Converting models to Caffe. about fb.resnet.torch

[colesbury]

Yes, you can reverse it:

running_std = running_var:add(self.eps):pow(0.5)

Or even better, add support for running_var to fb-caffe-exts 😃

[omkarparkhi]

Thanks. Ok. Let me try it first. If it works, we can think about a pull request on fb-caffe-exts :)

[omkarparkhi]

Sam, another quick question, how did you plot the training error curves? I couldn't find that being saved anywhere.

[colesbury]

I log a bunch of stats in JSON format and then plot using ipython and bokeh:

Around https://github.com/facebook/fb.resnet.torch/blob/master/main.lua#L61:

my_json_log(epoch, opt, trainer.optimState, {
      train_err = trainTop1,
      train_err5 = trainTop5,
      train_accuracy = 100 - trainTop1,
      train_loss = trainLoss,
      test_err = testTop1,
      test_err5 = testTop5,
      test_accuracy = 100 - testTop1,
      best_err = bestTop1,
      best_err5 = bestTop5,
      best_accuracy = 100 - bestTop1,
      finished = false,
   })

I'm not sure how the Torch community typically plots these curves. If there's a good way, I'd be glad to support it directly in this project.

[omkarparkhi]

May be @soumith, knows about torch plotting, but atleast I can save things using your input. Thanks!!

[szagoruyko]

The best way is to parse the log file in ipython notebook converting to pandas frame, then it's trivial to visualize it with bokeh. I can post an example if there is interest.

[omkarparkhi]

Sure. Please do.

[omkarparkhi]

Sam, have you tried converting a model to caffe? I fixed the std/var problem but there are still errors in the code I am not able to get around. E.g. Code errors while converting nn.Identity layer as it isn't handled. I am trying to see if I can add support for that myself. But if you have tried this before, would be great to have your input.

[colesbury]

No, I haven't tried converting the models to caffe

[miquelmarti]

@omkarparkhi Did you finally have any luck with that?

I want to convert ResNet-18 and Resnet-34.
There is a tool to create the prototxt files for ResNets but I am not sure they will match exactly the models here and I don't want to lose too much time tinkering around with the weights and I really want to avoid training the models from scratch.

[KevinYuk]

@miquelmarti
Hi miquelmarti,

How about the tool from jay-mahadeokar/pynetbuilder? Have you successfully got the good accuracy by that tool?

Thanks.