Files
ManimML/manim_ml/neural_network/neural_network.py

298 lines
12 KiB
Python

"""Neural Network Manim Visualization
This module is responsible for generating a neural network visualization with
manim, specifically a fully connected neural network diagram.
Example:
# Specify how many nodes are in each node layer
layer_node_count = [5, 3, 5]
# Create the object with default style settings
NeuralNetwork(layer_node_count)
"""
import textwrap
import numpy as np
from manim import *
from manim_ml.neural_network.layers.embedding import EmbeddingLayer
from manim_ml.neural_network.layers.feed_forward import FeedForwardLayer
from manim_ml.neural_network.layers.parent_layers import ConnectiveLayer, ThreeDLayer
from manim_ml.neural_network.layers.util import get_connective_layer
from manim_ml.list_group import ListGroup
from manim_ml.neural_network.neural_network_transformations import (
InsertLayer,
RemoveLayer,
)
class NeuralNetwork(Group):
"""Neural Network Visualization Container Class"""
def __init__(
self,
input_layers,
edge_color=WHITE,
layer_spacing=0.2,
animation_dot_color=RED,
edge_width=2.5,
dot_radius=0.03,
title=" ",
layout="linear",
layout_direction="left_to_right",
):
super(Group, self).__init__()
self.input_layers = ListGroup(*input_layers)
self.edge_width = edge_width
self.edge_color = edge_color
self.layer_spacing = layer_spacing
self.animation_dot_color = animation_dot_color
self.dot_radius = dot_radius
self.title_text = title
self.created = False
self.layout = layout
self.layout_direction = layout_direction
# TODO take layer_node_count [0, (1, 2), 0]
# and make it have explicit distinct subspaces
# Construct all of the layers
self._construct_input_layers()
# Place the layers
self._place_layers(layout=layout, layout_direction=layout_direction)
# Make the connective layers
self.connective_layers, self.all_layers = self._construct_connective_layers()
# Make overhead title
self.title = Text(self.title_text, font_size=DEFAULT_FONT_SIZE / 2)
self.title.next_to(self, UP, 1.0)
self.add(self.title)
# Place layers at correct z index
self.connective_layers.set_z_index(2)
self.input_layers.set_z_index(3)
# Center the whole diagram by default
self.all_layers.move_to(ORIGIN)
self.add(self.all_layers)
# Print neural network
print(repr(self))
def _construct_input_layers(self):
"""Constructs each of the input layers in context
of their adjacent layers"""
prev_layer = None
next_layer = None
# Go through all the input layers and run their construct method
print("Constructing layers")
for layer_index in range(len(self.input_layers)):
current_layer = self.input_layers[layer_index]
print(f"Current layer: {current_layer}")
if layer_index < len(self.input_layers) - 1:
next_layer = self.input_layers[layer_index + 1]
if layer_index > 0:
prev_layer = self.input_layers[layer_index - 1]
# Run the construct layer method for each
current_layer.construct_layer(prev_layer, next_layer)
def _place_layers(self, layout="linear", layout_direction="top_to_bottom"):
"""Creates the neural network"""
# TODO implement more sophisticated custom layouts
# Default: Linear layout
for layer_index in range(1, len(self.input_layers)):
previous_layer = self.input_layers[layer_index - 1]
current_layer = self.input_layers[layer_index]
current_layer.move_to(previous_layer.get_center())
if layout_direction == "left_to_right":
x_shift = previous_layer.get_width() / 2 \
+ current_layer.get_width() / 2 \
+ self.layer_spacing
shift_vector = np.array([x_shift, 0, 0])
elif layout_direction == "top_to_bottom":
y_shift = -((
previous_layer.get_width() / 2 \
+ current_layer.get_width() / 2
) + self.layer_spacing)
shift_vector = np.array([0, y_shift, 0])
else:
raise Exception(
f"Unrecognized layout direction: {layout_direction}"
)
current_layer.shift(shift_vector)
# After all layers have been placed place their activation functions
for current_layer in self.input_layers:
# Place activation function
if hasattr(current_layer, "activation_function"):
if not current_layer.activation_function is None:
up_movement = np.array(
[
0,
current_layer.get_height() / 2
+ current_layer.activation_function.get_height() / 2
+ 0.5 * self.layer_spacing,
0,
]
)
current_layer.activation_function.move_to(
current_layer,
)
current_layer.activation_function.shift(up_movement)
self.add(current_layer.activation_function)
def _construct_connective_layers(self):
"""Draws connecting lines between layers"""
connective_layers = ListGroup()
all_layers = ListGroup()
for layer_index in range(len(self.input_layers) - 1):
current_layer = self.input_layers[layer_index]
# Add the layer to the list of layers
all_layers.add(current_layer)
next_layer = self.input_layers[layer_index + 1]
