search

apache / mxnet

Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more
https://mxnet.apache.org
Apache License 2.0
20.74k stars 6.81k forks source link

Guided BackPropagation for visualization #5812

Closed bhokaal2k closed 6 years ago

bhokaal2k commented 7 years ago

I am trying to develop a minimalist, simple to use and effective visualization program for visualizing what the deep network learns on the lines of the work presented in the paper https://arxiv.org/pdf/1412.6806.pdf and named as "Guided Backpropagation". I am aiming to make a notebook example and submit a PR for the same. These are the rough steps I need -

  1. Forward pass an image to a network (keep the internal activations and weights)
  2. Choose a neuron (n_i) in some higher layer (L) for which we want to visualize the pattern in the image that gave rise to it.
  3. Obtain the gradient of n_i w.r.t the image with "Guided ReLU" during backpropagation, explained in the paper.

The issue I have now is first how to compute the gradient of one neuron w.r.t. the image, I can for example easily do it in Theano/Keras/Tensorflow by picking that one neuron and computing the grad w.r.t. the image. Is it also possible in MxNet Feedforward/Module class, I know that I can do it in MxNet with hand-defined symbols and binding, but does the same applies to the Feedforward/Module as well? Secondly, is it possible to have the normal ReLU during forward propagation and Guided ReLU during backpropagation for the same image? Remember I will be using a network that is trained using ReLU and it is only during the visualization that I would need the "Guided ReLU" that too only during back-propagation and not during forward propagation

Possible solution -

  1. I write my own "Guided ReLU" and convert all the network's ReLU activation to Guided ReLU through network surgery. I am not quite sure how to proceed with this approach, can someone please provide some pointers, a simple example of network surgery (I am familiar with adding/removing layers) with different activations?
sxjscience commented 7 years ago

Regarding the first question, In module, we have mod.get_input_grads() https://github.com/dmlc/mxnet/blob/master/python/mxnet/module/module.py#L595-L614 .

bhokaal2k commented 7 years ago

Thanks @sxjscience I just got my familiar with the Module api and it seems like it can take care of the first problem. What about the Guided ReLU? @piiswrong Is there a way to change the ReLU activations of a pre-trained model to GuidedReLU and save the new model to be used for network visualization?

sxjscience commented 6 years ago

Close it now because there is no discussion for a long time. @bhokaal2k Feel free to share your implementation of Guided BackPropagation.

ZhichengHuang commented 6 years ago

@bhokaal2k can you share your implementation of Guided BackPropagation, I have met the same problem, and can't find a proper solution.

sandareka commented 6 years ago

@ZhichengHuang Please check if this code helps you. I know this is a very late reply. But I put it if you are still looking for the answer and also for other's benefit.