def testFromStringHandle(self):
test_cases = [{
'shape': tensor_shape.TensorShape([])
}, {
'shape': tensor_shape.TensorShape([3])
}, {
'shape': tensor_shape.TensorShape([1, 2])
}, {
'shape': tensor_shape.TensorShape([1, 2, 3])
}]
for test_case in test_cases:
with ops.Graph().as_default() as g:
iterator = iterator_ops.Iterator.from_structure(dtypes.int64)
handle = iterator.string_handle()
iterator = iterator_ops.Iterator.from_string_handle(
handle, dtypes.int64, output_shapes=test_case['shape'])
get_next = iterator.get_next()
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(get_next)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
self.assertEqual(test_case['shape'],
op_properties['IteratorGetNext'][0].shape)
def run(self):
"""Generates a placement graph.
Returns:
a tuple of placement graph and a dictionary that maps an operator
name to a node index in the placement graph.
"""
op_graph = nx.DiGraph()
op_index = self._add_nodes(op_graph)
self._add_edges(op_graph, op_index)
op_graph, op_index = placer_utils.prune_dangling_ops(op_graph)
if self._only_important_ops:
metagraph = tf.train.export_meta_graph(
graph=self._tf_graph, clear_extraneous_savers=True)
item = gitem.Item(metagraph, ignore_colocation=False)
important_ops = item.IdentifyImportantOps()
_LOGGER.info("Use only important ops. # important ops=%d",
len(important_ops))
op_graph, op_index = placer_utils.prune_non_important_ops(
op_graph, important_ops)
# assign topo order
placer_utils.assign_topo_order(op_graph)
return op_graph, op_index
def testMap(self):
test_cases = [{
'tensor': 0,
'shape': tensor_shape.TensorShape([])
}, {
'tensor': np.array([1, 2, 3]),
'shape': tensor_shape.TensorShape([3])
}, {
'tensor': np.array([[1, 2, 3]]),
'shape': tensor_shape.TensorShape([3, 1])
}, {
'tensor': np.array([[[1, 2, 3], [4, 5, 6]]]),
'shape': tensor_shape.TensorShape([3, 2, 1])
}]
for test_case in test_cases:
with ops.Graph().as_default() as g:
dataset = dataset_ops.Dataset.from_tensors(test_case['tensor'])
dataset = dataset.map(array_ops.transpose)
iterator = dataset.make_one_shot_iterator()
get_next = iterator.get_next()
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(get_next)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
self.assertEqual(test_case['shape'],
op_properties['IteratorGetNext'][0].shape)
def GenerateMemoryReport(metagraph, detailed_report=True, cluster=None):
"""Analyze the peak memory usage for the provided metagraph.
Args:
metagraph: A TensorFlow MetaGraphDef.
detailed_report: print the live tensors in addition to the peak memory
usage.
cluster: Analyze the memory using the specified cluster, or the local
machine if no cluster was specified.
Returns:
A string with the formatted memory usage.
