def testSlotsUniqueEager(self):
v1 = tf.Variable(1.0)
v2 = tf.Variable(1.0)
opt = rmsprop.RMSprop(1.0, momentum=0.0, centered=False)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and one unique slot variable for v1 and v2.
self.assertLen(set({id(v) for v in opt.variables()}), 3)
self.assertEqual(
self.evaluate(opt.variables()[0]), self.evaluate(opt.iterations)
)
opt = rmsprop.RMSprop(learning_rate=1.0, momentum=0.2, centered=False)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and two unique slot variables for v1 and
# v2.
self.assertLen(set({id(v) for v in opt.variables()}), 5)
self.assertEqual(
self.evaluate(opt.variables()[0]), self.evaluate(opt.iterations)
)
opt = rmsprop.RMSprop(learning_rate=1.0, momentum=0.2, centered=True)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and three unique slot variables for v1 and
# v2
self.assertLen(set({id(v) for v in opt.variables()}), 7)
self.assertEqual(
self.evaluate(opt.variables()[0]), self.evaluate(opt.iterations)
)
def testConstructRMSpropWithLR(self):
opt = rmsprop.RMSprop(lr=1.0)
opt_2 = rmsprop.RMSprop(learning_rate=0.1, lr=1.0)
opt_3 = rmsprop.RMSprop(learning_rate=0.1)
self.assertIsInstance(opt.lr, tf.Variable)
self.assertIsInstance(opt_2.lr, tf.Variable)
self.assertIsInstance(opt_3.lr, tf.Variable)
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertAllClose(self.evaluate(opt.lr), (1.0))
self.assertAllClose(self.evaluate(opt_2.lr), (1.0))
self.assertAllClose(self.evaluate(opt_3.lr), (0.1))
def test_generator_input_to_fit_eval_predict(self):
val_data = np.ones([10, 10], np.float32), np.ones([10, 1], np.float32)
def ones_generator():
while True:
yield np.ones([10, 10], np.float32), np.ones([10, 1], np.float32)
model = test_utils.get_small_mlp(
num_hidden=10, num_classes=1, input_dim=10)
model.compile(
rmsprop.RMSprop(0.001),
'binary_crossentropy',
run_eagerly=test_utils.should_run_eagerly())
model.fit(
ones_generator(),
steps_per_epoch=2,
validation_data=val_data,
epochs=2)
model.evaluate(ones_generator(), steps=2)
model.predict(ones_generator(), steps=2)
# Test with a changing batch size
model = test_utils.get_small_mlp(
num_hidden=3, num_classes=4, input_dim=2)
model.compile(
loss='mse',
optimizer=rmsprop.RMSprop(1e-3),
metrics=['mae', metrics_module.CategoricalAccuracy()])
model.fit_generator(custom_generator_changing_batch_size(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False)
model.fit_generator(custom_generator_changing_batch_size(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False,
validation_data=custom_generator_changing_batch_size(),
validation_steps=10)
model.fit(
custom_generator_changing_batch_size(),
steps_per_epoch=5,
validation_data=custom_generator_changing_batch_size(),
validation_steps=10,
epochs=2)
model.evaluate(custom_generator_changing_batch_size(), steps=5)
model.predict(custom_generator_changing_batch_size(), steps=5)
def test_generator_methods(self):
model = test_utils.get_small_mlp(10, 4, 3)
optimizer = rmsprop.RMSprop(learning_rate=0.001)
model.compile(
optimizer,
loss="mse",
metrics=["mae", metrics_module.CategoricalAccuracy()],
run_eagerly=True,
)
x = np.random.random((10, 3))
y = np.random.random((10, 4))
def numpy_iterator():
while True:
yield x, y
model.fit_generator(numpy_iterator(), steps_per_epoch=3, epochs=1)
model.evaluate_generator(numpy_iterator(), steps=3)
def inference_numpy_iterator():
while True:
yield x
out = model.predict_generator(inference_numpy_iterator(), steps=3)
self.assertEqual(out.shape, (30, 4))
def test_saving_sequential_model(self):
with self.cached_session():
model = keras.models.Sequential()
model.add(keras.layers.Dense(2, input_shape=(3, )))
model.add(keras.layers.RepeatVector(3))
model.add(keras.layers.TimeDistributed(keras.layers.Dense(3)))
model.compile(
loss=keras.losses.MSE,
optimizer=rmsprop.RMSprop(lr=0.0001),
metrics=[keras.metrics.categorical_accuracy],
sample_weight_mode="temporal",
)
x = np.random.random((1, 3))
y = np.random.random((1, 3, 3))
model.train_on_batch(x, y)
ref_y = model.predict(x)
saved_model_dir = self._save_model_dir()
keras_saved_model.export_saved_model(model, saved_model_dir)
loaded_model = keras_saved_model.load_from_saved_model(
saved_model_dir)
