def build_graph(tower_devices, tensor_shapes, variable_mgr, num_iters):
"""Builds the graph for the benchmark.
Args:
tower_devices: A list of device strings of the devices to run the all-reduce
benchmark on.
tensor_shapes: A list of shapes of the tensors that will be aggregated for
the all-reduce.
variable_mgr: The VariableMgr to perform the all-reduce.
num_iters: Number of iterations to aggregate tensors for.
Returns:
An op that runs the benchmark.
"""
all_device_tensors = []
for i, tower_device in enumerate(tower_devices):
with tf.device(tower_device):
device_tensors = []
for j, shape in enumerate(tensor_shapes):
tensor = tf.Variable(tf.random_normal(shape, dtype=tf.float32),
name='tensor_%d_on_device_%d' % (j, i))
device_tensors.append(tensor)
all_device_tensors.append(device_tensors)
log_fn('Building all-reduce ops')
benchmark_op = build_all_reduce_iterations(all_device_tensors,
tower_devices, variable_mgr,
num_iters)
log_fn('Done building all-reduce ops')
return benchmark_op
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s' %
positional_arguments[1:])
options = make_options_from_flags(FLAGS)
params = benchmark_cnn.make_params_from_flags()
params = params._replace(batch_size=options.batch_size)
params = params._replace(model='MY_GTSRB')
params = params._replace(num_epochs=options.num_epochs)
params = params._replace(num_gpus=options.num_gpus)
params = params._replace(data_format='NHWC')
params = params._replace(train_dir=options.checkpoint_folder)
params = params._replace(allow_growth=True)
params = params._replace(variable_update='replicated')
params = params._replace(local_parameter_device='gpu')
params = params._replace(use_tf_layers=False)
# params = params._replace(all_reduce_spec='nccl')
# params = params._replace(bottom_file=options.bottom_file)
# params = params._replace(affine_files=options.affine_files)
# params = params._replace(affine_classes=options.affine_classes)
params = params._replace(optimizer=options.optimizer)
params = params._replace(weight_decay=options.weight_decay)
#params = params._replace(print_training_accuracy=True)
params = params._replace(backbone_model_path=options.backbone_model_path)
# Summary and Save & load checkpoints.
# params = params._replace(summary_verbosity=1)
# params = params._replace(save_summaries_steps=10)
# params = params._replace(save_model_secs=3600) # save every 1 hour
params = params._replace(save_model_secs=60) #save every 5 min
params = benchmark_cnn.setup(params)
#testtest(params)
#exit(0)
if 'test' in options.data_dir:
dataset = GTSRBTestDataset(options)
else:
dataset = GTSRBDataset(options)
model = Model_Builder(options.model_name, dataset.num_classes, options,
params)
bench = benchmark_cnn.BenchmarkCNN(params, dataset=dataset, model=model)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
tf.reset_default_graph()
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s' %
positional_arguments[1:])
params = benchmark_cnn.make_params_from_flags()
params = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
with log_context(LOGGER_URL,
LOGGER_USRENAME,
LOGGER_PASSWORD,
LOGGER_DB,
LOGGER_SERIES,
machine=LOGGER_VM):
bench.run()
def postprocess(self, results):
"""Postprocess results returned from model in Python."""
probs = results[self.PROBABILITY_TENSOR]
total_wer, total_cer = 0, 0
speech_labels = " abcdefghijklmnopqrstuvwxyz'-"
greedy_decoder = DeepSpeechDecoder(speech_labels)
# Evaluate the performance using WER (Word Error Rate) and CER (Character
# Error Rate) as metrics.
targets = results[self.LABEL_TENSOR] # The ground truth transcript
for i in range(self.batch_size):
# Decode string.
predicted_str = greedy_decoder.decode_logits(probs[i])
expected_str = greedy_decoder.decode(targets[i])
# Compute CER.
total_cer += (greedy_decoder.cer(predicted_str, expected_str) /
len(expected_str))
# Compute WER.
total_wer += (greedy_decoder.wer(predicted_str, expected_str) /
len(expected_str.split()))
# Get mean value
total_cer /= self.batch_size
total_wer /= self.batch_size
log_fn('total CER: {:f}; total WER: {:f}; total example: {:d}.'.format(
total_cer, total_wer, self.batch_size))
# TODO(laigd): get rid of top_N_accuracy bindings in benchmark_cnn.py
return {'top_1_accuracy': 0., 'top_5_accuracy': 0.}
def run_benchmark(bench_cnn, num_iters):
"""Runs the all-reduce benchmark.
