add ResNet graph construction and demo code

[BigRedT]

-examples/resnet_demo.py
-vis/pretrained_models/resnet

[kevjshih]

Since the demo is loading the pretrained model, shouldn't it be initializing the color mode to 'BGR'

[kevjshih]

Also I think a lot of the inference_fully_conv functions are identical to the ones in the inference.py, specifically bn and _max_pool. It would be better to simply import these from inference.py than to have duplicate implementations. You can also add a switch to the conv function to alternate between conv and atrous_conv to further reduce duplicate code.

[BigRedT]

color_mode refers to the format in which data is fed in. The pretrained model is wrapped in a class which reverses the channels if RGB is fed in. Look at models.py

[kevjshih]

Ah, I misread that line in models.py.

[BigRedT]

Yeah i can import bn, max_pool etc. Making a single conv with switch would require me to modify the code in many places and it's probably not worth the trouble.

[kevjshih]

That's fair. Also the only difference between model_fully_conv and model is a few lines pertaining to avg_pooling. It's easier to just add a flag in model.py to simply disable the avg_pool and return an error if get_avg_pool is called. You would also need to add a switch to choose which version of inference to call.

[BigRedT]

not really. there are more differences because i had to control the number of holes in atrous convs at each scale and also classification layer is made fully convolutional.

[kevjshih]

Right, but that doesn't affect model.py and model_fully_conv.py.

[kevjshih]

I don't know if this is bad practice, but you could basically do

if fully_conv:
import inference_fully_conv as inference
else:
import inference as inference

in the init function of the class if you had to

[BigRedT]

Ah i see. That would remove lot of duplication. I will fix it.

[kevjshih]

Sounds good. Another option is to use some class inheritance if you think these two files will severely diverge in the future but still share a subset of functions.

[BigRedT]

That's even better! Good call