y = loaded_model.predict(x)
self.assertAllClose(ref_y, y, atol=1e-05)
def test_saving_functional_model(self):
with self.cached_session():
inputs = keras.layers.Input(shape=(3, ))
x = keras.layers.Dense(2)(inputs)
output = keras.layers.Dense(3)(x)
model = keras.models.Model(inputs, output)
model.compile(
loss=keras.losses.MSE,
optimizer=rmsprop.RMSprop(lr=0.0001),
metrics=[keras.metrics.categorical_accuracy],
)
x = np.random.random((1, 3))
y = np.random.random((1, 3))
model.train_on_batch(x, y)
ref_y = model.predict(x)
saved_model_dir = self._save_model_dir()
keras_saved_model.export_saved_model(model, saved_model_dir)
loaded_model = keras_saved_model.load_from_saved_model(
saved_model_dir)
y = loaded_model.predict(x)
self.assertAllClose(ref_y, y, atol=1e-05)
def testMinimizeSparseResourceVariableCentered(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
with tf.Graph().as_default():
for dtype in _DATA_TYPES:
var0 = tf.Variable([[1.0, 2.0]], dtype=dtype)
x = tf.constant([[4.0], [5.0]], dtype=dtype)
def loss():
pred = tf.matmul(
tf.compat.v1.nn.embedding_lookup([var0], [0]), x
)
return pred * pred
# loss = lambda: pred * pred
# disable=cell-var-from-loop
sgd_op = rmsprop.RMSprop(
learning_rate=1.0,
rho=0.0,
momentum=0.0,
epsilon=1.0,
centered=True,
).minimize(loss, var_list=[var0])
self.evaluate(tf.compat.v1.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllCloseAccordingToType(
[[1.0, 2.0]], self.evaluate(var0)
)
# Run 1 step of sgd
self.evaluate(sgd_op)
# Validate updated params
self.assertAllCloseAccordingToType(
[[-111, -138]], self.evaluate(var0), atol=0.01
)
def test_fit_generator_method(self):
model = test_utils.get_small_mlp(
num_hidden=3, num_classes=4, input_dim=2)
model.compile(
loss='mse',
optimizer=rmsprop.RMSprop(1e-3),
metrics=['mae', metrics_module.CategoricalAccuracy()])
model.fit_generator(custom_generator_threads(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
workers=4,
use_multiprocessing=True)
model.fit_generator(custom_generator(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False)
model.fit_generator(custom_generator(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False,
validation_data=custom_generator(),
validation_steps=10)
model.fit_generator(custom_generator(),
steps_per_epoch=5,
validation_data=custom_generator(),
validation_steps=1,
workers=0)
def test_training_with_sequences(self):
class DummySequence(data_utils.Sequence):
def __getitem__(self, idx):
return np.zeros([10, 2]), np.ones([10, 4])
def __len__(self):
return 10
model = test_utils.get_small_mlp(
num_hidden=3, num_classes=4, input_dim=2)
model.compile(loss='mse', optimizer=rmsprop.RMSprop(1e-3))
model.fit_generator(DummySequence(),
steps_per_epoch=10,
validation_data=custom_generator(),
validation_steps=1,
max_queue_size=10,
workers=0,
use_multiprocessing=True)
model.fit_generator(DummySequence(),
steps_per_epoch=10,
validation_data=custom_generator(),
validation_steps=1,
max_queue_size=10,
workers=0,
use_multiprocessing=False)
def test_generator_methods_with_sample_weights(self):
model = test_utils.get_small_mlp(
num_hidden=3, num_classes=4, input_dim=2)
model.compile(
loss='mse',
optimizer=rmsprop.RMSprop(1e-3),
metrics=['mae', metrics_module.CategoricalAccuracy()],
run_eagerly=test_utils.should_run_eagerly())
model.fit_generator(custom_generator(mode=3),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False)
model.fit_generator(custom_generator(mode=3),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False,
validation_data=custom_generator(mode=3),
validation_steps=10)
model.predict_generator(custom_generator(mode=3),
steps=5,
max_queue_size=10,
use_multiprocessing=False)
model.evaluate_generator(custom_generator(mode=3),
steps=5,
max_queue_size=10,
use_multiprocessing=False)
def test_model_fit_and_validation_with_missing_arg_errors(self):
model = test_utils.get_small_mlp(10, 4, 3)
model.compile(
optimizer=rmsprop.RMSprop(learning_rate=0.001),
loss="mse",
run_eagerly=True,
)
x = tf.zeros(shape=(10, 3))
y = tf.zeros(shape=(10, 4))
dataset = tf.data.Dataset.from_tensor_slices(
(x, y)).repeat(10).batch(5)
validation_dataset = (tf.data.Dataset.from_tensor_slices(
(x, y)).repeat().batch(5)) # Infinite dataset.
