simplified-keras 0.0.22

Common used code in Keras

simplified keras

Common used actions in keras

Table of contents

• General info
• Libraries
• Setup
• Documentation
• PyPi
• TODO
• Development
• Status
• Contact

General info

This package is a set of common used actions in keras. At this moment includes (may be outdated):

• Main package
  • Default callbacks
  • Restore callbacks
  • Generators
• Plots
  • Accuracy and Loss plot
  • Predictions with image plot
  • Histogram with CDF and image plot
  • Confusion matrix
• Metrics
  • Confusion matrix
  • Model Statistics
  • Folder Statistics
  • Model memory usage
• Transformations
  • Convert predictions to classes array
  • Convert one hot encoding to sparse
  • Unfreeze model
  • Stretch histogram
  • Normalize histogram clahe
  • Replace activations
  • Layers

Libraries

• Matplotlib - version 3.3.3
• NumPy - version 1.19.4
• Tensorflow - version 2.4.1
• Pandas - version 1.1.5
• Seaborn - version 0.11.1

Setup

• Install from PyPi: pip install simplified-keras

Documentation

Main package

Generators

default generators

from keras.preprocessing.image import ImageDataGenerator
from simplified_keras.generators import get_train_val_generators, get_val_test_generators

img_size = (48, 48)
img_datagen = ImageDataGenerator(rescale=1/255)

# same for get_val_test_generators
# default: data_dir='../data', color_mode='rgb', batch_size=128, class_mode='categorical'
train_generator, validation_generator = get_train_val_generators(img_datagen, data_dir='../data/normal',
                                                                 color_mode='grayscale', target_size=img_size)
val_generator, test_generator = get_val_test_generators(img_datagen, batch_size=32)

numpy_memmap_generator

from simplified_keras.generators import numpy_memmap_generator

# default batch_size=128, shuffle_array=True
train_gen = numpy_memmap_generator('imgs.npy', 'labels.npy', batch_size=64, shuffle_array=False)

Default callbacks

from simplified_keras.callbacks import get_default_callbacks

callbacks = get_default_callbacks('models/vgg16_calssifier.h5', monitor='val_loss', verbose=0)

hist = model.fit(train_generator, steps_per_epoch=train_steps, validation_data=validation_generator,
                 validation_steps=valid_steps, epochs=100, callbacks=callbacks, verbose=2)

Signature:

def get_default_callbacks(model_path, monitor='val_acc', base_patience=3, lr_reduce_factor=0.5, min_lr=1e-7, verbose=1):
    return [
        ReduceLROnPlateau(monitor=monitor, factor=lr_reduce_factor, min_lr=min_lr, patience=base_patience, verbose=verbose),
        EarlyStopping(monitor=monitor, patience=(2 * base_patience + 1), verbose=verbose),
        ModelCheckpoint(monitor=monitor, filepath=model_path, save_best_only=True, verbose=verbose)
    ]

Restore callbacks

Used to restore callback after paused learning. Model should come from last checkpoint.

from simplified_keras.callbacks import restore_callbacks, get_default_callbacks

callbacks = get_default_callbacks('models/vgg16_calssifier.h5', monitor='val_loss', verbose=0)
acc, loss = model.evaluate(val_ds)

# acc or loss depending on the compiled model metrics
restore_callbacks(callbacks, acc)

Plots

Accuracy and Loss plot

from simplified_keras.plots.history_plots import plot_acc_and_loss

history = model.fit(train_gen, teps_per_epoch=train_steps, epochs=5, validation_data=val_gen, 
                    validation_steps=val_steps, callbacks=callbacks)

fig = plot_acc_and_loss(history)

Result:

Predictions with image plot

from keras.models import load_model
from keras.preprocessing.image import ImageDataGenerator
from simplified_keras.generators import get_train_val_generators
from simplified_keras.plots import plot_predictions_with_img

img_size = (48, 48)
img_datagen = ImageDataGenerator(rescale=1 / 255)

_, validation_generator = get_train_val_generators(img_datagen, data_dir='../data/normal',
                                                   color_mode='grayscale', target_size=img_size)
model = load_model('../models/standard_model.h5')

batch, labels = validation_generator.next()
preds = model.predict(batch)

named_labels = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
fig = plot_predictions_with_img(1, preds, labels, batch, named_labels, grayscale=True)

Result:

Histogram with CDF and image plot

import cv2
from simplified_keras.plots import plot_gray_img_with_histogram

img = cv2.imread(f'{src_train_path}/0/241.png', 0)
fig1 = plot_gray_img_with_histogram(img)
img2 = stretch_histogram(img)
fig2 = plot_gray_img_with_histogram(img2)

Result:

Confusion matrix plot

from simplified_keras.transformations import predictions_to_classes, one_hot_to_sparse
from simplified_keras.metrics import get_confusion_matrixes
from simplified_keras.plots import plot_confusion_matrix

predictions = model.predict(validation_images)
predicted_classes = predictions_to_classes(predictions)
sparse_labels = one_hot_to_sparse(validation_labels)

cm, cm_normalized = get_confusion_matrixes(predicted_classes, sparse_labels)
classes = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
f1 = plot_confusion_matrix(cm, classes)
f2 = plot_confusion_matrix(cm_normalized, classes, figsize=(10, 8))

Results:

Metrics

Confusion matrix

from simplified_keras.transformations import predictions_to_classes, one_hot_to_sparse
from simplified_keras.metrics import get_confusion_matrixes

predictions = model.predict(validation_images)
predicted_classes = predictions_to_classes(predictions)
sparse_labels = one_hot_to_sparse(validation_labels)

# Returns two numpy arrays: standard and normalized
cm, cm_normalized = get_confusion_matrixes(predicted_classes, sparse_labels)

Model Statistics

Calculates:

• FP
• FN
• TP
• TN
• TPR # Sensitivity/true positive rate
• TNR # Specificity/true negative rate
• PPV # Precision/positive predictive value
• NPV # Negative predictive value
• FPR # Fall out or false positive rate
• FNR # False negative rate
• FDR # False discovery rate
• ACC # Overall accuracy for each class
• Much more and still increasing

from simplified_keras.transformations import predictions_to_classes, one_hot_to_sparse
from simplified_keras.metrics import get_confusion_matrixes
from simplified_keras.metrics import get_model_statistics

predictions = model.predict(validation_images)
predicted_classes = predictions_to_classes(predictions)
sparse_labels = one_hot_to_sparse(validation_labels)

cm, cm_normalized = get_confusion_matrixes(predicted_classes, sparse_labels)

stats = get_model_statistics(cm)
classes = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
fig = stats.visualize(classes)
print(stats.TN) #[2890 3530 2874 2361 2591 3000 2661]

Visualization:

Folder statistics

from simplified_keras.metrics import get_folders_statistics

stat = get_folders_statistics('../data/normal/train')
print(stat.nr_of_elements, stat.info) # 28709 {'0': 3995, '1': 436, '2': 4097, '3': 7215, '4': 4830, '5': 3171, '6': 4965}
fig = stat.bar_plot()

Result:

Model memory usage

from simplified_keras.metrics import get_model_memory_usage

batch_size = 64
# outputs usage in GB
usage = get_model_memory_usage(batch_size, model)
print(usage, 'GB') # 8.34 GB

Transformations

Convert predictions to classes array

from simplified_keras.transformations import predictions_to_classes

predictions = model.predict(validation_images)
predicted_classes = predictions_to_classes(predictions)
print(pedicted_classes) #[6 3 3 ... 6 2 0]

Convert one hot encoding to sparse

from simplified_keras.transformations import one_hot_to_sparse

sparse_labels = one_hot_to_sparse(validation_labels)
print(sprase_labels) #[6 6 6 ... 6 2 0]

Stretch histogram

from simplified_keras.transformations import stretch_histogram

# default color_mode='rgb'
stretch_histogram(image, color_mode='grayscale')

Unfreeze model

from simplified_keras.transformations import unfreeze_model
from tensorflow.keras.optimizers import Adam

# default params: optimizer=Adam(learning_rate=1e-5), metrics="acc"
unfreeze_model(model, optimizer=Adam(learning_rate=1e-4), metrics="loss")

Normalize histogram clahe

from simplified_keras.transformations import normalize_histogram_clahe

# default clip_limit=2.0, tile_grid_size=(8, 8), color_mode='rgb'
normalize_histogram_clahe(image)

Replace activations

Replaces all activation functions in given model

from simplified_keras.transformations import replace_activations
from tensorflow.keras.layers import LeakyReLU

l_relu = LeakyReLU()
replace_activations(model, l_relu)

Layers

Augumentation layers build on tensorflow image operations

from simplified_keras.transformations import RandomSaturation, RandomHue, RandomBrightness
from tensorflow.keras import Sequential

# for more informations about parameters see tf.image docs
augument_layers = Sequential([
  RandomSaturation(0.5, 1.5),
  # must be [0 - 0,5]
  RandomHue(0.2),
  RandomBrightness(0.2)
])

PyPi

simplified-keras

TODO

• Add tests and a lot of features :)

Development

Want to contribute? Great!

To fix a bug or enhance an existing module, follow these steps:

• Fork the repo
• Create a new branch (git checkout -b improve-feature)
• Make the appropriate changes in the files
• Verify if they are correct
• Add changes to reflect the changes made
• Commit changes
• Push to the branch (git push origin improve-feature)
• Create a Pull Request

Status

Library is: in progress

Contact

[e-mail] - feel free to contact me!