Args:
bench_cnn: The BenchmarkCNN where params, the variable manager, and other
attributes are obtained.
num_iters: Number of iterations to do all-reduce for for.
Raises:
ValueError: Invalid params of bench_cnn.
"""
if bench_cnn.params.variable_update != 'replicated':
raise ValueError('--variable_update=replicated must be specified to use'
'the all-reduce benchmark')
if bench_cnn.params.variable_consistency == 'relaxed':
raise ValueError('--variable_consistency=relaxed is not supported')
benchmark_op = build_graph(bench_cnn.raw_devices,
get_var_shapes(bench_cnn.model),
bench_cnn.variable_mgr, num_iters)
init_ops = [
tf.global_variables_initializer(),
bench_cnn.variable_mgr.get_post_init_ops()
]
loss_op = tf.no_op()
if bench_cnn.graph_file:
path, filename = os.path.split(bench_cnn.graph_file)
as_text = filename.endswith('txt')
log_fn('Writing GraphDef as %s to %s' % (
'text' if as_text else 'binary', bench_cnn.graph_file))
tf.train.write_graph(tf.get_default_graph().as_graph_def(add_shapes=True),
path, filename, as_text)
run_graph(benchmark_op, bench_cnn, init_ops, loss_op)
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s' %
positional_arguments[1:])
params = benchmark_cnn.make_params_from_flags()
# Print ENV Variables
tf.logging.debug('=' * 20 + ' Environment Variables ' + '=' * 20)
for k, v in os.environ.items():
tf.logging.debug('{}: {}'.format(k, v))
with mlperf.mlperf_logger(absl_flags.FLAGS.ml_perf_compliance_logging,
params.model):
params = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s' %
positional_arguments[1:])
params = benchmark_cnn.make_params_from_flags()
params = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
print('num_inter_threads: ' + str(params.num_inter_threads))
print('num_intra_threads: ' + str(params.num_intra_threads))
print('datasets_num_private_threads: ' +
str(params.datasets_num_private_threads))
print('datasets_use_prefetch: ' + str(params.datasets_use_prefetch))
print('datasets_prefetch_buffer_size: ' +
str(params.datasets_prefetch_buffer_size))
bench.run()
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s'
% positional_arguments[1:])
params = benchmark_cnn.make_params_from_flags()
params = benchmark_cnn.setup(params)
import sys
if params.enable_dmo == True:
if LoadFileSystem() == False:
sys.exit(-1)
else :
print("\n*******DMO enabled********\n")
# sys.exit(0)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
def build_network(self, images, phase_train=True, nclass=1001, image_depth=3,
data_type=tf.float32, data_format='NCHW',
use_tf_layers=True, fp16_vars=False):
"""Returns logits and aux_logits from images."""
if data_format == 'NCHW':
images = tf.transpose(images, [0, 3, 1, 2])
var_type = tf.float32
if data_type == tf.float16 and fp16_vars:
var_type = tf.float16
network = convnet_builder.ConvNetBuilder(
images, image_depth, phase_train, use_tf_layers,
data_format, data_type, var_type)
with tf.variable_scope('cg', custom_getter=network.get_custom_getter()):
self.add_inference(network)
log_fn("Number of parameters: %d" % network.n_parameters)
# Add the final fully-connected class layer
logits = (network.affine(nclass, activation='linear')
if not self.skip_final_affine_layer()
else network.top_layer)
aux_logits = None
if network.aux_top_layer is not None:
with network.switch_to_aux_top_layer():
aux_logits = network.affine(
nclass, activation='linear', stddev=0.001)
if data_type == tf.float16:
