Manim routines for neural networks

manim_nn.py

library of routines to draw neural networks using manim

manim_nn_mnist.py

implementation of the manim_nn routines to illustrate mnist at work, with animations for training and prediction - but SLOW, depending on your hardware. i'm using this on an M1 macbook air.

User commands for `manim_nn_mnist.py`

# load mnist dataset
In [1]: self.usr_load_data()

# train neural net, animating only the last training data element at every (epochs % epoch_anim_interval) ... because SLOW
In [2]: self.usr_train(epochs = 3, animate_weights = True, epoch_anim_interval = 1)

# predict using mnist test dataset
In [21]: self.usr_predict(self.x_test[15],self.y_test[15])

Interactively addressing neurons and edges

Examples:

# animate the 784th neuron of the 1st layer
In [1]: self.play(myNetwork.layers[0].real_neurons[783].animate.shift(LEFT*3))

# animate the edge (weight) between the 784th neuron of the 1st layer and the 4th neuron of the next (2nd) layer
In [2]: self.play(myNetwork.layers[0].real_neurons[783].edge_to_neuron[3].animate.shift(LEFT*3))

Manim and Environment Setup and Issues

Key requirements:

miniforge 3.24.0 (python 3.10 default)
ManimGL v1.6.1
ffmpeg (via brew)
latex (i'm using mactex via pkg installer)
for mnist demo (manimgl_nn_mnist.py), you'll need my neural-net repository Neural-Network

NOTE - Had to reinstall miniforge 3.24.0 in order to solve environment and dependency issues including:

python version issues (suggested use of python==3.10 which is default for miniforge 3.24.0 )
manimpango and cairo.h issues
numpy issues (suggested version 1.24 to resolve ValueError: operands could not be broadcast together with shapes (24,3) (0,3))

Create miniforge conda environment

conda create --name manimgl --clone base
conda activate manimgl
conda install numpy==1.24
conda install keras
conda install tensorflow
pip install manimgl

mkiiim/manimgl-nn