Keras Data Augmentation for Scikit-learn
Here is a method to integrate a preprocessing utility from Keras with a model from Scikit-learn. The complete Jupyter notebook is in the reference section below.
We will be working with the MNIST data set:
- 60,000 Examples for the training set.
- 10,000 Examples for the test set.
- 784 Features (Images of handwritten digits, 28 * 28 pixels).
We will use a KNeighborsClassifier, a lazy learner! It memories the data during training time and learns during runtime.
Here we will use the ImageDataGenerator class from Keras. First, we create an object:
datagen = ImageDataGenerator(
width_shift_range= 2.0,
height_shift_range= 2.0,
rotation_range = 20,
fill_mode = 'constant',
)This object will:
- Shift the image by 2 pixels vertically and horizontally.
- Rotates the image by 20 degrees.
Calling the fit method:
datagen.fit(X_train.reshape(X_train.shape[0], 28, 28, 1))The fit method expects a rank 4 sample as an input (Examples, width, height, channels {1 for grayscale, 3 for RGB…}).
Then we call the flow method to generate the new data set:
data_generator = datagen.flow(X_train.reshape(X_train.shape[0], 28, 28, 1),shuffle=False, batch_size=1)We need to preserve the order of the data so that it matches the labels. The batch size should be 1 to get the whole data set.
type(data_generator) # => NumpyArrayIteratorThe new data is an iterator, ideal for incremental learners, but our model expects an array-like list.
We have to materialize the iterator to using the comprehension list:
X_train_aug = [data_generator.next() for i in range(0, m * 4)]This step will turn the iterator into a list. The 4 is a multiplier for how many sets to generate (4 * 60k = 240k).
The last step is to reshape the data from rank 4 to its original shape of rank 2:
X_train_240k = np.asarray(X_train_aug).reshape(m * 4, 28 * 28)When we plot, we get this.
Now the data is ready for a second run!
