def inference_gru_block_vs_gru_cell(batch_size,
cell_size,
input_size,
time_steps,
use_gpu=False,
iters=30):
"""Benchmark inference speed between GRUBlockCell vs GRUCell."""
ops.reset_default_graph()
with session.Session(graph=ops.Graph()) as sess:
with benchmarking.device(use_gpu):
# Random initializers.
seed = 1994
initializer = init_ops.random_uniform_initializer(-1, 1, seed=seed)
np.random.seed(seed)
# Inputs
concat_x = vs.get_variable("concat_x",
[time_steps, batch_size, input_size])
h = vs.get_variable("h", [batch_size, cell_size])
# Output from the basic GRU cell implementation.
with vs.variable_scope("basic", initializer=initializer):
cell = rnn_cell.GRUCell(cell_size)
outputs_dynamic, _ = rnn.dynamic_rnn(cell,
inputs=concat_x,
initial_state=h,
time_major=True,
dtype=dtypes.float32)
sess.run([variables.global_variables_initializer()])
basic_time_inference = benchmarking.seconds_per_run(
outputs_dynamic, sess, iters)
# Output from the block GRU cell implementation.
with vs.variable_scope("block", initializer=initializer):
cell = gru_ops.GRUBlockCell(cell_size)
outputs_dynamic, _ = rnn.dynamic_rnn(cell,
inputs=concat_x,
initial_state=h,
time_major=True,
dtype=dtypes.float32)
sess.run([variables.global_variables_initializer()])
block_time_inference = benchmarking.seconds_per_run(
outputs_dynamic, sess, iters)
performance_inference = (basic_time_inference - block_time_inference
) * 100 / basic_time_inference
print(",".join([
str(batch_size),
str(cell_size),
str(input_size),
str(time_steps),
str(use_gpu),
str(basic_time_inference),
str(block_time_inference),
str(performance_inference)
]))
return basic_time_inference, block_time_inference
def inference_gru_block_vs_gru_cell(batch_size,
cell_size,
input_size,
time_steps,
use_gpu=False,
iters=30):
"""Benchmark inference speed between GRUBlockCell vs GRUCell."""
ops.reset_default_graph()
with session.Session(graph=ops.Graph()) as sess:
with benchmarking.device(use_gpu):
# Random initializers.
seed = 1994
initializer = init_ops.random_uniform_initializer(-1, 1, seed=seed)
np.random.seed(seed)
# Inputs
concat_x = vs.get_variable("concat_x",
[time_steps, batch_size, input_size])
h = vs.get_variable("h", [batch_size, cell_size])
# Output from the basic GRU cell implementation.
with vs.variable_scope("basic", initializer=initializer):
cell = rnn_cell.GRUCell(cell_size)
outputs_dynamic, _ = rnn.dynamic_rnn(
cell,
inputs=concat_x,
initial_state=h,
time_major=True,
dtype=dtypes.float32)
sess.run([variables.global_variables_initializer()])
basic_time_inference = benchmarking.seconds_per_run(
outputs_dynamic, sess, iters)
# Output from the block GRU cell implementation.
with vs.variable_scope("block", initializer=initializer):
cell = gru_ops.GRUBlockCell(cell_size)
outputs_dynamic, _ = rnn.dynamic_rnn(
cell,
inputs=concat_x,
initial_state=h,
time_major=True,
dtype=dtypes.float32)
sess.run([variables.global_variables_initializer()])
block_time_inference = benchmarking.seconds_per_run(
outputs_dynamic, sess, iters)
performance_inference = (basic_time_inference - block_time_inference
) * 100 / basic_time_inference
print(",".join([
str(batch_size), str(cell_size), str(input_size), str(time_steps), str(
use_gpu), str(basic_time_inference), str(block_time_inference), str(
performance_inference)
]))
return basic_time_inference, block_time_inference
def benchmarkLSTMBlockCellBpropWithDynamicRNN(self):
print("BlockLSTMCell backward propagation via dynamic_rnn().")
print("--------------------------------------------------------------")
print("LSTMBlockCell Seconds per inference.")
