模型静态量化时出错

[ZEROICEWANG]

我在对模型进行静态量化时，程序报错

报错信息如下

2021-05-20 15:02:44,389-INFO: Load model and set data loader ...
2021-05-20 15:02:47,652-INFO: Collect quantized variable names ...
2021-05-20 15:02:47,692-INFO: Preparation stage ...
2021-05-20 15:02:54,276-INFO: Run batch: 0
2021-05-20 15:02:54,277-INFO: Finish preparation stage, all batch:1
2021-05-20 15:02:54,348-INFO: Sampling stage ...
2021-05-20 15:03:58,629-INFO: Run batch: 0
2021-05-20 15:03:58,630-INFO: Finish sampling stage, all batch: 1
2021-05-20 15:03:58,637-INFO: Calculate KL threshold ...
2021-05-20 15:19:03,373-INFO: Update the program ...
Traceback (most recent call last):
  File "F:/graduation_project_data/PyCharm Project/CPD_MNV3_Paddle/slim/slim.py", line 136, in <module>
    batch_size=20)
  File "E:\Anaconda3\envs\Paddle\lib\site-packages\paddleslim\quant\quanter.py", line 407, in quant_post_static
    post_training_quantization.quantize()
  File "E:\Anaconda3\envs\Paddle\lib\site-packages\paddle\fluid\contrib\slim\quantization\post_training_quantization.py", line 382, in quantize
    self._update_program()
  File "E:\Anaconda3\envs\Paddle\lib\site-packages\paddle\fluid\contrib\slim\quantization\post_training_quantization.py", line 735, in _update_program
    freeze_pass.apply(graph)
  File "E:\Anaconda3\envs\Paddle\lib\site-packages\paddle\fluid\contrib\slim\quantization\quantization_pass.py", line 1167, in apply
    self._insert_post_channel_dequant_op(graph, op_node)
  File "E:\Anaconda3\envs\Paddle\lib\site-packages\paddle\fluid\contrib\slim\quantization\quantization_pass.py", line 1211, in _insert_post_channel_dequant_op
    original_var_name)
AssertionError: The scale of parameter conv2d_209.w_0 is not a list.

Process finished with exit code -1073740791 (0xC0000409)

在使用Nettron打开模型结构图时，对比发现时我写的一个多层特征融合处发生了错误，错误对应的卷积为self.conv5
模型静态图推理和动态图推理都没有异常，只有在量化的时候出现了这个问题，请问是哪里设置的有问题吗？

代码如下

class aggregation(nn.Layer):
    def __init__(self, channel):
        super(aggregation, self).__init__()
        self.relu = paddle.nn.ReLU(True)

        self.upsample = paddle.nn.UpsamplingBilinear2D(scale_factor=2)
        self.conv_upsample1 = nn.Conv2D(channel, channel, 3, padding=1)
        self.conv_upsample2 = nn.Conv2D(channel, channel, 3, padding=1)
        self.conv_upsample3 = nn.Conv2D(channel, channel, 3, padding=1)
        self.conv_upsample4 = nn.Conv2D(channel, channel, 3, padding=1)
        self.conv_upsample5 = nn.Conv2D(2 * channel, 2 * channel, 3, padding=1)

        self.conv_concat2 = nn.Conv2D(2 * channel, 2 * channel, 3, padding=1)
        self.conv_concat3 = nn.Conv2D(3 * channel, 3 * channel, 3, padding=1)
        self.conv4 = nn.Conv2D(3 * channel, 3 * channel, 3, padding=1)
        self.conv5 = nn.Conv2D(3 * channel, 1, 1)

    def forward(self, x1, x2, x3):
        # x1: 1/32 x2: 1/16 x3: 1/8
        x1_1 = x1
        x2_1 = self.conv_upsample1(self.upsample(x1)) * x2
        x3_1 = self.conv_upsample2(self.upsample(self.upsample(x1))) \
               * self.conv_upsample3(self.upsample(x2)) * x3

        x2_2 = paddle.concat((x2_1, self.conv_upsample4(self.upsample(x1_1))), 1)
        x2_2 = self.conv_concat2(x2_2)

        x3_2 = paddle.concat((x3_1, self.conv_upsample5(self.upsample(x2_2))), 1)
        x3_2 = self.conv_concat3(x3_2)

        x = self.conv4(x3_2)
        x = self.conv5(x)

        return x

[juncaipeng]

@ZEROICEWANG 从log看不出来具体问题，请问使用哪个版本的paddleslim和paddlepaddle？同时请提供一下复现单侧。

[ZEROICEWANG]

已经排查到问题出现的原因，在进行卷及操作时，若是out_channel为1，则会出现这个问题，大于1则不会出现这个问题，希望你们能够尽快修复这个bug

[ZEROICEWANG]

paddlepaddle-gpu 2.0.2.post100
paddleslim 2.0.0

[XGZhang11]

您好，bug已经修复，PaddlePaddle/Paddle#33753

[ZEROICEWANG]

您好，bug已经修复，PaddlePaddle/Paddle#33753

在paddle2.0+paddleslim2.0与paddle2.1+paddleslim2.1上都做了测试，还是存在该问题