model.fit(dataset, epochs=1, verbose=0)
# Step argument is required for infinite datasets.
with self.assertRaises(ValueError):
model.fit(
dataset,
steps_per_epoch=2,
epochs=1,
verbose=0,
validation_data=validation_dataset,
)
with self.assertRaises(ValueError):
model.fit(
dataset,
steps_per_epoch=2,
epochs=1,
verbose=0,
validation_data=validation_dataset,
)
def test_evaluate_generator_method(self):
model = test_utils.get_small_mlp(num_hidden=3,
num_classes=4,
input_dim=2)
model.compile(
loss="mse",
optimizer=rmsprop.RMSprop(1e-3),
metrics=["mae", metrics_module.CategoricalAccuracy()],
run_eagerly=test_utils.should_run_eagerly(),
)
model.evaluate_generator(
custom_generator_threads(),
steps=5,
max_queue_size=10,
workers=2,
verbose=1,
use_multiprocessing=True,
)
model.evaluate_generator(
custom_generator(),
steps=5,
max_queue_size=10,
use_multiprocessing=False,
)
model.evaluate_generator(
custom_generator(),
steps=5,
max_queue_size=10,
use_multiprocessing=False,
workers=0,
)
def test_sequence_input_to_fit_eval_predict(self):
val_data = np.ones([10, 10], np.float32), np.ones([10, 1], np.float32)
class CustomSequence(data_utils.Sequence):
def __getitem__(self, idx):
return np.ones([10, 10], np.float32), np.ones([10, 1],
np.float32)
def __len__(self):
return 2
class CustomSequenceChangingBatchSize(data_utils.Sequence):
def __getitem__(self, idx):
batch_size = 10 - idx
return (
np.ones([batch_size, 10], np.float32),
np.ones([batch_size, 1], np.float32),
)
def __len__(self):
return 2
model = test_utils.get_small_mlp(num_hidden=10,
num_classes=1,
input_dim=10)
model.compile(rmsprop.RMSprop(0.001), "binary_crossentropy")
model.fit(CustomSequence(), validation_data=val_data, epochs=2)
model.evaluate(CustomSequence())
model.predict(CustomSequence())
with self.assertRaisesRegex(ValueError,
"`y` argument is not supported"):
model.fit(CustomSequence(), y=np.ones([10, 1]))
with self.assertRaisesRegex(
ValueError, "`sample_weight` argument is not supported"):
model.fit(CustomSequence(), sample_weight=np.ones([10, 1]))
model.compile(rmsprop.RMSprop(0.001), "binary_crossentropy")
model.fit(
CustomSequenceChangingBatchSize(),
validation_data=val_data,
epochs=2,
)
model.evaluate(CustomSequenceChangingBatchSize())
model.predict(CustomSequenceChangingBatchSize())
def test_control_flow_in_deferred_sequential_model(self):
model = keras.Sequential(
[ControlFlowLayer1(),
keras.layers.Dense(3),
ControlFlowLayer2()])
model.compile(rmsprop.RMSprop(0.001), loss='mse')
model.train_on_batch(np.random.random((2, 3)), np.random.random(
(2, 3)))
def testMomentumProperValue(self):
with self.assertRaisesRegex(
ValueError,
r"`momentum` must be between \[0, 1\]. "
r"Received: momentum=2.5 \(of type <class "
r"\'float\'>\).",
):
rmsprop.RMSprop(1.0, momentum=2.5, centered=False)
def testTrainAndServeWithKPL(self, distribution):
use_adapt = False
test_utils_obj = kpl_test_utils.DistributeKplTestUtils()
with distribution.scope():
(
feature_mapper,
label_mapper,
) = test_utils_obj.define_kpls_for_training(use_adapt)
model = test_utils_obj.define_model()
optimizer = rmsprop.RMSprop(learning_rate=0.1)
accuracy = keras.metrics.Accuracy()
def dataset_fn(_):
return test_utils_obj.dataset_fn(feature_mapper, label_mapper)
@tf.function
def train_step(iterator):
"""The step function for one training step."""
