def FocalLoss_model(input_shape, output_sequence_length, source_vocab_size,
target_vocab_size):
"""
Build and train a RNN model using word embedding on x and y
:param input_shape: Tuple of input shape
:param output_sequence_length: Length of output sequence
:param english_vocab_size: Number of unique English words in the dataset
:param french_vocab_size: Number of unique French words in the dataset
:return: Keras model built, but not trained
"""
input_seq = keras.Input(input_shape[1:])
if output_sequence_length > input_shape[1]:
expanded_seq = keras.backend.squeeze(keras.layers.ZeroPadding1D(
(0, output_sequence_length - input_shape[1]))(keras.layers.Reshape(
(input_shape[1], 1))(input_seq)),
axis=-1)
else:
expanded_seq = input_seq
embedded_seq = keras.layers.TimeDistributed(
keras.layers.BatchNormalization(axis=-1))(keras.layers.Embedding(
source_vocab_size,
embeddings_units,
input_length=output_sequence_length)(expanded_seq))
rnn = keras.layers.TimeDistributed(keras.layers.BatchNormalization(
axis=-1))(keras.layers.GRU(gru_units,
return_sequences=True)(embedded_seq))
probabilities = keras.layers.TimeDistributed(
keras.layers.Dense(target_vocab_size, activation='softmax'))(rnn)
model = keras.Model(input_seq, probabilities)
model.compile(loss=focal_loss(alpha=.25, gamma=2),
optimizer=keras.optimizers.Adam(learning_rate, clipnorm=3.0),
metrics=['accuracy'])
return model
def get_nested_model_3(input_dim, num_classes):
# A functional-API model with a subclassed model inside.
# NOTE: this requires the inner subclass to implement `compute_output_shape`.
inputs = keras.Input(shape=(input_dim, ))
x = keras.layers.Dense(32, activation='relu')(inputs)
x = keras.layers.BatchNormalization()(x)
class Inner(keras.Model):
def __init__(self):
super(Inner, self).__init__()
self.dense1 = keras.layers.Dense(32, activation='relu')
self.dense2 = keras.layers.Dense(5, activation='relu')
self.bn = keras.layers.BatchNormalization()
def call(self, inputs):
x = self.dense1(inputs)
x = self.dense2(x)
return self.bn(x)
test_model = Inner()
x = test_model(x)
outputs = keras.layers.Dense(num_classes)(x)
return keras.Model(inputs, outputs, name='nested_model_3')
def test_getitem_slice_with_stop_and_ellipsis_only(self):
if not context.executing_eagerly():
self.skipTest('Complex slicing like this fails in v1')
inp = keras.Input(shape=(4, 3, 8))
slice_stop = keras.Input(shape=(), dtype='int32')
out = inp[..., :slice_stop[0]]
model = keras.Model(
inputs=[inp, slice_stop],
outputs=out)
model.compile(
adam.Adam(0.001),
'mse',
run_eagerly=testing_utils.should_run_eagerly())
batch_size = 7
stop = 6
x = array_ops.stack([
math_ops.range(8) for _ in range(batch_size)])
args = [x, constant_op.constant(stop, shape=(batch_size,))]
expected = array_ops.stack([
math_ops.range(8)[:stop] for _ in range(batch_size)])
if keras_tensor.keras_tensors_enabled():
self.assertIn('tf.__operators__.getitem', (
x.name for x in model.layers))
self.assertNotIn('tf.strided_slice', (
x.name for x in model.layers))
self.assertAllEqual(model(args), expected)
self.assertAllEqual(model.predict(args, batch_size=batch_size), expected)
# Make sure it can be successfully saved and loaded
config = model.get_config()
model = keras.Model.from_config(config)
self.assertAllEqual(model(args), expected)
self.assertAllEqual(model.predict(args, batch_size=batch_size), expected)
def test_sequential_as_downstream_of_masking_layer(self):
inputs = keras.layers.Input(shape=(3, 4))
x = keras.layers.Masking(mask_value=0., input_shape=(3, 4))(inputs)
s = keras.Sequential()
s.add(keras.layers.Dense(5, input_shape=(4, )))
x = keras.layers.wrappers.TimeDistributed(s)(x)
model = keras.Model(inputs=inputs, outputs=x)
model.compile(optimizer=rmsprop.RMSPropOptimizer(1e-3), loss='mse')
model_input = np.random.randint(low=1, high=5,
size=(10, 3, 4)).astype('float32')
for i in range(4):
model_input[i, i:, :] = 0.
