def __init__(self,
model_name: Text,
logdir: Text,
tensorrt: Text = False,
use_xla: bool = False,
ckpt_path: Text = None,
export_ckpt: Text = None,
saved_model_dir: Text = None,
tflite_path: Text = None,
batch_size: int = 1,
hparams: Text = ''):
self.model_name = model_name
self.logdir = logdir
self.tensorrt = tensorrt
self.use_xla = use_xla
self.ckpt_path = ckpt_path
self.export_ckpt = export_ckpt
self.saved_model_dir = saved_model_dir
self.tflite_path = tflite_path
model_config = hparams_config.get_detection_config(model_name)
model_config.override(hparams) # Add custom overrides
model_config.is_training_bn = False
model_config.image_size = utils.parse_image_size(
model_config.image_size)
# If batch size is 0, then build a graph with dynamic batch size.
self.batch_size = batch_size or None
self.labels_shape = [batch_size, model_config.num_classes]
height, width = model_config.image_size
if model_config.data_format == 'channels_first':
self.inputs_shape = [batch_size, 3, height, width]
else:
self.inputs_shape = [batch_size, height, width, 3]
self.model_config = model_config
def __init__(self,
model_name: Text,
ckpt_path: Text,
image_size: Union[int, Tuple[int, int]] = None,
num_classes: int = None,
enable_ema: bool = True,
data_format: Text = None,
label_id_mapping: Dict[int, Text] = None):
"""Initialize the inference driver.
Args:
model_name: target model name, such as efficientdet-d0.
ckpt_path: checkpoint path, such as /tmp/efficientdet-d0/.
image_size: user specified image size. If None, use the default image size
defined by model_name.
num_classes: number of classes. If None, use the default COCO classes.
enable_ema: whether to enable moving average.
data_format: data format such as 'channel_last'.
label_id_mapping: a dictionary from id to name. If None, use the default
coco_id_mapping (with 90 classes).
"""
self.model_name = model_name
self.ckpt_path = ckpt_path
self.label_id_mapping = label_id_mapping or coco_id_mapping
self.params = hparams_config.get_detection_config(
self.model_name).as_dict()
self.params.update(dict(is_training_bn=False, use_bfloat16=False))
if image_size:
self.params.update(dict(image_size=image_size))
if num_classes:
self.params.update(dict(num_classes=num_classes))
if data_format:
self.params.update(dict(data_format=data_format))
self.disable_pyfun = True
self.enable_ema = enable_ema
def __init__(self,
model_name: Text,
ckpt_path: Text,
image_size: int = None,
label_id_mapping: Dict[int, Text] = None):
"""Initialize the inference driver.
Args:
model_name: target model name, such as efficientdet-d0.
ckpt_path: checkpoint path, such as /tmp/efficientdet-d0/.
image_size: user specified image size. If None, use the default image size
defined by model_name.
label_id_mapping: a dictionary from id to name. If None, use the default
coco_id_mapping (with 90 classes).
"""
self.model_name = model_name
self.ckpt_path = ckpt_path
self.label_id_mapping = label_id_mapping or coco_id_mapping
self.params = hparams_config.get_detection_config(
self.model_name).as_dict()
self.params.update(dict(is_training_bn=False, use_bfloat16=False))
if image_size:
self.params.update(dict(image_size=image_size))
def main(_):
tf.config.run_functions_eagerly(FLAGS.debug)
devices = tf.config.list_physical_devices('GPU')
for device in devices:
tf.config.experimental.set_memory_growth(device, True)
model_config = hparams_config.get_detection_config(FLAGS.model_name)
model_config.override(FLAGS.hparams) # Add custom overrides
model_config.is_training_bn = False
if FLAGS.image_size != -1:
model_config.image_size = FLAGS.image_size
model_config.image_size = utils.parse_image_size(model_config.image_size)
model_params = model_config.as_dict()
ckpt_path_or_file = FLAGS.ckpt_path
if tf.io.gfile.isdir(ckpt_path_or_file):
ckpt_path_or_file = tf.train.latest_checkpoint(ckpt_path_or_file)
driver = inference.ServingDriver(FLAGS.model_name, ckpt_path_or_file,
FLAGS.batch_size or None, model_params)
if FLAGS.mode == 'export':
if not FLAGS.saved_model_dir:
raise ValueError('Please specify --saved_model_dir=')
model_dir = FLAGS.saved_model_dir
if tf.io.gfile.exists(model_dir):
tf.io.gfile.rmtree(model_dir)
driver.export(model_dir, FLAGS.tensorrt, FLAGS.tflite)
print('Model are exported to %s' % model_dir)
elif FLAGS.mode == 'infer':
if FLAGS.saved_model_dir:
driver.load(FLAGS.saved_model_dir)
image_file = tf.io.read_file(FLAGS.input_image)
image_arrays = tf.io.decode_image(image_file)
image_arrays.set_shape((None, None, 3))
image_arrays = tf.expand_dims(image_arrays, axis=0)
detections_bs = driver.serve(image_arrays)
boxes, scores, classes, _ = tf.nest.map_structure(
np.array, detections_bs)
raw_image = Image.open(FLAGS.input_image)
img = driver.visualize(
raw_image,
boxes[0],
classes[0],
scores[0],
min_score_thresh=model_config.nms_configs.score_thresh,
max_boxes_to_draw=model_config.nms_configs.max_output_size)
output_image_path = os.path.join(FLAGS.output_image_dir, '0.jpg')
Image.fromarray(img).save(output_image_path)
print('writing file to %s' % output_image_path)
elif FLAGS.mode == 'benchmark':
if FLAGS.saved_model_dir:
driver.load(FLAGS.saved_model_dir)
batch_size = FLAGS.batch_size or 1
if FLAGS.input_image:
image_file = tf.io.read_file(FLAGS.input_image)
image_arrays = tf.image.decode_image(image_file)
image_arrays.set_shape((None, None, 3))
image_arrays = tf.expand_dims(image_arrays, 0)
if batch_size > 1:
image_arrays = tf.tile(image_arrays, [batch_size, 1, 1, 1])
else:
# use synthetic data if no image is provided.
image_arrays = tf.ones((batch_size, *model_config.image_size, 3),
dtype=tf.uint8)
driver.benchmark(image_arrays, FLAGS.bm_runs, FLAGS.trace_filename)
elif FLAGS.mode == 'dry':
# transfer to tf2 format ckpt
driver.build()
if FLAGS.export_ckpt:
driver.model.save_weights(FLAGS.export_ckpt)
elif FLAGS.mode == 'video':
import cv2 # pylint: disable=g-import-not-at-top
if FLAGS.saved_model_dir:
driver.load(FLAGS.saved_model_dir)
cap = cv2.VideoCapture(FLAGS.input_video)
if not cap.isOpened():
print('Error opening input video: {}'.format(FLAGS.input_video))
out_ptr = None
if FLAGS.output_video:
frame_width, frame_height = int(cap.get(3)), int(cap.get(4))
out_ptr = cv2.VideoWriter(
FLAGS.output_video, cv2.VideoWriter_fourcc('m', 'p', '4', 'v'),
25, (frame_width, frame_height))
while cap.isOpened():
# Capture frame-by-frame
ret, frame = cap.read()
if not ret:
break
raw_frames = np.array([frame])
detections_bs = driver.serve(raw_frames)
boxes, scores, classes, _ = tf.nest.map_structure(
np.array, detections_bs)
new_frame = driver.visualize(
raw_frames[0],
boxes[0],
scores[0],
classes[0],
min_score_thresh=model_config.nms_configs.score_thresh,
max_boxes_to_draw=model_config.nms_configs.max_output_size)
if out_ptr:
# write frame into output file.
out_ptr.write(new_frame)
else:
# show the frame online, mainly used for real-time speed test.
cv2.imshow('Frame', new_frame)
# Press Q on keyboard to exit
if cv2.waitKey(1) & 0xFF == ord('q'):
break
def main(argv):
del argv # Unused.
hvd_try_init()
set_env(use_amp=FLAGS.use_amp)
# Check data path
if FLAGS.mode in (
'train', 'train_and_eval') and FLAGS.training_file_pattern is None:
raise RuntimeError(
'You must specify --training_file_pattern for training.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('You must specify --validation_file_pattern '
'for evaluation.')
if not FLAGS.val_json_file and not FLAGS.testdev_dir:
raise RuntimeError(
'You must specify --val_json_file or --testdev for evaluation.'
