liveplotlib 0.3.3

Library for plotting (visualizing) cost function changes during model training (in real time)

LIVE PLOT LIBrary (liveplotlib)

• Library for plotting (visualizing) cost function changes during model training (in real time)

• please, check compatibility to see if this library meets your needs and environment

Notations

Specific

• J - cost/loss function
  It measures how well does your model performs and used for optimization of your model

• J_history - python list (of numbers), that contains previous values of J
  (for example, in previous epochs)

For subsets

• J_train_history - J_history, where J was calculated on train subset
• J_val_history - J_history, where J was calculated on val subset
• J_test_history - J_history, where J was calculated on test subset

General

• plotting - visualizing, making graph

• train/val/test - subsets of your original dataset

• train - train subset.
  Used for optimizing the model parameters (weights and biases)

• val/valid - validation subset
  (aka Cross-Validation (cv) or DEVeleopment (dev) subset). Used for optimizing the hyperparameters of your model (for example: learning_rate or regularization term)

• test - test subset, used for testing how your model performs on new data

• epoch - 1 iteration through the whole dataset (or subset).
  

Installation

>> pip install liveplotlib
# all dependencies will be installed automatically

Usage

• Only 5 steps!

• "foo" in names means "this is just for example, it means nothing"

• Basic functionality:

  from liveplotlib import LivePlotOnlyTrain
  
  J_train_history = []
  
  # begin session
  live_plot = LivePlotOnlyTrain()
  
  # update during training
  # J_train_history.append(new_J_train)
  live_plot.update(J_train_history)
  
  # end session
  live_plot.close()
  
  # (See explanations below)
  

Option 1: plot only J_train_history

# STEP 1: IMPORT

from liveplotlib import LivePlotOnlyTrain  



# ...............
# Some operations with data
# ................



# STEP 2: CREATE HISTORY LIST

# Along with model initialization, create empty J_train_history as well
model = FooSomeModelClass() 
J_train_history = []



# STEP 3: INITIALIZE LIVE PLOT

# Right before training
live_plot = LivePlotOnlyTrain()



# STEP 4: UPDATE DURING TRAINING

# -----Inside train function loop-----
# ...
# new_J_train = ...
# ...
J_train_history.append(new_J_train)
# ...
live_plot.update(J_train_history)
# ...
# ------------------------------------



# STEP 5: END SESSION

# In the end (especially important in jupyter notebooks)
live_plot.close()

Option 2: plot J_train_history and J_val_history

# STEP 1: IMPORT

from liveplotlib import LivePlotTrainAndVal 



# ...............
# Some operations with data
# ................



# STEP 2: CREATE HISTORY LISTS

# Along with model initialization, create empty J_train_history and J_val_history as well
model = FooSomeModelClass() 
J_train_history = []
J_val_history = []



# STEP 3: INITIALIZE LIVE PLOT (START SESSION)

# Right before training
live_plot = LivePlotTrainAndVal()



# STEP 4: UPDATE DURING TRAINING

# -----Inside train function loop-----
# ...
# new_J_train = ...
# new_J_val = ...
# ...
J_train_history.append(new_J_train)
J_val_history.append(new_J_val)
# ...
live_plot.update(J_train_history, J_val_history)
# ...
# ------------------------------------



# STEP 5: END SESSION

# In the end (especially important in jupyter notebooks)
live_plot.close()

About

What does it do

• This tool plots Cost (loss) function changes in real time in separated window (during training, not after its done) so you could see tendencies and diagnose/manage your optimization process more easily.

Who is it created for

• This python library is created for data scientists / ML engineers / data analysts or anyone else interested

Under the hood

• written on top of matplotlib library, using its figures, subplots and lines mechanics

Compatibility

• it works only if you are running it on your local machine.
  (It can't run in Google colab, because google colab's notebooks run on a remote server, so you can't exit inline mode => it will give you an error if you try).
  If you have any ideas how to fix this, please, feel free to propose such important improvement ))
• compatible with .py and .ipynb (jupyter notebook) files
• It automatically determines a caller file's format and takes appropriate actions

How to

Change slice size

Slice size is calculated in each .update() by formula:

slice_size = slice_fraction * len(J_history) + slice_bias

where slice_fraction and slice_bias are parameters of .__init__() function (used for initializing live_plot).

So, if you want to add some fixed number to amount of steps to plot, increase slice_bias. If you want to make a scalable, dynamic change (like increase fraction) - then, of corse increase slice_fraction

Recomendations

If you are using LivePlotTrainAndVal

• Ensure, that your J_train_history and J_val_history have the same length.
• Update once in epoch
  If you are using minibaches, then calculate J_train and J_val across all minibatches (average). Then, as always:

  # In the end of epoch
  J_train_history.append(J_train)
  J_val_history.append(J_val)
  live_plot.update(J_train_history, J_val_history)
  

Comments

• name "liveplotlib" comes from its "parent" library - matplotlib. (liveplotlib is written on top of matplotlib)
• please, feel free to give me feedback, propose some improvements/new functionality or report a bug.