model.fit(model_input,
np.random.random((10, 3, 5)),
epochs=1,
batch_size=6)
if not context.executing_eagerly():
# Note: this doesn't work in eager due to DeferredTensor/ops compatibility
# issue.
mask_outputs = [
model.layers[1].compute_mask(model.layers[1].input)
]
mask_outputs += [
model.layers[2].compute_mask(model.layers[2].input,
mask_outputs[-1])
]
func = keras.backend.function([model.input], mask_outputs)
mask_outputs_val = func([model_input])
self.assertAllClose(mask_outputs_val[0],
np.any(model_input, axis=-1))
self.assertAllClose(mask_outputs_val[1],
np.any(model_input, axis=-1))
def test_sparse_int_input_multi_bucket(self):
vocab_data = np.array([10, 11, 12, 13], dtype=np.int64)
input_array = sparse_tensor.SparseTensor(
indices=[[0, 0], [1, 2]],
values=np.array([13, 133], dtype=np.int64),
dense_shape=[3, 4])
expected_indices = [[0, 0], [1, 2]]
expected_values = [6, 2]
expected_dense_shape = [3, 4]
input_data = keras.Input(shape=(None, ),
dtype=dtypes.int64,
sparse=True)
layer = get_layer_class()(max_tokens=None,
dtype=dtypes.int64,
num_oov_tokens=2)
layer.set_vocabulary(vocab_data)
int_data = layer(input_data)
model = keras.Model(inputs=input_data, outputs=int_data)
output_data = model.predict(input_array, steps=1)
self.assertAllEqual(expected_indices, output_data.indices)
self.assertAllEqual(expected_values, output_data.values)
self.assertAllEqual(expected_dense_shape, output_data.dense_shape)
def test_batchnorm_non_trainable_with_fit(self):
inputs = keras.Input((3, ))
bn = normalization_v2.BatchNormalization()
outputs = bn(inputs)
model = keras.Model(inputs, outputs)
model.compile('rmsprop',
'mse',
run_eagerly=testing_utils.should_run_eagerly(),
experimental_run_tf_function=testing_utils.
should_run_tf_function())
model.fit(np.random.random((100, 3)), np.random.random((100, 3)))
test_data = np.random.random((10, 3))
test_targets = np.random.random((10, 3))
test_loss = model.evaluate(test_data, test_targets)
bn.trainable = False
model.compile('rmsprop',
'mse',
run_eagerly=testing_utils.should_run_eagerly(),
experimental_run_tf_function=testing_utils.
should_run_tf_function())
train_loss = model.train_on_batch(test_data, test_targets)
self.assertAlmostEqual(test_loss, train_loss)
def get_mnist_model(input_shape):
"""Define a deterministically-initialized CNN model for MNIST testing."""
inputs = keras.Input(shape=input_shape)
x = keras.layers.Conv2D(
32,
kernel_size=(3, 3),
activation="relu",
kernel_initializer=keras.initializers.TruncatedNormal(seed=99))(inputs)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Flatten()(x) + keras.layers.Flatten()(x)
x = keras.layers.Dense(
10,
activation="softmax",
kernel_initializer=keras.initializers.TruncatedNormal(seed=99))(x)
model = keras.Model(inputs=inputs, outputs=x)
# TODO(yuefengz): optimizer with slot variables doesn't work because of
# optimizer's bug.
# TODO(yuefengz): we should not allow non-v2 optimizer.
model.compile(
loss=keras.losses.sparse_categorical_crossentropy,
optimizer=gradient_descent.SGD(learning_rate=0.001),
metrics=["accuracy"])
return model
def get_model_from_layers_with_input(layers,
input_shape=None,
input_dtype=None,
model_input=None):
"""Builds a model from a sequence of layers."""