)
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError(
'--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
# Spatial partition requires that the to-be-partitioned tensors must have a
# dimension that is a multiple of `partition_dims`. Depending on the
# `partition_dims` and the `image_size` and the `max_level` in config, some
# high-level anchor labels (i.e., `cls_targets` and `box_targets`) cannot
# be partitioned. For example, when `partition_dims` is [1, 4, 2, 1], image
# size is 1536, `max_level` is 9, `cls_targets_8` has a shape of
# [batch_size, 6, 6, 9], which cannot be partitioned (6 % 4 != 0). In this
# case, the level-8 and level-9 target tensors are not partition-able, and
# the highest partition-able level is 7.
image_size = config.get('image_size')
for level in range(config.get('min_level'),
config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(
FLAGS.input_partition_dims)
spatial_dim = image_size // (2**level)
if _can_partition(spatial_dim):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [
FLAGS.input_partition_dims, labels_partition_dims
]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
if hvd is not None:
num_shards = hvd.size()
params = dict(
config.as_dict(),
model_name=FLAGS.model_name,
num_epochs=FLAGS.num_epochs,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_shards=num_shards,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
use_tpu=FLAGS.use_tpu,
backbone_ckpt=FLAGS.backbone_ckpt,
ckpt=FLAGS.ckpt,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
mode=FLAGS.mode,
)
run_config = tf.estimator.RunConfig(
session_config=get_session_config(use_xla=FLAGS.use_xla),
save_checkpoints_steps=600)
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
# TPU Estimator
logging.info(params)
if FLAGS.mode == 'train':
# train_estimator = tf.estimator.tpu.TPUEstimator(
# model_fn=model_fn_instance,
# use_tpu=FLAGS.use_tpu,
# train_batch_size=FLAGS.train_batch_size,
# config=run_config,
# params=params)
params['batch_size'] = FLAGS.train_batch_size
train_estimator = HorovodEstimator(model_fn=model_fn_instance,
model_dir=FLAGS.model_dir,
config=run_config,
params=params)
input_fn = dataloader.InputReader(FLAGS.training_file_pattern,
is_training=True,
params=params,
use_fake_data=FLAGS.use_fake_data)
max_steps = int((FLAGS.num_epochs * FLAGS.num_examples_per_epoch) /
(FLAGS.train_batch_size * num_shards)) + 1
train_estimator.train(input_fn=input_fn, max_steps=max_steps)
# if FLAGS.eval_after_training:
# # Run evaluation after training finishes.
# eval_params = dict(
# params,
# use_tpu=False,
# input_rand_hflip=False,
# is_training_bn=False,
# use_bfloat16=False,
# )
# eval_estimator = tf.estimator.tpu.TPUEstimator(
# model_fn=model_fn_instance,
# use_tpu=False,
# train_batch_size=FLAGS.train_batch_size,
# eval_batch_size=FLAGS.eval_batch_size,
# config=run_config,
# params=eval_params)
# eval_results = eval_estimator.evaluate(
# input_fn=dataloader.InputReader(FLAGS.validation_file_pattern,
# is_training=False),
# steps=FLAGS.eval_samples//FLAGS.eval_batch_size)
# logging.info('Eval results: %s', eval_results)
# ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
# utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
elif FLAGS.mode == 'eval':
config_proto = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
if FLAGS.use_xla and not FLAGS.use_tpu:
config_proto.graph_options.optimizer_options.global_jit_level = (
tf.OptimizerOptions.ON_1)
tpu_config = tf.estimator.tpu.TPUConfig(
FLAGS.iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
input_partition_dims=input_partition_dims,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig.
PER_HOST_V2)
run_config = tf.estimator.tpu.RunConfig(
cluster=None,
evaluation_master=FLAGS.eval_master,
model_dir=FLAGS.model_dir,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
tpu_config=tpu_config,
)
# Eval only runs on CPU or GPU host with batch_size = 1.
# Override the default options: disable randomization in the input pipeline
# and don't run on the TPU.
# Also, disable use_bfloat16 for eval on CPU/GPU.
eval_params = dict(
params,
use_tpu=False,
input_rand_hflip=False,
is_training_bn=False,
use_bfloat16=False,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=False,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
def terminate_eval():
logging.info('Terminating eval after %d seconds of no checkpoints',
FLAGS.eval_timeout)
return True
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout,
timeout_fn=terminate_eval):
logging.info('Starting to evaluate.')
try:
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern, is_training=False),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size)
logging.info('Eval results: %s', eval_results)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
total_step = int(
(FLAGS.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size)
if current_step >= total_step:
logging.info('Evaluation finished after training step %d',
current_step)
break
except tf.errors.NotFoundError:
# Since the coordinator is on a different job than the TPU worker,
# sometimes the TPU worker does not finish initializing until long after
# the CPU job tells it to start evaluating. In this case, the checkpoint
# file could have been deleted already.
logging.info(
'Checkpoint %s no longer exists, skipping checkpoint',
ckpt)
elif FLAGS.mode == 'train_and_eval':
for cycle in range(FLAGS.num_epochs):
logging.info('Starting training cycle, epoch: %d.', cycle)
train_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=FLAGS.use_tpu,
train_batch_size=FLAGS.train_batch_size,
config=run_config,
params=params)
train_estimator.train(input_fn=dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data),
steps=int(FLAGS.num_examples_per_epoch /
FLAGS.train_batch_size))
logging.info('Starting evaluation cycle, epoch: %d.', cycle)
# Run evaluation after every epoch.
eval_params = dict(
params,
use_tpu=False,
input_rand_hflip=False,
is_training_bn=False,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=False,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(FLAGS.validation_file_pattern,
is_training=False),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size)
logging.info('Evaluation results: %s', eval_results)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
pass
else:
logging.info('Mode not found.')
