def test_save_restore(self):
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
m = metrics.Mean()
checkpoint = tf.train.Checkpoint(mean=m)
self.evaluate(tf.compat.v1.variables_initializer(m.variables))
# update state
self.evaluate(m(100.0))
self.evaluate(m(200.0))
# save checkpoint and then add an update
save_path = checkpoint.save(checkpoint_prefix)
self.evaluate(m(1000.0))
# restore to the same checkpoint mean object
checkpoint.restore(save_path).assert_consumed().run_restore_ops()
self.evaluate(m(300.0))
self.assertEqual(200.0, self.evaluate(m.result()))
# restore to a different checkpoint mean object
restore_mean = metrics.Mean()
restore_checkpoint = tf.train.Checkpoint(mean=restore_mean)
status = restore_checkpoint.restore(save_path)
restore_update = restore_mean(300.0)
status.assert_consumed().run_restore_ops()
self.evaluate(restore_update)
self.assertEqual(200.0, self.evaluate(restore_mean.result()))
self.assertEqual(3, self.evaluate(restore_mean.count))
def test_multiple_instances(self):
m = metrics.Mean()
m2 = metrics.Mean()
assert m.name == 'mean'
assert m2.name == 'mean'
# check initial state
assert K.eval(m.total) == 0
assert K.eval(m.count) == 0
assert K.eval(m2.total) == 0
assert K.eval(m2.count) == 0
# check __call__()
assert K.eval(m(100)) == 100
assert K.eval(m.total) == 100
assert K.eval(m.count) == 1
assert K.eval(m2.total) == 0
assert K.eval(m2.count) == 0
assert K.eval(m2([63, 10])) == 36.5
assert K.eval(m2.total) == 73
assert K.eval(m2.count) == 2
assert K.eval(m.result()) == 100
assert K.eval(m.total) == 100
assert K.eval(m.count) == 1
def create_mean_metric(value, name=None):
# import keras will import base_layer and then this module, and metric
# relies on base_layer, which result into a cyclic dependency.
from keras import metrics as metrics_module
metric_obj = metrics_module.Mean(name=name, dtype=value.dtype)
return metric_obj, metric_obj(value)
def test_mean(self):
m = metrics.Mean(name='my_mean')
# check config
assert m.name == 'my_mean'
assert m.stateful
assert m.dtype == 'float32'
assert len(m.weights) == 2
# check initial state
assert K.eval(m.total) == 0
assert K.eval(m.count) == 0
# check __call__()
assert K.eval(m(100)) == 100
assert K.eval(m.total) == 100
assert K.eval(m.count) == 1
# check update_state() and result()
result = m([1, 5])
assert np.isclose(K.eval(result), 106. / 3)
assert K.eval(m.total) == 106 # 100 + 1 + 5
assert K.eval(m.count) == 3
# check reset_states()
m.reset_states()
assert K.eval(m.total) == 0
assert K.eval(m.count) == 0
# Check save and restore config
m2 = metrics.Mean.from_config(m.get_config())
assert m2.name == 'my_mean'
assert m2.stateful
assert m2.dtype == 'float32'
assert len(m2.weights) == 2
def test_mean_with_sample_weight(self):
m = metrics.Mean(dtype='float64')
assert m.dtype == 'float64'
# check scalar weight
result_t = m(100, sample_weight=0.5)
assert K.eval(result_t) == 50. / 0.5
assert K.eval(m.total) == 50
assert K.eval(m.count) == 0.5
# check weights not scalar and weights rank matches values rank
result_t = m([1, 5], sample_weight=[1, 0.2])
result = K.eval(result_t)
assert np.isclose(result, 52. / 1.7)
assert np.isclose(K.eval(m.total), 52) # 50 + 1 + 5 * 0.2
assert np.isclose(K.eval(m.count), 1.7) # 0.5 + 1.2
# check weights broadcast
result_t = m([1, 2], sample_weight=0.5)
assert np.isclose(K.eval(result_t), 53.5 / 2.7, rtol=3)
assert np.isclose(K.eval(m.total), 53.5, rtol=3) # 52 + 0.5 + 1
assert np.isclose(K.eval(m.count), 2.7, rtol=3) # 1.7 + 0.5 + 0.5
# check weights squeeze
result_t = m([1, 5], sample_weight=[[1], [0.2]])
assert np.isclose(K.eval(result_t), 55.5 / 3.9, rtol=3)
assert np.isclose(K.eval(m.total), 55.5, rtol=3) # 53.5 + 1 + 1
assert np.isclose(K.eval(m.count), 3.9, rtol=3) # 2.7 + 1.2
# check weights expand
result_t = m([[1], [5]], sample_weight=[1, 0.2])
assert np.isclose(K.eval(result_t), 57.5 / 5.1, rtol=3)
assert np.isclose(K.eval(m.total), 57.5, rtol=3) # 55.5 + 1 + 1
assert np.isclose(K.eval(m.count), 5.1, rtol=3) # 3.9 + 1.2
def test_reset_state_existing_metric_before_built(self):
metric = metrics_mod.Mean()
metric.update_state([2.0, 4.0])
self.assertEqual(metric.result().numpy(), 3.0)
metric_container = compile_utils.MetricsContainer(metric)
metric_container.reset_state()
self.assertEqual(metric.result().numpy(), 0.0)
def test_multiple_keras_metrics_experimental_run(self, distribution):
with distribution.scope():
loss_metric = metrics.Mean("loss", dtype=np.float32)
loss_metric_2 = metrics.Mean("loss_2", dtype=np.float32)