if model_input is not None and input_shape is not None:
raise ValueError("Cannot specify a model_input and an input shape.")
model_type = testing_utils.get_model_type()
if model_type == "subclass":
return _SubclassModel(layers, model_input)
if model_type == "sequential":
model = keras.models.Sequential()
if model_input is not None:
model.add(model_input)
elif input_shape is not None:
model.add(keras.Input(shape=input_shape, dtype=input_dtype))
for layer in layers:
model.add(layer)
return model
if model_type == "functional":
if model_input is not None:
inputs = model_input
else:
if not input_shape:
raise ValueError("Cannot create a functional model from layers with no "
"input shape.")
inputs = keras.Input(shape=input_shape, dtype=input_dtype)
outputs = inputs
for layer in layers:
outputs = layer(outputs)
return keras.Model(inputs, outputs)
raise ValueError("Unknown model type {}".format(model_type))
def test_saved_weights_keras(self):
input_data = [[1], [2], [3]]
expected_output = [[0], [1], [2]]
cls = discretization.Discretization
inputs = keras.Input(shape=(1,), dtype=dtypes.int32)
layer = cls(num_bins=3)
layer.adapt(input_data)
outputs = layer(inputs)
model = keras.Model(inputs=inputs, outputs=outputs)
output_data = model.predict(input_data)
self.assertAllClose(output_data, expected_output)
# Save the model to disk.
output_path = os.path.join(self.get_temp_dir(), "tf_keras_saved_weights")
model.save_weights(output_path, save_format="tf")
new_model = keras.Model.from_config(
model.get_config(), custom_objects={"Discretization": cls})
new_model.load_weights(output_path)
# Validate correctness of the new model.
new_output_data = new_model.predict(input_data)
self.assertAllClose(new_output_data, expected_output)
def test_model_saving_to_pre_created_h5py_file(self):
if h5py is None:
self.skipTest('h5py required to run this test')
with self.test_session():
inputs = keras.Input(shape=(3,))
x = keras.layers.Dense(2)(inputs)
outputs = keras.layers.Dense(3)(x)
model = keras.Model(inputs, outputs)
model.compile(loss=keras.losses.MSE,
optimizer=keras.optimizers.Adam(),
metrics=[keras.metrics.categorical_accuracy])
x = np.random.random((1, 3))
y = np.random.random((1, 3))
model.train_on_batch(x, y)
out = model.predict(x)
fd, fname = tempfile.mkstemp('.h5')
with h5py.File(fname, mode='r+') as h5file:
keras.models.save_model(model, h5file)
loaded_model = keras.models.load_model(h5file)
out2 = loaded_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
# Test non-default options in h5
with h5py.File('_', driver='core',
backing_store=False) as h5file:
keras.models.save_model(model, h5file)
loaded_model = keras.models.load_model(h5file)
out2 = loaded_model.predict(x)
self.assertAllClose(out, out2, atol=1e-05)
# Cleanup
os.close(fd)
os.remove(fname)
def test_train_eval_with_steps(self):
# See b/142880049 for more details.
inp = keras.Input(shape=(4,), name='inp1')
out = keras.layers.Dense(2)(inp)
model = keras.Model(inp, out)
model.compile(
'rmsprop', loss='mse', run_eagerly=testing_utils.should_run_eagerly())
inputs = np.zeros((100, 4), dtype=np.float32)
targets = np.random.randint(0, 2, size=100, dtype=np.int32)
training_ds = dataset_ops.Dataset.from_tensor_slices(
(inputs, targets)).repeat().batch(10)
