def benchmarkEagerLinearRegression(self):
num_epochs = 10
num_batches = 200
batch_size = 64
dataset = linear_regression.synthetic_dataset(w=tf.random_uniform(
[3, 1]),
b=tf.random_uniform([1]),
noise_level=0.01,
batch_size=batch_size,
num_batches=num_batches)
burn_in_dataset = dataset.take(10)
model = linear_regression.LinearModel()
with tf.device(device()):
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
# Perform burn-in.
linear_regression.fit(model, burn_in_dataset, optimizer)
start_time = time.time()
for _ in range(num_epochs):
linear_regression.fit(model, dataset, optimizer)
wall_time = time.time() - start_time
examples_per_sec = num_epochs * num_batches * batch_size / wall_time
self.report_benchmark(
name="eager_train_%s" %
("gpu" if tfe.num_gpus() > 0 else "cpu"),
iters=num_epochs * num_batches,
extras={"examples_per_sec": examples_per_sec},
wall_time=wall_time)
def benchmarkGraphLinearRegression(self):
num_epochs = 10
num_batches = 200
batch_size = 64
dataset = linear_regression.synthetic_dataset_helper(
w=tf.random_uniform([3, 1]),
b=tf.random_uniform([1]),
num_features=3,
noise_level=0.01,
batch_size=batch_size,
num_batches=num_batches)
iterator = tf.compat.v1.data.make_initializable_iterator(dataset)
x, y = iterator.get_next()
model = linear_regression.LinearModel()
if tf.test.is_gpu_available():
use_gpu = True
device = "/device:GPU:0"
else:
use_gpu = False
device = "/device:CPU:0"
with tf.device(device):
loss = linear_regression.mean_square_loss(model, x, y)
optimization_step = tf.train.GradientDescentOptimizer(
learning_rate=0.1).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
def train(num_epochs):
for _ in range(num_epochs):
sess.run(iterator.initializer)
try:
while True:
_, _ = sess.run([optimization_step, loss])
except tf.errors.OutOfRangeError:
pass
# Warmup: a single epoch.
train(1)
start_time = time.time()
train(num_epochs)
wall_time = time.time() - start_time
examples_per_sec = num_epochs * num_batches * batch_size / wall_time
self.report_benchmark(
name="graph_train_%s" %
("gpu" if use_gpu else "cpu"),
iters=num_epochs * num_batches,
extras={"examples_per_sec": examples_per_sec},
wall_time=wall_time)
def testLinearRegression(self):
true_w = [[1.0], [-0.5], [2.0]]
true_b = [1.0]
model = linear_regression.LinearModel()
dataset = linear_regression.synthetic_dataset(
true_w, true_b, noise_level=0., batch_size=64, num_batches=40)
with tf.device(device()):
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
linear_regression.fit(model, dataset, optimizer, logdir=self._tmp_logdir)
self.assertAllClose(true_w, model.variables[0].numpy(), rtol=1e-2)
self.assertAllClose(true_b, model.variables[1].numpy(), rtol=1e-2)
self.assertTrue(glob.glob(os.path.join(self._tmp_logdir, "events.out.*")))