Pytorch量化深度可分离卷积的坑

使用pytorch自带的量化方法。

动态量化方法无法量化卷积层，代码不会报错，只是不会对做量化操作

quantized_model = torch.quantization.quantize_dynamic(model, {nn.Conv2d}, dtype=torch.qint8)

 只能量化以下层

        if dtype == torch.qint8:
            qconfig_spec = {
                nn.Linear : default_dynamic_qconfig,
                nn.LSTM : default_dynamic_qconfig,
                nn.GRU : default_dynamic_qconfig,
                nn.LSTMCell : default_dynamic_qconfig,
                nn.RNNCell : default_dynamic_qconfig,
                nn.GRUCell : default_dynamic_qconfig,
            }

采用静态量化方法到转化模型这一步

model_int8 = torch.quantization.convert(model_fused_prepared)

由于深度可分离卷积模型定义如下

        def conv_dw(inp, oup, stride):
            return nn.Sequential(
                nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=True),
                nn.ReLU(inplace=True),
                nn.Conv2d(inp, oup, 1, 1, 0, bias=True),
                nn.ReLU(inplace=True),
            )

报错，深度卷积层的groups为3，所以这样也不行。

RuntimeError: Quantized cudnn conv2d is currenty limited to groups = 1; received groups =3