# TODO(reedwm): Determine if we should do this cast here.
logits = tf.cast(logits, tf.float32)
if aux_logits is not None:
aux_logits = tf.cast(aux_logits, tf.float32)
return logits, aux_logits
def main(_):
params = benchmark_cnn.make_params_from_flags()
params = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
def _eval_once(self, saver, image_producer_ops, fetches,
local_var_init_op_group, nth_ckpt):
with tf.Session(config=create_config_proto(self.params)) as sess:
coord = tf.train.Coordinator()
if self.params.checkpoint_dir is None:
raise ValueError(
'Checkpoint directory for evaluation is not specified')
try:
global_step = load_checkpoint(saver, sess,
self.params.checkpoint_dir,
nth_ckpt)
except CheckpointNotFoundException:
log_fn(
'Checkpoint not found in %s' % self.params.checkpoint_dir)
sys.exit(-1)
return
log_fn('[Evaluation] START')
sess.run(local_var_init_op_group)
assert not self.use_synthetic_gpu_images
dummy_queue = tf.FIFOQueue(1, [tf.bool], shapes=[[]],
name='dummy_queue',
shared_name='dummy_queue')
qr = tf.train.QueueRunner(dummy_queue, image_producer_ops)
tf.add_to_collection(tf.GraphKeys.QUEUE_RUNNERS, qr)
enqueue_threads = qr.create_threads(sess=sess, coord=coord,
daemon=True)
for thread in enqueue_threads:
thread.start()
top_1_accuracy_sum = 0.0
top_5_accuracy_sum = 0.0
total_eval_count = self.num_batches_for_eval * self.batch_size
for step in xrange(self.num_batches_for_eval):
results = sess.run(fetches)
top_1_accuracy_sum += results['top_1_accuracy']
top_5_accuracy_sum += results['top_5_accuracy']
if (step + 1) % self.params.display_every_for_eval == 0:
log_fn('%i\ttop_1_accuracy: %.4f' % (
step + 1, top_1_accuracy_sum / step))
log_fn('%i\ttop_5_accuracy: %.4f' % (
step + 1, top_5_accuracy_sum / step))
accuracy_at_1 = top_1_accuracy_sum / self.num_batches_for_eval
accuracy_at_5 = top_5_accuracy_sum / self.num_batches_for_eval
log_fn(
'[SUMMARY] Global step: %d Accuracy @ 1 = %.4f Accuracy @ 5 = %.4f [%d examples]' %
(global_step, accuracy_at_1, accuracy_at_5, total_eval_count))
sess.run(dummy_queue.close(cancel_pending_enqueues=True))
coord.request_stop()
def main(_):
# Build benchmark_cnn model
params = benchmark_cnn.make_params_from_flags()
params, sess_config = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
# Print informaton
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
# Build single-GPU benchmark_cnn model
with tf.Graph().as_default() as single_gpu_graph:
bench.build_model()
def run(sess, num_iters, tensor_or_op_name_to_replica_names, num_workers,
worker_id, num_replicas_per_worker):
fetches = {
'global_step':
tensor_or_op_name_to_replica_names[bench.global_step.name][0],
'cost':
tensor_or_op_name_to_replica_names[bench.cost.name][0],
'train_op':
tensor_or_op_name_to_replica_names[bench.train_op.name][0],
}
if isinstance(bench.lr, tf.Tensor):
fetches['lr'] = tensor_or_op_name_to_replica_names[
bench.lr.name][0]
start = time.time()
for i in range(num_iters):
results = sess.run(fetches)
if i % FLAGS.log_frequency == 0:
end = time.time()
throughput = float(FLAGS.log_frequency) / float(end - start)
parallax.log.info(
"global step: %d, lr: %f, loss: %f, "
"throughput: %f steps/sec" %
(results['global_step'], results['lr'] if 'lr' in results
else bench.lr, results['cost'], throughput))
start = time.time()
config = parallax_config.build_config()
config.sess_config = sess_config
parallax.parallel_run(single_gpu_graph,
run,
FLAGS.resource_info_file,
FLAGS.max_steps,
sync=FLAGS.sync,
parallax_config=config)
def main(positional_arguments):
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s'
% positional_arguments[1:])
options = make_options_from_flags(FLAGS)
params = benchmark_cnn.make_params_from_flags()
params = params._replace(batch_size=options.batch_size)
params = params._replace(model='MY_GTSRB')
params = params._replace(num_epochs=options.num_epochs)
params = params._replace(num_gpus=options.num_gpus)
params = params._replace(data_format='NHWC')
params = params._replace(train_dir=options.checkpoint_folder)
params = params._replace(allow_growth=True)
params = params._replace(variable_update='replicated')
params = params._replace(local_parameter_device='gpu')
params = params._replace(use_tf_layers=False)
# params = params._replace(all_reduce_spec='nccl')
# params = params._replace(bottom_file=options.bottom_file)
# params = params._replace(affine_files=options.affine_files)
# params = params._replace(affine_classes=options.affine_classes)
params = params._replace(optimizer=options.optimizer)
params = params._replace(weight_decay=options.weight_decay)
params = params._replace(print_training_accuracy=True)
params = params._replace(backbone_model_path=options.backbone_model_path)
