anetomy 0.2.0

A neat instrument designed for visualizing neural networks.

aNETomy

A neat instrument designed for visualizing neural networks.
Experience seamless visualization without the burden of conversion.

---

🚀 Spotlight

🎯 Effortless Visualization

Given that PyTorch is an imperative framework, to visualize on the fly could be very useful for developers. aNETomy sidesteps the prep and conversion while keeping the accessibility and compatibility.

The fully expanded graph looks like below.

---

📊 Comparison

The following comparison are concluded according to hands-on practical experience. It could be wrong somewhere.

Feature	aNETomy	Netron	TensorBoard	torchviz	torchinfo	torchview	torchlens	torchexplorer
Less Preparation	✅	❌	✅	✅	✅	✅	✅	✅
Handles Dynamic Models	✅	❌	⚠️	❌	⚠️	✅	✅	❌
Interactive Graphs	✅	✅	✅	❌	❌	❌	❌	✅
Tensor Shapes	✅	✅	❌	✅	✅	✅	✅	✅
Structure Completeness	✅	✅	✅	❌ based on autograd	⚠️ some functions are missing	⚠️	✅	❌ based on autograd
Compatibility	✅	⚠️ unfriendly to some ops	⚠️ limited	✅	❌ No	✅	✅	✅

---

⚡ Quick Start

Add several lines to your python code for visualizing and saving graphs.

from anetomy import Inspector
import torch
from torch import nn
import torch.nn.functional as F

# define the toy model
class Dense(nn.Module):
    def __init__(self):
        super().__init__()
        self.lin = nn.Linear(28*28*2, 10)

    def forward(self, x):
        x = F.relu(x, inplace=True)
        x = x.reshape(-1,)
        y = self.lin(x)
        return y
          
class NeuralNetwork(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 8, 3, padding=1)
        self.mpool1 = nn.MaxPool2d(3, stride=2, ceil_mode=True)
        self.dense = Dense()

    def forward(self, x, scale=1):
        x *= scale
        y = self.conv1(x)
        y = self.mpool1(y)
        y = self.dense(y)
        return y

# input dummy data to visualize
net = NeuralNetwork()
dummy_x = torch.randn(1, 3, 28, 28)
dummy_s = 2
ant = Inspector(graph_path='./toynn.png')
ant.dissect(net, dummy_x, scale=dummy_s)
# use render() to save network image directly
ant.render(3) # 3 is the max depth to expand
# use launch() to run a web server for inspection
ant.launch('127.0.0.1', port=7880)

---

📦 Installation

At first, install graphviz beforehand.

🍎 macOS

Use Homebrew for a quick installation:

brew install graphviz

🐧 Linux

For Debian-based systems (Ubuntu, Debian):

sudo apt update && sudo apt install graphviz

For Arch-based systems (Arch, Manjaro):

sudo pacman -S graphviz

For RHEL-based systems (Fedora, CentOS):

sudo dnf install graphviz

🐍 Python

Then install the bindings.

pip install graphviz

Finally, install anetomy via pip.

pip install anetomy

---

❤️ Support

If you find aNETomy helpful, please give it a ⭐ on GitHub! ▶️ https://github.com/SeriaQ/aNETomy