def generate(self, X, y, reject=True, same_size=True):
inputs = X.copy()
targets = one_hot_encoding(y, self.num_classes)
#noise = self.noise_generator(size=(inputs.shape[0], self.prior_dim))
noise = self.noise_generator.generate(X, y, self.block)
feed_dict = {}
feed_dict[self.inputs] = inputs
feed_dict[self.targets] = targets
feed_dict[self.prior_noise] = noise
feed_dict[self.masks] = np.broadcast_to(self.mask, shape=inputs.shape)
feed_dict[self.is_training] = False
feed_dict[self.dis_l2_scale] = self.cur_dis_l2_scale
return self.session.run(self.gen_inputs, feed_dict=feed_dict)
"""
def pre_training(self, X, y):
batch_size = X.shape[0]
feed_dict = {}
feed_dict[self.inputs] = X
#noise = self.noise_generator(size=(batch_size, self.prior_dim))
noise = np.random.uniform(size=(batch_size, self.prior_dim))
feed_dict[self.prior_noise] = noise
feed_dict[self.targets] = one_hot_encoding(y, self.num_classes)
feed_dict[self.masks] = np.broadcast_to(self.mask, shape=X.shape)
#feed_dict[self.is_training] = True
feed_dict[self.dis_l2_scale] = self.cur_dis_l2_scale
feed_dict[self.is_gen] = 1.
#feed_dict[self.source_targets] = np.zeros((batch_size, 1))
feed_dict[self.source_targets] = np.ones(
(batch_size,
1)) # min log(1-D) has vanishing gradients, use max logD
self.session.run(self.pretraining_optimizer, feed_dict=feed_dict)
return self.session.run(self.pretraining_error, feed_dict=feed_dict)
def train_discriminator_step(self, X, y):
batch_size = X.shape[0]
feed_dict = {}
feed_dict[self.inputs] = X
#noise = self.noise_generator(size=(batch_size, self.prior_dim))
noise = self.noise_generator.generate(X, y, self.block)
feed_dict[self.prior_noise] = noise
feed_dict[self.targets] = one_hot_encoding(y, self.num_classes)
feed_dict[self.masks] = np.broadcast_to(self.mask, shape=X.shape)
feed_dict[self.is_training] = True
feed_dict[self.dis_l2_scale] = self.cur_dis_l2_scale
feed_dict[self.is_gen] = 1.
feed_dict[self.source_targets] = np.zeros((batch_size, 1))
self.session.run(self.discriminator_optimizer, feed_dict=feed_dict)
feed_dict[self.is_gen] = 0.
feed_dict[self.source_targets] = np.ones((batch_size, 1))
#feed_dict[self.source_targets] = np.concatenate([np.zeros((batch_size, 1)), np.ones((batch_size, 1))], axis=0)
self.session.run(self.discriminator_optimizer, feed_dict=feed_dict)
return self.session.run([self.source_error, self.classification_error],
feed_dict=feed_dict)
def evaluate(self, X, y, batch_size=100):
y = one_hot_encoding(y)
return self.model.evaluate(X, y, batch_size=batch_size)
def fit(self, X, y, max_num_epochs=500, validation_split=0.2, batch_size=100, verbose=1):
y = one_hot_encoding(y)
early_stopping = EarlyStopping(monitor='val_loss', patience=5)
self.model.fit(X, y, validation_split=validation_split, callbacks=[early_stopping], epochs=max_num_epochs, batch_size=batch_size, verbose=verbose)