Incubator-mxnet: Request: Visualization for Gluon neural networks

Created on 19 Sep 2017 · 6Comments · Source: apache/incubator-mxnet

Here shows how to visualize mx.Symbol networks.

mx.viz.plot_network(net)

The same thing does not work for initialized gluon.Block networks.

Call for Contribution

Source

SumNeuron

Most helpful comment

You're able to plot if you've got a HybridBlock, since you can convert to a Symbol by hybridizing and passing an input Symbol. A more direct method might be nice though.

import mxnet as mx

net = mx.gluon.nn.HybridSequential()
with net.name_scope():
    net.add(mx.gluon.nn.Dense(256, activation="relu"))
    net.add(mx.gluon.nn.Dense(256, activation="relu"))
    net.add(mx.gluon.nn.Dense(10))

# Option 1    
print(net)

# Option 2 (net must be HybridSequential, if want to plot whole graph)
net.hybridize()
net.collect_params().initialize()
x = mx.sym.var('data')
sym = net(x)
mx.viz.plot_network(sym)

thomelane on 26 Jan 2018

👍9 ❤1

All 6 comments

It's a good idea to have visualization tools. Currently, for standard blocks, you can print them directly, though the string representation doesn't reflect the computation graph.

szha on 23 Sep 2017

@szha I just find it weird that it exists in symbol and not in gluon.
I understand the underlying differences between the two, but the visualization is nice regardless

SumNeuron on 23 Sep 2017

You're able to plot if you've got a HybridBlock, since you can convert to a Symbol by hybridizing and passing an input Symbol. A more direct method might be nice though.

import mxnet as mx

net = mx.gluon.nn.HybridSequential()
with net.name_scope():
    net.add(mx.gluon.nn.Dense(256, activation="relu"))
    net.add(mx.gluon.nn.Dense(256, activation="relu"))
    net.add(mx.gluon.nn.Dense(10))

# Option 1    
print(net)

# Option 2 (net must be HybridSequential, if want to plot whole graph)
net.hybridize()
net.collect_params().initialize()
x = mx.sym.var('data')
sym = net(x)
mx.viz.plot_network(sym)

thomelane on 26 Jan 2018

👍9 ❤1

We have open-sourced MXBoard which enables users to visualize MXNet data (including network structures) 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.