@tf.function
def train_step():
def step_fn():
loss = tf.constant(5.0, dtype=np.float32)
loss_metric.update_state(loss)
loss_metric_2.update_state(loss)
distribution.run(step_fn)
train_step()
self.assertEqual(loss_metric.result().numpy(),
loss_metric_2.result().numpy())
self.assertEqual(loss_metric.result().numpy(), 5.0)
def test_update_keras_metric_outside_strategy_scope_cross_replica(
self, distribution):
metric = metrics.Mean("test_metric", dtype=np.float32)
with distribution.scope():
for i in range(10):
metric.update_state(i)
# This should be the mean of integers 0-9 which has a sum of 45 and a count
# of 10 resulting in mean of 4.5.
self.assertEqual(metric.result().numpy(), 4.5)
def _create_metrics(self):
"""Creates per-output loss metrics, but only for multi-output Models."""
if len(self._output_names) == 1:
self._per_output_metrics = [None]
else:
self._per_output_metrics = []
for loss_obj, output_name in zip(self._losses, self._output_names):
if loss_obj is None:
self._per_output_metrics.append(None)
else:
self._per_output_metrics.append(
metrics_mod.Mean(output_name + '_loss'))
def test_function_wrapped_reset_state(self):
m = metrics.Mean(name="my_mean")
# check reset_state in function.
@tf.function
def reset_in_fn():
m.reset_state()
return m.update_state(100)
for _ in range(5):
self.evaluate(reset_in_fn())
self.assertEqual(self.evaluate(m.count), 1)
def __init__(self, losses, loss_weights=None, output_names=None):
super(LossesContainer, self).__init__(output_names=output_names)
# Keep user-supplied values untouched for recompiling and serialization.
self._user_losses = losses
self._user_loss_weights = loss_weights
self._losses = losses
self._loss_weights = loss_weights
self._per_output_metrics = None # Per-output losses become metrics.
self._loss_metric = metrics_mod.Mean(name='loss') # Total loss.
self._built = False
def test_multiple_instances(self):
m = metrics.Mean()
m2 = metrics.Mean()
self.assertEqual(m.name, "mean")
self.assertEqual(m2.name, "mean")
self.assertEqual(
[v.name for v in m.variables],
test_utils.get_expected_metric_variable_names(["total", "count"]),
)
self.assertEqual(
[v.name for v in m2.variables],
test_utils.get_expected_metric_variable_names(["total", "count"],
name_suffix="_1"),
)
self.evaluate(tf.compat.v1.variables_initializer(m.variables))
self.evaluate(tf.compat.v1.variables_initializer(m2.variables))
# check initial state
self.assertEqual(self.evaluate(m.total), 0)
self.assertEqual(self.evaluate(m.count), 0)
self.assertEqual(self.evaluate(m2.total), 0)
self.assertEqual(self.evaluate(m2.count), 0)
# check __call__()
self.assertEqual(self.evaluate(m(100)), 100)
self.assertEqual(self.evaluate(m.total), 100)
self.assertEqual(self.evaluate(m.count), 1)
self.assertEqual(self.evaluate(m2.total), 0)
self.assertEqual(self.evaluate(m2.count), 0)
self.assertEqual(self.evaluate(m2([63, 10])), 36.5)
self.assertEqual(self.evaluate(m2.total), 73)
self.assertEqual(self.evaluate(m2.count), 2)
self.assertEqual(self.evaluate(m.result()), 100)
self.assertEqual(self.evaluate(m.total), 100)
self.assertEqual(self.evaluate(m.count), 1)
def test_deepcopy_of_metrics(self):
m = metrics.Mean(name="my_mean")
m.reset_state()
m.update_state(100)
m_copied = copy.deepcopy(m)
m_copied.update_state(200)
self.assertEqual(self.evaluate(m.result()), 100)
self.assertEqual(self.evaluate(m_copied.result()), 150)
m.reset_state()
self.assertEqual(self.evaluate(m.result()), 0)
self.assertEqual(self.evaluate(m_copied.result()), 150)
def test_duplicated_metric_instance(self):
mean_obj = metrics_mod.Mean()
metric = mean_obj
with self.assertRaisesRegex(ValueError, "Found duplicated metrics"):
compile_utils.MetricsContainer(metrics=metric,
weighted_metrics=metric)
# duplicated string should be fine
metric = "acc"
compile_utils.MetricsContainer(metrics=metric, weighted_metrics=metric)
# complicated structure
metric = [mean_obj, "acc"]
weighted_metric = {"output1": mean_obj, "output2": "acc"}
with self.assertRaisesRegex(ValueError, "Found duplicated metrics"):
compile_utils.MetricsContainer(metrics=metric,
weighted_metrics=weighted_metric)
def test_update_keras_metric_declared_in_strategy_scope(self, distribution):
with distribution.scope():
metric = metrics.Mean("test_metric", dtype=np.float32)
dataset = tf.data.Dataset.range(10).batch(2)
dataset = distribution.experimental_distribute_dataset(dataset)
@tf.function
def step_fn(i):
metric.update_state(i)
for i in dataset:
distribution.run(step_fn, args=(i,))