# Create eval dataset with generator, so that dataset won't contain the
# overall size metadata. Without eval_steps, we expect to run through all
# the data in this dataset every epoch.
def gen():
for _ in range(100):
yield (np.zeros(4, dtype=np.float32),
np.random.randint(0, 2, size=1, dtype=np.int32))
eval_ds = dataset_ops.Dataset.from_generator(
generator=gen,
output_types=('float64', 'int32'),
output_shapes=([4], [1])).batch(100)
batch_counter = BatchCounterCallback()
model.fit(
training_ds,
steps_per_epoch=10,
epochs=10,
validation_data=eval_ds,
callbacks=[batch_counter])
# Expect 10 batch from training per epoch.
self.assertEqual(batch_counter.batch_end_count, 100)
def create_model(self,
model_name,
input_size,
embedding_size,
num_classes,
include_top=False):
base_network = model_sets[model_name](input_shape=(input_size,
input_size, 3),
include_top=include_top,
weights='imagenet')
base_network.trainable = True
inputs = base_network.input
x = base_network(inputs)
x = layers.GlobalAveragePooling2D()(x)
embedding = layers.Dense(embedding_size, name='embedding')(x)
logits = layers.Dense(num_classes, name='logits')(embedding)
model = keras.Model(inputs=inputs,
outputs={
'embedding': embedding,
'logits': logits
})
model.summary()
return model
def get_model(self,
max_words=10,
initial_weights=None,
distribution=None,
experimental_run_tf_function=None,
input_shapes=None):
del input_shapes
if tf2.enabled():
if not context.executing_eagerly():
self.skipTest("LSTM v2 and legacy graph mode don't work together.")
lstm = rnn_v2.LSTM
else:
lstm = rnn_v1.LSTM
with keras_correctness_test_base.MaybeDistributionScope(distribution):
word_ids = keras.layers.Input(
shape=(max_words,), dtype=np.int32, name='words')
word_embed = keras.layers.Embedding(input_dim=20, output_dim=10)(word_ids)
lstm_embed = lstm(units=4, return_sequences=False)(
word_embed)
preds = keras.layers.Dense(2, activation='softmax')(lstm_embed)
model = keras.Model(inputs=[word_ids], outputs=[preds])
if initial_weights:
model.set_weights(initial_weights)
optimizer_fn = gradient_descent_keras.SGD
model.compile(
optimizer=optimizer_fn(learning_rate=0.1),
loss='sparse_categorical_crossentropy',
metrics=['sparse_categorical_accuracy'],
experimental_run_tf_function=experimental_run_tf_function)
return model
def test_one_hot_output_rank_one_input(self):
input_data = np.array([3, 2, 0, 1])
expected_output = [
[0, 0, 0, 1],
[0, 0, 1, 0],
[1, 0, 0, 0],
[0, 1, 0, 0],
]
num_tokens = 4
expected_output_shape = [None, num_tokens]
# Test call on layer directly.
layer = category_encoding.CategoryEncoding(
num_tokens=num_tokens, output_mode=category_encoding.ONE_HOT)
output_data = layer(input_data)
self.assertAllEqual(expected_output, output_data)