# Summary and Save & load checkpoints.
# params = params._replace(summary_verbosity=1)
# params = params._replace(save_summaries_steps=10)
params = params._replace(save_model_secs=3600) # save every 1 hour
# params = params._replace(save_model_secs=300) #save every 5 min
params = benchmark_cnn.setup(params)
dataset = CifarDataset(options)
model = Model_Builder(options.model_name, dataset.num_classes, options, params)
bench = benchmark_cnn.BenchmarkCNN(params, dataset=dataset, model=model)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
def postprocess(self, results):
"""Postprocess results returned from model."""
try:
import coco_metric # pylint: disable=g-import-not-at-top
except ImportError:
raise ImportError('To use the COCO dataset, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models and tensorflow/models/research to '
'the PYTHONPATH, and compile the protobufs by '
'following https://github.com/tensorflow/models/blob/'
'master/research/object_detection/g3doc/installation.md'
'#protobuf-compilation ; To evaluate using COCO'
'metric, download and install Python COCO API from'
'https://github.com/cocodataset/cocoapi')
pred_boxes = results[ssd_constants.PRED_BOXES]
pred_scores = results[ssd_constants.PRED_SCORES]
# TODO(haoyuzhang): maybe use these values for visualization.
# gt_boxes = results['gt_boxes']
# gt_classes = results['gt_classes']
source_id = results[ssd_constants.SOURCE_ID]
raw_shape = results[ssd_constants.RAW_SHAPE]
for i in range(self.get_batch_size()):
self.predictions[int(source_id[i])] = {
ssd_constants.PRED_BOXES: pred_boxes[i],
ssd_constants.PRED_SCORES: pred_scores[i],
ssd_constants.SOURCE_ID: source_id[i],
ssd_constants.RAW_SHAPE: raw_shape[i]
}
# COCO metric calculates mAP only after a full epoch of evaluation. Return
# dummy results for top_N_accuracy to be compatible with benchmar_cnn.py.
if len(self.predictions) >= ssd_constants.COCO_NUM_VAL_IMAGES:
annotation_file = os.path.join(self.params.data_dir,
ssd_constants.ANNOTATION_FILE)
eval_results = coco_metric.compute_map(self.predictions.values(),
annotation_file)
ret = {'top_1_accuracy': 0., 'top_5_accuracy': 0.}
for metric_key, metric_value in eval_results.items():
ret['simple_value:' + metric_key] = metric_value
return ret
log_fn('Got {:d} out of {:d} eval examples.'
