如何在Keras / Tensorflow中返回增强数据

Question

I have a pre-trained network in augmented data, but I want to extract feature vectors from the last but one layer and train another classifier (eg, svm). To do that I need to extract the output on the augmented training data and from the test data.

However, I am quite noob in Keras/tensorflow and I just need to have the augmented training data in a numpy array to use it in my feature extractor code. I can do this if I dont use augmented training data with no problem.

Here is what I tried so far:

#train on augmented data
model.fit_generator(datagen.flow(x_train, y_train,
                                     batch_size=batch_size),
                                     steps_per_epoch=x_train.shape[0] // batch_size,
                                     epochs=epochs,
                                     validation_data=(x_test, y_test))

#extract augmented data. Is this correct?
x_train_augmented, y_train_augmented=datagen.flow(x_train, y_train, batch_size=batch_size)

According to Keras Documentation function flow(X, y): Takes numpy data & label arrays, and generates batches of augmented/normalized data. Yields batches indefinitely, in an infinite loop.

So, how can I loop and return the augmented data in a matrix of shape (num_images, width, height, channels)?

Answer 1

Suppose you have N number of grayscale images of size 28x28, then you can use

for x_train_augmented, y_train_augmented in datagen.flow(x_train, y_train, batch_size=batch_size):
        x_train_data = x_train_augmented.reshape(-1, 28, 28, 1)

Here, the shape of the x_train_data will be [N, 28, 28, 1].