# Test call on model.
inputs = keras.Input(shape=(1,), dtype=dtypes.int32)
outputs = layer(inputs)
model = keras.Model(inputs=inputs, outputs=outputs)
output_data = model(input_data)
self.assertAllEqual(expected_output_shape, outputs.shape.as_list())
self.assertAllEqual(expected_output, output_data)
def test_distribution_strategy_output(self, distribution):
vocab_data = ["earth", "wind", "and", "fire"]
input_array = np.array([["earth", "wind", "and", "fire"],
["fire", "and", "earth", "michigan"]])
input_dataset = dataset_ops.Dataset.from_tensor_slices(input_array).batch(
2, drop_remainder=True)
expected_output = [[2, 3, 4, 5], [5, 4, 2, 1]]
config.set_soft_device_placement(True)
with distribution.scope():
input_data = keras.Input(shape=(None,), dtype=dtypes.string)
layer = get_layer_class()(
max_tokens=None,
standardize=None,
split=None,
output_mode=text_vectorization.INT)
layer.set_vocabulary(vocab_data)
int_data = layer(input_data)
model = keras.Model(inputs=input_data, outputs=int_data)
output_dataset = model.predict(input_dataset)
self.assertAllEqual(expected_output, output_dataset)
def test_multiple_ngram_values(self):
input_array = np.array([["earth", "wind", "and", "fire"],
["fire", "and", "earth", "michigan"]])
# pyformat: disable
expected_output = [[
b"earth wind", b"wind and", b"and fire", b"earth wind and",
b"wind and fire"
],
[
b"fire and", b"and earth", b"earth michigan",
b"fire and earth", b"and earth michigan"
]]
# pyformat: enable
input_data = keras.Input(shape=(4, ), dtype=dtypes.string)
layer = get_layer_class()(max_tokens=None,
standardize=None,
split=None,
ngrams=(2, 3),
output_mode=None)
int_data = layer(input_data)
model = keras.Model(inputs=input_data, outputs=int_data)
output_dataset = model.predict(input_array)
self.assertAllEqual(expected_output, output_dataset)
def load_keras_model(model_dir, max_seq_len):
from tensorflow.python import keras
from bert import BertModelLayer
from bert.loader import StockBertConfig, load_stock_weights
bert_config_file = os.path.join(model_dir, "bert_config.json")
bert_ckpt_file = os.path.join(model_dir, "bert_model.ckpt")
l_bert = None
with tf.io.gfile.GFile(bert_config_file, "r") as reader:
bc = StockBertConfig.from_json_string(reader.read())
l_bert = BertModelLayer.from_params(bc.to_bert_model_layer_params(), name="bert")
l_input_ids = keras.layers.Input(shape=(max_seq_len,), dtype='int32', name="input_ids")
l_token_type_ids = keras.layers.Input(shape=(max_seq_len,), dtype='int32', name="token_type_ids")
output = l_bert([l_input_ids, l_token_type_ids])
model = keras.Model(inputs=[l_input_ids, l_token_type_ids], outputs=output)
model.build(input_shape=[(None, max_seq_len),
(None, max_seq_len)])
load_stock_weights(l_bert, bert_ckpt_file)
return model
def test_if_training_pattern_update(self):
class MyLayer(keras.layers.Layer):
def build(self, input_shape):
self.counter = self.add_weight(shape=(),
trainable=False,
initializer='zeros')
def call(self, inputs, training=None):
if training:
increment = 1.
else:
increment = 0.
self.counter.assign_add(increment)
return inputs
inputs = keras.Input((3, ))
layer = MyLayer()
outputs = layer(inputs)
model = keras.Model(inputs, outputs)
model.compile('sgd',
'mse',
run_eagerly=testing_utils.should_run_eagerly())
model.train_on_batch(np.ones((2, 3)), np.ones((2, 3)))
self.assertEqual(keras.backend.get_value(layer.counter), 1.)
def get_model(self, initial_weights=None, distribution=None):
with keras_correctness_test_base.MaybeDistributionScope(distribution):
image = keras.layers.Input(shape=(28, 28, 3), name='image')
c1 = keras.layers.Conv2D(
name='conv1',
filters=16,
kernel_size=(3, 3),
strides=(4, 4),
kernel_regularizer=keras.regularizers.l2(1e-4))(image)
if self.with_batch_norm:
c1 = keras.layers.BatchNormalization(name='bn1')(c1)
c1 = keras.layers.MaxPooling2D(pool_size=(2, 2))(c1)
logits = keras.layers.Dense(10, activation='softmax', name='pred')(
keras.layers.Flatten()(c1))
model = keras.Model(inputs=[image], outputs=[logits])
if initial_weights:
model.set_weights(initial_weights)
model.compile(optimizer=gradient_descent.SGD(learning_rate=0.1),
loss='sparse_categorical_crossentropy',
metrics=['sparse_categorical_accuracy'])
return model
def test_if_training_pattern_metric(self):
class MyLayer(keras.layers.Layer):
def call(self, inputs, training=None):
if training:
metric = math_ops.reduce_sum(inputs)
else:
metric = 0.
