AdaptiveAvgPool2D 不支持 onnx 导出，自定义一个类代替 AdaptiveAvgPool2D

article/2025/8/28 17:04:51

AdaptiveAvgPool2D 不支持 onnx 导出，导出过程会告诉你，onnx不支持那个动态操作巴拉巴拉

我用的是 pp_liteseg 导出为 onnx 模型，都一样，paddle和Torch的 Adaptive Pool2D 都是动态的，onnx暂时都不支持，我根据下述公式，将 AdaptiveAvgPool2D 改了，可以用onnx导出了，但是需要指定一下原图的大小，和输出图的大小

$\begin{aligned} hstart &= floor(i * H_{in} / H_{out})\\ hend &= ceil((i + 1) * H_{in} / H_{out})\\ wstart &= floor(j * W_{in} / W_{out}) \\ wend &= ceil((j + 1) * W_{in} / W_{out}) \\ Output(i ,j) &= \frac{\sum Input[hstart:hend, wstart:wend]}{(hend - hstart) * (wend - wstart)} \end{aligned}$
参考自：
AdaptiveAvgPool2D

class CostomAdaptiveAvgPool2D(nn.Layer):def __init__(self, output_size, input_size):super(CostomAdaptiveAvgPool2D, self).__init__()self.output_size = output_sizeself.input_size  = input_sizedef forward(self, x):H_in,  W_in  = self.input_sizeH_out, W_out = [self.output_size, self.output_size] \if isinstance(self.output_size, int) \else self.output_sizeout_i = []for i in range(H_out):out_j = []for j in range(W_out):hs = int(np.floor(i * H_in / H_out))he = int(np.ceil((i+1) * H_in / H_out))ws = int(np.floor(j * W_in / W_out))we = int(np.ceil((j+1) * W_in / W_out))# print(hs, he, ws, we)kernel_size = [he-hs, we-ws]out = F.avg_pool2d(x[:, :, hs:he, ws:we], kernel_size) out_j.append(out)out_j = paddle.concat(out_j, -1)out_i.append(out_j)out_i = paddle.concat(out_i, -2)return out_i

用 CostomAdaptiveAvgPool2D 调换 AdaptiveAvgPool2D，并且指定原图和输出图的大小即可

在这里插入图片描述

[14,14] 就是输入图的大小，我这个pp_liteseg 输入shape是 $224\times 224$ ，所以上一层的shape是

[-1, 512, 14, 14]

然后一次过三个AdaptiveAvgPool2D 变成

[-1, 512,  1,   1]
[-1, 512,  2,   2]
[-1, 512,  4,   4]

然后 paddle2onnx 导出即可

paddle2onnx --model_dir . \--model_filename model.pdmodel \--params_filename model.pdiparams \--opset_version 11 \--save_file output.onnx

最后测试一下：

import paddle
import numpy as np
import paddle.nn as nn
import paddle.nn.functional as Fclass CostomAdaptiveAvgPool2D(nn.Layer):def __init__(self, output_size, input_size):super(CostomAdaptiveAvgPool2D, self).__init__()self.output_size = output_sizeself.input_size  = input_sizedef forward(self, x):H_in,  W_in  = self.input_sizeH_out, W_out = [self.output_size, self.output_size] \if isinstance(self.output_size, int) \else self.output_sizeout_i = []for i in range(H_out):out_j = []for j in range(W_out):hs = int(np.floor(i * H_in / H_out))he = int(np.ceil((i+1) * H_in / H_out))ws = int(np.floor(j * W_in / W_out))we = int(np.ceil((j+1) * W_in / W_out))# print(hs, he, ws, we)kernel_size = [he-hs, we-ws]out = F.avg_pool2d(x[:, :, hs:he, ws:we], kernel_size) out_j.append(out)out_j = paddle.concat(out_j, -1)out_i.append(out_j)out_i = paddle.concat(out_i, -2)return out_i# 输入
input_data = np.random.rand(1, 3, 128, 128)
x = paddle.to_tensor(input_data)adaptive_avg_pool = paddle.nn.AdaptiveAvgPool2D(output_size=1)
pool_out = adaptive_avg_pool(x = x)
costom_avg_pool = CostomAdaptiveAvgPool2D(output_size=(1, 1), input_size=(128, 128))
costom_pool_out = costom_avg_pool(x = x)

>>> costom_pool_out
Tensor(shape=[1, 3, 1, 1], dtype=float64, place=Place(gpu:0), stop_gradient=True,[[[[0.49809119]],[[0.49674173]],[[0.50111742]]]])>>> pool_out
Tensor(shape=[1, 3, 1, 1], dtype=float64, place=Place(gpu:0), stop_gradient=True,[[[[0.49809119]],[[0.49674173]],[[0.50111742]]]])