Incubator-mxnet: How to visualise filters of MXNet model?

[lixiangchun]

How to visualise filters of MXNet model like what has been done in keras - https://blog.keras.io/how-convolutional-neural-networks-see-the-world.html?

[reminisce]

Thanks for the reference. We have been working on enabling visualization of MXNet data in the TensorBoard. You can see the design proposal here: https://cwiki.apache.org/confluence/display/MXNET/Logging+MXNet+Data+for+Visualization+in+TensorBoard

After the PR is merged, this might be a good tutorial to write about how to use the visualization tool in MXNet.

[feevos]

Hi this solution applies to the gluon framework. I am not sure if you want to visualize the weights or the output of the convolution for some input (the latter is what I usually do).

Assuming you have a model, in the forward call you can return more than the last layer and visualize them:

%pylab inline

import mxnet as mx
from mxnet import gluon, nd

class tNet(gluon.nn.HybridBlock):

    def __init__(self,**kwards):
        gluon.nn.HybridBlock.__init__(self,**kwards)

        with self.name_scope():

            self.layer1 = gluon.nn.Conv2D(12,kernel_size=3,padding=1)
            self.layer2 = gluon.nn.Conv2D(12,kernel_size=3,padding=1)

    def hybrid_forward(self,F,_input):

        out1 = self.layer1(_input)
        out2 = self.layer2(out1)

        return out1, out2

mynet = tNet()
mynet.initialize(mx.init.Xavier())
xx = nd.random_uniform(shape=[4,12,128,128])
out1, out2 = mynet(xx)
imshow(out1[0,0].asnumpy())

this produces the following image:

Now say you want to visualize the weights only of a layer, without the convolution. For example you can obtain the weights of self.layer2 or self.layer1 like this:

layer2_weight = mynet.layer2.weight.data() # This is an NDArray object
layer2_bias = mynet.layer2.bias.data() # similarly, the bias terms, NDArray object

mynet.layer2.weight.data().shape 
# prints (12,12,3,3) for this particular example. 

so you can create histograms etc. Hope this helps.

[reminisce]

We have open-sourced MXBoard which enables users to visualize MXNet data (including network structures) in TensorBoard. In your case, you can just simply pass the weights/filters to the logging API and view it in TensorBoard.

The current version on PyPI is 0.1.0rc3. You can either install from source or pip install. The GitHub site has rich examples and tutorials on installation and how to use it in MXNet. Could you try it and give us feedbacks? Thanks.

[thomelane]

@sandeep-krishnamurthy good answer on this, so good to close. thanks!