def main(_):
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
if FLAGS.use_xla and FLAGS.strategy != 'tpu':
tf.config.optimizer.set_jit(True)
for gpu in tf.config.list_physical_devices('GPU'):
tf.config.experimental.set_memory_growth(gpu, True)
if FLAGS.debug:
tf.debugging.set_log_device_placement(True)
os.environ['TF_DETERMINISTIC_OPS'] = '1'
tf.random.set_seed(FLAGS.tf_random_seed)
logging.set_verbosity(logging.DEBUG)
if FLAGS.strategy == 'tpu':
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tf.config.experimental_connect_to_cluster(tpu_cluster_resolver)
tf.tpu.experimental.initialize_tpu_system(tpu_cluster_resolver)
ds_strategy = tf.distribute.TPUStrategy(tpu_cluster_resolver)
logging.info('All devices: %s', tf.config.list_logical_devices('TPU'))
elif FLAGS.strategy == 'gpus':
gpus = tf.config.list_physical_devices('GPU')
if FLAGS.batch_size % len(gpus):
raise ValueError(
'Batch size divide gpus number must be interger, but got %f' %
(FLAGS.batch_size / len(gpus)))
if platform.system() == 'Windows':
# Windows doesn't support nccl use HierarchicalCopyAllReduce instead
# TODO(fsx950223): investigate HierarchicalCopyAllReduce performance issue
cross_device_ops = tf.distribute.HierarchicalCopyAllReduce()
else:
cross_device_ops = None
ds_strategy = tf.distribute.MirroredStrategy(
cross_device_ops=cross_device_ops)
logging.info('All devices: %s', gpus)
else:
if tf.config.list_physical_devices('GPU'):
ds_strategy = tf.distribute.OneDeviceStrategy('device:GPU:0')
else:
ds_strategy = tf.distribute.OneDeviceStrategy('device:CPU:0')
steps_per_epoch = FLAGS.num_examples_per_epoch // FLAGS.batch_size
params = dict(profile=FLAGS.profile,
model_name=FLAGS.model_name,
steps_per_execution=FLAGS.steps_per_execution,
model_dir=FLAGS.model_dir,
steps_per_epoch=steps_per_epoch,
strategy=FLAGS.strategy,
batch_size=FLAGS.batch_size,
tf_random_seed=FLAGS.tf_random_seed,
debug=FLAGS.debug,
val_json_file=FLAGS.val_json_file,
eval_samples=FLAGS.eval_samples,
num_shards=ds_strategy.num_replicas_in_sync)
config.override(params, True)
# set mixed precision policy by keras api.
precision = utils.get_precision(config.strategy, config.mixed_precision)
policy = tf.keras.mixed_precision.Policy(precision)
tf.keras.mixed_precision.set_global_policy(policy)
def get_dataset(is_training, config):
file_pattern = (FLAGS.train_file_pattern
if is_training else FLAGS.val_file_pattern)
if not file_pattern:
raise ValueError('No matching files.')
return dataloader.InputReader(
file_pattern,
is_training=is_training,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=config.max_instances_per_image,
debug=FLAGS.debug)(config.as_dict())
with ds_strategy.scope():
if config.model_optimizations:
tfmot.set_config(config.model_optimizations.as_dict())
if FLAGS.hub_module_url:
model = train_lib.EfficientDetNetTrainHub(
config=config, hub_module_url=FLAGS.hub_module_url)
else:
model = train_lib.EfficientDetNetTrain(config=config)
model = setup_model(model, config)
if FLAGS.debug:
tf.config.run_functions_eagerly(True)
if FLAGS.pretrained_ckpt and not FLAGS.hub_module_url:
ckpt_path = tf.train.latest_checkpoint(FLAGS.pretrained_ckpt)
util_keras.restore_ckpt(model,
ckpt_path,
config.moving_average_decay,
exclude_layers=['class_net'])
init_experimental(config)
if 'train' in FLAGS.mode:
val_dataset = get_dataset(False,
config) if 'eval' in FLAGS.mode else None
model.fit(
get_dataset(True, config),
epochs=config.num_epochs,
steps_per_epoch=steps_per_epoch,
callbacks=train_lib.get_callbacks(config.as_dict(),
val_dataset),
validation_data=val_dataset,
validation_steps=(FLAGS.eval_samples // FLAGS.batch_size))
else:
# Continuous eval.
for ckpt in tf.train.checkpoints_iterator(FLAGS.model_dir,
min_interval_secs=180):
logging.info('Starting to evaluate.')
# Terminate eval job when final checkpoint is reached.
try:
current_epoch = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
current_epoch = 0
val_dataset = get_dataset(False, config)
logging.info('start loading model.')
model.load_weights(tf.train.latest_checkpoint(FLAGS.model_dir))
logging.info('finish loading model.')
coco_eval = train_lib.COCOCallback(val_dataset, 1)
coco_eval.set_model(model)
eval_results = coco_eval.on_epoch_end(current_epoch)
logging.info('eval results for %s: %s', ckpt, eval_results)
try:
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
except tf.errors.NotFoundError:
# Checkpoint might be not already deleted by the time eval finished.
logging.info('Checkpoint %s no longer exists, skipping.',
ckpt)
if current_epoch >= config.num_epochs or not current_epoch:
logging.info('Eval epoch %d / %d', current_epoch,
config.num_epochs)
break
def main(_):
if FLAGS.strategy == 'tpu':
tf.disable_eager_execution()
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tpu_grpc_url = tpu_cluster_resolver.get_master()
tf.Session.reset(tpu_grpc_url)
else:
tpu_cluster_resolver = None
# Check data path
if FLAGS.mode in ('train', 'train_and_eval'):
if FLAGS.training_file_pattern is None:
raise RuntimeError(
'Must specify --training_file_pattern for train.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError(
'Must specify --validation_file_pattern for eval.')
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError(
'--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
'image_masks': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
feat_sizes = utils.get_feat_sizes(config.get('image_size'),
config.get('max_level'))
for level in range(config.get('min_level'),
config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(
FLAGS.input_partition_dims)
spatial_dim = feat_sizes[level]
if _can_partition(spatial_dim['height']) and _can_partition(
spatial_dim['width']):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [
FLAGS.input_partition_dims, labels_partition_dims
]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
params = dict(config.as_dict(),
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_shards=num_shards,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
strategy=FLAGS.strategy,
backbone_ckpt=FLAGS.backbone_ckpt,
ckpt=FLAGS.ckpt,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
mode=FLAGS.mode)
config_proto = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
if FLAGS.strategy != 'tpu':
if FLAGS.use_xla:
config_proto.graph_options.optimizer_options.global_jit_level = (
tf.OptimizerOptions.ON_1)
config_proto.gpu_options.allow_growth = True
model_dir = FLAGS.model_dir
strategy = None
if FLAGS.strategy == 'tpu':
tpu_config = tf.estimator.tpu.TPUConfig(
FLAGS.iterations_per_loop if FLAGS.strategy == 'tpu' else 1,
num_cores_per_replica=num_cores_per_replica,
input_partition_dims=input_partition_dims,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig.
PER_HOST_V2)
run_config = tf.estimator.tpu.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=model_dir,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
tpu_config=tpu_config,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
tf_random_seed=FLAGS.tf_random_seed,
)
else:
if FLAGS.strategy == 'gpus':
strategy = tf.distribute.MirroredStrategy()
run_config = tf.estimator.RunConfig(
model_dir=model_dir,
train_distribute=strategy,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
tf_random_seed=FLAGS.tf_random_seed,
)
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
max_instances_per_image = config.max_instances_per_image
eval_steps = int(FLAGS.eval_samples // FLAGS.eval_batch_size)
total_examples = int(config.num_epochs * FLAGS.num_examples_per_epoch)
train_steps = total_examples // FLAGS.train_batch_size
logging.info(params)
with tf.io.gfile.GFile(os.path.join(model_dir, 'config.yaml'), 'w') as f:
f.write(str(config))
train_input_fn = dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image)
eval_input_fn = dataloader.InputReader(
FLAGS.validation_file_pattern,
is_training=False,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image)
if FLAGS.strategy == 'tpu':
estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=params)
else:
params['batch_size'] = (FLAGS.train_batch_size //
getattr(strategy, 'num_replicas_in_sync', 1))
params['num_shards'] = getattr(strategy, 'num_replicas_in_sync', 1)
estimator = tf.estimator.Estimator(model_fn=model_fn_instance,
config=run_config,
params=params)
# start train/eval flow.
if FLAGS.mode == 'train':
estimator.train(input_fn=train_input_fn, max_steps=train_steps)
if FLAGS.eval_after_training:
estimator.evaluate(input_fn=eval_input_fn, steps=eval_steps)
elif FLAGS.mode == 'eval':
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout):
logging.info('Starting to evaluate.')