def step_fn(inputs):
"""The computation to run on each replica(GPU)."""
features, labels = inputs
with tf.GradientTape() as tape:
pred = model(features, training=True)
loss = keras.losses.binary_crossentropy(labels, pred)
loss = tf.nn.compute_average_loss(loss)
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(
list(zip(grads, model.trainable_variables))
)
actual_pred = tf.cast(tf.greater(pred, 0.5), tf.int64)
accuracy.update_state(labels, actual_pred)
distribution.run(step_fn, args=(next(iterator),))
distributed_dataset = (
distribution.distribute_datasets_from_function(dataset_fn)
)
distributed_iterator = iter(distributed_dataset)
num_epochs = 4
num_steps = 7
for _ in range(num_epochs):
accuracy.reset_state()
for _ in range(num_steps):
train_step(distributed_iterator)
self.assertGreater(accuracy.result().numpy(), 0.5)
self.assertEqual(
optimizer.iterations.numpy(), num_epochs * num_steps
)
# Test save/load/serving the trained model.
test_utils_obj.test_save_load_serving_model(
model, feature_mapper, test_utils_obj.define_reverse_lookup_layer()
)
def testCallableParams(self):
for dtype in _DATA_TYPES:
var0 = tf.Variable([1.0, 2.0], dtype=dtype)
var1 = tf.Variable([3.0, 4.0], dtype=dtype)
grads0 = tf.constant([0.1, 0.1], dtype=dtype)
grads1 = tf.constant([0.01, 0.01], dtype=dtype)
learning_rate = lambda: 2.0
rho = lambda: 0.9
momentum = lambda: 0.0
epsilon = 1.0
opt = rmsprop.RMSprop(learning_rate, rho, momentum, epsilon)
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
# Step 1: the rms accumulators where 1. So we should see a normal
# update: v -= grad * learning_rate
opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
# Check the parameters.
self.assertAllCloseAccordingToType(
np.array([
1.0 - (0.1 * 2.0 / math.sqrt(0.001 + 1.0)),
2.0 - (0.1 * 2.0 / math.sqrt(0.001 + 1.0)),
]),
self.evaluate(var0),
)
self.assertAllCloseAccordingToType(
np.array([
3.0 - (0.01 * 2.0 / math.sqrt(0.00001 + 1.0)),
4.0 - (0.01 * 2.0 / math.sqrt(0.00001 + 1.0)),
]),
self.evaluate(var1),
)
# Step 2: the root mean square accumulators contain the previous
# update.
opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
# Check the parameters.
self.assertAllCloseAccordingToType(
np.array([
1.0 - (0.1 * 2.0 / math.sqrt(0.001 + 1.0)) -
(0.1 * 2.0 / math.sqrt(0.001 * 0.9 + 0.001 + 1.0)),
2.0 - (0.1 * 2.0 / math.sqrt(0.001 + 1.0)) -
(0.1 * 2.0 / math.sqrt(0.001 * 0.9 + 0.001 + 1.0)),
]),
self.evaluate(var0),
)
self.assertAllCloseAccordingToType(
np.array([
3.0 - (0.01 * 2.0 / math.sqrt(0.00001 + 1.0)) -
(0.01 * 2.0 / math.sqrt(0.00001 * 0.9 + 1e-5 + 1.0)),
4.0 - (0.01 * 2.0 / math.sqrt(0.00001 + 1.0)) -
(0.01 * 2.0 / math.sqrt(0.00001 * 0.9 + 1e-5 + 1.0)),
]),
self.evaluate(var1),
)
def test_generator_methods_invalid_use_case(self):
def invalid_generator():
while 1:
yield (0, 0, 0, 0)
model = test_utils.get_small_mlp(num_hidden=3,
num_classes=4,
input_dim=2)
model.compile(
loss="mse",
optimizer=rmsprop.RMSprop(1e-3),
run_eagerly=test_utils.should_run_eagerly(),
)
with self.assertRaises(ValueError):
model.fit_generator(
invalid_generator(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False,
)
with self.assertRaises(ValueError):
model.fit_generator(
custom_generator(),
steps_per_epoch=5,
epochs=1,
verbose=1,
max_queue_size=10,
use_multiprocessing=False,
validation_data=invalid_generator(),
validation_steps=10,
)
with self.assertRaises(ValueError):
model.predict_generator(
invalid_generator(),
steps=5,
max_queue_size=10,
use_multiprocessing=False,
)
with self.assertRaises(ValueError):
model.evaluate_generator(
invalid_generator(),
steps=5,
max_queue_size=10,
use_multiprocessing=False,
)
def testRunMinimizeOnGPUForCPUVariables(self, use_resource):
with tf.device("/device:CPU:0"):
if use_resource:
var0 = tf.Variable([1.0, 2.0], dtype=tf.float32)
var1 = tf.Variable([3.0, 4.0], dtype=tf.float32)
else:
var0 = tf.Variable([1.0, 2.0], dtype=tf.float32)
var1 = tf.Variable([3.0, 4.0], dtype=tf.float32)
def loss():
return 5 * var0 + 3 * var1
opt = rmsprop.RMSprop(learning_rate=1.0,
decay=0.9,
momentum=0.5,
epsilon=1.0)
# Fetch params to validate initial values
self.evaluate(tf.compat.v1.global_variables_initializer())
