onnx截取子模型，并修改模型的输出

最新推荐文章于 2024-10-17 16:43:54 发布

wozwdaqian

最新推荐文章于 2024-10-17 16:43:54 发布

阅读量2.1k

点赞数 1

文章标签：深度学习 pytorch 人工智能

本文链接：https://blog.youkuaiyun.com/qq_42001765/article/details/131766687

版权

该文展示了如何从一个预先训练的ONNX模型中提取部分，并添加新的卷积层。首先定义了一个包含多个卷积层的PyTorch模型并导出为ONNX格式，然后使用ONNX库提取特定节点，创建新的卷积节点，并将修改后的模型保存为新的ONNX文件。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

import torch 
import onnx
from onnx import helper, TensorProto
import numpy as np
 
class Model(torch.nn.Module): 
 
    def __init__(self): 
        super().__init__() 
        self.convs1 = torch.nn.Sequential(torch.nn.Conv2d(3, 3, 3), 
                                          torch.nn.Conv2d(3, 3, 3), 
                                          torch.nn.Conv2d(3, 3, 3)) 
        self.convs2 = torch.nn.Sequential(torch.nn.Conv2d(3, 3, 3), 
                                          torch.nn.Conv2d(3, 3, 3)) 
        self.convs3 = torch.nn.Sequential(torch.nn.Conv2d(3, 3, 3), 
                                          torch.nn.Conv2d(3, 3, 3)) 
        self.convs4 = torch.nn.Sequential(torch.nn.Conv2d(3, 3, 3), 
                                          torch.nn.Conv2d(3, 3, 3), 
                                          torch.nn.Conv2d(3, 3, 3)) 
    def forward(self, x): 
        x = self.convs1(x) 
        x1 = self.convs2(x) 
        x2 = self.convs3(x) 
        x = x1 + x2 
        x = self.convs4(x) 
        return x 
 
model = Model() 
input = torch.randn(1, 3, 20, 20) 
 
torch.onnx.export(model, input, 'whole_model.onnx') 

onnx.utils.extract_model('/home/zhengwei/my_jupyter/lab8-lpr/Models/lprnet.onnx', 'partial_model.onnx', ['input.1'], ['/Div_output_0'])

model = onnx.load('partial_model.onnx')

div_node = None

for i in model.graph.node:
    if i.op_type == "Div":
        div_node = i


conv_w_weight = np.random.random_sample((1, 64, 1, 1))
conv_w = helper.make_tensor("conv_w", TensorProto.FLOAT, (1, 64, 1, 1), conv_w_weight)
conv_node = onnx.helper.make_node(
    "Conv",
    inputs=[div_node.output[0], "conv_w"],
    outputs=["conv_node_output"],
    kernel_shape=[1, 1],
    # Default values for other attributes: strides=[1, 1], dilations=[1, 1], groups=1
)

model.graph.initializer.append(conv_w)
model.graph.node.append(conv_node)

new_output = onnx.helper.make_tensor_value_info('conv_node_output', onnx.TensorProto.FLOAT, [1, 1, 4, 18])
model.graph.output.extend([new_output])
model.graph.output.remove(model.graph.output[0])

onnx.save(model, 'tmp.onnx')