print("batch_size,cell_size,input_size,time_steps,use_gpu,wall_time")
iters = 10
for config in benchmarking.dict_product({
"batch_size": [1, 8, 13, 32, 67, 128],
"cell_size": [128, 250, 512, 650, 1024, 1350],
"time_steps": [40],
"use_gpu": [True, False]
}):
with ops.Graph().as_default():
with benchmarking.device(use_gpu=config["use_gpu"]):
time_steps = config["time_steps"]
batch_size = config["batch_size"]
cell_size = input_size = config["cell_size"]
inputs = variable_scope.get_variable(
"x", [time_steps, batch_size, cell_size],
trainable=False,
dtype=dtypes.float32)
with variable_scope.variable_scope(
"rnn", reuse=variable_scope.AUTO_REUSE):
w = variable_scope.get_variable(
"rnn/lstm_cell/kernel",
shape=[input_size + cell_size, cell_size * 4],
dtype=dtypes.float32)
b = variable_scope.get_variable(
"rnn/lstm_cell/bias",
shape=[cell_size * 4],
dtype=dtypes.float32,
initializer=init_ops.zeros_initializer())
cell = lstm_ops.LSTMBlockCell(cell_size)
outputs = rnn.dynamic_rnn(
cell, inputs, time_major=True, dtype=dtypes.float32)
grads = gradients_impl.gradients(outputs, [inputs, w, b])
init_op = variables.global_variables_initializer()
with session.Session() as sess:
sess.run(init_op)
wall_time = benchmarking.seconds_per_run(grads, sess, iters)
# Print to stdout. If the TEST_REPORT_FILE_PREFIX environment variable
# is set, this will produce a copy-paste-able CSV file.
print(",".join(
map(str, [
batch_size, cell_size, cell_size, time_steps, config["use_gpu"],
wall_time
])))
benchmark_name_template = "_".join([
"LSTMBlockCell_bprop", "BS%(batch_size)i", "CS%(cell_size)i",
"IS%(cell_size)i", "TS%(time_steps)i", "gpu_%(use_gpu)s"
])
self.report_benchmark(
name=benchmark_name_template % config,
iters=iters,
wall_time=wall_time,
extras=config)
def benchmarkLSTMBlockCellBpropWithDynamicRNN(self):
print("BlockLSTMCell backward propagation via dynamic_rnn().")
print("--------------------------------------------------------------")
print("LSTMBlockCell Seconds per inference.")
print("batch_size,cell_size,input_size,time_steps,use_gpu,wall_time")
iters = 10
for config in benchmarking.dict_product({
"batch_size": [1, 8, 13, 32, 67, 128],
"cell_size": [128, 250, 512, 650, 1024, 1350],
"time_steps": [40],
"use_gpu": [True, False]
}):
with ops.Graph().as_default():
with benchmarking.device(use_gpu=config["use_gpu"]):
time_steps = config["time_steps"]
batch_size = config["batch_size"]
cell_size = input_size = config["cell_size"]
inputs = variable_scope.get_variable(
"x", [time_steps, batch_size, cell_size],
trainable=False,
dtype=dtypes.float32)
with variable_scope.variable_scope(
"rnn", reuse=variable_scope.AUTO_REUSE):
w = variable_scope.get_variable(
"rnn/lstm_cell/kernel",
shape=[input_size + cell_size, cell_size * 4],
dtype=dtypes.float32)
b = variable_scope.get_variable(
"rnn/lstm_cell/bias",
shape=[cell_size * 4],
dtype=dtypes.float32,
initializer=init_ops.zeros_initializer())
cell = lstm_ops.LSTMBlockCell(cell_size)
outputs = rnn.dynamic_rnn(
cell, inputs, time_major=True, dtype=dtypes.float32)
grads = gradients_impl.gradients(outputs, [inputs, w, b])
init_op = variables.global_variables_initializer()
with session.Session() as sess:
sess.run(init_op)
wall_time = benchmarking.seconds_per_run(grads, sess, iters)
# Print to stdout. If the TEST_REPORT_FILE_PREFIX environment variable
# is set, this will produce a copy-paste-able CSV file.
print(",".join(
map(str, [
batch_size, cell_size, cell_size, time_steps, config["use_gpu"],
wall_time
])))
benchmark_name_template = "_".join([
"LSTMBlockCell_bprop", "BS%(batch_size)i", "CS%(cell_size)i",
"IS%(cell_size)i", "TS%(time_steps)i", "gpu_%(use_gpu)s"
])
self.report_benchmark(
name=benchmark_name_template % config,
iters=iters,
wall_time=wall_time,
extras=config)
def single_bprop_step_gru_block_vs_gru_cell(batch_size,
cell_size,
input_size,
use_gpu=False,
iters=30):
"""Benchmark single bprop step speed between GRUBlockCell vs GRUCell."""