# Check if layer is actually a nested NeuralNetwork
if isinstance(current_layer, NeuralNetwork):
# Last layer of the current layer
current_layer = current_layer.all_layers[-1]
if isinstance(next_layer, NeuralNetwork):
# First layer of the next layer
next_layer = next_layer.all_layers[0]
# Find connective layer with correct layer pair
connective_layer = get_connective_layer(current_layer, next_layer)
connective_layers.add(connective_layer)
# Construct the connective layer
connective_layer.construct_layer(current_layer, next_layer)
# Add the layer to the list of layers
all_layers.add(connective_layer)
# Add final layer
all_layers.add(self.input_layers[-1])
# Handle layering
return connective_layers, all_layers
def insert_layer(self, layer, insert_index):
"""Inserts a layer at the given index"""
neural_network = self
insert_animation = InsertLayer(layer, insert_index, neural_network)
return insert_animation
def remove_layer(self, layer):
"""Removes layer object if it exists"""
neural_network = self
return RemoveLayer(layer, neural_network, layer_spacing=self.layer_spacing)
def replace_layer(self, old_layer, new_layer):
"""Replaces given layer object"""
raise NotImplementedError()
remove_animation = self.remove_layer(insert_index)
insert_animation = self.insert_layer(layer, insert_index)
# Make the animation
animation_group = AnimationGroup(
FadeOut(self.all_layers[insert_index]), FadeIn(layer), lag_ratio=1.0
)
return animation_group
def make_forward_pass_animation(
self, run_time=None, passing_flash=True, layer_args={}, **kwargs
):
"""Generates an animation for feed forward propagation"""
all_animations = []
per_layer_runtime = (
run_time / len(self.all_layers) if not run_time is None else None
)
for layer_index, layer in enumerate(self.all_layers):
# Get the layer args
if isinstance(layer, ConnectiveLayer):
"""
NOTE: By default a connective layer will get the combined
layer_args of the layers it is connecting and itself.
"""
before_layer_args = {}
current_layer_args = {}
after_layer_args = {}
if layer.input_layer in layer_args:
before_layer_args = layer_args[layer.input_layer]
if layer in layer_args:
current_layer_args = layer_args[layer]
if layer.output_layer in layer_args:
after_layer_args = layer_args[layer.output_layer]
# Merge the two dicts
current_layer_args = {
**before_layer_args,
**current_layer_args,
**after_layer_args,
}
else:
current_layer_args = {}
if layer in layer_args:
current_layer_args = layer_args[layer]
# Perform the forward pass of the current layer
layer_forward_pass = layer.make_forward_pass_animation(
layer_args=current_layer_args, run_time=per_layer_runtime, **kwargs
)
all_animations.append(layer_forward_pass)
# Make the animation group
animation_group = Succession(*all_animations, lag_ratio=1.0)
return animation_group
@override_animation(Create)
def _create_override(self, **kwargs):
"""Overrides Create animation"""
# Stop the neural network from being created twice
if self.created:
return AnimationGroup()
self.created = True
animations = []
# Create the overhead title
animations.append(Create(self.title))
# Create each layer one by one
for layer in self.all_layers:
layer_animation = Create(layer)
# Make titles
create_title = Create(layer.title)
# Create layer animation group
animation_group = AnimationGroup(layer_animation, create_title)
animations.append(animation_group)
animation_group = AnimationGroup(*animations, lag_ratio=1.0)
return animation_group
def set_z_index(self, z_index_value: float, family=False):
"""Overriden set_z_index"""
# Setting family=False stops sub-neural networks from inheriting parent z_index
for layer in self.all_layers:
if not isinstance(NeuralNetwork):
layer.set_z_index(z_index_value)
def scale(self, scale_factor, **kwargs):
"""Overriden scale"""
for layer in self.all_layers:
layer.scale(scale_factor, **kwargs)
# Place layers with scaled spacing
self.layer_spacing *= scale_factor
self._place_layers(layout=self.layout, layout_direction=self.layout_direction)
def filter_layers(self, function):
"""Filters layers of the network given function"""
layers_to_return = []
for layer in self.all_layers:
func_out = function(layer)
assert isinstance(
func_out, bool
), "Filter layers function returned a non-boolean type."
if func_out:
layers_to_return.append(layer)
return layers_to_return
def __repr__(self, metadata=["z_index", "title_text"]):
"""Print string representation of layers"""
inner_string = ""
for layer in self.all_layers:
inner_string += f"{repr(layer)}("
for key in metadata:
value = getattr(layer, key)
if not value is "":
inner_string += f"{key}={value}, "
inner_string += "),\n"
inner_string = textwrap.indent(inner_string, " ")
string_repr = "NeuralNetwork([\n" + inner_string + "])"
return string_repr