Metadata-Version: 2.4
Name: lcurvetools
Version: 1.2.0
Summary: Simple Python tools for plotting learning curves of neural network models trained with the Keras, Ultralytics YOLO or scikit-learn framework in a single figure in an easy-to-understand format.
Home-page: https://github.com/kamua/lcurvetools
Author: Andriy Konovalov
Author-email: kandriy74@gmail.com
License: BSD 3-Clause License
Keywords: learning curve,keras,ultralytics yolo,scikit-learn,loss_curve,validation_score
Classifier: Development Status :: 5 - Production/Stable
Classifier: License :: OSI Approved :: BSD License
Classifier: Intended Audience :: Science/Research
Classifier: Intended Audience :: Education
Classifier: Topic :: Scientific/Engineering
Classifier: Operating System :: Microsoft :: Windows
Classifier: Operating System :: POSIX
Classifier: Operating System :: Unix
Classifier: Operating System :: MacOS
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3 :: Only
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: numpy
Requires-Dist: matplotlib
Requires-Dist: scikit-learn
Dynamic: author
Dynamic: author-email
Dynamic: classifier
Dynamic: description
Dynamic: description-content-type
Dynamic: home-page
Dynamic: keywords
Dynamic: license
Dynamic: license-file
Dynamic: requires-dist
Dynamic: requires-python
Dynamic: summary

# LCurveTools

Simple Python tools for plotting learning curves of neural network models trained with the Keras, Ultralytics YOLO or scikit-learn framework in a single figure in an easy-to-understand format.

Currently the `lcurvetools` package provides three basic functions: `lcurves`, `history_concatenate` and `lcurves_by_MLP_estimator`.

**NOTE:** All of the plotting examples below are for [interactive Python mode](https://matplotlib.org/stable/users/explain/figure/interactive.html#interactive-mode) in Jupyter-like environments. If you are in non-interactive mode you may need to explicitly call [`plt.show()`](https://matplotlib.org/stable/api/_as_gen/matplotlib.pyplot.show.html) after calling the `lcurves` or `lcurves_by_MLP_estimator` function to display the window with built plots on your screen.

## Installation

The easiest way to install the `lcurvetools` package for the first time is using pip:

```sh
pip install lcurvetools
```

To update a previously installed package to the latest version, use the command:

```sh
pip install lcurvetools --upgrade
```

## The `lcurves` function to plot learning curves based on model training history dictionaries

The `lcurves` function and its alias `lcurves_by_history` plot learning curves based on a dictionary or a list of dictionaries with model training histories. Each dictionary with a history should contain keys with training and validation values of losses and metrics, as well as learning rate values at successive epochs, for example:
`{"loss": [0.5, 0.3, 0.2], "val_loss": [0.6, 0.4, 0.25], "accuracy": [0.7, 0.8, 0.85], "val_accuracy": [0.65, 0.75, 0.8], "lr": [0.01, 0.01, 0.001]}`. This training history format is adopted in the Keras library and is also widely used when working with other libraries.

Neural network model training with Keras is performed using the [fit](https://keras.io/api/models/model_training_apis/#fit-method) method. The method returns the `History` object with the `history` attribute which is dictionary and contains keys with training and validation values of losses and metrics, as well as learning rate values at successive epochs. The `lcurves` function and its alias `lcurves_by_history` use the `History.history` dictionary to plot the learning curves as the dependences of the above values on the epoch index.

