Inference of densenet161-float32.onnx and densenet161-float16.onnx got two problems.
The first one refers to this issue #386
Since opset 11, 'pads' and 'value' have been moved from attributes to inputs, and currently the TranslatePadOp of nnfusion onnx frontend parses onnx opset as 11, but the opset version of superbench to generate the model is 9.
Solution: processing the tranlatepadop differently depending on the size of the attribute.
The second one refers to this issue #434
cuda does't implement operator overload for datatype half.
Solution: rewrite the cuda::BatchNormNCHW::emit_function_body() to solve this problem.
Inference of densenet161-float32.onnx and densenet161-float16.onnx got two problems.
The first one refers to this issue #386
Since opset 11, 'pads' and 'value' have been moved from attributes to inputs, and currently the TranslatePadOp of nnfusion onnx frontend parses onnx opset as 11, but the opset version of superbench to generate the model is 9.
Solution: processing the tranlatepadop differently depending on the size of the attribute.
The second one refers to this issue #434
cuda does't implement operator overload for datatype half.
Solution: rewrite the cuda::BatchNormNCHW::emit_function_body() to solve this problem.
Now these two network already tested pass.