try:
eval_results = estimator.evaluate(eval_input_fn,
steps=eval_steps)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
if current_step >= train_steps:
logging.info('Eval finished step %d/%d', current_step,
train_steps)
break
except tf.errors.NotFoundError:
# Checkpoint might be not already deleted by the time eval finished.
# We simply skip ssuch case.
logging.info('Checkpoint %s no longer exists, skipping.', ckpt)
elif FLAGS.mode == 'train_and_eval':
train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn,
max_steps=train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_input_fn,
steps=eval_steps,
throttle_secs=600)
tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)
else:
logging.info('Invalid mode: %s', FLAGS.mode)
def main(_):
if FLAGS.strategy == 'horovod':
import horovod.tensorflow as hvd # pylint: disable=g-import-not-at-top
logging.info('Use horovod with multi gpus')
hvd.init()
os.environ['CUDA_VISIBLE_DEVICES'] = str(hvd.local_rank())
import tensorflow.compat.v1 as tf # pylint: disable=g-import-not-at-top
tf.enable_v2_tensorshape()
tf.disable_eager_execution()
if FLAGS.strategy == 'tpu':
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tpu_grpc_url = tpu_cluster_resolver.get_master()
tf.Session.reset(tpu_grpc_url)
else:
tpu_cluster_resolver = None
# Check data path
if FLAGS.mode in (
'train', 'train_and_eval') and FLAGS.training_file_pattern is None:
raise RuntimeError(
'You must specify --training_file_pattern for training.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('You must specify --validation_file_pattern '
'for evaluation.')
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError(
'--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
# Spatial partition requires that the to-be-partitioned tensors must have a
# dimension that is a multiple of `partition_dims`. Depending on the
# `partition_dims` and the `image_size` and the `max_level` in config, some
# high-level anchor labels (i.e., `cls_targets` and `box_targets`) cannot
# be partitioned. For example, when `partition_dims` is [1, 4, 2, 1], image
# size is 1536, `max_level` is 9, `cls_targets_8` has a shape of
# [batch_size, 6, 6, 9], which cannot be partitioned (6 % 4 != 0). In this
# case, the level-8 and level-9 target tensors are not partition-able, and
# the highest partition-able level is 7.
feat_sizes = utils.get_feat_sizes(config.get('image_size'),
config.get('max_level'))
for level in range(config.get('min_level'),
config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(
FLAGS.input_partition_dims)
spatial_dim = feat_sizes[level]
if _can_partition(spatial_dim['height']) and _can_partition(
spatial_dim['width']):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [
FLAGS.input_partition_dims, labels_partition_dims
]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
params = dict(config.as_dict(),
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_shards=num_shards,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
strategy=FLAGS.strategy,
backbone_ckpt=FLAGS.backbone_ckpt,
ckpt=FLAGS.ckpt,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
mode=FLAGS.mode)
config_proto = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
if FLAGS.strategy != 'tpu':
if FLAGS.use_xla:
config_proto.graph_options.optimizer_options.global_jit_level = (
tf.OptimizerOptions.ON_1)
config_proto.gpu_options.allow_growth = True
tpu_config = tf.estimator.tpu.TPUConfig(
FLAGS.iterations_per_loop if FLAGS.strategy == 'tpu' else 1,
num_cores_per_replica=num_cores_per_replica,
input_partition_dims=input_partition_dims,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig.
PER_HOST_V2)
if FLAGS.strategy == 'horovod':
model_dir = FLAGS.model_dir if hvd.rank() == 0 else None
else:
model_dir = FLAGS.model_dir
run_config = tf.estimator.tpu.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=model_dir,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
tpu_config=tpu_config,
tf_random_seed=FLAGS.tf_random_seed,
)
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
max_instances_per_image = config.max_instances_per_image
eval_steps = int(FLAGS.eval_samples // FLAGS.eval_batch_size)
use_tpu = (FLAGS.strategy == 'tpu')
logging.info(params)
def _train(steps):
"""Build train estimator and run training if steps > 0."""
train_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
config=run_config,
params=params)
train_estimator.train(input_fn=dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image),
max_steps=steps)
def _eval(steps):
"""Build estimator and eval the latest checkpoint if steps > 0."""
eval_params = dict(
params,
strategy=FLAGS.strategy,
input_rand_hflip=False,
is_training_bn=False,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
eval_results = eval_estimator.evaluate(input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern,
is_training=False,
max_instances_per_image=max_instances_per_image),
steps=steps,
name=FLAGS.eval_name)
logging.info('Evaluation results: %s', eval_results)
return eval_results
# start train/eval flow.
if FLAGS.mode == 'train':
total_examples = int(config.num_epochs * FLAGS.num_examples_per_epoch)
_train(total_examples // FLAGS.train_batch_size)
if FLAGS.eval_after_training:
_eval(eval_steps)
elif FLAGS.mode == 'eval':
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout):
logging.info('Starting to evaluate.')
try:
eval_results = _eval(eval_steps)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
total_step = int(
(config.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size)
if current_step >= total_step:
logging.info('Evaluation finished after training step %d',
current_step)
break
except tf.errors.NotFoundError:
# Since the coordinator is on a different job than the TPU worker,
# sometimes the TPU worker does not finish initializing until long after
# the CPU job tells it to start evaluating. In this case, the checkpoint
# file could have been deleted already.
logging.info('Checkpoint %s no longer exists, skipping.', ckpt)
elif FLAGS.mode == 'train_and_eval':
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
try:
step = int(os.path.basename(ckpt).split("-")[1])
current_epoch = (step * FLAGS.train_batch_size //
FLAGS.num_examples_per_epoch)
logging.info('found ckpt at step %d (epoch %d)', step,
current_epoch)
except (IndexError, TypeError):
logging.info("Folder has no ckpt with valid step.",
FLAGS.model_dir)
current_epoch = 0
epochs_per_cycle = 1 # higher number has less graph construction overhead.
for e in range(current_epoch + 1, config.num_epochs + 1,
epochs_per_cycle):
print('-----------------------------------------------------\n'
'=====> Starting training, epoch: %d.' % e)
_train(e * FLAGS.num_examples_per_epoch // FLAGS.train_batch_size)
print('-----------------------------------------------------\n'
'=====> Starting evaluation, epoch: %d.' % e)
eval_results = _eval(eval_steps)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
else:
logging.info('Invalid mode: %s', FLAGS.mode)
def main(_):
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
if FLAGS.use_xla and FLAGS.strategy != 'tpu':
tf.config.optimizer.set_jit(True)
for gpu in tf.config.list_physical_devices('GPU'):
tf.config.experimental.set_memory_growth(gpu, True)
if FLAGS.debug:
tf.config.experimental_run_functions_eagerly(True)
tf.debugging.set_log_device_placement(True)
os.environ['TF_DETERMINISTIC_OPS'] = '1'
tf.random.set_seed(FLAGS.tf_random_seed)
logging.set_verbosity(logging.DEBUG)
if FLAGS.strategy == 'tpu':
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tf.config.experimental_connect_to_cluster(tpu_cluster_resolver)
tf.tpu.experimental.initialize_tpu_system(tpu_cluster_resolver)
ds_strategy = tf.distribute.TPUStrategy(tpu_cluster_resolver)
logging.info('All devices: %s', tf.config.list_logical_devices('TPU'))
elif FLAGS.strategy == 'gpus':
ds_strategy = tf.distribute.MirroredStrategy()
logging.info('All devices: %s', tf.config.list_physical_devices('GPU'))
else:
if tf.config.list_physical_devices('GPU'):
ds_strategy = tf.distribute.OneDeviceStrategy('device:GPU:0')
else:
ds_strategy = tf.distribute.OneDeviceStrategy('device:CPU:0')
steps_per_epoch = FLAGS.num_examples_per_epoch // FLAGS.batch_size
params = dict(profile=FLAGS.profile,
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
steps_per_epoch=steps_per_epoch,
strategy=FLAGS.strategy,
batch_size=FLAGS.batch_size,
tf_random_seed=FLAGS.tf_random_seed,
debug=FLAGS.debug,
val_json_file=FLAGS.val_json_file,
eval_samples=FLAGS.eval_samples,
num_shards=ds_strategy.num_replicas_in_sync)
config.override(params, True)