Import the `keras` module and the `lcurves` function:

```python
import keras
from lcurvetools import lcurves
```

### Plotting learning curves of one model based on one dictionary

#### Usage scheme

1. [Create](https://keras.io/api/models/), [compile](https://keras.io/api/models/model_training_apis/#compile-method)
and [fit](https://keras.io/api/models/model_training_apis/#fit-method) the keras model:

   ```python
   model = keras.Model(...) # or keras.Sequential(...)
   model.compile(...)
   hist = model.fit(...)
   ```

2. Use `hist.history` dictionary to plot the learning curves as the dependences of values of all keys in the dictionary on an epoch index with automatic recognition of keys of losses, metrics and learning rate:

   ```python
   lcurves(hist.history);
   ```

#### Typical appearance of the output figure

The appearance of the output figure depends on the list of keys in the `hist.history` dictionary, which is determined by the parameters of the `compile` and `fit` methods of the model. For example, for a typical usage of these methods, the list of keys would be `['loss', 'accuracy', 'val_loss', 'val_accuracy']` and the output figure will contain 2 subplots with loss and metrics vertical axes and might look like this:

```python
model.compile(loss="categorical_crossentropy", metrics=["accuracy"])
hist = model.fit(x_train, y_train, validation_split=0.1, epochs=50)
lcurves(hist.history);
```

![typical plot of learning curves](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/typical_plot.png)

Of course, if the `metrics` parameter of the `compile` method is not specified, then the output figure will not contain a metric subplot.

Minimum values of loss curves and the best values of metric curves are marked by points. The best values of metrics are determined based on the specified optimization mode for each metric (see `optimization_modes` parameter). By default, the optimization modes are determined automatically based on metric name with the `lcurves.utils.get_mode_by_metric_name` function.

Usage of callbacks for the `fit` method can add new keys to the `hist.history` dictionary. For example, the [ReduceLROnPlateau](https://keras.io/api/callbacks/reduce_lr_on_plateau/) callback adds the `lr` key with learning rate values for successive epochs. In this case the output figure will contain additional subplot with learning rate vertical axis in a logarithmic scale and might look like this:

```python
hist = model.fit(x_train, y_train, validation_split=0.1, epochs=50,
    callbacks=[keras.callbacks.ReduceLROnPlateau()],
)
lcurves(hist.history);
```

![figure with learning rate subplot](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/learning_rate_subplot.png)

#### Customizing appearance of the output figure

The `lcurves` function has optional parameters to customize the appearance of the output figure. For example, the `epoch_range_to_scale` option allows to specify the epoch index range within which the subplots of the losses and metrics are scaled.

- If `epoch_range_to_scale` is a list or a tuple of two int values, then they specify the epoch index limits of the scaling range in the form `[start, stop)`, i.e. as for `slice` and `range` objects.
- If `epoch_range_to_scale` is an int value, then it specifies the lower epoch index `start` of the scaling range, and the losses and metrics subplots are scaled by epochs with indices from `start` to the last.

So, you can exclude the first 5 epochs from the scaling range as follows:

```python
lcurves(hist.history, epoch_range_to_scale=5);
```

![figure with custom scaling](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/custom_scaling.png)

For a description of other optional parameters of the `lcurves` function to customize the appearance of the output figure, see its docstring.

The `lcurves` function returns a numpy array or a list of the [`matplotlib.axes.Axes`](https://matplotlib.org/stable/api/_as_gen/matplotlib.axes.Axes.html) objects corresponded to the built subplots from top to bottom. So, you can use the methods of these objects to customize the appearance of the output figure.

```python
axs = lcurves(history, epoch_range_to_scale=6)
axs[0].tick_params(axis="x", labeltop=True)
axs[-1].set_xlabel('number of passed epochs')
axs[-1].legend().remove()
```

![figure with adjusted axes](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/adjusted_axes.png)

### Plotting learning curves based on a list of dictionaries with fitting histories of several models

#### Usage scheme

1. [Create](https://keras.io/api/models/), [compile](https://keras.io/api/models/model_training_apis/#compile-method)
and [fit](https://keras.io/api/models/model_training_apis/#fit-method) the several keras models:

   ```python
   model_1 = keras.Model(...) # or keras.Sequential(...)
   model_1.compile(...)
   hist_1 = model.fit(...)

   model_2 = keras.Model(...) # or keras.Sequential(...)
   model_2.compile(...)
   hist_2 = model.fit(...)

   <...>
   ```