self.add_metric(metric, name='my_metric', aggregation='mean')
return inputs
inputs = keras.Input((3, ))
outputs = MyLayer()(inputs)
model = keras.Model(inputs, outputs)
model.compile('sgd',
'mse',
run_eagerly=testing_utils.should_run_eagerly())
for _ in range(3):
_, train_metric = model.train_on_batch(np.ones((2, 3)),
np.ones((2, 3)))
self.assertEqual(train_metric, 2 * 3)
_, test_metric = model.test_on_batch(np.ones((2, 3)),
np.ones((2, 3)))
self.assertEqual(test_metric, 0)
def test_learning_phase_value(self):
# TODO(anjalisridhar): Modify this test to use Lambdas since we can compare
# meaningful values. Currently we don't pass the learning phase if the
# Lambda layer uses the learning phase.
with self.cached_session():
x = keras.layers.Input(shape=(1,), name='input')
y = keras.layers.Dense(1, kernel_initializer='ones')(x)
z = keras.layers.Dropout(0.9999)(y)
model = keras.Model(x, z)
initial_weights = model.get_weights()
optimizer = gradient_descent.GradientDescentOptimizer(0.005)
loss = 'mse'
metrics = ['acc']
strategy = mirrored_strategy.MirroredStrategy(
['/device:GPU:0', '/device:GPU:1'])
model.compile(optimizer, loss, metrics=metrics, distribute=strategy)
inputs = np.ones((10, 1), dtype=np.float32)
targets = np.ones((10, 1), dtype=np.float32)
dataset = dataset_ops.Dataset.from_tensor_slices((inputs, targets))
dataset = dataset.repeat().batch(8)
hist = model.fit(dataset, epochs=1, steps_per_epoch=20, verbose=1)
self.assertAlmostEqual(hist.history['acc'][0], 0, 0)
model.set_weights(initial_weights)
evaluate_output = model.evaluate(dataset, steps=20)
self.assertAlmostEqual(evaluate_output[1], 1, 0)
inputs = np.ones((10, 1), dtype=np.float32)
predict_dataset = dataset_ops.Dataset.from_tensor_slices(inputs)
predict_dataset = predict_dataset.repeat().batch(5)
output = model.predict(predict_dataset, steps=10)
ref_output = np.ones((50, 1), dtype=np.float32)
self.assertArrayNear(output[0], ref_output, 1e-1)
def test_sparse_adapt(self):
vocab_data = sparse_ops.from_dense(
np.array([[1, 1, 0, 1, 1, 2, 2, 0, 2, 3, 3, 0, 4]], dtype=np.int64))
vocab_dataset = dataset_ops.Dataset.from_tensors(vocab_data)
input_array = sparse_ops.from_dense(
np.array([[1, 2, 3, 0], [0, 3, 1, 0]], dtype=np.int64))
# pyformat: disable
expected_output = [[0, 1, 1, 1, 0],
[0, 1, 0, 1, 0]]
# pyformat: enable
max_tokens = 5
expected_output_shape = [None, max_tokens]
input_data = keras.Input(shape=(None,), dtype=dtypes.int64, sparse=True)
layer = get_layer_class()(
max_tokens=None, output_mode=category_encoding.BINARY)
layer.adapt(vocab_dataset)
int_data = layer(input_data)
self.assertAllEqual(expected_output_shape, int_data.shape.as_list())
model = keras.Model(inputs=input_data, outputs=int_data)
output_dataset = model.predict(input_array, steps=1)
self.assertAllEqual(expected_output, output_dataset)
def testSingleInputCase(self):
class LayerWithOneInput(keras.layers.Layer):
def build(self, input_shape):
self.w = array_ops.ones(shape=(3, 4))
def call(self, inputs):
return keras.backend.dot(inputs, self.w)
inputs = input_layer_lib.Input(shape=(3, ))
layer = LayerWithOneInput()
if context.executing_eagerly():
self.assertEqual(
layer.compute_output_shape((None, 3)).as_list(), [None, 4])
