CODEWITHSUNDEEP
pythontensorflowkerasdata-augmentationconv-neural-network
Bram Van Soomeren
Bram Van Soomeren

Reputation: 11

why does data augmentation not improve my performance (cnn)?

I'm relatively new to deep learning. I am trying to train a CNN model to classify spectograms of EEG data. When applying data augmentation, the model performs worse than without... What am I missing? Normally our model runs with an accuracy of 0.84 and a loss of 0.5 for both training and validation. 

datagen = ImageDataGenerator(
    featurewise_center=True,
    featurewise_std_normalization=True,
    rotation_range=0,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True)

datagen.fit(X_train)

model.fit_generator(datagen.flow(X_train, y_train, batch_size=128), validation_data= (X_test, y_test), steps_per_epoch=len(X_train) / 32, epochs=100)

after training with generated data

Upvotes: 1

Views: 2199

Answers (1)

Sebastian Dziadzio
Sebastian Dziadzio

Reputation: 530

This type of data augmentation makes sense when you're dealing with actual images. A shifted or flipped image of a dog is still an image of a dog. A flipped EEG spectrogram is a completely different signal. See here for data augmentation techniques that might be applicable in your case.

Upvotes: 1

Related Questions