visunn 0.1.1

visunn: aesthetic visualization of neural networks

Visunn: Aesthetic Visualizations of Neural Networks for Deep Learning

Visunn is a visualization tool that leverages functional and modular visualizations to provide an visual understanding of neural network architectures. Currently, torch>=1.4.0 backend is supported.

Setup

The backend and user API is in Python (all pip dependencies can be found in requirements.txt) and serves a frontend consisting of a fusion of React and Three.js (all npm dependencies can be found in visunn/frontend/package.json).

Usage

The following examples will use the model models/ThreeLayerMLP.

To pip install

1. Install python package

pip install visunn

2. Initialize Visu in Python script

from visunn import Visu
visu = Visu(model, dataloader, logdir='logs', name='ThreeLayerMLP')

3. Launch web app

visu -l logs -n ThreeLayerMLP -p 5000

To use source

1. Build frontend (requires npm)

sh build.sh

2. Initialize Visu

python samples/train.py -l logs -n ThreeLayerMLP

3. Launch web app

python samples/serve.py -l logs -n ThreeLayerMLP -p 5000

Metrics

Below are time and space metrics for a few sample models, averaged over 5 runs. These can be obtained by running python samples/debug.py, which prints metrics of the 5 steps required to create a modularized topology (run on CIFAR-10):

Step 1 is the largest bottlenecks in the algorithm.

1. Convert model to protobuf (built in Pytorch function)
2. Convert protobuf to dict
3. Prune trivial nodes
4. Prune trivial modules
5. Build modularized topology

Model	Step 1	Step 2	Step 3	Step 4	Step 5	Space
ThreeLayerMLP	0.042 s	0.001 s	0.000 s	0.000 s	0.000 s	32.37 kb
ThreeLayerConvNet	0.051 s	0.002 s	0.000 s	0.000 s	0.000 s	35.33 kb
resnet18	1.282 s	0.016 s	0.002 s	0.001 s	0.002 s	335.40 kb
resnet152	6.819 s	0.115 s	0.065 s	0.005 s	0.013 s	2.14 Mb
densenet121	4.143 s	0.103 s	0.045 s	0.008 s	0.017 s	1.94 Mb
densenet201	7.528 s	0.174 s	0.125 s	0.015 s	0.034 s	3.56 Mb
googlenet	2.241 s	0.046 s	0.011 s	0.001 s	0.011 s	941.52 kb
shufflenet_v2_x2_0	2.289 s	0.070 s	0.023 s	0.007 s	0.010 s	1.07 Mb
mobilenet_v2	1.685 s	0.045 s	0.009 s	0.003 s	0.005 s	674.06 kb
resnext101_32x8d	11.744 s	0.085 s	0.033 s	0.004 s	0.009 s	1.35 Mb
wide_resnet101_2	13.495 s	0.002 s	0.165 s	0.007 s	0.054 s	1.10 Mb
mnasnet1_3	1.562 s	0.042 s	0.008 s	0.003 s	0.005 s	658.74 kb

Notes

This section is dedicated to address nuances that come with the Pytorch backend.

Recycled Layers

This problem is exemplified with the following following variation of ThreeLayerMLP:

class Model(nn.Module):
    def __init__(self, num_classes=10, **kwargs):
        self.flatten = nn.Flatten(1, -1)
        self.linear1 = nn.Linear(32*32*3, 1024, bias=True)
        self.linear2 = nn.Linear(1024, 256, bias=True)
        self.linear3 = nn.Linear(256, num_classes, bias=True)
        self.relu = nn.ReLU()

    def forward(self, x):
        x = self.flatten(x)
        x = self.linear1(x)
        x = self.relu(x)
        x = self.linear2(x)
        x = self.relu(x)
        x = self.linear3(x)
        return x

Note how self.relu = nn.ReLU() is reused multiple times. While this is not incorrect in any way. However, from a graph perspective, there is one self.relu = ReLU() node, which has multiple input and output nodes. Rendering this topology declaration will show cycles in the graph due to the multiple passes through the recycled node. A potential solution would be to split all nodes that consist of >1 unconnected graph.

Weights and Biases

With how torch>=1.4.0 constructs the graph protobuf, it is not possible to acquire the _output_shapes attribute from the weights and biases nodes (recoverable for all other nodes). Those fields empty when aggregating attributes.