Created by: gewen

As described in issue #42, the caffe2mdl script does not support input with height NOT EQUAL TO width. But does mdl self support this forward setting? Can it produce the right result? I have tested different input sizes on one fully convolutional network. I found that the generated data.bin is the same. And only the input/output shape in model.json is different. So if the answer to the above question is yes, can I modify the model.json by editing the input/output shape and use the same data.bin so that achieves a caffe2mdl conversion with height NOT EQUAL TO width?