"""
if cluster is None:
cluster = gcluster.Cluster(
disable_detailed_stats=True, disable_timeline=True)
item = gitem.Item(metagraph)
peak_usage = cluster.DeterminePeakMemoryUsage(item)
report = ""
for device, snapshot in peak_usage.items():
peak_usage = snapshot[0]
report += "Peak usage for device " + device + ": " + str(
peak_usage) + " bytes\n"
if detailed_report:
live_tensors = snapshot[1]
for tensor in live_tensors:
op_name = tensor[0]
output_id = tensor[1]
mem_used = tensor[2]
report += " " + str(op_name) + ":" + str(output_id) + " uses " + str(
mem_used) + " bytes\n"
return report
def testPaddedBatch(self):
test_cases = [{
'tensor': 0,
'shape': tensor_shape.TensorShape([None])
}, {
'tensor': np.array([1, 2, 3]),
'shape': tensor_shape.TensorShape([None, 4])
}, {
'tensor': np.array([[1, 2, 3]]),
'shape': tensor_shape.TensorShape([None, 2, 4])
}]
for test_case in test_cases:
with ops.Graph().as_default() as g:
dataset = dataset_ops.Dataset.from_tensors(test_case['tensor'])
dataset = dataset.padded_batch(
42, padded_shapes=test_case['shape'][1:])
iterator = dataset.make_one_shot_iterator()
get_next = iterator.get_next()
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(get_next)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
inferred_shape = self.as_tensor_shape(
op_properties['IteratorGetNext'][0].shape)
self.assertTrue(test_case['shape'].dims[0].is_compatible_with(
inferred_shape[0]))
self.assertEqual(test_case['shape'][1:], inferred_shape[1:])
def testVirtualCluster(self):
with ops.Graph().as_default() as g:
with ops.device('/device:GPU:0'):
a = random_ops.random_uniform(shape=[1024, 1024])
b = random_ops.random_uniform(shape=[1024, 1024])
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type='GPU',
frequency=1000,
num_cores=60,
environment={'architecture': '7'})
named_device = device_properties_pb2.NamedDevice(
properties=device_properties, name='/device:GPU:0')
grappler_cluster = cluster.Cluster(
disable_detailed_stats=False,
disable_timeline=False,
devices=[named_device])
op_perfs, run_time, _ = grappler_cluster.MeasureCosts(grappler_item)
self.assertEqual(run_time, 0.000545)
self.assertEqual(len(op_perfs), 15)
estimated_perf = grappler_cluster.EstimatePerformance(named_device)
self.assertEqual(7680.0, estimated_perf)
def testUpdates(self):
with ops.Graph().as_default() as g:
a = constant_op.constant(10)
b = constant_op.constant(20)
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
initial_tf_item = grappler_item.tf_item
no_change_tf_item = grappler_item.tf_item
self.assertEqual(initial_tf_item, no_change_tf_item)
# Modify the placement.
for node in grappler_item.metagraph.graph_def.node:
node.device = '/cpu:0'
new_tf_item = grappler_item.tf_item
self.assertNotEqual(initial_tf_item, new_tf_item)
# Assign the same placement.
for node in grappler_item.metagraph.graph_def.node:
node.device = '/cpu:0'
newest_tf_item = grappler_item.tf_item
self.assertEqual(new_tf_item, newest_tf_item)
def testFromGenerator(self):
test_cases = [{
'tensor': 0,
'shape': tensor_shape.TensorShape([])
}, {
'tensor': np.array([1, 2, 3]),
'shape': tensor_shape.TensorShape([3])
}, {
'tensor': np.array([[1, 2, 3]]),
'shape': tensor_shape.TensorShape([1, 3])
}]
for test_case in test_cases:
def make_generator(tensor):
def generator():
yield tensor
return generator
with ops.Graph().as_default() as g:
dataset = dataset_ops.Dataset.from_generator(
make_generator(test_case['tensor']),
dtypes.int64,
output_shapes=test_case['shape'])
iterator = dataset.make_one_shot_iterator()
get_next = iterator.get_next()
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(get_next)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
self.assertEqual(test_case['shape'],
op_properties['IteratorGetNext'][0].shape)
def testInterleave(self):
test_cases = [{
'tensor': 0,
'shape': tensor_shape.TensorShape([])
}, {
'tensor': np.array([1, 2, 3]),
'shape': tensor_shape.TensorShape([3])
}, {
'tensor': np.array([[1, 2, 3]]),
'shape': tensor_shape.TensorShape([1, 3])
}]
for test_case in test_cases:
with ops.Graph().as_default() as g:
dataset = dataset_ops.Dataset.range(42)
def make_dataset(tensor):
def dataset_fn(n):
return dataset_ops.Dataset.from_tensors(tensor).repeat(
n)
return dataset_fn
dataset = dataset.interleave(make_dataset(test_case['tensor']),
cycle_length=42)
iterator = dataset.make_one_shot_iterator()
get_next = iterator.get_next()
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(get_next)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
self.assertEqual(test_case['shape'],
op_properties['IteratorGetNext'][0].shape)
def testInvalidItem(self):
with ops.Graph().as_default() as g:
a = constant_op.constant(10)
b = constant_op.constant(20)
c = a + b # pylint: disable=unused-variable
mg = meta_graph.create_meta_graph_def(graph=g)
