Image Pipelines

Here we will be showing some examples using MLBlocks to resolve image problems.

Image Classification

For the image classification examples we will be using the USPS Dataset, which we will load using the mlblocks.dataset.load_usps function.

The data of this dataset is a 3d numpy array vector with shape (224, 224, 3) containing 9298 224x224 RGB photos of handwritten digits, and the target is a 1d numpy integer array containing the label of the digit represented in the image.

OpenCV GaussianBlur + Scikit-image HOG + Scikit-Learn RandomForestClassifier

In this first example, we will attempt to resolve the problem using some basic preprocessing with the OpenCV GaussianBlur function, to later on calculate the Histogram of Oriented Gradients using the corresponding scikit-image function to later on use a simple RandomForestClassifier from scikit-learn on the generated features.

from mlblocks import MLPipeline
from mlblocks.datasets import load_usps

dataset = load_usps()
dataset.describe()

X_train, X_test, y_train, y_test = dataset.get_splits(1)

primitives = [
    'cv2.GaussianBlur',
    'skimage.feature.hog',
    'sklearn.ensemble.RandomForestClassifier'
]
init_params = {
    'skimage.feature.hog': {
        'multichannel': True,
        'visualize': False
    }
}
pipeline = MLPipeline(primitives, init_params)

pipeline.fit(X_train, y_train)

predictions = pipeline.predict(X_test)

dataset.score(y_test, predictions)

OpenCV GaussianBlur + Keras Single Layer CNN

In this example, we will preprocess the images using the OpenCV GaussianBlur function and directly after go into a Single Layer CNN Classifier built on Keras using the corresponding MLPrimitives primitive.

from mlblocks import MLPipeline
from mlblocks.datasets import load_usps

dataset = load_usps()
dataset.describe()

X_train, X_test, y_train, y_test = dataset.get_splits(1)

primitives = [
    'cv2.GaussianBlur',
    'keras.Sequential.SingleLayerCNNImageClassifier'
]
init_params = {
    'keras.Sequential.SingleLayerCNNImageClassifier': {
        'dense_units': 11,
        'epochs': 5
    }
}
pipeline = MLPipeline(primitives, init_params)

pipeline.fit(X_train, y_train)

predictions = pipeline.predict(X_test)

dataset.score(y_test, predictions)

Image Regression

For the image regression examples we will be using the Handgeometry Dataset, which we will load using the mlblocks.dataset.load_handgeometry function.

The data of this dataset is a 3d numpy array vector with shape (224, 224, 3) containing 112 224x224 RGB photos of hands, and the target is a 1d numpy float array containing the width of the wrist in centimeters.

Keras MobileNet + XGBRegressor

Here we will introduce the usage of the Pretrained Networks from Keras. In particular, we will be using the MobileNet for feature extraction, and pass its features to an XGBRegressor primitive.

from mlblocks import MLPipeline
from mlblocks.datasets import load_handgeometry

dataset = load_handgeometry()
dataset.describe()

X_train, X_test, y_train, y_test = dataset.get_splits(1)

primitives = [
    'keras.applications.mobilenet.preprocess_input',
    'keras.applications.mobilenet.MobileNet',
    'xgboost.XGBRegressor'
]
init_params = {
    'xgboost.XGBRegressor': {
        'n_estimators': 300,
        'learning_rate': 0.1
    }
}
pipeline = MLPipeline(primitives, init_params)

pipeline.fit(X_train, y_train)

predictions = pipeline.predict(X_test)

dataset.score(y_test, predictions)