2. Use a list of dictionaries `[hist_1.history, hist_2.history, ...]` to plot all learning curves of all models in a single figure:

   ```python
   lcurves([hist_1.history, hist_2.history, ...]);
   ```

### Plotting learning curves based on a list of dictionaries with independent refitting histories of one model

#### Usage scheme

1. Organize a loop of multiple independent retraining of the model and create a list of dictionaries with fitting histories:

   ```python
   histories = []
   for i in range(5):
       model = keras.Model(...) # or keras.Sequential(...)
       model.compile(...)
       hist = model.fit(...)
       histories.append(hist.history)
   ```

2. Use the `histories` dictionary list to plot all learning curves of the model in a single figure:

   ```python
   lcurves(histories);
   ```

**Note:** The ability to plot all of a model's learning curves in a single figure is useful for k-fold cross-validation analysis.

## The `history_concatenate` function to concatenate two `History.history` dictionaries

This function is useful for combining histories of a model fitting with two or more consecutive runs into a single history to plot full learning curves.

### Usage scheme

- Import the `keras` module and the `history_concatenate`, `lcurves_by_history` function:

```python
import keras
from lcurvetools import history_concatenate, lcurves
```

- [Create](https://keras.io/api/models/), [compile](https://keras.io/api/models/model_training_apis/#compile-method)
and [fit](https://keras.io/api/models/model_training_apis/#fit-method) the keras model:

```python
model = keras.Model(...) # or keras.Sequential(...)
model.compile(...)
hist1 = model.fit(...)
```

- Compile as needed and fit using possibly other parameter values:

```python
model.compile(...) # optional
hist2 = model.fit(...)
```

- Concatenate the `.history` dictionaries into one:

```python
full_history = history_concatenate(hist1.history, hist2.history)
```

- Use `full_history` dictionary to plot full learning curves:

```python
lcurves(full_history);
```

## The `lcurves_by_MLP_estimator` function to plot learning curves of the scikit-learn MLP estimator

The scikit-learn library provides 2 classes for building multi-layer perceptron (MLP) models of classification and regression: [`MLPClassifier`](https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPClassifier.html) and [`MLPRegressor`](https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPRegressor.html). After creation and fitting of these MLP estimators with using `early_stopping=True` the `MLPClassifier` and `MLPRegressor` objects have the `loss_curve_` and `validation_scores_` attributes with train loss and validation score values at successive epochs. The `lcurves_by_MLP_estimator` function uses the `loss_curve_` and `validation_scores_` attributes to plot the learning curves as the dependences of the above values on the epoch index.

### Usage scheme

- Import the `MLPClassifier` (or `MLPRegressor`) class and the `lcurves_by_MLP_estimator` function:

```python
from sklearn.neural_network import MLPClassifier
from lcurvetools import lcurves_by_MLP_estimator
```

- [Create](https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPClassifier.html#sklearn.neural_network.MLPClassifier) and [fit](https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPClassifier.html#sklearn.neural_network.MLPClassifier.fit) the scikit-learn MLP estimator:

```python
clf = MLPClassifier(..., early_stopping=True)
clf.fit(...)
```

- Use `clf` object with `loss_curve_` and `validation_scores_` attributes to plot the learning curves as the dependences of loss and validation score values on epoch index:

```python
lcurves_by_MLP_estimator(clf);
```

### Typical appearance of the output figure

The `lcurves_by_MLP_estimator` function with default value of the parameter `on_separate_subplots=False` shows the learning curves of loss and validation score on one plot with two vertical axes scaled independently. Loss values are plotted on the left axis and validation score values are plotted on the right axis. The output figure might look like this:

![lcurves_by_MLP_estimator on one plot](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/lcurves_by_MLP_estimator-1_plot.png)

