def benchmark_cnn_cifar10_bs_1024_gpu_2(self):
"""Measure performance with batch_size=1024, gpu=2 and
distribution_strategy=`mirrored`.
"""
batch_size = 1024
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.x_train,
y=self.y_train,
batch_size=batch_size,
num_gpus=2,
distribution_strategy='mirrored',
epochs=self.epochs,
optimizer=tf.keras.optimizers.RMSprop(learning_rate=0.0001,
decay=1e-6),
loss='categorical_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'cnn', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time,
metrics=metrics,
extras=extras)
def benchmark_antirectifier_bs_512(self):
"""Measure performance with batch_size=512."""
batch_size = 512
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.x_train,
y=self.y_train,
batch_size=batch_size,
optimizer="rmsprop",
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=["sparse_categorical_accuracy"])
metadata = benchmark_util.get_keras_examples_metadata(
"antirectifier", batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time, metrics=metrics, extras=extras)
def benchmark_text_classification_bs_512(self):
"""Measure performance with batch_size=512."""
batch_size = 512
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.imdb_x,
y=self.imdb_y,
batch_size=batch_size,
optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'transformer', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time, metrics=metrics, extras=extras)
def benchmark_hrnn_mnist_bs_1024(self):
"""Measure performance with batch_size=1024."""
batch_size = 1024
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.x_train,
y=self.y_train,
batch_size=batch_size,
optimizer='rmsprop',
loss='categorical_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'hierarchical_rnn', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time,
metrics=metrics,
extras=extras)
def benchmark_bidirect_lstm_imdb_bs_512(self):
"""Measure performance with batch_size=512."""
batch_size = 512
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.imdb_x,
y=self.imdb_y,
batch_size=batch_size,
optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'bidirectional_lstm', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time,
metrics=metrics,
extras=extras)
def benchmark_irnn_mnist_bs_512(self):
"""Measure performance with batch_size=512."""
batch_size = 512
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.x_train,
y=self.y_train,
batch_size=batch_size,
optimizer=tf.keras.optimizers.RMSprop(
learning_rate=self.learning_rate),
loss='categorical_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'irnn', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time,
metrics=metrics,
extras=extras)
def benchmark_mlp_reuters_bs_256(self):
"""Measure performance with batch_size=256."""
batch_size = 256
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.x_train,
y=self.y_train,
batch_size=batch_size,
epochs=self.epochs,
optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'mlp', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time,
metrics=metrics,
extras=extras)
def benchmark_custom_training_mnist_bs_512(self):
"""Measure performance with batch_size=512 and run_iters=10."""
batch_size = 512
run_iters = 5
train_dataset = self.train_dataset.shuffle(
buffer_size=1024).batch(batch_size)
# Instantiate a loss function.
loss_fn = tf.keras.losses.CategoricalCrossentropy(
reduction=tf.keras.losses.Reduction.NONE)
# Instantiate an optimizer to train the model.
optimizer = tf.keras.optimizers.Adam()
model = self._build_model()
metrics, wall_time = self.measure_performance(model, train_dataset, loss_fn,
optimizer, batch_size,
run_iters, self.epochs)
extras = benchmark_util.get_keras_examples_metadata('conv', batch_size,
'.keras.ctl_graph')
self.report_benchmark(
iters=run_iters, wall_time=wall_time, metrics=metrics, extras=extras)
def benchmark_custom_training_mnist_bs_512_gpu_2(self):
"""Measure performance with batch_size=512, run_iters=10, gpu=2 and
distribution_strategy='mirrored'.
"""
batch_size = 512
run_iters = 10
train_dataset = self.train_dataset.shuffle(
buffer_size=1024).batch(batch_size)
distribution_strategy = 'mirrored'
strategy = distribution_util.get_distribution_strategy(
distribution_strategy=distribution_strategy, num_gpus=2)
if distribution_strategy != 'off':
train_dataset = strategy.experimental_distribute_dataset(
train_dataset)
strategy_scope = distribution_util.get_strategy_scope(strategy)
with strategy_scope:
# Instantiate a loss function.
loss_fn = tf.keras.losses.CategoricalCrossentropy(
reduction=tf.keras.losses.Reduction.NONE)
# Instantiate an optimizer to train the model.
optimizer = tf.keras.optimizers.Adam()
model = self._build_model()
metrics, wall_time = self.measure_performance(model, train_dataset,
loss_fn, optimizer,
batch_size, run_iters,
self.epochs, strategy)
extras = benchmark_util.get_keras_examples_metadata(
'conv', batch_size, '.keras.ctl_graph')
self.report_benchmark(iters=run_iters,
wall_time=wall_time,
metrics=metrics,
extras=extras)
def benchmark_mlp_reuters_bs_512_gpu_2(self):
"""Measure performance with batch_size=512, gpu=2 and
distribution_strategy='mirrored'
"""
batch_size = 512
metrics, wall_time, extras = benchmark_util.measure_performance(
self._build_model,
x=self.x_train,
y=self.y_train,
batch_size=batch_size,
num_gpus=2,
distribution_strategy='mirrored',
epochs=self.epochs,
optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
metadata = benchmark_util.get_keras_examples_metadata(
'mlp', batch_size)
extras.update(metadata)
self.report_benchmark(wall_time=wall_time,
metrics=metrics,
extras=extras)