def testDistributedReplicatedSavableVars(self):
test_util.monkey_patch_base_cluster_manager()
params = benchmark_cnn.make_params(
variable_update='distributed_replicated',
model='inception4',
data_name='imagenet',
data_dir=os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data'),
job_name='worker',
worker_hosts='w1,w2,w3,w4',
ps_hosts='p1')
bench = benchmark_cnn.BenchmarkCNN(params)
with tf.Graph().as_default():
bench._build_model()
savable_vars = bench.variable_mgr.savable_variables()
# Assert all global variables are in savable_vars
for v in tf.global_variables():
if not v.name.startswith(
variable_mgr_util.PS_SHADOW_VAR_PREFIX + '/v0'):
self.assertEqual(v.name, 'global_step:0')
name = bench.variable_mgr._strip_port(v.name)
if name.startswith(variable_mgr_util.PS_SHADOW_VAR_PREFIX):
name = name[len(variable_mgr_util.PS_SHADOW_VAR_PREFIX +
'/'):]
self.assertIn(name, savable_vars)
self.assertIn(savable_vars[name], tf.global_variables())
# Assert all local variables on the first tower are in savable_vars
for v in tf.local_variables():
if v.name.startswith('v0/'):
name = bench.variable_mgr._strip_port(v.name)
self.assertIn(name, savable_vars)
def testPreprocessingTrain(self):
test_data_dir = os.path.join(platforms_util.get_test_data_dir(),
'images')
black_file = os.path.join(test_data_dir, 'black_image.jpg')
with open(black_file, 'r') as f:
black_jpg_buffer = f.read()
white_file = os.path.join(test_data_dir, 'white_image.jpg')
with open(white_file, 'r') as f:
white_jpg_buffer = f.read()
bbox = tf.zeros((1, 0, 4), dtype=tf.float32)
batch_position = 0
# Each size config is (output_height, output_width, resize_method)
size_configs = [(100, 100, 'round_robin'), (150, 10, 'bilinear'),
(10, 150, 'nearest')]
# Each image config is (image_buf, image_color)
image_configs = [(white_jpg_buffer, 255), (black_jpg_buffer, 0)]
for (image_buf, image_color) in image_configs:
for output_height, output_width, resize_method in size_configs:
for distortions in [True, False]:
for summary_verbosity in [0, 2]:
for fuse_decode_and_crop in [True, False]:
self._test_preprocessing_traing(
image_buf, image_color, output_height,
output_width, bbox, batch_position,
resize_method, distortions, summary_verbosity,
fuse_decode_and_crop)
def testImagenetPreprocessorNoDistortions(self):
imagenet_dir = os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data')
params = test_util.get_params(
'testImagenetPreprocessorNoDistortions')._replace(
data_dir=imagenet_dir, data_name='imagenet', distortions=False)
self._train_and_eval_local(params, use_test_preprocessor=False)
def testReplicatedRealData(self):
test_name = 'testReplicatedRealData'
imagenet_dir = os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data')
params = test_util.get_params(test_name)._replace(
variable_update='distributed_replicated',
data_dir=imagenet_dir,
data_name='imagenet')
self._test_distributed(test_name, 2, 2, params)
def __init__(self, output_dir=None):
# Load default values if the benchmark is not run with absl.app.run()
if not flags.FLAGS.is_parsed():
flags.FLAGS.mark_as_parsed()
self.fake_data_dir = os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data')
self.data_dir = ('/readahead/200M/placer/prod/home/distbelief/'
'imagenet-tensorflow/imagenet-2012-tfrecord')
self.output_dir = output_dir
def testOfficialImagenetPreprocessorFp16(self):
imagenet_dir = os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data')
params = test_util.get_params(
'testOfficialImagenetPreprocessorFp16')._replace(
data_dir=imagenet_dir,
data_name='imagenet',
input_preprocessor='official_models_imagenet',
use_fp16=True)
self._train_and_eval_local(params, use_test_preprocessor=False)
def testShiftRatio(self):
test_util.monkey_patch_base_cluster_manager()
params = benchmark_cnn.make_params(
data_name='imagenet',
data_dir=os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data'),
job_name='worker',
worker_hosts='w1,w2,w3,w4',
ps_hosts='p1',
task_index=0)
self.assertEqual(
benchmark_cnn.BenchmarkCNN(params).image_preprocessor.shift_ratio, 0.0)
params = params._replace(task_index=3)
self.assertEqual(
benchmark_cnn.BenchmarkCNN(params).image_preprocessor.shift_ratio, 0.75)
def __init__(self, output_dir=None, root_data_dir=None):
"""Base class for all benchmarks in this file.
Args:
output_dir: directory where to output e.g. log files
root_data_dir: directory under which to look for dataset
"""
# Load default values if the benchmark is not run with absl.app.run()
if not flags.FLAGS.is_parsed():
flags.FLAGS.mark_as_parsed()
self.fake_data_dir = os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data')
if root_data_dir is None:
self.data_dir = ('/readahead/200M/placer/prod/home/distbelief/'
'imagenet-tensorflow/imagenet-2012-tfrecord')
else:
self.data_dir = os.path.join(root_data_dir, 'imagenet')
self.output_dir = output_dir
def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
"""Base class for all benchmarks in this file.
Args:
output_dir: directory where to output e.g. log files
root_data_dir: directory under which to look for dataset
**kwargs: arbitrary named arguments. This is needed to make the
constructor forward compatible in case PerfZero provides more
named arguments before updating the constructor.
"""
# Load default values if the benchmark is not run with absl.app.run()
if not flags.FLAGS.is_parsed():
flags.FLAGS.mark_as_parsed()
self.fake_data_dir = os.path.join(platforms_util.get_test_data_dir(),
'fake_tf_record_data')
self.output_dir = output_dir
if root_data_dir is None:
self.data_dir = ('/readahead/200M/placer/prod/home/distbelief/'
'imagenet-tensorflow/imagenet-2012-tfrecord')
else:
self.data_dir = os.path.join(root_data_dir, 'imagenet')
