Adds support for emitting darknet models

[fran6co]

Implements the following layers for darknet:

• Convolution
• Relu
• LeakyRelu
• PRelu (not currently supported by upstream)
• Max pooling
• Avg pooling
• Concat
• Softmax
• Squeeze
• Flatten
• Dropout
• FullyConnected

Having some issues with the testing side, some cases the SNR doesn't agree with the threshold.

[namizzz]

Hi @fran6co ,thanks very much! Have you test the accuracy of the conversion to darknet? We have implemented a set of test tools and error tolerance . Could you please add some test scripts to test the accuracy of your code ?

[fran6co]

Hi @namizzz, I couldn't find a darknet python package for testing it using MMdnn scripts. It's a simple C library so in theory is just enough to have the library somewhere and use ctypes, do you have suggestions how to integrate it to MMdnn?

[namizzz]

Hi @fran6co ,yes! we have https://github.com/Microsoft/MMdnn/blob/master/mmdnn/conversion/examples/darknet/. You can use it.

[kitstar]

Hi @fran6co Thanks for the contributeing! Do you have time to finish some unimplement ops in darknet emitter, like "Flatten", "FullyConnected" and "Dropout"? Thanks!

[fran6co]

@kitstar yes, I'm slowly working through the errors in the tests. I'm not sure why it's failing on the SNR for some of them.

[fran6co]

@kitstar just implemented those extra layers but it's still failing, it seems to be failing even worse now but because of some network issues instead.