Jupyter Notebook kernel 在编译神经网络时死机

Question

While creating and compiling a keras Dense Neural Network my jupyter notebook kernel always dies. The terminal gives me a message that it couldn't allocate space and that CUDA is out of memory. My GPU (2060 Super) has already run this model many times, just not on jupyter. I have already done a lot of searching whithout an answer that actually works. Some of the things I tried were changing my kernel, using numba device.reset() and reinstalling both conda and jupyter but nothing seems to work.

Here is the code block that always gets the error:

inputs = keras.Input(shape=(69,))
fc1 = keras.layers.Dense(100, activation='relu')(inputs)
d1 = keras.layers.Dropout(0.1)(fc1, training=True)
bn1 = keras.layers.BatchNormalization()(d1)
fc2 = keras.layers.Dense(150, activation='relu')(bn1)
d2 = keras.layers.Dropout(0.1)(fc2, training=True)
bn2 = keras.layers.BatchNormalization()(d2)
fc3 = keras.layers.Dense(200, activation='relu')(bn2)
d3 = keras.layers.Dropout(0.1)(fc3, training=True)
bn3 = keras.layers.BatchNormalization()(d3)
fc4 = keras.layers.Dense(120, activation='relu')(bn3)
d4 = keras.layers.Dropout(0.1)(fc4, training=True)
bn4 = keras.layers.BatchNormalization()(d4)
fc5 = keras.layers.Dense(60, activation='relu')(bn4)
bn5 = keras.layers.BatchNormalization()(fc5)
outputs = keras.layers.Dense(2, activation='softmax')(bn5)
model = keras.Model(inputs, outputs)

model.summary()
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])

Answer 1

Try adding this function to your notebook. I am not certain but it may help.

def setup_gpus():
gpus = tf.config.list_physical_devices('GPU')
if gpus:
    try:
        tf.config.experimental.set_visible_devices(gpus[0],'GPU')
        tf.config.experimental.set_virtual_device_configuration(gpus[0],[tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1500)])
    except RuntimeError as e:
        print(e)