def __init__(self, min_level, max_level, num_scales, aspect_ratios,
anchor_scale, image_size):
"""Constructs multiscale anchors.
Args:
min_level: integer number of minimum level of the output feature pyramid.
max_level: integer number of maximum level of the output feature pyramid.
num_scales: integer number representing intermediate scales added
on each level. For instances, num_scales=2 adds two additional
anchor scales [2^0, 2^0.5] on each level.
aspect_ratios: list of representing the aspect ratio anchors added
on each level. For instances, aspect_ratios = [1.0, 2.0, 0..5]
adds three anchors on each level.
anchor_scale: float number representing the scale of size of the base
anchor to the feature stride 2^level. Or a list, one value per layer.
image_size: integer number or tuple of integer number of input image size.
"""
self.min_level = min_level
self.max_level = max_level
self.num_scales = num_scales
self.aspect_ratios = aspect_ratios
if isinstance(anchor_scale, (list, tuple)):
assert len(anchor_scale) == max_level - min_level + 1
self.anchor_scales = anchor_scale
else:
self.anchor_scales = [anchor_scale] * (max_level - min_level + 1)
self.image_size = utils.parse_image_size(image_size)
self.feat_sizes = utils.get_feat_sizes(image_size, max_level)
self.config = self._generate_configs()
self.boxes = self._generate_boxes()
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.train_file_pattern is None:
raise RuntimeError('Must specify --train_file_pattern for train.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.val_file_pattern is None:
raise RuntimeError('Must specify --val_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,
profile=FLAGS.profile,
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
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
max_instances_per_image = config.max_instances_per_image
if FLAGS.eval_samples:
eval_steps = int((FLAGS.eval_samples + FLAGS.eval_batch_size - 1) //
FLAGS.eval_batch_size)
else:
eval_steps = None
total_examples = int(config.num_epochs * FLAGS.num_examples_per_epoch)
train_steps = total_examples // FLAGS.train_batch_size
logging.info(params)
if not tf.io.gfile.exists(model_dir):
tf.io.gfile.makedirs(model_dir)
config_file = os.path.join(model_dir, 'config.yaml')
if not tf.io.gfile.exists(config_file):
tf.io.gfile.GFile(config_file, 'w').write(str(config))
train_input_fn = dataloader.InputReader(
FLAGS.train_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.val_file_pattern,
is_training=False,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image)
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,
)
# TPUEstimator can do both train and eval.
train_est = 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)
eval_est = train_est
else:
strategy = None
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,
)
def get_estimator(global_batch_size):
params['num_shards'] = getattr(strategy, 'num_replicas_in_sync', 1)
params['batch_size'] = global_batch_size // params['num_shards']
return tf.estimator.Estimator(model_fn=model_fn_instance,
config=run_config,
params=params)
# train and eval need different estimator due to different batch size.
train_est = get_estimator(FLAGS.train_batch_size)
eval_est = get_estimator(FLAGS.eval_batch_size)
# start train/eval flow.
if FLAGS.mode == 'train':
train_est.train(input_fn=train_input_fn, max_steps=train_steps)
if FLAGS.eval_after_train:
eval_est.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 = eval_est.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':
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 %s has no ckpt with valid step.',
FLAGS.model_dir)
current_epoch = 0
def run_train_and_eval(e):
print('\n =====> Starting training, epoch: %d.' % e)
train_est.train(input_fn=train_input_fn,
max_steps=e * FLAGS.num_examples_per_epoch //
FLAGS.train_batch_size)
print('\n =====> Starting evaluation, epoch: %d.' % e)
eval_results = eval_est.evaluate(input_fn=eval_input_fn,
steps=eval_steps)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
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):
if FLAGS.run_epoch_in_child_process:
p = multiprocessing.Process(target=run_train_and_eval,
args=(e, ))
p.start()
p.join()
if p.exitcode != 0:
return p.exitcode
else:
tf.compat.v1.reset_default_graph()
run_train_and_eval(e)
else:
logging.info('Invalid mode: %s', FLAGS.mode)
def build_feature_network(features, config):
"""Build FPN input features.
Args:
features: input tensor.
config: a dict-like config, including all parameters.
Returns:
A dict from levels to the feature maps processed after feature network.
"""
feat_sizes = utils.get_feat_sizes(config.image_size, config.max_level)
feats = []
if config.min_level not in features.keys():
raise ValueError(
'features.keys ({}) should include min_level ({})'.format(
features.keys(), config.min_level))
# Build additional input features that are not from backbone.
for level in range(config.min_level, config.max_level + 1):
if level in features.keys():
feats.append(features[level])
else:
h_id, w_id = (2,
3) if config.data_format == 'channels_first' else (1,
2)
# Adds a coarser level by downsampling the last feature map.
feats.append(
resample_feature_map(
feats[-1],
name='p%d' % level,
target_height=(feats[-1].shape[h_id] - 1) // 2 + 1,
target_width=(feats[-1].shape[w_id] - 1) // 2 + 1,
target_num_channels=config.fpn_num_filters,
apply_bn=config.apply_bn_for_resampling,
is_training=config.is_training_bn,
conv_after_downsample=config.conv_after_downsample,
strategy=config.strategy,
data_format=config.data_format,
batch_norm_trainable=config.batch_norm_trainable))
utils.verify_feats_size(feats,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
with tf.variable_scope('fpn_cells'):
for rep in range(config.fpn_cell_repeats):
with tf.variable_scope('cell_{}'.format(rep)):
logging.info('building cell %d', rep)
new_feats = build_bifpn_layer(feats, feat_sizes, config)
feats = [
new_feats[level]
for level in range(config.min_level, config.max_level + 1)
]
utils.verify_feats_size(feats,
feat_sizes=feat_sizes,
min_level=config.min_level,
max_level=config.max_level,
data_format=config.data_format)
return new_feats