# The train op isn't specified: this should raise an InvalidArgumentError
# exception.
with self.assertRaises(errors_impl.InvalidArgumentError):
item.Item(mg)
def testImportantOps(self):
with ops.Graph().as_default() as g:
a = constant_op.constant(10)
b = constant_op.constant(20)
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_list = grappler_item.IdentifyImportantOps()
self.assertItemsEqual(['Const', 'Const_1', 'add'], op_list)
def testRange(self):
with ops.Graph().as_default() as g:
dataset = dataset_ops.Dataset.range(42)
iterator = dataset.make_one_shot_iterator()
get_next = iterator.get_next()
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(get_next)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
self.assertEqual(tensor_shape.scalar(),
op_properties['IteratorGetNext'][0].shape)
def testColocationContraints(self):
with ops.Graph().as_default() as g:
c = constant_op.constant([10])
v = variables.VariableV1([3], dtype=dtypes.int32)
i = gen_array_ops.ref_identity(v)
a = state_ops.assign(i, c)
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(a)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
groups = grappler_item.GetColocationGroups()
self.assertEqual(len(groups), 1)
self.assertItemsEqual(
groups[0], ['Assign', 'RefIdentity', 'Variable', 'Variable/Assign'])
def testContext(self):
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=())
b = random_ops.random_uniform(shape=())
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
with cluster.Provision(
disable_detailed_stats=False, disable_timeline=False) as gcluster:
op_perfs, run_time, step_stats = gcluster.MeasureCosts(grappler_item)
self.assertTrue(run_time > 0)
self.assertEqual(len(op_perfs), 4)
self.assertTrue(step_stats.dev_stats)
def testNoDetailedStats(self):
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=())
b = random_ops.random_uniform(shape=())
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
grappler_cluster = cluster.Cluster(disable_detailed_stats=True)
op_perfs, run_time, step_stats = grappler_cluster.MeasureCosts(
grappler_item)
self.assertTrue(run_time > 0)
self.assertEqual(len(op_perfs), 0)
self.assertEqual(len(step_stats.dev_stats), 0)
def testSupportDevices(self):
gpu_type = test_util.gpu_device_type()
gpu_name = test_util.gpu_device_name()
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=(2, 3))
b = random_ops.random_uniform(shape=(2, 3))
c = a + b
dims = math_ops.range(0, array_ops.rank(c), 1)
d = math_ops.reduce_sum(a, axis=dims)
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(d)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type=gpu_type, frequency=1000, num_cores=60)
named_gpu = device_properties_pb2.NamedDevice(
properties=device_properties, name=gpu_name)
device_properties = device_properties_pb2.DeviceProperties(
type='CPU', frequency=3000, num_cores=6)
named_cpu = device_properties_pb2.NamedDevice(
properties=device_properties, name='/CPU:0')
virtual_cluster = cluster.Cluster(devices=[named_cpu, named_gpu])
supported_dev = virtual_cluster.GetSupportedDevices(grappler_item)
self.assertEqual(supported_dev['add'], ['/CPU:0', gpu_name])
self.assertEqual(supported_dev['Sum'], ['/CPU:0', gpu_name])
self.assertEqual(supported_dev['range'], ['/CPU:0', gpu_name])
real_cluster = cluster.Cluster()
supported_dev = real_cluster.GetSupportedDevices(grappler_item)
if test.is_gpu_available():
self.assertEqual(supported_dev['add'], [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0' + gpu_name
])
self.assertEqual(supported_dev['Sum'], [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0' + gpu_name
])
# The axis tensor must reside on the host
self.assertEqual(
supported_dev['range'],
['/job:localhost/replica:0/task:0/device:CPU:0'])
else:
self.assertEqual(
supported_dev['add'],
['/job:localhost/replica:0/task:0/device:CPU:0'])
def testOpProperties(self):
with ops.Graph().as_default() as g:
a = constant_op.constant(10)
b = constant_op.constant(20)