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
# Run 1 step through optimizer on GPU.
# Slot variables are created the first time optimizer is used on some
# variable. This tests that slot variables will be colocated with the
# base variable.
with tf.device("/device:GPU:0"):
# Note that for eager execution, minimize expects a function instead
# of a Tensor.
opt_op = opt.minimize(loss, [var0, var1])
self.evaluate(tf.compat.v1.global_variables_initializer())
self.evaluate(opt_op)
# Validate updated params, All variables should have decreased.
self.assertTrue(
all(v < 0.0 for v in self.evaluate(var0)),
msg=f"updated variables: {self.evaluate(var0)}",
)
self.assertTrue(
all(v < 2.0 for v in self.evaluate(var1)),
msg=f"updated variables: {self.evaluate(var1)}",
)
def test_loss_correctness(self, optimizer_kwargs):
# Test that training loss is the same in eager and graph
# (by comparing it to a reference value in a deterministic case)
layers = [
keras.layers.Dense(3, activation='relu',
kernel_initializer='ones'),
keras.layers.Dense(2,
activation='softmax',
kernel_initializer='ones')
]
model = test_utils.get_model_from_layers(layers, input_shape=(4, ))
model.compile(loss='sparse_categorical_crossentropy',
optimizer=rmsprop.RMSprop(learning_rate=0.001,
**optimizer_kwargs),
run_eagerly=test_utils.should_run_eagerly())
x = np.ones((100, 4))
np.random.seed(123)
y = np.random.randint(0, 1, size=(100, 1))
history = model.fit(x, y, epochs=1, batch_size=10)
self.assertAlmostEqual(history.history['loss'][-1], 0.5836, 4)
def test_model_methods_with_eager_tensors_single_io(self):
if not tf.executing_eagerly():
