替换算子：图拓扑有误或额外算子增加

### What was the cause of the issue？
操作：将onnx图中的node替换为另外一种类型，之后使用onxxsim简化模型报错。
例：col2im替换为slice算子。
### Steps to reproduce
```
#### 1
Col2im_slice = graph.add_node(f'Slice_{node.name}',
               'Slice')
slice_shape = node['output_size']
starts_init = graph.add_initializer(f'{node.name}_starts', np.array([0, 0, 0, 0]))
ends_init = graph.add_initializer(f'{node.name}_ends', np.array([1, 192, slice_shape[0], slice_shape[1]]))
axes_init = graph.add_initializer(f'{node.name}_axes', np.array([0, 1, 2, 3]))
Col2im_slice.inputs = [node.inputs[0], f'{node.name}_starts', f'{node.name}_ends', f'{node.name}_axes']
Col2im_slice.outputs = [node.outputs[0]]
graph[f'{node.name}'] = Col2im_slice
```
```
#### 2
Col2im_slice = graph.add_node(f'Slice_{node.name}',
               'Slice')
slice_shape = node['output_size']
starts_init = graph.add_initializer(f'{node.name}_starts', np.array([0, 0, 0, 0]))
ends_init = graph.add_initializer(f'{node.name}_ends', np.array([1, 192, slice_shape[0], slice_shape[1]]))
axes_init = graph.add_initializer(f'{node.name}_axes', np.array([0, 1, 2, 3]))
Col2im_slice.inputs = [node.inputs[0], f'{node.name}_starts', f'{node.name}_ends', f'{node.name}_axes']
Col2im_slice.outputs = [node.outputs[0]]
graph[node.name].node.CopyFrom(Col2im_slice.node)
```
### Error information
```
#### 1
onnx.onnx_cpp2py_export.checker.ValidationError: Nodes in a graph must be topologically sorted, however input '388' of node:
name: Transpose_99 OpType: Transpose
 is not output of any previous nodes.
```
```
#### 2
onnx.onnx_cpp2py_export.checker.ValidationError: Nodes in a graph must be topologically sorted, however input 'Null' of node:
name: Slice_Col2im_98 OpType: Slice
 is not output of any previous nodes.
onnx的图中增加了额外node，其输入输出会报错。
```
相关onnx文件和代码链接如下：方法1的导出为volo_slice.onnx；方法2导出为zl_debug.onnx
链接：https://pan.baidu.com/s/1rH_g8JxhKuyICr8-hsSDow 
提取码：9z6g

镇亮/MagicONNX

内容风险标识

评论 (0)

镇亮/MagicONNX .gitee-modal { width: 500px !important; }

内容风险标识