when training= false still dropout

from keras import layers<br /> from keras import models<br /> from keras import backend as K<br /> import numpy as np</p> <p>inp = layers.Input(shape=(10,))<br /> out = layers.Dropout(0.5)(inp)</p> <p>model = models.Model(inp, out)<br /> model.layers[-1].trainable = False # set dropout layer as non-trainable<br /> model.compile(optimizer=”adam”, loss=”mse”) # IMPORTANT: we must always compile model after changing `trainable` attribute</p> <p># create a custom backend function so that we can control the learning phase<br /> func = K.function(model.inputs + [K.learning_phase()], model.outputs)</p> <p>x = np.ones((1,10))<br /> # learning phase = 1, i.e. training mode<br /> print(func([x, 1]))<br /> # the output will be:<br /> [array([[2., 2., 2., 0., 0., 2., 2., 2., 0., 0.]], dtype=float32)]<br /> # as you can see some of the neurons have been dropped</p> <p># now set learning phase = 0, i.e test mode<br /> print(func([x, 0]))<br /> # the output will be:<br /> [array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.]], dtype=float32)]<br /> # unsurprisingly, no neurons have been dropped in test phase