Fix parse bug

[Sundrops]

#882

[lxgyChen]

@Sundrops 好像不起作用啊#882，还是报一样的错误。

[Sundrops]

It's ok on torch1.6.0

[lxgyChen]

我刚试过，在torch1.5.1中，当output_shape_str=['%500 : Float', '1, 256, 20, 20', ' ']时，在torch1.6.0中是['%500 : Float', '1:102400, 256:400, 20:20, 20:1', ' ``']，这种情况下在目前的mmdnn版本下也是能正确解析的。
但在我转换yolov5s的情况，见#882，输出的output_shape_str：

['%498 : Float', '1:102400, 256:400, 20:20, 20:1', ' '] ['%499 : Float', '1:102400, 256:400, 20:20, 20:1', ' '] ['%500 : Double', ' '] Traceback (most recent call last): File "./convertToIR.py", line 202, in <module> _main() File "./convertToIR.py", line 197, in _main ret = _convert(args) File "./convertToIR.py", line 97, in _convert parser = PytorchParser151(model, inputshape[0]) File "/home/user/anaconda3/envs/pytorch/lib/python3.7/site-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 533, in __init__ self.build_graph(input_shape) File "/home/user/anaconda3/envs/pytorch/lib/python3.7/site-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 92, in build_graph self.pytorch_graph.build(self.input_shape) File "/home/user/anaconda3/envs/pytorch/lib/python3.7/site-packages/mmdnn/conversion/pytorch/pytorch_graph.py", line 136, in build output_shape = [int(x.replace('!', '').split(':')[0]) for x in output_shape_str[1].split(',')] File "/home/user/anaconda3/envs/pytorch/lib/python3.7/site-packages/mmdnn/conversion/pytorch/pytorch_graph.py", line 136, in <listcomp> output_shape = [int(x.replace('!', '').split(':')[0]) for x in output_shape_str[1].split(',')] ValueError: invalid literal for int() with base 10: ' '

当output_shape_str=['%500 : Double', ' ']时，解析出来就是空字符串‘ ’了，转换为int型就报错了。这种情况怎么处理呢？
我把conversion/pytorch/pytorch_graph.py 大约在134行的

if len(output_shape_str) > 1 改成 if len(output_shape_str) > 2不会报错了，但是网络没有完全转换成功，只转换了第一层，而且提示PyTorch parser has not supported operator [onnx::Constant]. IR network strucuture may lost info.

[Sundrops]

@lxgyChen do you use Upsample layer in yolov5s? Upsample layer will be converted to

, %224 : Double() = onnx::Constant[value={2}](), scope: CustomizedNet/Upsample[Sampling30]
, %225 : Double() = onnx::Constant[value={2}](), scope: CustomizedNet/Upsample[Sampling30]
, %226 : Tensor = onnx::Unsqueeze[axes=[0]](%224), scope: CustomizedNet/Upsample[Sampling30]
, %227 : Tensor = onnx::Cast[to=1](%226), scope: CustomizedNet/Upsample[Sampling30]
, %228 : Tensor = onnx::Unsqueeze[axes=[0]](%225), scope: CustomizedNet/Upsample[Sampling30]
, %229 : Tensor = onnx::Cast[to=1](%228), scope: CustomizedNet/Upsample[Sampling30]
, %230 : Tensor = onnx::Constant[value= 1  1 [ CPUFloatType{2} ]](), scope: CustomizedNet/Upsample[Sampling30]
, %231 : Tensor = onnx::Concat[axis=0](%230, %227, %229), scope: CustomizedNet/Upsample[Sampling30]
, %232 : Tensor = onnx::Constant[value=[ CPUFloatType{0} ]](), scope: CustomizedNet/Upsample[Sampling30]
, %233 : Float(1:60768, 422:144, 12:12, 12:1) = onnx::Resize[coordinate_transformation_mode="asymmetric", cubic_coeff_a=-0.75, mode="nearest", nearest_mode="floor"](%223, %232, %231), scope: CustomizedNet/Upsampl
e[Sampling30] # /mnt/WXRC0020/release/gluoncvplus/.local/opt/anaconda3/envs/gp_v0.0.1/lib/python3.7/site-packages/torch/nn/functional.py:3143:0

And mmdnn does not support Constant op.

[lxgyChen]

@Sundrops 上面的output_shape_str=['%500 : Double', ' ']还没到Upsample,但也是一个Constant。YOLO v3 v4 v5都有Upsample层，那这样都不能转换了啊？

[Sundrops]

@lxgyChen As mentioned above, Upsample layer will be converted to onnx:: Constant and other layers. If you want to use Upsample, you should modify code to merge onnx:: Constant, ... onnx::Resize in onnx graph(https://github.com/microsoft/MMdnn/blob/master/mmdnn/conversion/pytorch/pytorch_graph.py#L231)

[lxgyChen]

@Sundrops 好的，我研究一下。怎么避免Constant Op呢，用已知的固定值代替可以吗？PyTorch导出为ONNX时可以选择Constant Folding，但是MMdnn不支持ONNX to IR。