根据下面的reference将动态图模型转为静态图模型后

https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/dygraph_cn/TracedLayer_cn.html#tracedlayer

使用paddle2onnx转换时无法通过onnx的checker:

onnx.onnx_cpp2py_export.checker.ValidationError: Nodes in a graph must be topologically sorted, however input 't_14' of node:
input: "t_14" input: "t_15@reshape_y" output: "t_15" op_type: "Add"

代码如下：

import paddle
import paddle.fluid as fluid
import numpy as np
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, Linear

class DCNN(fluid.dygraph.Layer):
    def __init__(self, name_scope, num_classes=10):
        super(DCNN, self).__init__(name_scope)
        name_scope = self.full_name()
        # AlexNet与LeNet一样也会同时使用卷积和池化层提取图像特征
        # 与LeNet不同的是激活函数换成了‘relu’
        
        self.conv1 = Conv2D(num_channels=3, num_filters=32, filter_size=3, stride=1, padding=1, act='relu')
      

        self.pool1 = Pool2D(pool_size=2, pool_stride=2, pool_type='max')

        
        self.conv2 = Conv2D(num_channels=32, num_filters=16, filter_size=3, stride=1, padding=2, act='relu')

        self.pool2 = Pool2D(pool_size=2, pool_stride=2, pool_type='max')

        self.conv3 = Conv2D(num_channels=16, num_filters=8, filter_size=3, stride=1, padding=1, act='relu')
    
        self.pool3 = Pool2D(pool_size=2, pool_stride=2, pool_type='max')

        self.fc1 = Linear(input_dim=6272, output_dim=1568, act='relu')

        self.drop_ratio1 = 0.5

        self.fc2 = Linear(input_dim=1568, output_dim=128, act='relu')

        self.drop_ratio2 = 0.5

        self.fc3 = Linear(input_dim=128, output_dim=num_classes, act='relu')

        
    def forward(self, inputs,label=None):
        x = self.conv1(inputs)

        x = self.pool1(x)
        #print(x.numpy().shape)
        x = self.conv2(x)
        x = self.pool2(x)
        #print(x.numpy().shape)
        x = self.conv3(x)
        x = self.pool3(x)
        #print(x.numpy().shape)
        x = fluid.layers.reshape(x, [x.shape[0], -1])

        x = self.fc1(x)
        x= fluid.layers.dropout(x, self.drop_ratio1)
        x = self.fc2(x)
        x= fluid.layers.dropout(x, self.drop_ratio2)
        x = self.fc3(x)
        x = fluid.layers.softmax(x)
        if label is not None:
            acc = fluid.layers.accuracy(input=x, label=label)
            return x, acc
        else:
            return x

with fluid.dygraph.guard():
    model = DCNN("DCNN",num_classes = 39)
    #print(model) 
    #train(model)
    in_np = np.random.random([10, 3, 224, 224]).astype('float32')
    # 将numpy的ndarray类型的数据转换为Variable类型
    input_var = fluid.dygraph.to_variable(in_np)
    # 通过 TracerLayer.trace 接口将动态图模型转换为静态图模型
    out_dygraph, static_layer = TracedLayer.trace(model, inputs=[input_np])
    out_static_graph = static_layer([input_np])
    print(len(out_static_graph)) # 1
    print(out_static_graph[0].shape) # (2, 10)
    save_dirname = './saved_infer_model'
    # 将转换后的模型保存
    static_layer.save_inference_model(dirname='./saved_infer_model')