def testMultilabelOneErrorOutputCorrectness(self):
y_true = np.array([[1, 0, 1], [0, 0, 1]])
y_pred = np.array([[1, 0.5, 0], [1, 0.2, 0.1]])
expected_output = 0.5
output = RankMetrics().one_error(y_true, y_pred)
self.assertAllEqual(expected_output, output)
def testRankLossOutputCorrectness(self):
y_true = np.array([[1, 0, 0], [0, 0, 1]])
y_pred = np.array([[0.75, 0.5, 1], [1, 0.2, 0.1]])
expected_output = 5.
output = RankMetrics().rank_loss(y_true, y_pred)
self.assertAllEqual(expected_output, output)
def testCoverageErrorOutputCorrectness(self):
y_true = np.array([[1, 0, 0], [0, 0, 1]])
y_pred = np.array([[0.75, 0.5, 1], [1, 0.2, 0.1]])
expected_output = 2.5
output = RankMetrics().coverage_error(y_true, y_pred)
self.assertAllEqual(expected_output, output)
def evaluation(self, model, weight_file, gen_logs = True):
metrics = MetricsAtTopK(k=3)
rank_metrics = RankMetrics()
losses = LossFunctions()
optimiser = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss=losses.weighted_binary_crossentropy, optimizer=optimiser, metrics=[losses.hamming_loss,
rank_metrics.one_error,
metrics.recall_at_k,
metrics.precision_at_k,
metrics.f1_at_k])
mcp_save = ModelCheckpoint('weights/' + weight_file + '.hdf5', save_best_only=True, monitor='val_f1_at_k', mode='max')
es = EarlyStopping(monitor='val_f1_at_k', mode='max', patience=10)
reduce_lr = ReduceLROnPlateau(monitor='val_f1_at_k', mode='max', factor=0.2, patience=5, verbose=1)
callbacks = [mcp_save, es, reduce_lr]
if gen_logs:
logdir = os.path.join("logs", datetime.now().strftime("%Y%m%d-%H%M%S"))
tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir,
histogram_freq=1,
write_graph=True,
write_images=True,
embeddings_freq=0)
callbacks.append(tensorboard_callback)
model.fit_generator(
self.train_data,
steps_per_epoch=self.train_data.samples // self.batch_size,
epochs=self.epochs,
validation_data=self.test_data,
validation_steps=self.test_data.samples // self.batch_size,
validation_freq=5,
callbacks=callbacks)