' Waiting for the remaining to calculate mAP...'.format(
len(self.predictions), ssd_constants.COCO_NUM_VAL_IMAGES))
return {'top_1_accuracy': 0., 'top_5_accuracy': 0.}
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s' %
positional_arguments[1:])
params = benchmark_cnn.make_params_from_flags()
params = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
run_benchmark(bench, absl_flags.FLAGS.iters_per_step)
def main(extra_flags):
# extra_flags is a list of command line arguments, excluding those defined
# in tf.flags.FLAGS. extra_flags[0] is always the program name. It is an error
# to supply flags not defined with tf.flags.FLAGS, so we raise an ValueError
# in that case.
assert len(extra_flags) >= 1
if len(extra_flags) > 1:
raise ValueError('Received unknown flags: %s' % extra_flags[1:])
params = benchmark_cnn.make_params_from_flags()
benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError("Received unknown positional arguments: %s" % positional_arguments[1:])
params = benchmark_cnn.make_params_from_flags()
with mlperf.mlperf_logger(absl_flags.FLAGS.ml_perf_compliance_logging, params.model):
params = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn("TensorFlow: %i.%i" % (tfversion[0], tfversion[1]))
bench.print_info()
bench.run()
def main(_):
# Build benchmark_cnn model
params = benchmark_cnn.make_params_from_flags()
params, sess_config = benchmark_cnn.setup(params)
bench = benchmark_cnn.BenchmarkCNN(params)
# Print informaton
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
# Build single-GPU benchmark_cnn model
single_gpu_graph = tf.Graph()
with single_gpu_graph.as_default():
bench.build_model()
config = parallax_config.build_config()
config.sess_config = sess_config
sess, num_workers, worker_id, num_replicas_per_worker = \
parallax.parallel_run(single_gpu_graph,
FLAGS.resource_info_file,
sync=FLAGS.sync,
parallax_config=config)
fetches = {
'global_step': bench.global_step,
'cost': bench.cost,
'train_op': bench.train_op,
}
start = time.time()
for i in range(FLAGS.max_steps):
results = sess.run(fetches)
if (i + 1) % FLAGS.log_frequency == 0:
end = time.time()
throughput = float(FLAGS.log_frequency) / float(end - start)
parallax.log.info(
"global step: %d, loss: %f, throughput: %f steps/sec" %
(results['global_step'][0] + 1, results['cost'][0],
throughput))
start = time.time()
def load_checkpoint(saver, sess, checkpoint_dir, nth_ckpt):
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if nth_ckpt >= len(ckpt.all_model_checkpoint_paths):
raise CheckpointNotFoundException('No more checkpoint file.')
log_fn("Evaluate checkpoint file [%d/%d]" % (
nth_ckpt, len(ckpt.all_model_checkpoint_paths)))
if ckpt and ckpt.all_model_checkpoint_paths[nth_ckpt]:
if os.path.isabs(ckpt.all_model_checkpoint_paths[nth_ckpt]):
model_checkpoint_path = ckpt.all_model_checkpoint_paths[nth_ckpt]
else:
raise ValueError('Checkpoint path should be absolute path.')
global_step = \
ckpt.all_model_checkpoint_paths[nth_ckpt].split('/')[-1].split('-')[-1]
if not global_step.isdigit():
global_step = 0
else:
global_step = int(global_step)
saver.restore(sess, model_checkpoint_path)
log_fn('Successfully loaded model from %s.' % model_checkpoint_path)
return global_step
else:
raise CheckpointNotFoundException('No checkpoint file found.')
def run_graph(benchmark_op, bench_cnn, init_ops, dummy_loss_op):
"""Runs the graph for the benchmark.
Args:
benchmark_op: An op that runs the benchmark.
bench_cnn: The BenchmarkCNN where params and other attributes are obtained.
init_ops: A list of ops that are run before `benchmark_op` for
initialization.
dummy_loss_op: Any op. We must pass a loss op to
`benchmark_cnn.benchmark_one_step`, but the result of the op is never
actually used.
"""
config = benchmark_cnn.create_config_proto(bench_cnn.params)
with tf.Session(config=config) as sess:
for op in init_ops:
sess.run(op)
step_train_times = []
fetches = {'average_loss': dummy_loss_op, 'benchmark_op': benchmark_op}
log_fn('Running warmup')
for i in range(-bench_cnn.num_warmup_batches, bench_cnn.num_batches):
if i == 0:
log_fn('Running all-reduce ops')
start = time.time()
if i > 0 and i % bench_cnn.params.display_every == 0:
log_fn('Iteration: %d. Average time per step so far: %s' %
(i, (time.time() - start) / i))
