def gen_dataset(self,
model_spec,
batch_size=None,
is_training=True,
use_fake_data=False):
"""Generate a batched tf.data.Dataset for training/evaluation.
Args:
model_spec: Specification for the model.
batch_size: A integer, the returned dataset will be batched by this size.
is_training: A boolean, when True, the returned dataset will be optionally
shuffled and repeated as an endless dataset.
use_fake_data: Use fake input.
Returns:
A TF dataset ready to be consumed by Keras model.
"""
reader = det_dataloader.InputReader(
self.tfrecord_file_patten,
is_training=is_training,
use_fake_data=use_fake_data,
max_instances_per_image=model_spec.config.max_instances_per_image,
debug=model_spec.config.debug)
self._dataset = reader(model_spec.config.as_dict(),
batch_size=batch_size)
return self._dataset
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())
def __init__(self, config):
"""Initializes RecordInspect with passed config.
Args:
config: config file to initialize input_fn.
"""
self.input_fn = dataloader.InputReader(
FLAGS.file_pattern,
is_training=not FLAGS.eval,
use_fake_data=False,
max_instances_per_image=config.max_instances_per_image)
self.params = dict(
config.as_dict(), batch_size=FLAGS.samples, model_name=FLAGS.model_name)
logging.info(self.params)
self.cls_to_label = config.label_map
os.makedirs(FLAGS.save_samples_dir, exist_ok=True)
def test_parser(self):
tf.random.set_seed(111111)
params = hparams_config.get_detection_config(
'efficientdet-d0').as_dict()
input_anchors = anchors.Anchors(params['min_level'],
params['max_level'],
params['num_scales'],
params['aspect_ratios'],
params['anchor_scale'],
params['image_size'])
anchor_labeler = anchors.AnchorLabeler(input_anchors,
params['num_classes'])
example_decoder = tf_example_decoder.TfExampleDecoder(
regenerate_source_id=params['regenerate_source_id'])
tfrecord_path = test_util.make_fake_tfrecord(self.get_temp_dir())
dataset = tf.data.TFRecordDataset([tfrecord_path])
value = next(iter(dataset))
reader = dataloader.InputReader(tfrecord_path, True)
result = reader.dataset_parser(value, example_decoder, anchor_labeler,
params)
self.assertEqual(len(result), 11)
def main(_):
config = hparams_config.get_efficientdet_config(FLAGS.model_name)
config.override(FLAGS.hparams)
config.val_json_file = FLAGS.val_json_file
config.nms_configs.max_nms_inputs = anchors.MAX_DETECTION_POINTS
config.drop_remainder = False # eval all examples w/o drop.
config.image_size = utils.parse_image_size(config['image_size'])
# Evaluator for AP calculation.
label_map = label_util.get_label_map(config.label_map)
evaluator = coco_metric.EvaluationMetric(filename=config.val_json_file,
label_map=label_map)
# dataset
batch_size = 1
ds = dataloader.InputReader(
FLAGS.val_file_pattern,
is_training=False,
max_instances_per_image=config.max_instances_per_image)(
config, batch_size=batch_size)
eval_samples = FLAGS.eval_samples
if eval_samples:
ds = ds.take((eval_samples + batch_size - 1) // batch_size)
# Network
lite_runner = LiteRunner(FLAGS.tflite_path, FLAGS.only_network)
eval_samples = FLAGS.eval_samples or 5000
pbar = tf.keras.utils.Progbar(
(eval_samples + batch_size - 1) // batch_size)
for i, (images, labels) in enumerate(ds):
if not FLAGS.only_network:
nms_boxes_bs, nms_classes_bs, nms_scores_bs, _ = lite_runner.run(
images)
nms_classes_bs += postprocess.CLASS_OFFSET
height, width = utils.parse_image_size(config.image_size)
normalize_factor = tf.constant([height, width, height, width],
dtype=tf.float32)
nms_boxes_bs *= normalize_factor
if labels['image_scales'] is not None:
scales = tf.expand_dims(
tf.expand_dims(labels['image_scales'], -1), -1)
nms_boxes_bs = nms_boxes_bs * tf.cast(scales,
nms_boxes_bs.dtype)
detections = postprocess.generate_detections_from_nms_output(
nms_boxes_bs, nms_classes_bs, nms_scores_bs,
labels['source_ids'])
else:
cls_outputs, box_outputs = lite_runner.run(images)
detections = postprocess.generate_detections(
config,
cls_outputs,
box_outputs,
labels['image_scales'],
labels['source_ids'],
pre_class_nms=FLAGS.pre_class_nms)
detections = postprocess.transform_detections(detections)
evaluator.update_state(labels['groundtruth_data'].numpy(),
detections.numpy())
pbar.update(i)