# set mixed precision policy by keras api.
precision = utils.get_precision(config.strategy, config.mixed_precision)
policy = tf.keras.mixed_precision.experimental.Policy(precision)
tf.keras.mixed_precision.experimental.set_policy(policy)
def get_dataset(is_training, config):
file_pattern = (FLAGS.training_file_pattern
if is_training else FLAGS.val_file_pattern)
if not file_pattern:
raise ValueError('No matching files.')
return dataloader.InputReader(
file_pattern,
is_training=is_training,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=config.max_instances_per_image,
debug=FLAGS.debug)(config.as_dict())
with ds_strategy.scope():
if config.model_optimizations:
tfmot.set_config(config.model_optimizations.as_dict())
model = setup_model(config)
if FLAGS.pretrained_ckpt:
ckpt_path = tf.train.latest_checkpoint(FLAGS.pretrained_ckpt)
util_keras.restore_ckpt(model, ckpt_path)
init_experimental(config)
val_dataset = get_dataset(False, config).repeat()
model.fit(get_dataset(True, config),
epochs=config.num_epochs,
steps_per_epoch=steps_per_epoch,
callbacks=train_lib.get_callbacks(config.as_dict(),
val_dataset),
validation_data=val_dataset,
validation_steps=(FLAGS.eval_samples // FLAGS.batch_size))
model.save_weights(os.path.join(FLAGS.model_dir, 'ckpt-final'))
def test_train(self):
tf.random.set_seed(1111)
config = hparams_config.get_detection_config('efficientdet-d0')
config.batch_size = 1
config.num_examples_per_epoch = 1
config.model_dir = tempfile.mkdtemp()
x = tf.ones((1, 512, 512, 3))
labels = {
'box_targets_%d' % i: tf.ones((1, 512 // 2**i, 512 // 2**i, 36))
for i in range(3, 8)
}
labels.update({
'cls_targets_%d' % i: tf.ones((1, 512 // 2**i, 512 // 2**i, 9),
dtype=tf.int32) for i in range(3, 8)
})
labels.update({'mean_num_positives': tf.constant([10.0])})
params = config.as_dict()
params['num_shards'] = 1
model = train_lib.EfficientDetNetTrain(config=config)
model.build((1, 512, 512, 3))
model.compile(
optimizer=train_lib.get_optimizer(params),
loss={
'box_loss':
train_lib.BoxLoss(
params['delta'], reduction=tf.keras.losses.Reduction.NONE),
'box_iou_loss':
train_lib.BoxIouLoss(
params['iou_loss_type'],
reduction=tf.keras.losses.Reduction.NONE),
'class_loss':
train_lib.FocalLoss(
params['alpha'],
params['gamma'],
label_smoothing=params['label_smoothing'],
reduction=tf.keras.losses.Reduction.NONE)
})
# Test single-batch
outputs = model.train_on_batch(x, labels, return_dict=True)
expect_results = {'loss': 26278.25,
'det_loss': 26277.033203125,
'cls_loss': 5060.716796875,
'box_loss': 424.3263244628906,
'box_iou_loss': 0,
'gnorm': 5873.78759765625}
self.assertAllClose(outputs, expect_results, rtol=.1, atol=100.)
outputs = model.test_on_batch(x, labels, return_dict=True)
expect_results = {'loss': 26079.712890625,
'det_loss': 26078.49609375,
'cls_loss': 5063.3759765625,
'box_loss': 420.30242919921875,
'box_iou_loss': 0}
self.assertAllClose(outputs, expect_results, rtol=.1, atol=100.)
# Test fit.
hist = model.fit(
x,
labels,
steps_per_epoch=1,
epochs=1,
callbacks=train_lib.get_callbacks(params))
expect_results = {'loss': [26063.099609375],
'det_loss': [26061.8828125],
'cls_loss': [5058.1337890625],
'box_loss': [420.074951171875],
'box_iou_loss': [0],
'gnorm': [5107.46435546875]}
self.assertAllClose(
hist.history, expect_results, rtol=.1, atol=100.)
def test_train(self):
tf.random.set_seed(1111)
config = hparams_config.get_detection_config('efficientdet-d0')
config.heads = ['object_detection', 'segmentation']
config.batch_size = 1
config.num_examples_per_epoch = 1
config.model_dir = tempfile.mkdtemp()
config.steps_per_epoch = 1
x = tf.ones((1, 512, 512, 3))
labels = {
'box_targets_%d' % i: tf.ones((1, 512 // 2**i, 512 // 2**i, 36))
for i in range(3, 8)
}
labels.update({
'cls_targets_%d' % i: tf.ones((1, 512 // 2**i, 512 // 2**i, 9),
dtype=tf.int32)
for i in range(3, 8)
})
labels.update({'image_masks': tf.ones((1, 128, 128, 1))})
labels.update({'mean_num_positives': tf.constant([10.0])})
params = config.as_dict()
params['num_shards'] = 1
model = train_lib.EfficientDetNetTrain(config=config)
model.build((1, 512, 512, 3))
model.compile(
optimizer=train_lib.get_optimizer(params),
loss={
'box_loss':
train_lib.BoxLoss(params['delta'],
reduction=tf.keras.losses.Reduction.NONE),
'box_iou_loss':
train_lib.BoxIouLoss(params['iou_loss_type'],
params['min_level'],
params['max_level'],
params['num_scales'],
params['aspect_ratios'],
params['anchor_scale'],
params['image_size'],
reduction=tf.keras.losses.Reduction.NONE),
'class_loss':
train_lib.FocalLoss(params['alpha'],
params['gamma'],
label_smoothing=params['label_smoothing'],
reduction=tf.keras.losses.Reduction.NONE),
'seg_loss':
tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
})
# Test single-batch
outputs = model.train_on_batch(x, labels, return_dict=True)
expect_results = {
'loss': [26278.3, 5061.9, 425.5, 1.217],
'det_loss': 26277.033203125,
'cls_loss': 5060.716796875,
'box_loss': 424.3263244628906,
'gnorm': 5873.78759765625,
'seg_loss': 1.2215478420257568,
}
self.assertAllClose(outputs, expect_results, rtol=.1, atol=100.)
outputs = model.test_on_batch(x, labels, return_dict=True)
expect_results = {
'loss': [26278.3, 5061.9, 425.5, 1.217],
'det_loss': 26078.49609375,
'cls_loss': 5063.3759765625,
'box_loss': 420.30242919921875,
'seg_loss': 1.2299377918243408,
}
self.assertAllClose(outputs, expect_results, rtol=.1, atol=100.)
