 10, 10]         2,416
             Pool2D-2                [[1, 16, 10, 10]]      [1, 16, 5, 5]            0
     FakeQuantMovingAverage-3            [[1, 400]]            [1, 400]              3
FakeChannelWiseQuantDequantAbsMax-3     [[400, 120]]          [400, 120]            120
         QuantizedLinear-1               [[1, 400]]            [1, 120]           48,120
     FakeQuantMovingAverage-4            [[1, 120]]            [1, 120]              3
FakeChannelWiseQuantDequantAbsMax-4     [[120, 84]]           [120, 84]             84
         QuantizedLinear-2               [[1, 120]]            [1, 84]            10,164
     FakeQuantMovingAverage-5            [[1, 84]]             [1, 84]               3
FakeChannelWiseQuantDequantAbsMax-5      [[84, 10]]            [84, 10]             10
         QuantizedLinear-3               [[1, 84]]             [1, 10]              850
===============================================================================================
Total params: 61,861
Trainable params: 61,610
Non-trainable params: 251
-----------------------------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.55
Params size (MB): 0.24
Estimated Total Size (MB): 0.79
-----------------------------------------------------------------------------------------------

PACT QAT model:

-----------------------------------------------------------------------------------------------
           Layer (type)                 Input Shape          Output Shape         Param #
===============================================================================================
              PACT-1                  [[1, 1, 28, 28]]      [1, 1, 28, 28]           1
     FakeQuantMovingAverage-1         [[1, 1, 28, 28]]      [1, 1, 28, 28]           3
FakeChannelWiseQuantDequantAbsMax-1    [[6, 1, 3, 3]]        [6, 1, 3, 3]            6
         QuantizedConv2D-1            [[1, 1, 28, 28]]      [1, 6, 28, 28]          60
             Pool2D-1                 [[1, 6, 28, 28]]      [1, 6, 14, 14]           0
              PACT-2                  [[1, 6, 14, 14]]      [1, 6, 14, 14]           1
     FakeQuantMovingAverage-2         [[1, 6, 14, 14]]      [1, 6, 14, 14]           3
FakeChannelWiseQuantDequantAbsMax-2   [[16, 6, 5, 5]]       [16, 6, 5, 5]           16
         QuantizedConv2D-2            [[1, 6, 14, 14]]     [1, 16, 10, 10]         2,416
             Pool2D-2                [[1, 16, 10, 10]]      [1, 16, 5, 5]            0
              PACT-3                     [[1, 400]]            [1, 400]              1
     FakeQuantMovingAverage-3            [[1, 400]]            [1, 400]              3
FakeChannelWiseQuantDequantAbsMax-3     [[400, 120]]          [400, 120]            120
         QuantizedLinear-1               [[1, 400]]            [1, 120]           48,120
              PACT-4                     [[1, 120]]            [1, 120]              1
     FakeQuantMovingAverage-4            [[1, 120]]            [1, 120]              3
FakeChannelWiseQuantDequantAbsMax-4     [[120, 84]]           [120, 84]             84
         QuantizedLinear-2               [[1, 120]]            [1, 84]            10,164
              PACT-5                     [[1, 84]]             [1, 84]               1
     FakeQuantMovingAverage-5            [[1, 84]]             [1, 84]               3
FakeChannelWiseQuantDequantAbsMax-5      [[84, 10]]            [84, 10]             10
         QuantizedLinear-3               [[1, 84]]             [1, 10]              850
===============================================================================================
Total params: 61,866
Trainable params: 61,615
Non-trainable params: 251
-----------------------------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.57
Params size (MB): 0.24
Estimated Total Size (MB): 0.81
-----------------------------------------------------------------------------------------------

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[juncaipeng]

weight_preprocess and act_preprocess 会保存到预测模型中，建议尝试对量化的层注册forward hook进行预处理。

用户定义的weight_quantize和act_quantize没办法在paddlelite和paddleinference中支持，这两个参数是否需要、是否暴露给用户，值得商榷。

[baiyfbupt]

weight_preprocess and act_preprocess 会保存到预测模型中，建议尝试对量化的层注册forward hook进行预处理。

用户定义的weight_quantize和act_quantize没办法在paddlelite和paddleinference中支持，这两个参数是否需要、是否暴露给用户，值得商榷。

• forward_pre_hook中进行预处理无法在反向传播时更新可学习参数，PACT等带有可学习参数的方法不适用

• 这里是留一个快速验证新型量化方法是否有效的接口，用户可以快速写一个量化方式进行量化训练实验验证，如果有效再考虑新增op。这里还不需要预测库支持

[juncaipeng]

LGTM