ops.reset_default_graph()
with session.Session(graph=ops.Graph()) as sess:
with benchmarking.device(use_gpu):
initializer = init_ops.random_uniform_initializer(-1, 1, seed=1989)
# Inputs
x = vs.get_variable("x", [batch_size, input_size])
h = vs.get_variable("h", [batch_size, cell_size])
# Output from the basic GRU cell implementation.
with vs.variable_scope("basic", initializer=initializer):
output = rnn_cell.GRUCell(cell_size)(array_ops.identity(x),
array_ops.identity(h))
sess.run([variables.global_variables_initializer()])
grad_output_wrt_input = gradients_impl.gradients([output], h)
basic_time_bprop = benchmarking.seconds_per_run(
grad_output_wrt_input, sess, iters)
# Output from the block GRU cell implementation.
with vs.variable_scope("block", initializer=initializer):
output = gru_ops.GRUBlockCell(cell_size)(array_ops.identity(x),
array_ops.identity(h))
sess.run([variables.global_variables_initializer()])
grad_output_wrt_input = gradients_impl.gradients([output], h)
block_time_bprop = benchmarking.seconds_per_run(
grad_output_wrt_input, sess, iters)
performance_inference = (basic_time_bprop -
block_time_bprop) * 100 / basic_time_bprop
print(",".join([
str(batch_size),
str(cell_size),
str(input_size),
str(use_gpu),
str(basic_time_bprop),
str(block_time_bprop),
str(performance_inference)
]))
return basic_time_bprop, block_time_bprop
def benchmarkLSTMBlockCellFpropWithDynamicRNN(self):
print("BlockLSTMCell forward propagation via dynamic_rnn().")
print("--------------------------------------------------------------")
print("LSTMBlockCell Seconds per inference.")
print("batch_size,cell_size,input_size,time_steps,use_gpu,wall_time")
iters = 10
for config in benchmarking.dict_product({
"batch_size": [1, 8, 13, 32, 67, 128],
"cell_size": [128, 250, 512, 650, 1024, 1350],
"time_steps": [40],
"use_gpu": [True, False],
"dtype": ["float32", "float16"],
}):
dtype = dtypes.float32 if config[
"dtype"] == "float32" else dtypes.float16
with ops.Graph().as_default():
with benchmarking.device(use_gpu=config["use_gpu"]):
inputs = variable_scope.get_variable(
"x",
dtype=dtype,
shape=[
config["time_steps"], config["batch_size"],
config["cell_size"]
])
cell = lstm_ops.LSTMBlockCell(config["cell_size"],
dtype=dtype)
outputs = rnn.dynamic_rnn(cell,
inputs,
time_major=True,
dtype=dtype)
init_op = variables.global_variables_initializer()
with session.Session() as sess:
sess.run(init_op)
wall_time = benchmarking.seconds_per_run(
outputs, sess, iters)
# Print to stdout. If the TEST_REPORT_FILE_PREFIX environment variable
# is set, this will produce a copy-paste-able CSV file.
print(",".join(
map(str, [
config["dtype"], config["batch_size"],
config["cell_size"], config["cell_size"],
config["time_steps"], config["use_gpu"], wall_time
])))
benchmark_name_template = "_".join([
"LSTMBlockCell_fprop", "DT_%(dtype)s", "BS%(batch_size)i",
"CS%(cell_size)i", "IS%(cell_size)i", "TS%(time_steps)i",
"gpu_%(use_gpu)s"
])
self.report_benchmark(name=benchmark_name_template % config,
iters=iters,
wall_time=wall_time,
extras=config)
def single_bprop_step_gru_block_vs_gru_cell(batch_size,
cell_size,
input_size,
use_gpu=False,
iters=30):
"""Benchmark single bprop step speed between GRUBlockCell vs GRUCell."""
ops.reset_default_graph()
with session.Session(graph=ops.Graph()) as sess:
with benchmarking.device(use_gpu):
initializer = init_ops.random_uniform_initializer(-1, 1, seed=1989)
# Inputs
x = vs.get_variable("x", [batch_size, input_size])
h = vs.get_variable("h", [batch_size, cell_size])
# Output from the basic GRU cell implementation.
with vs.variable_scope("basic", initializer=initializer):
output = rnn_cell.GRUCell(cell_size)(array_ops.identity(x),
array_ops.identity(h))
sess.run([variables.global_variables_initializer()])
grad_output_wrt_input = gradients_impl.gradients([output], h)
basic_time_bprop = benchmarking.seconds_per_run(grad_output_wrt_input,
sess, iters)
# Output from the block GRU cell implementation.
with vs.variable_scope("block", initializer=initializer):
output = gru_ops.GRUBlockCell(cell_size)(array_ops.identity(x),
array_ops.identity(h))
sess.run([variables.global_variables_initializer()])
grad_output_wrt_input = gradients_impl.gradients([output], h)
block_time_bprop = benchmarking.seconds_per_run(grad_output_wrt_input,
sess, iters)
performance_inference = (
basic_time_bprop - block_time_bprop) * 100 / basic_time_bprop
print(",".join([
str(batch_size), str(cell_size), str(input_size), str(use_gpu), str(
basic_time_bprop), str(block_time_bprop), str(performance_inference)
]))
return basic_time_bprop, block_time_bprop
def benchmarkLSTMBlockCellFpropWithDynamicRNN(self):
print("BlockLSTMCell forward propagation via dynamic_rnn().")