# This should be the mean of integers 0-9 which has a sum of 45 and a count
# of 10 resulting in mean of 4.5.
self.assertEqual(metric.result().numpy(), 4.5)
def test_duplicated_metric_instance(self):
mean_obj = metrics_mod.Mean()
metric = mean_obj
with self.assertRaisesRegex(ValueError, 'Found duplicated metrics'):
compile_utils.MetricsContainer(metrics=metric,
weighted_metrics=metric)
# duplicated string should be fine
metric = 'acc'
compile_utils.MetricsContainer(metrics=metric, weighted_metrics=metric)
# complicated structure
metric = [mean_obj, 'acc']
weighted_metric = {'output1': mean_obj, 'output2': 'acc'}
with self.assertRaisesRegex(ValueError, 'Found duplicated metrics'):
compile_utils.MetricsContainer(metrics=metric,
weighted_metrics=weighted_metric)
def test_mean_with_sample_weight(self):
m = metrics.Mean(dtype=tf.float64)
self.assertEqual(m.dtype, tf.float64)
self.evaluate(tf.compat.v1.variables_initializer(m.variables))
# check scalar weight
result_t = m(100, sample_weight=0.5)
self.assertEqual(self.evaluate(result_t), 50 / 0.5)
self.assertEqual(self.evaluate(m.total), 50)
self.assertEqual(self.evaluate(m.count), 0.5)
# check weights not scalar and weights rank matches values rank
result_t = m([1, 5], sample_weight=[1, 0.2])
result = self.evaluate(result_t)
self.assertAlmostEqual(result, 52 / 1.7, 2)
self.assertAlmostEqual(self.evaluate(m.total), 52,
2) # 50 + 1 + 5 * 0.2
self.assertAlmostEqual(self.evaluate(m.count), 1.7, 2) # 0.5 + 1.2
# check weights broadcast
result_t = m([1, 2], sample_weight=0.5)
self.assertAlmostEqual(self.evaluate(result_t), 53.5 / 2.7, 2)
self.assertAlmostEqual(self.evaluate(m.total), 53.5, 2) # 52 + 0.5 + 1
self.assertAlmostEqual(self.evaluate(m.count), 2.7,
2) # 1.7 + 0.5 + 0.5
# check weights squeeze
result_t = m([1, 5], sample_weight=[[1], [0.2]])
self.assertAlmostEqual(self.evaluate(result_t), 55.5 / 3.9, 2)
self.assertAlmostEqual(self.evaluate(m.total), 55.5, 2) # 53.5 + 1 + 1
self.assertAlmostEqual(self.evaluate(m.count), 3.9, 2) # 2.7 + 1.2
# check weights expand
result_t = m([[1], [5]], sample_weight=[1, 0.2])
self.assertAlmostEqual(self.evaluate(result_t), 57.5 / 5.1, 2)
self.assertAlmostEqual(self.evaluate(m.total), 57.5, 2) # 55.5 + 1 + 1
self.assertAlmostEqual(self.evaluate(m.count), 5.1, 2) # 3.9 + 1.2
# check values reduced to the dimensions of weight
result_t = m([[[1.0, 2.0], [3.0, 2.0], [0.5, 4.0]]],
sample_weight=[0.5])
result = np.round(self.evaluate(result_t), decimals=2) # 58.5 / 5.6
self.assertEqual(result, 10.45)
self.assertEqual(np.round(self.evaluate(m.total), decimals=2), 58.54)
self.assertEqual(np.round(self.evaluate(m.count), decimals=2), 5.6)
def test_update_keras_metrics_dynamic_shape(self, distribution):
with distribution.scope():
metric = metrics.Mean("test_metric", dtype=np.float32)
dataset = tf.data.Dataset.range(10).batch(2, drop_remainder=False)
@tf.function
def train_fn(dataset):
weights = tf.constant([0.1, 0.1])
def step_fn(i):
metric.update_state(i, weights)
for i in dataset:
distribution.run(step_fn, args=(i, ))
train_fn(dataset)
