Out of memory (OOM) error of tensorflow/keras model

Question

when i tried to add dropout to the keras model it cause OOM error: tensorflow.python.framework.errors_impl.ResourceExhaustedError: OOM when allocating tensor with shape[128,128,176,216]...

the model suppose to be autoencoder that produce super resolution x4.

autoencoder = Sequential()
autoencoder.add(Conv2D(64*comlex, (3, 3), activation='relu', 
padding='same', input_shape=x_train[0].shape))
autoencoder.add(Dropout(0.25))
autoencoder.add(UpSampling2D((2, 2)))
autoencoder.add(Conv2D(64*comlex, (3, 3), activation='relu', padding='same'))
# autoencoder.add(Dropout(0.25))
autoencoder.add(UpSampling2D((2, 2)))
autoencoder.add(Conv2D(3, (3, 3), activation='sigmoid', padding='same'))
autoencoder.compile(optimizer='adam', loss='binary_crossentropy',metrics=['accuracy'])

the line in comment causes OOM.

why does dropout take so much memory?

It doesn't, its probably just that your feature maps are really big (as seen in the error msg) and you don't have enough RAM for it. Try reducing the batch size. — Dr. Snoopy

Geeocode Geeocode · Accepted Answer · 2018-12-06T23:50:36

Update:

Tensorflow/Keras OOM(Out of memory) error occures because of the excess amount of model parameters e.g. image size/feature number etc. with high batch size value, which can cause higher GPU memory consumption. Sometimes these can associate with residual memory from previous processes as well.

Out of memory (OOM) error of tensorflow/keras model

1 Answers