A module with a bunch of utilites for TensorFlow + Keras. These are some functions and classes that I found myself writing more than once while creating various projects. Hopefully, some of it is useful for others as well.
Install:
pip3 install git+https://github.com/am1tyadav/tfutils.git
Usage:
import tfutils
model.fit(
x, y,
epochs=1,
callbacks=[
tfutils.keras.callbacks.SimpleTrainingPlot(plt)
]
)What's available:
| Utility | Description |
|---|---|
tfutils.keras.callbacks.SimpleTrainingPlot(plt) |
Requires matplotlib.pyplot argument passed as plt, and returns a Keras callback that will plot training metrics loss, val_loss, accuracy, val_accuracy. Please note the full keyword accuracy is used and not acc. |
tfutils.keras.callbacks.PlotEmbedding(plt, embedding_model, x_test, y_test) |
Requires matplotlib.pyplot argument passed as plt, and returns a Keras callback that will plot a 2-dimensional representation of embedding from the embedding_model on the set x_test, colored with the values of labels y_test. |
| Utility | Description |
|---|---|
tfutils.keras.losses.triplet_loss(dim, alpha) |
Returns a Keras Loss function. Argument dim is the dimension of embedding, and alpha is the margin with default value set to 0.2 |
| Utility | Description |
|---|---|
tfutils.keras.plotting.plot_training_history(plt, history) |
Plots training history from the history object returned from Keras' model.fit() to the plt object passed as argument, and then returns the plt object. |
| Utility | Description |
|---|---|
tfutils.datasets.mnist.load_data() |
Returns the MNIST dataset after normalizing, and reshaping the examples, and one hot encoding the labels for both training and test sets. |
tfutils.datasets.mnist.plot_ten_random_examples(plt, x, y, p=None) |
Plots ten random examples from MNIST examples and labels x and y to the matplotlib.pyplot passed as plt, and returns the plt object. Optinally predictions can be passed as p. |