关于pytorch转onnx经常出现的问题

1. F.avg_pool2d以及F.max_pool2d的转换

直接转会出现以下的错误：

RuntimeError: Failed to export an ONNX attribute 'onnx::Gather', since it's not constant, please try to make things (e.g., kernel size) static if possible

解决办法：

• nn.AdaptiveAvgPool2d(1)替换 F.avg_pool2d
• nn.AdaptiveMaxPool2d(1)替换F.max_pool2d

2. 转换过程种不能有动态的非tensor的数值

我的模型有一段是需要将scale在channel维度split成不均等的份数，代码是这么写的：

        scale = F.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3)
        start = 0
        out = []
        for x in inputs:
            x = x * scale[:,start:start+x.size(1),:,:]
            start = start + x.size(1)
            out.append(x)

然后再转onnx的时候会出现：

RuntimeError: Unsupported: ONNX export of Slice with dynamic inputs. DynamicSlice is a deprecated experimental op. Please use statically allocated variables or export to a higher opset version.

主要就是start是一个int的数值，在for循环种是迭代的。最后修改代码为如下，即固定好每一个split的份数。

        x1, x2, x3, x4, x5 = inputs 
        x1 = x1 *scale[:,self.start[0]:self.end[0],:,:]
        x2 = x2* scale[:,self.start[1]:self.end[1],:,:]
        x3 = x3* scale[:,self.start[2]:self.end[2],:,:]
        x4 = x4* scale[:,self.start[3]:self.end[3],:,:]
        x5 = x5* scale[:,self.start[4]:self.end[4],:,:]
        return x1, x2, x3, x4, x5