# Only test V2 Function and V2 Eager modes, as V1 Graph mode with
# symbolic tensors has different requirements.
return
model = test_utils.get_small_mlp(10, 4, 3)
optimizer = rmsprop.RMSprop(learning_rate=0.001)
loss = "mse"
metrics = ["mae", metrics_module.CategoricalAccuracy()]
model.compile(
optimizer,
loss,
metrics=metrics,
run_eagerly=test_utils.should_run_eagerly(),
)
inputs = tf.zeros(shape=(10, 3))
targets = tf.zeros(shape=(10, 4))
model.fit(inputs, targets, epochs=1, batch_size=2, verbose=0)
model.fit(inputs,
targets,
epochs=1,
batch_size=3,
verbose=0,
shuffle=False)
model.fit(
inputs,
targets,
epochs=1,
batch_size=4,
verbose=0,
validation_data=(inputs, targets),
)
model.evaluate(inputs, targets, batch_size=2, verbose=0)
model.predict(inputs, batch_size=2)
model.train_on_batch(inputs, targets)
model.test_on_batch(inputs, targets)
def test_loss_correctness_with_iterator(self):
# Test that training loss is the same in eager and graph
# (by comparing it to a reference value in a deterministic case)
layers = [
keras.layers.Dense(3, activation='relu',
kernel_initializer='ones'),
keras.layers.Dense(2,
activation='softmax',
kernel_initializer='ones')
]
model = test_utils.get_model_from_layers(layers, input_shape=(4, ))
model.compile(loss='sparse_categorical_crossentropy',
optimizer=rmsprop.RMSprop(learning_rate=0.001),
run_eagerly=test_utils.should_run_eagerly())
x = np.ones((100, 4), dtype=np.float32)
np.random.seed(123)
y = np.random.randint(0, 1, size=(100, 1))
dataset = tf.data.Dataset.from_tensor_slices((x, y))
dataset = dataset.repeat(100)
dataset = dataset.batch(10)
history = model.fit(dataset, epochs=1, steps_per_epoch=10)
self.assertAlmostEqual(history.history['loss'][-1], 0.5836, 4)
def test_loss_correctness_clipvalue_zero(self):
# Test that training loss is the same in eager and graph
# (by comparing it to a reference value in a deterministic case)
# And confirm that setting clipvalue to zero stops all training
layers = [
keras.layers.Dense(3, activation="relu", kernel_initializer="ones"),
keras.layers.Dense(
2, activation="softmax", kernel_initializer="ones"
),
]
model = test_utils.get_model_from_layers(layers, input_shape=(4,))
model.compile(
loss="sparse_categorical_crossentropy",
optimizer=rmsprop.RMSprop(learning_rate=0.001, clipvalue=0.0),
run_eagerly=test_utils.should_run_eagerly(),
)
x = np.ones((100, 4))
np.random.seed(123)
y = np.random.randint(0, 1, size=(100, 1))
history = model.fit(x, y, epochs=3, batch_size=10)
self.assertAlmostEqual(history.history["loss"][-3], 0.6931, 4)
self.assertAlmostEqual(history.history["loss"][-2], 0.6931, 4)
self.assertAlmostEqual(history.history["loss"][-1], 0.6931, 4)
def testSparse(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for (
dtype,
learning_rate,
rho,
momentum,
epsilon,
centered,
) in _TESTPARAMS:
with tf.compat.v1.get_default_graph().as_default(), test_utils.use_gpu(): # noqa: E501
# Initialize variables for numpy implementation.
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01], dtype=dtype.as_numpy_dtype)
var0 = tf.Variable(var0_np)
var1 = tf.Variable(var1_np)
grads0_np_indices = np.array([0], dtype=np.int32)
grads0 = tf.IndexedSlices(
tf.constant(grads0_np),
tf.constant(grads0_np_indices),
tf.constant([1]),
)
grads1_np_indices = np.array([1], dtype=np.int32)
grads1 = tf.IndexedSlices(
tf.constant(grads1_np),
tf.constant(grads1_np_indices),
tf.constant([1]),
)
opt = rmsprop.RMSprop(
learning_rate=learning_rate,
rho=rho,
momentum=momentum,
epsilon=epsilon,
centered=centered,
)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1])
)
self.evaluate(tf.compat.v1.global_variables_initializer())
if centered:
mg0 = opt.get_slot(var0, "mg")
self.assertEqual(mg0 is not None, centered)
mg1 = opt.get_slot(var1, "mg")
self.assertEqual(mg1 is not None, centered)
else:
mg0 = None
mg1 = None
rms0 = opt.get_slot(var0, "rms")
self.assertIsNotNone(rms0)
rms1 = opt.get_slot(var1, "rms")
self.assertIsNotNone(rms1)
if momentum > 0.0:
mom0 = opt.get_slot(var0, "momentum")
mom1 = opt.get_slot(var1, "momentum")
else:
mom0 = None
mom1 = None
mg0_np = np.array([0.0, 0.0], dtype=dtype.as_numpy_dtype)