# Test fit.
hist = model.fit(x,
labels,
steps_per_epoch=1,
epochs=1,
callbacks=train_lib.get_callbacks(params))
self.assertAllClose(hist.history['loss'],
[[26067, 5057.5, 421.4, 1.2]],
rtol=.1,
atol=10.)
self.assertAllClose(hist.history['det_loss'], [26061.],
rtol=.1,
atol=10.)
self.assertAllClose(hist.history['cls_loss'], [5058.],
rtol=.1,
atol=10.)
self.assertAllClose(hist.history['box_loss'], [420.],
rtol=.1,
atol=100.)
self.assertAllClose(hist.history['seg_loss'], [1.2299],
rtol=.1,
atol=100.)
def main(_):
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
if FLAGS.use_xla and FLAGS.strategy != 'tpu':
tf.config.optimizer.set_jit(True)
for gpu in tf.config.list_physical_devices('GPU'):
tf.config.experimental.set_memory_growth(gpu, True)
if FLAGS.debug:
tf.config.experimental_run_functions_eagerly(True)
tf.debugging.set_log_device_placement(True)
tf.random.set_seed(111111)
logging.set_verbosity(logging.DEBUG)
if FLAGS.strategy == 'tpu':
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tf.config.experimental_connect_to_cluster(tpu_cluster_resolver)
tf.tpu.experimental.initialize_tpu_system(tpu_cluster_resolver)
ds_strategy = tf.distribute.TPUStrategy(tpu_cluster_resolver)
logging.info('All devices: %s', tf.config.list_logical_devices('TPU'))
elif FLAGS.strategy == 'gpus':
ds_strategy = tf.distribute.MirroredStrategy()
logging.info('All devices: %s', tf.config.list_physical_devices('GPU'))
else:
if tf.config.list_physical_devices('GPU'):
ds_strategy = tf.distribute.OneDeviceStrategy('device:GPU:0')
else:
ds_strategy = tf.distribute.OneDeviceStrategy('device:CPU:0')
# Check data path
if FLAGS.mode in (
'train', 'train_and_eval') and FLAGS.training_file_pattern is None:
raise RuntimeError(
'You must specify --training_file_pattern for training.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('You must specify --validation_file_pattern '
'for evaluation.')
params = dict(config.as_dict(),
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
strategy=FLAGS.strategy,
batch_size=FLAGS.batch_size //
ds_strategy.num_replicas_in_sync,
num_shards=ds_strategy.num_replicas_in_sync,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
mode=FLAGS.mode)
# set mixed precision policy by keras api.
precision = utils.get_precision(params['strategy'],
params['mixed_precision'])
policy = tf.keras.mixed_precision.experimental.Policy(precision)
tf.keras.mixed_precision.experimental.set_policy(policy)
def get_dataset(is_training, params):
file_pattern = (FLAGS.training_file_pattern
if is_training else FLAGS.validation_file_pattern)
return dataloader.InputReader(
file_pattern,
is_training=is_training,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=config.max_instances_per_image)(params)
with ds_strategy.scope():
model = train_lib.EfficientDetNetTrain(params['model_name'], config)
height, width = utils.parse_image_size(params['image_size'])
model.build((params['batch_size'], height, width, 3))
model.compile(
optimizer=train_lib.get_optimizer(params),
loss={
'box_loss':
train_lib.BoxLoss(params['delta'],
reduction=tf.keras.losses.Reduction.NONE),
'box_iou_loss':
train_lib.BoxIouLoss(params['iou_loss_type'],
params['min_level'],
params['max_level'],
params['num_scales'],
params['aspect_ratios'],
params['anchor_scale'],
params['image_size'],
reduction=tf.keras.losses.Reduction.NONE),
'class_loss':
train_lib.FocalLoss(params['alpha'],
params['gamma'],
label_smoothing=params['label_smoothing'],
reduction=tf.keras.losses.Reduction.NONE)
})
ckpt_path = tf.train.latest_checkpoint(FLAGS.model_dir)
if ckpt_path:
model.load_weights(ckpt_path)
model.freeze_vars(params['var_freeze_expr'])
model.fit(get_dataset(True, params=params),
steps_per_epoch=FLAGS.num_examples_per_epoch,
callbacks=train_lib.get_callbacks(params, FLAGS.profile),
validation_data=get_dataset(False, params=params),
validation_steps=FLAGS.eval_samples)
model.save_weights(os.path.join(FLAGS.model_dir, 'model'))
def main(_):
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
if FLAGS.use_xla and FLAGS.strategy != 'tpu':
tf.config.optimizer.set_jit(True)
for gpu in tf.config.list_physical_devices('GPU'):
tf.config.experimental.set_memory_growth(gpu, True)
if FLAGS.debug:
tf.config.experimental_run_functions_eagerly(True)
tf.debugging.set_log_device_placement(True)
tf.random.set_seed(111111)
logging.set_verbosity(logging.DEBUG)
if FLAGS.strategy == 'tpu':
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tf.config.experimental_connect_to_cluster(tpu_cluster_resolver)
tf.tpu.experimental.initialize_tpu_system(tpu_cluster_resolver)
ds_strategy = tf.distribute.TPUStrategy(tpu_cluster_resolver)
logging.info('All devices: %s', tf.config.list_logical_devices('TPU'))
elif FLAGS.strategy == 'gpus':
ds_strategy = tf.distribute.MirroredStrategy()
logging.info('All devices: %s', tf.config.list_physical_devices('GPU'))
else:
if tf.config.list_physical_devices('GPU'):
ds_strategy = tf.distribute.OneDeviceStrategy('device:GPU:0')
else:
ds_strategy = tf.distribute.OneDeviceStrategy('device:CPU:0')
steps_per_epoch = FLAGS.num_examples_per_epoch // FLAGS.batch_size
params = dict(config.as_dict(),
profile=FLAGS.profile,
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
steps_per_epoch=steps_per_epoch,
strategy=FLAGS.strategy,
batch_size=FLAGS.batch_size,
num_shards=ds_strategy.num_replicas_in_sync)
# set mixed precision policy by keras api.
precision = utils.get_precision(params['strategy'],
params['mixed_precision'])
policy = tf.keras.mixed_precision.experimental.Policy(precision)
tf.keras.mixed_precision.experimental.set_policy(policy)
def get_dataset(is_training, params):
file_pattern = (FLAGS.training_file_pattern
if is_training else FLAGS.validation_file_pattern)
if not file_pattern:
raise ValueError('No matching files.')
return dataloader.InputReader(
file_pattern,
is_training=is_training,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=config.max_instances_per_image)(params)
with ds_strategy.scope():
model = train_lib.EfficientDetNetTrain(params['model_name'], config)
model.compile(
optimizer=train_lib.get_optimizer(params),
loss={
'box_loss':
train_lib.BoxLoss(params['delta'],
reduction=tf.keras.losses.Reduction.NONE),
'box_iou_loss':
train_lib.BoxIouLoss(params['iou_loss_type'],
params['min_level'],
params['max_level'],
params['num_scales'],
params['aspect_ratios'],
params['anchor_scale'],
params['image_size'],
reduction=tf.keras.losses.Reduction.NONE),
'class_loss':
train_lib.FocalLoss(params['alpha'],
params['gamma'],
label_smoothing=params['label_smoothing'],
reduction=tf.keras.losses.Reduction.NONE),
'seg_loss':
tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True, reduction=tf.keras.losses.Reduction.NONE)
})
if FLAGS.pretrained_ckpt:
ckpt_path = tf.train.latest_checkpoint(FLAGS.pretrained_ckpt)
util_keras.restore_ckpt(model, ckpt_path,
params['moving_average_decay'])
tf.io.gfile.makedirs(FLAGS.model_dir)
if params['model_optimizations']:
model_optimization.set_config(params['model_optimizations'])
model.build((FLAGS.batch_size, *config.image_size, 3))
model.fit(get_dataset(True, params=params),
epochs=params['num_epochs'],
steps_per_epoch=steps_per_epoch,
callbacks=train_lib.get_callbacks(params),
validation_data=get_dataset(False, params=params).repeat(),
validation_steps=(FLAGS.eval_samples // FLAGS.batch_size))
model.save_weights(os.path.join(FLAGS.model_dir, 'ckpt-final'))
def main(argv):
del argv # Unused.
