MLVTK

A loss surface visualization tool

Simple DNN trained on MNIST data set, using Adamax optimizer

Simple DNN trained on MNIST, using SGD optimizer

Simple DNN trained on MNIST, using Adam optimizer

Simple DNN trained on MNIST, using SGD optimizer

Why?

• :shipit: Simple: A single line addition is all that is needed.
• :question: Informative: Gain insight into what your model is seeing.
• :notebook: Educational: See how your hyper parameters and architecture impact your models perception.

Quick Start

Install locally (Also works in google Colab!):

pip install mlvtk

Optionally for use with jupyter notebook/lab:

Notebook

pip install "notebook>=5.3" "ipywidgets==7.5"

Lab

pip install jupyterlab "ipywidgets==7.5"

# Basic JupyterLab renderer support
jupyter labextension install jupyterlab-plotly@4.10.0

# OPTIONAL: Jupyter widgets extension for FigureWidget support
jupyter labextension install @jupyter-widgets/jupyterlab-manager plotlywidget@4.10.0

Basic Example

from mlvtk.base import Vmodel
import tensorflow as tf
import numpy as np

# NN with 1 hidden layer
inputs = tf.keras.layers.Input(shape=(None,100))
dense_1 = tf.keras.layers.Dense(50, activation='relu')(inputs)
outputs = tf.keras.layers.Dense(10, activation='softmax')(dense_1)
_model = tf.keras.Model(inputs, outputs)

# Wrap with Vmodel
model = Vmodel(_model)
model.compile(optimizer=tf.keras.optimizers.SGD(),
loss=tf.keras.losses.CategoricalCrossentropy(), metrics=['accuracy'])

# All tf.keras.(Model/Sequential/Functional) methods/properties are accessible
# from Vmodel

model.summary()
model.get_config()
model.get_weights()
model.layers

# Create random example data
x = np.random.rand(3, 10, 100)
y = np.random.randint(9, size=(3, 10, 10))
xval = np.random.rand(1, 10, 100)
yval = np.random.randint(9, size=(1,10,10))

# Only difference, model.fit requires validation_data (tf.data.Dataset, or
# other container
history = model.fit(x, y, validation_data=(xval, yval), epochs=10, verbose=0)

# Calling model.surface_plot() returns a plotly.graph_objs.Figure
# model.surface_plot() will attempt to display the figure inline

fig = model.surface_plot()

# fig can save an interactive plot to an html file,
fig.write_html("surface_plot.html")

# or display the plot in jupyter notebook/lab or other compatible tool.
fig.show()