mg1_np = np.array([0.0, 0.0], dtype=dtype.as_numpy_dtype)
rms0_np = np.array([0.0, 0.0], dtype=dtype.as_numpy_dtype)
rms1_np = np.array([0.0, 0.0], dtype=dtype.as_numpy_dtype)
mom0_np = np.array([0.0, 0.0], dtype=dtype.as_numpy_dtype)
mom1_np = np.array([0.0, 0.0], dtype=dtype.as_numpy_dtype)
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
# Run 3 steps of RMSprop
for _ in range(1, 4):
self.evaluate(update)
(
var0_np,
mg0_np,
rms0_np,
mom0_np,
) = self._sparse_rmsprop_update_numpy(
var0_np,
grads0_np_indices,
grads0_np,
mg0_np,
rms0_np,
mom0_np,
learning_rate,
rho,
momentum,
epsilon,
centered,
)
(
var1_np,
mg1_np,
rms1_np,
mom1_np,
) = self._sparse_rmsprop_update_numpy(
var1_np,
grads1_np_indices,
grads1_np,
mg1_np,
rms1_np,
mom1_np,
learning_rate,
rho,
momentum,
epsilon,
centered,
)
# Validate updated params
if centered:
self.assertAllCloseAccordingToType(
mg0_np, self.evaluate(mg0)
)
self.assertAllCloseAccordingToType(
mg1_np, self.evaluate(mg1)
)
self.assertAllCloseAccordingToType(
rms0_np, self.evaluate(rms0)
)
self.assertAllCloseAccordingToType(
rms1_np, self.evaluate(rms1)
)
if momentum > 0.0:
self.assertAllCloseAccordingToType(
mom0_np, self.evaluate(mom0)
)
self.assertAllCloseAccordingToType(
mom1_np, self.evaluate(mom1)
)
self.assertAllCloseAccordingToType(
var0_np, self.evaluate(var0)
)
self.assertAllCloseAccordingToType(
var1_np, self.evaluate(var1)
)
def testDenseWithLearningRateInverseTimeDecay(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
with tf.Graph().as_default():
var0_np = np.array([1.0, 2.0])
grads0_np = np.array([0.1, 0.2])
var1_np = np.array([3.0, 4.0])
grads1_np = np.array([0.01, 0.2])
var0 = tf.Variable(var0_np)
var1 = tf.Variable(var1_np)
grads0 = tf.constant(grads0_np)
grads1 = tf.constant(grads1_np)
learning_rate = 0.01
rho = 0.9
momentum = 0.0
epsilon = 1e-7
centered = False
decay = 0.5
lr_schedule = learning_rate_schedule.InverseTimeDecay(
learning_rate, decay_steps=1.0, decay_rate=decay
)
opt = rmsprop.RMSprop(
learning_rate=lr_schedule,
rho=rho,
momentum=momentum,
epsilon=epsilon,
centered=centered,
)
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(tf.compat.v1.global_variables_initializer())
rms0 = opt.get_slot(var0, "rms")
self.assertIsNotNone(rms0)
rms1 = opt.get_slot(var1, "rms")
self.assertIsNotNone(rms1)
if momentum > 0.0:
mom0 = opt.get_slot(var0, "momentum")
mom1 = opt.get_slot(var1, "momentum")
else:
mom0 = None
mom1 = None
mg0_np = np.array([0.0, 0.0])
mg1_np = np.array([0.0, 0.0])
rms0_np = np.array([0.0, 0.0])
rms1_np = np.array([0.0, 0.0])
mom0_np = np.array([0.0, 0.0])
mom1_np = np.array([0.0, 0.0])
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
# Run 4 steps of RMSprop
for t in range(2):
self.evaluate(update)
lr = learning_rate / (1 + decay * t)
var0_np, mg0_np, rms0_np, mom0_np = self._rmsprop_update_numpy(
var0_np,
grads0_np,
mg0_np,
rms0_np,
mom0_np,
lr,
rho,
momentum,
epsilon,
centered,
)
var1_np, mg1_np, rms1_np, mom1_np = self._rmsprop_update_numpy(
var1_np,
grads1_np,
mg1_np,
rms1_np,
mom1_np,
lr,
rho,
momentum,
epsilon,
centered,
)
# Validate updated params
self.assertAllCloseAccordingToType(rms0_np, self.evaluate(rms0))
self.assertAllCloseAccordingToType(rms1_np, self.evaluate(rms1))
if momentum > 0.0:
self.assertAllCloseAccordingToType(
mom0_np, self.evaluate(mom0)
)
self.assertAllCloseAccordingToType(
mom1_np, self.evaluate(mom1)
)
self.assertAllCloseAccordingToType(var0_np, self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np, self.evaluate(var1))
def testRMSprop(self):
self._compare_numerical(rmsprop_new.RMSprop(), rmsprop_old.RMSprop())
def test_control_flow_model(self, model_class):
model = model_class()
model.compile(rmsprop.RMSprop(0.001), loss='mse')
model.train_on_batch(np.random.random((2, 3)), np.random.random(
(2, 3)))
def test_control_flow_layer(self, layer_class):
model = test_utils.get_model_from_layers([layer_class()],
input_shape=(3, ))
model.compile(rmsprop.RMSprop(0.001), loss='mse')
model.train_on_batch(np.random.random((2, 3)), np.random.random(
(2, 3)))
def test_model_methods_with_eager_tensors_multi_io(self):
if not tf.executing_eagerly():