# if given an efficentdet ckpt don't use default backbone ckpt
if FLAGS.backbone_ckpt == BACKBONE_CKPT_DEFAULT_DIR and FLAGS.ckpt is not None:
print("Using ckpt flag: {}, ignoring default backbone_ckpt: {}".format(
FLAGS.ckpt, FLAGS.backbone_ckpt))
FLAGS.backbone_ckpt = None
if FLAGS.use_horovod is not None:
if FLAGS.dump_all_ranks:
FLAGS.model_dir += "/worker_" + str(hvd.rank())
if not 'HOROVOD_CYCLE_TIME' in os.environ:
os.environ['HOROVOD_CYCLE_TIME'] = '0.5'
if not 'HABANA_HCCL_COMM_API' in os.environ:
os.environ['HABANA_HCCL_COMM_API'] = '0'
hvd_init()
if not FLAGS.no_hpu:
from habana_frameworks.tensorflow import load_habana_module
load_habana_module()
if FLAGS.use_horovod:
assert (horovod_enabled())
set_env(use_amp=FLAGS.use_amp)
# deterministic setting
if FLAGS.sbs_test or FLAGS.deterministic:
set_deterministic()
# Check data path
if FLAGS.mode in (
'train', 'train_and_eval') and FLAGS.training_file_pattern is None:
raise RuntimeError(
'You must specify --training_file_pattern for training.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('You must specify --validation_file_pattern '
'for evaluation.')
if not FLAGS.val_json_file and not FLAGS.testdev_dir:
raise RuntimeError(
'You must specify --val_json_file or --testdev for evaluation.'
)
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError(
'--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
# Spatial partition requires that the to-be-partitioned tensors must have a
# dimension that is a multiple of `partition_dims`. Depending on the
# `partition_dims` and the `image_size` and the `max_level` in config, some
# high-level anchor labels (i.e., `cls_targets` and `box_targets`) cannot
# be partitioned. For example, when `partition_dims` is [1, 4, 2, 1], image
# size is 1536, `max_level` is 9, `cls_targets_8` has a shape of
# [batch_size, 6, 6, 9], which cannot be partitioned (6 % 4 != 0). In this
# case, the level-8 and level-9 target tensors are not partition-able, and
# the highest partition-able level is 7.
image_size = config.get('image_size')
for level in range(config.get('min_level'),
config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(
FLAGS.input_partition_dims)
spatial_dim = image_size // (2**level)
if _can_partition(spatial_dim):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [
FLAGS.input_partition_dims, labels_partition_dims
]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
if horovod_enabled():
num_shards = hvd.size()
else:
num_shards = 1
params = build_estimator_params('train', config, num_shards)
# disabling input data scaling/flip manipulations.
if FLAGS.sbs_test:
sbs_params = dict(input_rand_hflip=False,
train_scale_min=1,
train_scale_max=1,
dropout_rate=0.0)
params.update(sbs_params)
tf_random_seed = 0 if FLAGS.deterministic else None
run_config = build_estimator_config('train', config, num_shards,
num_cores_per_replica,
input_partition_dims)
write_hparams_v1(FLAGS.model_dir, {
'batch_size': FLAGS.train_batch_size,
**FLAGS.flag_values_dict()
})
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
# TPU Estimator
logging.info(params)
if FLAGS.mode == 'train':
train_estimator = HorovodEstimator(model_fn=model_fn_instance,
model_dir=FLAGS.model_dir,
config=run_config,
params=params)
# for deterministic input, we pass to dataloader False for not manipulating input data
is_training = not FLAGS.deterministic
use_fake_data = FLAGS.use_fake_data or FLAGS.deterministic
input_fn = dataloader.InputReader(FLAGS.training_file_pattern,
is_training=is_training,
params=params,
use_fake_data=use_fake_data,
is_deterministic=FLAGS.deterministic)
max_steps = int((FLAGS.num_epochs * FLAGS.num_examples_per_epoch) /
(FLAGS.train_batch_size * num_shards)) + 1
# for sbs test, train under sbs callbacks
if FLAGS.sbs_test:
from TensorFlow.common.debug import dump_callback
SBS_TEST_CONFIG = os.path.join(
os.environ['TF_TESTS_ROOT'],
"tests/tf_training_tests/side_by_side/topologies/efficientdet/dump_config.json"
)
with dump_callback(SBS_TEST_CONFIG):
train_estimator.train(input_fn=input_fn, max_steps=max_steps)
else:
if FLAGS.ckpt is not None:
train_estimator.train(input_fn=input_fn, steps=max_steps)
else:
train_estimator.train(input_fn=input_fn, max_steps=max_steps)
elif FLAGS.mode == 'eval':
eval_params = build_estimator_params('eval', config, num_shards)
eval_config = build_estimator_config('eval', config, num_shards,
num_cores_per_replica,
input_partition_dims)
# Eval only runs on CPU or GPU host with batch_size = 1.
# Override the default options: disable randomization in the input pipeline
# and don't run on the TPU.
# Also, disable use_bfloat16 for eval on CPU/GPU.
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=False,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=eval_config,
params=eval_params)
def terminate_eval():
logging.info('Terminating eval after %d seconds of no checkpoints',
FLAGS.eval_timeout)
return True
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout,
timeout_fn=terminate_eval):
logging.info('Starting to evaluate.')
try:
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern, is_training=False),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size)
logging.info('Eval results: %s', eval_results)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
write_summary(eval_results, ckpt, current_step)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
total_step = int(
(FLAGS.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size)
if current_step >= total_step:
logging.info('Evaluation finished after training step %d',
current_step)
break
except tf.errors.NotFoundError:
# Since the coordinator is on a different job than the TPU worker,
# sometimes the TPU worker does not finish initializing until long after
# the CPU job tells it to start evaluating. In this case, the checkpoint
# file could have been deleted already.
logging.info(
'Checkpoint %s no longer exists, skipping checkpoint',
ckpt)
elif FLAGS.mode == 'train_and_eval':
train_params = build_estimator_params('train', config, num_shards)
train_config = build_estimator_config('train', config, num_shards,
num_cores_per_replica,
input_partition_dims)
train_estimator = HorovodEstimator(model_fn=model_fn_instance,
model_dir=FLAGS.model_dir,
config=train_config,
params=train_params)
eval_estimator = None
for cycle in range(FLAGS.num_epochs):
logging.info('Starting training cycle, epoch: %d.', cycle)
train_estimator.train(
input_fn=dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data),
max_steps=(cycle + 1) *
int(FLAGS.num_examples_per_epoch / FLAGS.train_batch_size))
# synchronization point for all ranks
if horovod_enabled():
hvd.allreduce(tf.constant(0))
logging.info('Starting evaluation cycle, epoch: %d.', cycle)
# Run evaluation after every epoch.
if eval_estimator is None:
eval_params = build_estimator_params('eval', config,
num_shards)
eval_config = build_estimator_config('eval', config,
num_shards,
num_cores_per_replica,
input_partition_dims)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=False,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=eval_config,
params=eval_params)
if is_rank0():
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern, is_training=False),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size)
checkpoint_path = Path(FLAGS.model_dir)
last_ckpt = tf.train.latest_checkpoint(str(checkpoint_path),
latest_filename=None)
current_step = int(os.path.basename(last_ckpt).split('-')[1])
write_summary(eval_results, FLAGS.model_dir, current_step)
logging.info('Evaluation results: %s', eval_results)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
pass
else:
logging.info('Mode not found.')