# Call benchmark_one_step instead of directly calling sess.run(...), to
# potentially get a trace file, partitioned graphs, etc.
benchmark_cnn.benchmark_one_step(
sess=sess,
fetches=fetches,
step=i,
# The batch size is only used for the images/sec calculation, which is
# not actually calculated because we pass show_images_per_sec=False.
batch_size=None,
step_train_times=step_train_times,
trace_filename=bench_cnn.trace_filename,
partitioned_graph_file_prefix=(
bench_cnn.params.partitioned_graph_file_prefix),
profiler=None,
image_producer=None,
params=bench_cnn.params,
show_images_per_sec=False)
log_fn('Average time per step: %s' %
((time.time() - start) / bench_cnn.num_batches))
def postprocess(self, results):
"""Postprocess results returned from model."""
try:
import coco_metric # pylint: disable=g-import-not-at-top
except ImportError:
raise ImportError(
'To use the COCO dataset, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models and tensorflow/models/research to '
'the PYTHONPATH, and compile the protobufs by '
'following https://github.com/tensorflow/models/blob/'
'master/research/object_detection/g3doc/installation.md'
'#protobuf-compilation ; To evaluate using COCO'
'metric, download and install Python COCO API from'
'https://github.com/cocodataset/cocoapi')
pred_boxes = results[ssd_constants.PRED_BOXES]
pred_scores = results[ssd_constants.PRED_SCORES]
# TODO(haoyuzhang): maybe use these values for visualization.
# gt_boxes = results['gt_boxes']
# gt_classes = results['gt_classes']
source_id = results[ssd_constants.SOURCE_ID]
raw_shape = results[ssd_constants.RAW_SHAPE]
# COCO evaluation requires processing COCO_NUM_VAL_IMAGES exactly once. Due
# to rounding errors (i.e., COCO_NUM_VAL_IMAGES % batch_size != 0), setting
# `num_eval_epochs` to 1 is not enough and will often miss some images. We
# expect user to set `num_eval_epochs` to >1, which will leave some unused
# images from previous steps in `predictions`. Here we check if we are doing
# eval at a new global step.
if results['global_step'] > self.eval_global_step:
self.eval_global_step = results['global_step']
self.predictions.clear()
for i, sid in enumerate(source_id):
self.predictions[int(sid)] = {
ssd_constants.PRED_BOXES: pred_boxes[i],
ssd_constants.PRED_SCORES: pred_scores[i],
ssd_constants.SOURCE_ID: source_id[i],
ssd_constants.RAW_SHAPE: raw_shape[i]
}
# COCO metric calculates mAP only after a full epoch of evaluation. Return
# dummy results for top_N_accuracy to be compatible with benchmar_cnn.py.
if len(self.predictions) >= ssd_constants.COCO_NUM_VAL_IMAGES:
log_fn('Got results for all {:d} eval examples. Calculate mAP...'.
format(ssd_constants.COCO_NUM_VAL_IMAGES))
annotation_file = os.path.join(self.params.data_dir,
ssd_constants.ANNOTATION_FILE)
eval_results = coco_metric.compute_map(self.predictions.values(),
annotation_file)
self.predictions.clear()
ret = {'top_1_accuracy': 0., 'top_5_accuracy': 0.}
for metric_key, metric_value in eval_results.items():
ret[constants.SIMPLE_VALUE_RESULT_PREFIX +
metric_key] = metric_value
return ret
log_fn('Got {:d} out of {:d} eval examples.'
