CODEWITHSUNDEEP
pythontensorflowkerasdeep-learning
François MENTEC
François MENTEC

Reputation: 1304

How to free GPU memory from keras model?

I'm loading a keras model that I previously trained, to initialize another network with his weights. Unfortunately, the model I load fills my entire memory making the training of the new model impossible. Here is the code :

import gc
import keras
from keras.models import model_from_json

def loadModel (path, loss=None, optimizer=None):
    with open(path + '/model.json', 'r') as f:
        model = model_from_json(f.read())
    model.load_weights(path + '/model.h5')
    if loss and optimizer:
        model.compile(loss=loss, optimizer=optimizer)
    return model

model = loadModel('the/path/to/my/model')

# The GPU memory is filled

keras.backend.clear_session()

# memory still filled

del model
gc.collect()

# memory still filled

I checked multiple posts, and usually gc.collect() or clear_session() does the trick, but for me, it doesn't work so far. Any idea?

PS: I'm using tensorflow as backend.

Upvotes: 0

Views: 5260

Answers (1)

François MENTEC
François MENTEC

Reputation: 1304

As @MatiasValdenegro said, tensorflow allocate the entire memory, that's why I couldn't see the difference after deleting the model. So I basically loaded my pre-trained model, created my new model and initialized his weights with those from the pre-trained model. After that, I deleted the pre-trained model using del model and gc.collect(). Since the new model has one more layer than the pre-trained one, I had to reduce my batch size to not run out of memory.

Upvotes: 1

Related Questions