labels['cls_targets_%d' % level] = cls_targets[level]
labels['box_targets_%d' % level] = box_targets[level]
# Concatenate groundtruth annotations to a tensor.
groundtruth_data = tf.concat([boxes, areas, classes], axis=2)
#labels['source_ids'] = source_ids
labels['groundtruth_data'] = groundtruth_data
labels['image_scales'] = image_scales
return images, labels
dataset = dataset.map(_process_example)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
#if self._use_fake_data:
# Turn this dataset into a semi-fake dataset which always loop at the
# first batch. This reduces variance in performance and is useful in
# testing.
#dataset = dataset.take(1).cache().repeat()
return dataset
if __name__ == '__main__':
Datatset = InputReader(
"F:/automl-master/efficientdet/dataset/tfrecord/-00000-of-00001.tfrecord",
is_training=True,
use_fake_data=False)
config = hparams_config.get_detection_config('efficientdet-d1')
params = dict(config.as_dict())
dataset = Datatset(params)
iter = dataset.make_one_shot_iterator()
images, labels = iter.get_next()
with tf.Session() as sess:
print(sess.run(images))
def main(argv):
assert len(argv) >= 1
if len(argv) > 1: # Do not accept unknown args.
raise ValueError('Received unknown arguments: {}'.format(argv[1:]))
if FLAGS.use_tpu:
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu,
zone=FLAGS.tpu_zone,
project=FLAGS.gcp_project)
tpu_grpc_url = tpu_cluster_resolver.get_master()
tf.Session.reset(tpu_grpc_url)
else:
tpu_cluster_resolver = None
# Check data path
if FLAGS.mode in ('train',
'train_and_eval') and FLAGS.training_file_pattern is None:
raise RuntimeError('You must specify --training_file_pattern for training.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('You must specify --validation_file_pattern '
'for evaluation.')
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError('--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
# Spatial partition requires that the to-be-partitioned tensors must have a
# dimension that is a multiple of `partition_dims`. Depending on the
# `partition_dims` and the `image_size` and the `max_level` in config, some
# high-level anchor labels (i.e., `cls_targets` and `box_targets`) cannot
# be partitioned. For example, when `partition_dims` is [1, 4, 2, 1], image
# size is 1536, `max_level` is 9, `cls_targets_8` has a shape of
# [batch_size, 6, 6, 9], which cannot be partitioned (6 % 4 != 0). In this
# case, the level-8 and level-9 target tensors are not partition-able, and
# the highest partition-able level is 7.
feat_sizes = utils.get_feat_sizes(
config.get('image_size'), config.get('max_level'))
for level in range(config.get('min_level'), config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(FLAGS.input_partition_dims)
spatial_dim = feat_sizes[level]
if _can_partition(spatial_dim['height']) and _can_partition(
spatial_dim['width']):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [
FLAGS.input_partition_dims, labels_partition_dims]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
params = dict(
config.as_dict(),
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_shards=num_shards,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
use_tpu=FLAGS.use_tpu,
backbone_ckpt=FLAGS.backbone_ckpt,
ckpt=FLAGS.ckpt,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
mode=FLAGS.mode,
)
config_proto = tf.ConfigProto(
allow_soft_placement=True, log_device_placement=False)
if FLAGS.use_xla and not FLAGS.use_tpu:
config_proto.graph_options.optimizer_options.global_jit_level = (
tf.OptimizerOptions.ON_1)
tpu_config = tf.estimator.tpu.TPUConfig(
FLAGS.iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
input_partition_dims=input_partition_dims,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig
.PER_HOST_V2)
run_config = tf.estimator.tpu.RunConfig(
cluster=tpu_cluster_resolver,
evaluation_master=FLAGS.eval_master,
model_dir=FLAGS.model_dir,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
tpu_config=tpu_config,
tf_random_seed=FLAGS.tf_random_seed,
)
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
# TPU Estimator
logging.info(params)
if FLAGS.mode == 'train':
train_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=FLAGS.use_tpu,
train_batch_size=FLAGS.train_batch_size,
config=run_config,
params=params)
train_estimator.train(
input_fn=dataloader.InputReader(FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data),
max_steps=int((config.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size))
if FLAGS.eval_after_training:
# Run evaluation after training finishes.
eval_params = dict(
params,
use_tpu=FLAGS.use_tpu,
input_rand_hflip=False,
is_training_bn=False,
precision=None,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=FLAGS.use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(FLAGS.validation_file_pattern,
is_training=False),
steps=FLAGS.eval_samples//FLAGS.eval_batch_size)
logging.info('Eval results: %s', eval_results)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
elif FLAGS.mode == 'eval':
# Eval only runs on CPU or GPU host with batch_size = 1.
# Override the default options: disable randomization in the input pipeline
# and don't run on the TPU.
eval_params = dict(
params,
use_tpu=FLAGS.use_tpu,
input_rand_hflip=False,
is_training_bn=False,
precision=None,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=FLAGS.use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
def terminate_eval():
logging.info('Terminating eval after %d seconds of no checkpoints',
FLAGS.eval_timeout)
return True
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout,
timeout_fn=terminate_eval):
logging.info('Starting to evaluate.')
try:
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(FLAGS.validation_file_pattern,
is_training=False),
steps=FLAGS.eval_samples//FLAGS.eval_batch_size)
logging.info('Eval results: %s', eval_results)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
total_step = int((config.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size)
if current_step >= total_step:
logging.info('Evaluation finished after training step %d',
current_step)
break
except tf.errors.NotFoundError:
# Since the coordinator is on a different job than the TPU worker,
# sometimes the TPU worker does not finish initializing until long after
# the CPU job tells it to start evaluating. In this case, the checkpoint
# file could have been deleted already.
logging.info('Checkpoint %s no longer exists, skipping checkpoint',
ckpt)
elif FLAGS.mode == 'train_and_eval':
# resumeした場合、global_stepにstep数が入っていて、cycleは関係がない
for cycle in range(config.num_epochs):
logging.info('Starting training cycle, epoch: %d.', cycle)
train_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=FLAGS.use_tpu,
train_batch_size=FLAGS.train_batch_size,
config=run_config,
params=params)
train_estimator.train(
input_fn=dataloader.InputReader(FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data),
steps=int(FLAGS.num_examples_per_epoch / FLAGS.train_batch_size))
logging.info('Starting evaluation cycle, epoch: %d.', cycle)
# Run evaluation after every epoch.
eval_params = dict(
params,
use_tpu=FLAGS.use_tpu,
input_rand_hflip=False,
is_training_bn=False,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=FLAGS.use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(FLAGS.validation_file_pattern,
is_training=False),
steps=FLAGS.eval_samples//FLAGS.eval_batch_size)
logging.info('Evaluation results: %s', eval_results)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
logging.info(f'save_checkpoint_start for epoch: {cycle}')
now = datetime.now().strftime('%Y%m%d%H%M%S')
with open('/tmp/main_inner.log', 'a') as f:
f.write(f'{now}: save_checkpoint_start {cycle}\n')
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
with open('/tmp/main_inner.log', 'a') as f:
f.write(f'{now}: save_checkpoint_end {cycle}\n')
logging.info(f'save_checkpoint_end for epoch: {cycle}')
else:
logging.info('Mode not found.')