# As a side-effect, compute_output_shape builds the layer.
self.assertTrue(layer.built)
# We can still query the layer's compute_output_shape with compatible
# input shapes.
self.assertEqual(
layer.compute_output_shape((6, 3)).as_list(), [6, 4])
outputs = layer(inputs)
model = keras.Model(inputs, outputs)
self._testShapeInference(model, (2, 3), (2, 4))
def test_tfidf_output_soft_maximum(self):
tfidf_data = [.05, .5, .25, .2, .125]
input_array = np.array([[1, 2, 3, 1], [0, 4, 1, 0]])
# pyformat: disable
# pylint: disable=bad-whitespace
expected_output = [[ 0, 1, .25, .2, 0],
[.1, .5, 0, 0, .125]]
# pylint: enable=bad-whitespace
# pyformat: enable
max_tokens = 5
expected_output_shape = [None, max_tokens]
input_data = keras.Input(shape=(None,), dtype=dtypes.int32)
layer = get_layer_class()(
max_tokens=None, output_mode=category_encoding.TFIDF)
layer.set_num_elements(max_tokens)
layer.set_tfidf_data(tfidf_data)
int_data = layer(input_data)
self.assertAllEqual(expected_output_shape, int_data.shape.as_list())
model = keras.Model(inputs=input_data, outputs=int_data)
output_dataset = model.predict(input_array)
self.assertAllClose(expected_output, output_dataset)
def test_embedding_with_ragged_input(self):
layer = keras.layers.Embedding(
input_dim=3,
output_dim=2,
weights=[np.array([[0., 0.], [1., 1.], [2., 2.]])])
inputs = keras.layers.Input(shape=(None, ),
dtype=dtypes.float32,
ragged=True)
# pylint: disable=unnecessary-lambda
outputs = keras.layers.Lambda(
lambda args: keras.backend.identity(args))(inputs)
# pylint: enable=unnecessary-lambda
outputs = layer(outputs)
model = keras.Model(inputs, outputs)
model.run_eagerly = testing_utils.should_run_eagerly()
outputs = model.predict(
ragged_factory_ops.constant([[1., 2., 2.], [0.], [1., 2.]],
ragged_rank=1))
self.assertAllClose(
outputs,
ragged_factory_ops.constant([[[1., 1.], [2., 2.], [2., 2.]],
[[0., 0.]], [[1., 1.], [2., 2.]]],
ragged_rank=1))
def test_calling_model_on_same_dataset(self):
with self.test_session():
x = keras.layers.Input(shape=(3, ), name='input')
y = keras.layers.Dense(4, name='dense')(x)
model = keras.Model(x, y)
optimizer = gradient_descent.GradientDescentOptimizer(0.001)
loss = 'mse'
metrics = ['mae']
strategy = mirrored_strategy.MirroredStrategy(
['/device:GPU:1', '/device:GPU:0'])
model.compile(optimizer,
loss,
metrics=metrics,
distribute=strategy)
inputs = np.zeros((10, 3), dtype=np.float32)
targets = np.zeros((10, 4), dtype=np.float32)
dataset = dataset_ops.Dataset.from_tensor_slices((inputs, targets))
dataset = dataset.repeat(100)
dataset = dataset.batch(10)
# Call fit with validation data
model.fit(dataset,
epochs=1,
steps_per_epoch=2,
verbose=0,
validation_data=dataset,
validation_steps=2)
model.fit(dataset,
epochs=1,
steps_per_epoch=2,
verbose=0,
validation_data=dataset,
validation_steps=2)
model.predict(dataset, steps=2)
def test_distribution(self, distribution):
