tensorRT 遇到的一个问题

onnx转TensorRT的问题

TensorRT7.0.1.1转换没出问题

TensorRT8.2.4.2一直报下面的错误

4: [shapeCompiler.cpp::nvinfer1::builder::DynamicSlotBuilder::evaluateShapeChecks::832] Error Code 4: Internal Error (kOPT values for profile 0 violate shape constraints: condition ‘==’ violated. IAssertionLayer (Unnamed Layer* 5) [Assertion]: condition[0] is false: (EQUAL (# 2 (SHAPE input0)) (# 2 (SHAPE input1))). For input: ‘input0’ all named dimensions that share the same name must be equal. Note: Named dimensions were present on the following axes: 2 (name: ‘height’), 2 (name: ‘height’))

原因是：

input0 input1 input2 在实际使用的时候，width height大小是/2递减的，而在保存onnx的时候，命名的时候是相同的（上一辈留下的问题，心里苦）

 TensorRT8.2.4.2中要求，相同name的维度需要一致，input0中叫height，如果在input1中也有相同名字的height，那么这两个大小必须一致，不然就会报错。

解决方法，生成的onnx的时候，分别命名即可：

增加onnx直接修改名字的代码：

import onnx
import torch
import argparse
import os


def Test_Onnx_Change_Param_Name(onnx_model):
    for input in onnx_model.graph.input:
            dim1 = input.type.tensor_type.shape.dim[2]
            dim2 = input.type.tensor_type.shape.dim[3]
            dim1.dim_param = "height" + input.name[-1]
            dim2.dim_param = "width" + input.name[-1]

    for output in onnx_model.graph.output:
            dim1 = output.type.tensor_type.shape.dim[2]
            dim2 = output.type.tensor_type.shape.dim[3]
            dim1.dim_param = "height" + output.name[-1]
            dim2.dim_param = "width" + output.name[-1]


def apply(transform, infile, outfile):
    model = onnx.load(infile)
    transform(model)
    onnx.save(model, outfile)





parser = argparse.ArgumentParser(description='Configurations for Change Onnx Name')
parser.add_argument('--onnx_path', type=str, default=r'D:\networks\deblur\rst/model.onnx', help='src onnx path')
parser.add_argument('--save_new_path', type=str, default=r'D:\networks\deblur\rst/modified.onnx', help='new onnx path')
args = parser.parse_args()


if __name__ == "__main__":
    apply(Test_Onnx_Change_Param_Name, args.onnx_path, args.save_new_path)