print("--------------------------------------------------------------")
print("LSTMBlockCell Seconds per inference.")
print("batch_size,cell_size,input_size,time_steps,use_gpu,wall_time")
iters = 10
for config in benchmarking.dict_product({
"batch_size": [1, 8, 13, 32, 67, 128],
"cell_size": [128, 250, 512, 650, 1024, 1350],
"time_steps": [40],
"use_gpu": [True, False],
"dtype": ["float32", "float16"],
}):
dtype = dtypes.float32 if config["dtype"] == "float32" else dtypes.float16
with ops.Graph().as_default():
with benchmarking.device(use_gpu=config["use_gpu"]):
inputs = variable_scope.get_variable(
"x",
dtype=dtype,
shape=[
config["time_steps"], config["batch_size"],
config["cell_size"]
])
cell = lstm_ops.LSTMBlockCell(config["cell_size"], dtype=dtype)
outputs = rnn.dynamic_rnn(cell, inputs, time_major=True, dtype=dtype)
init_op = variables.global_variables_initializer()
with session.Session() as sess:
sess.run(init_op)
wall_time = benchmarking.seconds_per_run(outputs, sess, iters)
# Print to stdout. If the TEST_REPORT_FILE_PREFIX environment variable
# is set, this will produce a copy-paste-able CSV file.
print(",".join(
map(str, [
config["dtype"], config["batch_size"], config["cell_size"],
config["cell_size"], config["time_steps"], config["use_gpu"],
wall_time
])))
benchmark_name_template = "_".join([
"LSTMBlockCell_fprop", "DT_%(dtype)s", "BS%(batch_size)i",
"CS%(cell_size)i", "IS%(cell_size)i", "TS%(time_steps)i",
"gpu_%(use_gpu)s"
])
self.report_benchmark(
name=benchmark_name_template % config,
iters=iters,
wall_time=wall_time,
extras=config)
def training_gru_block_vs_gru_cell(batch_size,
cell_size,
input_size,
time_steps,
use_gpu=False,
iters=30):
"""Benchmark training speed between GRUBlockCell vs GRUCell."""
ops.reset_default_graph()
with session.Session(graph=ops.Graph()) as sess:
# Specify the device which is been used.
with benchmarking.device(use_gpu):
# Random initializers.
seed = 1994
initializer = init_ops.random_uniform_initializer(-1, 1, seed=seed)
np.random.seed(seed)
# Inputs
concat_x = vs.get_variable("concat_x",
[time_steps, batch_size, input_size])
h = vs.get_variable("h", [batch_size, cell_size])
y = vs.get_variable("y", [time_steps, batch_size, cell_size])
# Output from the basic GRU cell implementation.
with vs.variable_scope("basic", initializer=initializer):
cell = rnn_cell.GRUCell(cell_size)
outputs_dynamic, _ = rnn.dynamic_rnn(
cell,
inputs=concat_x,
initial_state=h,
time_major=True,
dtype=dtypes.float32)
sess.run([variables.global_variables_initializer()])
cost = math_ops.reduce_mean(math_ops.square(outputs_dynamic - y))
learning_rate = 0.01
optimizer = gradient_descent.GradientDescentOptimizer(
learning_rate).minimize(cost)
# time for a training step.
basic_time_training = benchmarking.seconds_per_run(
optimizer, sess, iters)
# Output from the basic GRU cell implementation.
with vs.variable_scope("block", initializer=initializer):
cell = gru_ops.GRUBlockCell(cell_size)
outputs_dynamic, _ = rnn.dynamic_rnn(
cell,
inputs=concat_x,
initial_state=h,
time_major=True,
dtype=dtypes.float32)
sess.run([variables.global_variables_initializer()])
cost = math_ops.reduce_mean(math_ops.square(outputs_dynamic - y))
learning_rate = 0.01
optimizer = gradient_descent.GradientDescentOptimizer(
learning_rate).minimize(cost)
# time for a training step.
block_time_training = benchmarking.seconds_per_run(
optimizer, sess, iters)
performance_training = (
basic_time_training - block_time_training) * 100 / basic_time_training
print(",".join([
str(batch_size), str(cell_size), str(input_size), str(time_steps), str(
use_gpu), str(basic_time_training), str(block_time_training), str(
performance_training)
]))
return basic_time_training, block_time_training