# TODO(b/159738418): large image input causes OOM in ubuntu multi gpu.
np_images = np.random.random((32, 32, 32, 3)).astype(np.float32)
image_dataset = dataset_ops.Dataset.from_tensor_slices(
np_images).batch(16, drop_remainder=True)
with distribution.scope():
input_data = keras.Input(shape=(32, 32, 3), dtype=dtypes.float32)
image_preprocessor = keras.Sequential([
image_preprocessing.Resizing(height=256, width=256),
image_preprocessing.RandomCrop(height=224, width=224),
image_preprocessing.RandomTranslation(.1, .1),
image_preprocessing.RandomRotation(.2),
image_preprocessing.RandomFlip(),
image_preprocessing.RandomZoom(.2, .2)
])
preprocessed_image = image_preprocessor(input_data)
flatten_layer = keras.layers.Flatten(data_format="channels_last")
output = flatten_layer(preprocessed_image)
cls_layer = keras.layers.Dense(units=1, activation="sigmoid")
output = cls_layer(output)
model = keras.Model(inputs=input_data, outputs=output)
model.compile(loss="binary_crossentropy")
_ = model.predict(image_dataset)
def test_updates_and_losses_for_nested_models_in_subclassed_model(self):
# Case 1: deferred-build sequential nested in subclass.
class TestModel1(keras.Model):
def __init__(self):
super(TestModel1, self).__init__()
self.fc = keras.layers.Dense(10, input_shape=(784,),
activity_regularizer='l1')
self.bn = keras.Sequential([keras.layers.BatchNormalization(axis=1)])
def call(self, x):
return self.bn(self.fc(x))
with self.cached_session():
model = TestModel1()
x = array_ops.ones(shape=[100, 784], dtype='float32')
model(x)
self.assertEqual(len(model.get_updates_for(x)), 2)
self.assertEqual(len(model.get_losses_for(x)), 1)
# Case 2: placeholder-sequential nested in subclass.
class TestModel2(keras.Model):
def __init__(self):
super(TestModel2, self).__init__()
self.fc = keras.layers.Dense(10, input_shape=(784,),
activity_regularizer='l1')
self.bn = keras.Sequential(
[keras.layers.BatchNormalization(axis=1, input_shape=(10,))])
def call(self, x):
return self.bn(self.fc(x))
with self.cached_session():
model = TestModel2()
x = array_ops.ones(shape=[100, 784], dtype='float32')
model(x)
self.assertEqual(len(model.get_updates_for(x)), 2)
self.assertEqual(len(model.get_losses_for(x)), 1)
# Case 3: functional-API model nested in subclass.
inputs = keras.Input((10,))
outputs = keras.layers.BatchNormalization(axis=1)(inputs)
bn = keras.Model(inputs, outputs)
class TestModel3(keras.Model):
def __init__(self):
super(TestModel3, self).__init__()
self.fc = keras.layers.Dense(10, input_shape=(784,),
activity_regularizer='l1')
self.bn = bn
def call(self, x):
return self.bn(self.fc(x))
with self.cached_session():
model = TestModel3()
x = array_ops.ones(shape=[100, 784], dtype='float32')
model(x)
self.assertEqual(len(model.get_updates_for(x)), 2)
self.assertEqual(len(model.get_losses_for(x)), 1)
def test_functional_model(self):
inputs = keras.Input(10)
outputs = keras.layers.Dense(1)(inputs)
model = keras.Model(inputs, outputs)
self._check_model_class(model.__class__.__bases__[0])
self._check_layer_class(model)
def get_model():
x = keras.layers.Input(shape=(3, ), name='input')
y = keras.layers.Dense(4, name='dense')(x)
model = keras.Model(x, y)
return model