' Waiting for the remaining to calculate mAP...'.format(
len(self.predictions), ssd_constants.COCO_NUM_VAL_IMAGES))
return {'top_1_accuracy': 0., 'top_5_accuracy': 0.}
def postprocess(self, results):
"""Postprocess results returned from model."""
try:
import coco_metric # pylint: disable=g-import-not-at-top
except ImportError:
raise ImportError('To use the COCO dataset, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models and tensorflow/models/research to '
'the PYTHONPATH, and compile the protobufs by '
'following https://github.com/tensorflow/models/blob/'
'master/research/object_detection/g3doc/installation.md'
'#protobuf-compilation ; To evaluate using COCO'
'metric, download and install Python COCO API from'
'https://github.com/cocodataset/cocoapi')
pred_boxes = results[ssd_constants.PRED_BOXES]
pred_scores = results[ssd_constants.PRED_SCORES]
# TODO(haoyuzhang): maybe use these values for visualization.
# gt_boxes = results['gt_boxes']
# gt_classes = results['gt_classes']
source_id = results[ssd_constants.SOURCE_ID]
raw_shape = results[ssd_constants.RAW_SHAPE]
# COCO evaluation requires processing COCO_NUM_VAL_IMAGES exactly once. Due
# to rounding errors (i.e., COCO_NUM_VAL_IMAGES % batch_size != 0), setting
# `num_eval_epochs` to 1 is not enough and will often miss some images. We
# expect user to set `num_eval_epochs` to >1, which will leave some unused
# images from previous steps in `predictions`. Here we check if we are doing
# eval at a new global step.
if results['global_step'] > self.eval_global_step:
self.eval_global_step = results['global_step']
self.predictions.clear()
for i, sid in enumerate(source_id):
self.predictions[int(sid)] = {
ssd_constants.PRED_BOXES: pred_boxes[i],
ssd_constants.PRED_SCORES: pred_scores[i],
ssd_constants.SOURCE_ID: source_id[i],
ssd_constants.RAW_SHAPE: raw_shape[i]
}
# COCO metric calculates mAP only after a full epoch of evaluation. Return
# dummy results for top_N_accuracy to be compatible with benchmar_cnn.py.
if len(self.predictions) >= ssd_constants.COCO_NUM_VAL_IMAGES:
log_fn('Got results for all {:d} eval examples. Calculate mAP...'.format(
ssd_constants.COCO_NUM_VAL_IMAGES))
annotation_file = os.path.join(self.params.data_dir,
ssd_constants.ANNOTATION_FILE)
# Size of predictions before decoding about 15--30GB, while size after
# decoding is 100--200MB. When using async eval mode, decoding takes
# 20--30 seconds of main thread time but is necessary to avoid OOM during
# inter-process communication.
decoded_preds = coco_metric.decode_predictions(self.predictions.values())
self.predictions.clear()
if self.params.collect_eval_results_async:
def _eval_results_getter():
"""Iteratively get eval results from async eval process."""
while True:
step, eval_results = self.async_eval_results_queue.get()
self.eval_coco_ap = eval_results['COCO/AP']
mlperf.logger.log_eval_accuracy(
self.eval_coco_ap, step, self.batch_size * self.params.num_gpus,
ssd_constants.COCO_NUM_TRAIN_IMAGES)
if self.reached_target():
# Reached target, clear all pending messages in predictions queue
# and insert poison pill to stop the async eval process.
while not self.async_eval_predictions_queue.empty():
self.async_eval_predictions_queue.get()
self.async_eval_predictions_queue.put('STOP')
break
if not self.async_eval_process:
# Limiting the number of messages in predictions queue to prevent OOM.
# Each message (predictions data) can potentially consume a lot of
# memory, and normally there should only be few messages in the queue.
# If often blocked on this, consider reducing eval frequency.
self.async_eval_predictions_queue = multiprocessing.Queue(2)
self.async_eval_results_queue = multiprocessing.Queue()