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
op_properties = grappler_item.GetOpProperties()
# All the nodes in this model have one scalar output
for node in grappler_item.metagraph.graph_def.node:
node_prop = op_properties[node.name]
self.assertEqual(1, len(node_prop))
self.assertEqual(dtypes.int32, node_prop[0].dtype)
self.assertEqual(tensor_shape.scalar(), node_prop[0].shape)
def testVirtualCluster(self):
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=())
b = random_ops.random_uniform(shape=())
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type='GPU', environment={
'architecture': '7'
})
named_device = device_properties_pb2.NamedDevice(
properties=device_properties, name='/GPU:0')
grappler_cluster = cluster.Cluster(devices=[named_device])
op_perfs, run_time, _ = grappler_cluster.MeasureCosts(grappler_item)
self.assertGreater(run_time, 0)
self.assertEqual(len(op_perfs), 15)
def testLoops(self):
g = ops.Graph()
with g.as_default():
def _Cond(_, counter):
return counter < end
def _Body(buf, counter):
buf = array_ops.concat([buf, [counter]], 0)
counter += 1
return [buf, counter]
start = array_ops.placeholder(shape=[], dtype=dtypes.int32)
end = array_ops.placeholder(shape=[], dtype=dtypes.int32)
init_buf = array_ops.zeros(shape=[0], dtype=dtypes.int32)
loop_vars = [init_buf, start]
shape_inv = [
tensor_shape.TensorShape([None]),
tensor_shape.TensorShape([])
]
buf, _ = control_flow_ops.while_loop(_Cond, _Body, loop_vars,
shape_inv)
f = -array_ops.ones_like(buf, optimize=False) # pylint: disable=invalid-unary-operand-type
buf_shape = array_ops.shape(buf)
f_shape = array_ops.shape(f)
ops.add_to_collection('train_op', buf_shape)
ops.add_to_collection('train_op', f_shape)
# Optimize the graph.
mg = meta_graph.create_meta_graph_def(graph=g)
config = config_pb2.ConfigProto()
rewriter_config = config.graph_options.rewrite_options
rewriter_config.min_graph_nodes = -1
optimized_graph = tf_optimizer.OptimizeGraph(config, mg)
mg.graph_def.CopyFrom(optimized_graph)
# Check that the nodes referenced in various collections have been preserved
item = gitem.Item(mg)
props = item.GetOpProperties()
buf_prop = props[buf.op.name]
f_prop = props[f.op.name]
self.assertEqual(buf_prop, f_prop)
def testMemoryEstimates(self):
with ops.Graph().as_default() as g:
with ops.device('/job:localhost/replica:0/task:0/device:CPU:0'):
a = random_ops.random_uniform(shape=())
b = random_ops.random_uniform(shape=())
c = a + b
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(c)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
grappler_cluster = cluster.Cluster(
disable_detailed_stats=True, disable_timeline=True)
peak_mem = grappler_cluster.DeterminePeakMemoryUsage(grappler_item)
self.assertLessEqual(1, len(peak_mem))
snapshot = peak_mem['/job:localhost/replica:0/task:0/device:CPU:0']
peak_usage = snapshot[0]
self.assertEqual(52, peak_usage)
live_tensors = snapshot[1]
self.assertEqual(15, len(live_tensors))
def testSupportDevices(self):
with ops.Graph().as_default() as g:
a = random_ops.random_uniform(shape=(2, 3))
b = random_ops.random_uniform(shape=(2, 3))
c = a + b
dims = math_ops.range(0, array_ops.rank(c), 1)
d = math_ops.reduce_sum(a, axis=dims)
train_op = ops.get_collection_ref(ops.GraphKeys.TRAIN_OP)
train_op.append(d)
mg = meta_graph.create_meta_graph_def(graph=g)
grappler_item = item.Item(mg)
device_properties = device_properties_pb2.DeviceProperties(
type='GPU', frequency=1000, num_cores=60)
named_gpu = device_properties_pb2.NamedDevice(
properties=device_properties, name='/GPU:0')
device_properties = device_properties_pb2.DeviceProperties(
type='CPU', frequency=3000, num_cores=6)
named_cpu = device_properties_pb2.NamedDevice(
properties=device_properties, name='/CPU:0')
virtual_cluster = cluster.Cluster(devices=[named_cpu, named_gpu])
supported_dev = virtual_cluster.GetSupportedDevices(grappler_item)
self.assertEqual(supported_dev['add'], ['/CPU:0', '/GPU:0'])
self.assertEqual(supported_dev['Sum'], ['/CPU:0', '/GPU:0'])
