A collection of deep learning -- computer vision utility functions

Installation

pip install optical_toolkit

Visualize

• Visualize a dataset in a grid

from sklearn.datasets import load_digits
from optical_toolkit.visualize import plot_images

X, y = load_digits()

plot_images(X, targets=y)

• Summarize a dataset by classes

from sklearn.datasets import load_digits
from optical_toolkit.visualize import plot_images

X, y = load_digits()

summarize_images(X, targets=y, num_images_per_class=10, num_classes=10)

• Visualize the 2d and 3d embeddings of images

from sklearn.datasets import load_digits
from optical_toolkit.visualize.embeddings import get_embeddings

X, y = load_digits()

2d_embeddings, fig_2d = get_embeddings(X, y, dims=2, embedding_type="tsne", return_plot=True)
3d_embeddings, fig_3d = get_embeddings(X, y, dims=3, embedding_type="tsne", return_plot=True)

Insight

• Visualize the filters of a (trained) CNN model

from optical_toolkit.cnn_filters import display_filters, display_model_filters

model_name = "xception"

layer_names = [
    "block2_sepconv1",
    "block5_sepconv1",
    "block9_sepconv1",
    "block14_sepconv1",
]

for layer_name in layer_names:
    display_filters(
    model=model_name,
    layer_name=layer_name,
)

display_model_filters(model=model_name)

• Visualize the filters of your custom CNN with custom objects

import keras

model_name = "examples/custom_models/svdnet.keras"
dir_name = "examples/insights"

@keras.saving.register_keras_serializable()
class ResidualConvBlock(keras.layers.Layer):
    ...

display_model_filters(
    model_name,
    custom_layer_prefix="residual",
)

Analyze

• A high level function for image dataset analysis

from sklearn.datasets import load_digits

from optical_toolkit.analyze.analyze import analyze_image_dataset

digits = load_digits()
X = digits.images
y = digits.target

analyze_image_dataset(X, y, output_path="examples/analyze/analysis.pdf")

View full analysis (PDF)