# This should be the mean of integers 0-9 which has a sum of 45 and a count
# of 10 resulting in mean of 4.5.
self.assertEqual(metric.result().numpy(), 4.5)
def test_mean_graph_with_placeholder(self):
with tf.compat.v1.get_default_graph().as_default(
), self.cached_session() as sess: # noqa: E501
m = metrics.Mean()
v = tf.compat.v1.placeholder(tf.float32)
w = tf.compat.v1.placeholder(tf.float32)
self.evaluate(tf.compat.v1.variables_initializer(m.variables))
# check __call__()
result_t = m(v, sample_weight=w)
result = sess.run(result_t, feed_dict=({v: 100, w: 0.5}))
self.assertEqual(self.evaluate(m.total), 50)
self.assertEqual(self.evaluate(m.count), 0.5)
self.assertEqual(result, 50 / 0.5)
# check update_state() and result()
result = sess.run(result_t, feed_dict=({v: [1, 5], w: [1, 0.2]}))
self.assertAlmostEqual(self.evaluate(m.total), 52,
2) # 50 + 1 + 5 * 0.2
self.assertAlmostEqual(self.evaluate(m.count), 1.7, 2) # 0.5 + 1.2
self.assertAlmostEqual(result, 52 / 1.7, 2)
def test_mean(self):
m = metrics.Mean(name="my_mean")
# check config
self.assertEqual(m.name, "my_mean")
self.assertTrue(m.stateful)
self.assertEqual(m.dtype, tf.float32)
self.assertEqual(len(m.variables), 2)
self.evaluate(tf.compat.v1.variables_initializer(m.variables))
# check initial state
self.assertEqual(self.evaluate(m.total), 0)
self.assertEqual(self.evaluate(m.count), 0)
# check __call__()
self.assertEqual(self.evaluate(m(100)), 100)
self.assertEqual(self.evaluate(m.total), 100)
self.assertEqual(self.evaluate(m.count), 1)
# check update_state() and result() + state accumulation + tensor input
update_op = m.update_state(
[tf.convert_to_tensor(1),
tf.convert_to_tensor(5)])
self.evaluate(update_op)
self.assertAlmostEqual(self.evaluate(m.result()), 106 / 3, 2)
self.assertEqual(self.evaluate(m.total), 106) # 100 + 1 + 5
self.assertEqual(self.evaluate(m.count), 3)
# check reset_state()
m.reset_state()
self.assertEqual(self.evaluate(m.total), 0)
self.assertEqual(self.evaluate(m.count), 0)
# Check save and restore config
m2 = metrics.Mean.from_config(m.get_config())
self.assertEqual(m2.name, "my_mean")
self.assertTrue(m2.stateful)
self.assertEqual(m2.dtype, tf.float32)
self.assertEqual(len(m2.variables), 2)
# (X_train, Y_train), (X_test, Y_test) = mnist.load_data()
# X_train, X_test = X_train / 255.0, X_test / 255.0
# X_train = X_train[..., tf.newaxis].astype(np.float32)
# X_test = X_test[..., tf.newaxis].astype(np.float32)
# X와 Y로 데이터프레임을 만드는 법
train_ds = tf.data.Dataset.from_tensor_slices((X_train, Y_train)).shuffle(10000).batch(batch_size)
test_ds = tf.data.Dataset.from_tensor_slices((X_test, Y_test)).batch(batch_size)
model = ResNet()
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
optimizer = optimizers.Adam()
train_loss = metrics.Mean(name='train_loss')
train_accuracy = metrics.SparseCategoricalAccuracy(name='train_accuracy')
test_loss = metrics.Mean(name='test_loss')
test_accuracy = metrics.SparseCategoricalAccuracy(name='test_accuracy')
for epoch in range(epoches):
for images, labels in train_ds:
train_step(model, images, labels, loss_object, optimizer, train_loss, train_accuracy)
for test_images, test_labels in test_ds:
test_step(model, test_images, test_labels, loss_object, test_loss, test_accuracy)
template = "Epoch {}, Loss: {}, Accuracy: {}, Test Loss: {}, Test Accuracy: {}"
print(template.format(epoch+1,
train_loss.result(),
def __init__(self, **kwargs):
super(MyModel, self).__init__(**kwargs)
self.mean_obj = metrics.Mean(name='my_mean_obj')
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.mean_obj = metrics.Mean(name="my_mean_obj")