self.assertEqual(supported_dev['range'], ['/CPU:0', '/GPU:0'])
real_cluster = cluster.Cluster()
supported_dev = real_cluster.GetSupportedDevices(grappler_item)
#NCL 18.04 -- Hack to account for possible XLA devices
if test.is_gpu_available():
add_devices = [
d for d in supported_dev['add']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(add_devices, [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0/device:GPU:0'
])
Sum_devices = [
d for d in supported_dev['Sum']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(Sum_devices, [
'/job:localhost/replica:0/task:0/device:CPU:0',
'/job:localhost/replica:0/task:0/device:GPU:0'
])
# The axis tensor must reside on the host
range_devices = [
d for d in supported_dev['range']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(
range_devices,
['/job:localhost/replica:0/task:0/device:CPU:0'])
else:
add_devices = [
d for d in supported_dev['add']
if not d.split(':')[-2].startswith('XLA')
]
self.assertEqual(
add_devices,
['/job:localhost/replica:0/task:0/device:CPU:0'])
def PlaceGraph(metagraph,
cluster=None,
allotted_time=3600,
hparams=None,
verbose=False):
"""Place the provided metagraph.
Args:
metagraph: the metagraph to place.
cluster: an optional set of hardware resource to optimize the placement for.
If none is specified, we'll optimize the placement for the hardware
available on the local machine.
allotted_time: the maximum amount to time in seconds to spend optimizing
the placement.
hparams: hyperparameters used to fine tune the placer.
verbose: prints debug information if True.
Returns:
The placed metagraph.
"""
if cluster is None:
cluster = gcluster.Cluster()
# Optimize the metagraph to speedup the placement
rewriter_config = rewriter_config_pb2.RewriterConfig()
rewriter_config.optimizers.append("pruning")
rewriter_config.optimizers.append("constfold")
rewriter_config.optimizers.append("arithmetic")
rewriter_config.optimizers.append("dependency")
rewriter_config.optimizers.append("pruning")
optimized_graph = tf_optimizer.OptimizeGraph(rewriter_config,
metagraph,
verbose=verbose,
cluster=cluster)
optimized_metagraph = meta_graph_pb2.MetaGraphDef()
optimized_metagraph.CopyFrom(metagraph)
optimized_metagraph.graph_def.CopyFrom(optimized_graph)
item = gitem.Item(optimized_metagraph)
# Measure the runtime achievable with the original placement.
try:
_, original_run_time, _ = cluster.MeasureCosts(item)
if verbose:
print("Runtime for original placement: " + str(original_run_time))
except errors.OpError as e:
if verbose:
print("Original placement isn't feasible: " + str(e))
original_run_time = hparams.failing_signal
if hparams is None:
hparams = hierarchical_controller.hierarchical_controller_hparams()
# We run with a single child
hparams.num_children = 1
with tf_ops.Graph().as_default():
# Place all the nodes of the controller on the CPU. We don't want them to
# fight for accelerator memory with the model to optimize.
with tf_ops.device("/device:CPU:0"):
model = hierarchical_controller.HierarchicalController(
hparams, item, cluster)
ops = model.build_controller()
session_creator = training.ChiefSessionCreator()
with training.MonitoredSession(
session_creator=session_creator) as sess:
start_time = time.time()
current_time = start_time
while current_time - start_time < allotted_time:
grouping_actions = model.generate_grouping(sess)
input_to_seq2seq = model.create_group_embeddings(
grouping_actions, verbose=verbose)
model.generate_placement(input_to_seq2seq, sess)
try:
run_time = model.eval_placement(sess, verbose=verbose)
except errors.OpError as e:
if verbose:
print("Failed to run graph:" + str(e))
run_time = hparams.failing_signal
updated = model.update_reward(sess,
run_time,
verbose=verbose)
if updated and run_time < original_run_time:
if verbose:
print("Found better placement, with runtime " +
str(run_time))
model.export_placement(metagraph)
model.process_reward(sess)
current_time = time.time()
return metagraph