# Only test V2 Function and V2 Eager modes, as V1 Graph mode with
# symbolic tensors has different requirements.
return
input_a = keras.layers.Input(shape=(3, ), name="input_a")
input_b = keras.layers.Input(shape=(3, ), name="input_b")
dense = keras.layers.Dense(4, name="dense")
dropout = keras.layers.Dropout(0.5, name="dropout")
model = test_utils.get_multi_io_model([input_a, dense],
[input_b, dense, dropout])
optimizer = rmsprop.RMSprop(learning_rate=0.001)
loss = "mse"
loss_weights = [1.0, 0.5]
metrics = ["mae", metrics_module.CategoricalAccuracy()]
model.compile(
optimizer,
loss,
metrics=metrics,
loss_weights=loss_weights,
run_eagerly=test_utils.should_run_eagerly(),
sample_weight_mode=None,
)
input_a = tf.zeros(shape=(10, 3))
input_b = tf.zeros(shape=(10, 3))
target_a = tf.zeros(shape=(10, 4))
target_b = tf.zeros(shape=(10, 4))
model.fit(
[input_a, input_b],
[target_a, target_b],
epochs=1,
batch_size=5,
verbose=0,
)
# Test: no shuffle.
model.fit(
[input_a, input_b],
[target_a, target_b],
epochs=1,
batch_size=5,
verbose=0,
shuffle=False,
)
# Test: validation data.
model.fit(
[input_a, input_b],
[target_a, target_b],
epochs=1,
batch_size=2,
verbose=0,
validation_data=([input_a, input_b], [target_a, target_b]),
)
model.train_on_batch([input_a, input_b], [target_a, target_b])
model.predict([input_a, input_b], batch_size=5)
model.evaluate([input_a, input_b], [target_a, target_b],
batch_size=2,
verbose=0)
model.test_on_batch([input_a, input_b], [target_a, target_b])
# Test: mix np and tensors.
input_b = np.zeros(shape=(10, 3)).astype("float32")
target_b = np.zeros(shape=(10, 4)).astype("float32")
model.fit(
[input_a, input_b],
[target_a, target_b],
epochs=1,
batch_size=5,
verbose=0,
)
model.fit(
[input_a, input_b],
[target_a, target_b],
epochs=1,
batch_size=2,
verbose=0,
validation_data=([input_a, input_b], [target_a, target_b]),
)
model.fit(
[input_a, input_b],
[target_a, target_b],
epochs=1,
batch_size=5,
verbose=0,
shuffle=False,
)
model.train_on_batch([input_a, input_b], [target_a, target_b])
model.predict([input_a, input_b], batch_size=5)
model.evaluate([input_a, input_b], [target_a, target_b],
batch_size=2,
verbose=0)
model.test_on_batch([input_a, input_b], [target_a, target_b])
adamax_optimizer_keras_v2_fn = tf.__internal__.test.combinations.NamedObject(
"AdamaxKerasV2", lambda: adamax_keras_v2.Adamax(0.001, epsilon=1.0)
)
nadam_optimizer_keras_v2_fn = tf.__internal__.test.combinations.NamedObject(
"NadamKerasV2", lambda: nadam_keras_v2.Nadam(0.001, epsilon=1.0)
)
ftrl_optimizer_keras_v2_fn = tf.__internal__.test.combinations.NamedObject(
"FtrlKerasV2", lambda: ftrl_keras_v2.Ftrl(0.001)
)
gradient_descent_optimizer_keras_v2_fn = (
tf.__internal__.test.combinations.NamedObject(
"GradientDescentKerasV2", lambda: gradient_descent_keras_v2.SGD(0.001)
)
)
rmsprop_optimizer_keras_v2_fn = tf.__internal__.test.combinations.NamedObject(
"RmsPropKerasV2", lambda: rmsprop_keras_v2.RMSprop(0.001)
)
# TODO(shiningsun): consider adding the other v2 optimizers
optimizers_v2 = [
gradient_descent_optimizer_keras_v2_fn,
adagrad_optimizer_keras_v2_fn,
]
optimizers_v1_and_v2 = optimizers_v1 + optimizers_v2
def distributions_and_v1_optimizers():
"""A common set of combination with DistributionStrategies and
Optimizers."""
return tf.__internal__.test.combinations.combine(