# Reason to use a Process as opposed to Thread is mainly the
# computationally intensive eval runner. Python multithreading is not
# truly running in parallel, a runner thread would get significantly
# delayed (or alternatively delay the main thread).
self.async_eval_process = multiprocessing.Process(
target=coco_metric.async_eval_runner,
args=(self.async_eval_predictions_queue,
self.async_eval_results_queue,
annotation_file))
self.async_eval_process.daemon = True
self.async_eval_process.start()
self.async_eval_results_getter_thread = threading.Thread(
target=_eval_results_getter, args=())
self.async_eval_results_getter_thread.daemon = True
self.async_eval_results_getter_thread.start()
self.async_eval_predictions_queue.put(
(self.eval_global_step, decoded_preds))
return {'top_1_accuracy': 0, 'top_5_accuracy': 0.}
eval_results = coco_metric.compute_map(decoded_preds, annotation_file)
self.eval_coco_ap = eval_results['COCO/AP']
ret = {'top_1_accuracy': self.eval_coco_ap, 'top_5_accuracy': 0.}
for metric_key, metric_value in eval_results.items():
ret[constants.SIMPLE_VALUE_RESULT_PREFIX + metric_key] = metric_value
mlperf.logger.log_eval_accuracy(self.eval_coco_ap, self.eval_global_step,
self.batch_size * self.params.num_gpus,
ssd_constants.COCO_NUM_TRAIN_IMAGES)
return ret
log_fn('Got {:d} out of {:d} eval examples.'
' Waiting for the remaining to calculate mAP...'.format(
len(self.predictions), ssd_constants.COCO_NUM_VAL_IMAGES))
return {'top_1_accuracy': self.eval_coco_ap, 'top_5_accuracy': 0.}
def main(positional_arguments):
# Command-line arguments like '--distortions False' are equivalent to
# '--distortions=True False', where False is a positional argument. To prevent
# this from silently running with distortions, we do not allow positional
# arguments.
# For DGX servers use hierarchical_copy=True argument
assert len(positional_arguments) >= 1
if len(positional_arguments) > 1:
raise ValueError('Received unknown positional arguments: %s' %
positional_arguments[1:])
tests_models = [
{
'num_gpus': None,
'batch_size': 64,
'variable_update': 'parameter_server',
'model': 'inception3'
},
{
'num_gpus': None,
'batch_size': 64,
'variable_update': 'parameter_server',
'model': 'resnet50'
},
{
'num_gpus': None,
'batch_size': 32,
'variable_update': 'parameter_server',
'model': 'resnet152'
}, #batch=64 crashes
{
'num_gpus': None,
'batch_size': 64,
'variable_update': 'replicated',
'model': 'vgg16'
},
{
'num_gpus': None,
'batch_size': 512,
'variable_update': 'replicated',
'model': 'alexnet'
}
]
test_gpus = [1, 2, 4, 8]
stats = []
for test in tests_models:
for num_gpus in test_gpus:
test['num_gpus'] = num_gpus
params = benchmark_cnn.make_params_from_flags()
params = benchmark_cnn.setup(params)
# force --hierarchical_copy to False when using 1 GPU
if num_gpus == 1:
params = params._replace(hierarchical_copy=False)
params = params._replace(num_gpus=test['num_gpus'],
batch_size=test['batch_size'],
model=test['model'],
variable_update=test['variable_update'])
bench = benchmark_cnn.BenchmarkCNN(params)
tfversion = cnn_util.tensorflow_version_tuple()
log_fn('TensorFlow: %i.%i' % (tfversion[0], tfversion[1]))
bench.print_info()
results = bench.run()
# result
# {
# 'average_wall_time': 0.6646941304206848,
# 'images_per_sec': 385.1395525908701,
# 'last_average_loss': 7.256145,
# 'num_steps': 100,
# 'num_workers': 1
# }
stats.append({'test': test.copy(), 'result': results})
# summary
print('summary:')
print('==========')
pprint.pprint(stats)
print('==========')
s = ''
for i in range(len(test_gpus)):
for j in range(len(tests_models)):
s += str(stats[i + j * len(test_gpus)]['result']['images_per_sec'])
s += ', '
s += '\n'
print(s)
print('==========')
def print_info(self):
"""Print basic information."""
log_fn('Model: %s' % self.model.get_model())
dataset_name = self.dataset.name
if self.dataset.use_synthetic_gpu_images():
dataset_name += ' (synthetic)'
log_fn('Dataset: %s' % dataset_name)
log_fn('Mode: %s' % get_mode_from_params(self.params))
log_fn('Batch size: %s per device' % self.batch_size)
if self.batch_group_size > 1:
log_fn(' %d batches per prepocessing group' %
self.batch_group_size)
log_fn('Data format: %s' % self.data_format)
log_fn('Optimizer: %s' % self.params.optimizer)
log_fn('==========')