**Note:** the minimum value of loss curve and the maximum value of validation score curve are marked by points.

### Customizing appearance of the output figure

The `lcurves_by_MLP_estimator` function has optional parameters to customize the appearance of the output figure. For example,the `lcurves_by_MLP_estimator` function with `on_separate_subplots=True` shows the learning curves of loss and validation score on two separated subplots:

```python
lcurves_by_MLP_estimator(clf, on_separate_subplots=True);
```

![lcurves_by_MLP_estimator on two subplot](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/lcurves_by_MLP_estimator-2_subplots.png)

And the `epoch_range_to_scale` option allows to specify the epoch index range within which the subplots of the losses and metrics are scaled (see details about this option in the docstring of the `lcurves_by_MLP_estimator` function).

```python
lcurves_by_MLP_estimator(clf, epoch_range_to_scale=10);
```

![lcurves_by_MLP_estimator - custom scaling](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/lcurves_by_MLP_estimator-custom_scaling.png)

For a description of other optional parameters of the `lcurves_by_MLP_estimator` function to customize the appearance of the output figure, see its docstring.

The `lcurves_by_MLP_estimator` function returns a numpy array or a list of the [`matplotlib.axes.Axes`](https://matplotlib.org/stable/api/_as_gen/matplotlib.axes.Axes.html) objects of an output figure (see additional details in the `Returns` section of the `lcurves_by_MLP_estimator` function docstring). So, you can use the methods of these objects to customize the appearance of the output figure.

```python
axs = lcurves_by_MLP_estimator(clf, epoch_range_to_scale=11)
axs[0].grid(axis='y', visible=False)
axs[1].grid(axis='y', visible=False)
axs[0].set_xlabel('number of passed epochs')
axs[1].set_ylabel('validation accuracy');
```

![lcurves_by_MLP_estimator - adjusted axes](https://raw.githubusercontent.com/kamua/lcurvetools/main/img/lcurves_by_MLP_estimator-adjusted_axes.png)


# CHANGELOG

All notable changes to this project will be documented in this file.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/).

## [1.2.0] - 2026-05-05

### Added

- A new boolean parameter `optimum_values_in_legend` has been added to the `lcurves()` function. The parameter specifies whether to display the optimum values of epochs and metrics in the legends of the the losses and metrics subplots.

## [1.1.1] - 2026-02-26

### Changed

- Improved x-axis tick placement logic in `lcurves()` function for better readability of epoch numbers on the plot.

### Fixed

- Small bug fixes.

## [1.1.0] - 2026-01-24

### Changed

- The `lcurves_by_history()` function has been renamed to `lcurves()` to make the code easier to write and understand, but `lcurves_by_history()` remains as an alias for backward compatibility.
- The `history` parameter of the `lcurves()` function has been renamed to `histories`. It can now accept a list of dictionaries with several fitting histories of keras models.

### Added

- Support for correct recognition of key names in learning history of model training with the Ultralytics YOLO library.
- Support for plotting multiple histories on a single figure using `lcurves()` function.
- The default value of the `initial_epoch` parameter of the `lcurves` function has been changed from 0 to 1.
- New parameters in `lcurves()`:
  - `color_grouping_by`: Controls curve color grouping in subplots.
  - `model_names`: Customizes legend labels for each history.
  - `optimization_modes`: Sets metric optimization direction ("min"/"max")
- New `utils` module with utility functions:
  - `get_mode_by_metric_name()`: Determines metric optimization mode.
  - `get_best_epoch_value()`: Finds optimal epoch value.
  - `history_concatenate()`: Concatenates two histories into one. The function was moved from the `lcurvetools` module to the `utils` module, but it remained available for import from the main `lcurvetools` package module.

## [1.0.1] - 2025-01-23

### Changed

The default value for the `figsize` parameter of the `lcurves_by_history()` and `lcurves_by_MLP_estimator()` functions has been changed to `None`.

## [1.0.0] - 2024-01-22

Initial release.
