一、问题描述
单算子模型转换样例太少，遇到未提供样例的算子该如何书写规范的描述文件呢？
https://www.hiascend.com/document/detail/zh/canncommercial/80RC2/devaids/auxiliarydevtool/atlasatc_16_0034.html?sub_id=%2Fzh%2Fcanncommercial%2F80RC2%2Fdevaids%2Fauxiliarydevtool%2Fatlasatc_16_0055.html
（1）Concat算子描述文件该如何规范化书写？
由该图所示，Concat的有两个输入，一个是list of tensor，另一个是记录拼接维度的常数
而在书写算子描述文件中input_desc这一键值对时，如何构造list of tensor类型的TensorDesc数组？给的样例似乎都是tensor，TensorDesc数组type值的规范未看到符合条件的
（2）Reshape算子描述文件该如何规范化书写？
(3)疑问：
a. 遍历onnx模型中的所有算子，每个单算子对应生成Ascend算子描述文件，并且都能够成功转换和推理，那么该onnx模型是否就能适配到Ascend设备？
b. 如果遇到未公开书写规范的算子，应该如何准确地书写Ascend算子描述文件呢？只参照Ascend算子描述文档吗？有没有一些自动化的方法？
二、测试步骤
（1）将Concat算子“x”输入写成两个TensorDesc数组，则会报输入个数未对齐的错误
（2）将一个[1,255,80,80]reshape成[1,3，85，80，80]
reshape算子描述文件书写如下：
三、日志信息
(1)concat单算子转换报错如下：
ATC start working now, please wait for a moment.
...
ATC run failed, Please check the detail log, Try 'atc --help' for more information
E14001: 2024-09-06-10:11:42.192.405 Argument [inputs size[2]] for Op [exp_concat, optype [Concat]], is invalid. Reason: tensor size is [3].
(2)reshape单算子转换报错如下：
ATC start working now, please wait for a moment.
...
ATC run failed, Please check the detail log, Try 'atc --help' for more information
EZ9999: Inner Error!
EZ9999: 2024-09-06-10:26:44.852.873 Call InferShapeAndType for node:Reshape(Reshape) failed[FUNC:Infer][FILE:infershape_pass.cc][LINE:113]
TraceBack (most recent call last):
process pass InferShapePass on node:Reshape failed, ret:4294967295[FUNC:RunPassesOnNode][FILE:base_pass.cc][LINE:563]
build graph failed, graph id:0, ret:1343242270[FUNC:BuildModelWithGraphId][FILE:ge_generator.cc][LINE:1608]
点赞
收藏
回复
分享
举报
已解决
一、问题描述
单算子模型转换样例太少，遇到未提供样例的算子该如何书写规范的描述文件呢？
https://www.hiascend.com/document/detail/zh/canncommercial/80RC2/devaids/auxiliarydevtool/atlasatc_16_0034.html?sub_id=%2Fzh%2Fcanncommercial%2F80RC2%2Fdevaids%2Fauxiliarydevtool%2Fatlasatc_16_0055.html
（1）Concat算子描述文件该如何规范化书写？
由该图所示，Concat的有两个输入，一个是list of tensor，另一个是记录拼接维度的常数
而在书写算子描述文件中input_desc这一键值对时，如何构造list of tensor类型的TensorDesc数组？给的样例似乎都是tensor，TensorDesc数组type值的规范未看到符合条件的
（2）Reshape算子描述文件该如何规范化书写？
(3)疑问：
a. 遍历onnx模型中的所有算子，每个单算子对应生成Ascend算子描述文件，并且都能够成功转换和推理，那么该onnx模型是否就能适配到Ascend设备？
b. 如果遇到未公开书写规范的算子，应该如何准确地书写Ascend算子描述文件呢？只参照Ascend算子描述文档吗？有没有一些自动化的方法？
二、测试步骤
（1）将Concat算子“x”输入写成两个TensorDesc数组，则会报输入个数未对齐的错误
[ { "op": "Concat", "name": "exp_concat", "input_desc": [ { "name": "x", "format": "NCHW", "shape": [1, 32, 160, 160], "type": "float32" }, { "name": "x", "format": "NCHW", "shape": [1, 32, 160, 160], "type": "float32" }, { "name": "concat_dim", "format": "ND", "is_const": true, "const_value": 1, "shape" : [], "type": "int32" } ], "output_desc": [ { "format": "NCHW", "shape": [1, 64, 160, 160], "type": "float32" } ], "attr": [ { "name": "N", "type": "int", "value": 1 } ] } ]（2）将一个[1,255,80,80]reshape成[1,3，85，80，80]
reshape算子描述文件书写如下：
[ { "op": "Reshape", "input_desc": [ { "format": "NCHW", "shape": [1, 255, 80, 80], "type": "float32" }, { "format": "ND", "shape": [5], "type": "int32", "const_value": [1, 3, 85, 80, 80], "is_const": true } ], "output_desc": [ { "format": "NC1HWC0", "shape": [1, 3, 85, 80, 80], "type": "float32" } ], "attr": [ { "name": "axis", "type": "int", "value": 1 }, { "name": "num_axes", "type": "int", "value": 1 } ] } ]三、日志信息
(1)concat单算子转换报错如下：
ATC start working now, please wait for a moment.
...
ATC run failed, Please check the detail log, Try 'atc --help' for more information
E14001: 2024-09-06-10:11:42.192.405 Argument [inputs size[2]] for Op [exp_concat, optype [Concat]], is invalid. Reason: tensor size is [3].
(2)reshape单算子转换报错如下：
ATC start working now, please wait for a moment.
...
ATC run failed, Please check the detail log, Try 'atc --help' for more information
EZ9999: Inner Error!
EZ9999: 2024-09-06-10:26:44.852.873 Call InferShapeAndType for node:Reshape(Reshape) failed[FUNC:Infer][FILE:infershape_pass.cc][LINE:113]
TraceBack (most recent call last):
process pass InferShapePass on node:Reshape failed, ret:4294967295[FUNC:RunPassesOnNode][FILE:base_pass.cc][LINE:563]
build graph failed, graph id:0, ret:1343242270[FUNC:BuildModelWithGraphId][FILE:ge_generator.cc][LINE:1608]