# compute the final eval results.
metrics = evaluator.result()
metric_dict = {}
for i, name in enumerate(evaluator.metric_names):
metric_dict[name] = metrics[i]
if label_map:
for i, cid in enumerate(sorted(label_map.keys())):
name = 'AP_/%s' % label_map[cid]
metric_dict[name] = metrics[i + len(evaluator.metric_names)]
print(FLAGS.model_name, metric_dict)
def main(_):
config = hparams_config.get_efficientdet_config(FLAGS.model_name)
config.override(FLAGS.hparams)
config.val_json_file = FLAGS.val_json_file
config.nms_configs.max_nms_inputs = anchors.MAX_DETECTION_POINTS
config.drop_remainder = False # eval all examples w/o drop.
config.image_size = utils.parse_image_size(config['image_size'])
if config.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 config.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')
with ds_strategy.scope():
# Network
model = efficientdet_keras.EfficientDetNet(config=config)
model.build((None, *config.image_size, 3))
util_keras.restore_ckpt(model,
tf.train.latest_checkpoint(FLAGS.model_dir),
config.moving_average_decay,
skip_mismatch=False)
@tf.function
def model_fn(images, labels):
cls_outputs, box_outputs = model(images, training=False)
detections = postprocess.generate_detections(
config, cls_outputs, box_outputs, labels['image_scales'],
labels['source_ids'])
tf.numpy_function(evaluator.update_state, [
labels['groundtruth_data'],
postprocess.transform_detections(detections)
], [])
# Evaluator for AP calculation.
label_map = label_util.get_label_map(config.label_map)
evaluator = coco_metric.EvaluationMetric(filename=config.val_json_file,
label_map=label_map)
# dataset
batch_size = FLAGS.batch_size # global batch size.
ds = dataloader.InputReader(
FLAGS.val_file_pattern,
is_training=False,
max_instances_per_image=config.max_instances_per_image)(
config, batch_size=batch_size)
if FLAGS.eval_samples:
ds = ds.take((FLAGS.eval_samples + batch_size - 1) // batch_size)
ds = ds_strategy.experimental_distribute_dataset(ds)
# evaluate all images.
eval_samples = FLAGS.eval_samples or 5000
pbar = tf.keras.utils.Progbar(
(eval_samples + batch_size - 1) // batch_size)
for i, (images, labels) in enumerate(ds):
ds_strategy.run(model_fn, (images, labels))
pbar.update(i)
# compute the final eval results.
metrics = evaluator.result()
metric_dict = {}
for i, name in enumerate(evaluator.metric_names):
metric_dict[name] = metrics[i]
if label_map:
for i, cid in enumerate(sorted(label_map.keys())):
name = 'AP_/%s' % label_map[cid]
metric_dict[name] = metrics[i + len(evaluator.metric_names)]
print(FLAGS.model_name, metric_dict)
def main(_):
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'):
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.reset_default_graph()
run_train_and_eval(e)
else:
logging.info('Invalid mode: %s', FLAGS.mode)
def export(self,
output_dir: Text = None,
tensorrt: Text = None,
tflite: Text = None,
file_pattern: Text = None,
num_calibration_steps: int = 2000):
"""Export a saved model, frozen graph, and potential tflite/tensorrt model.
Args:
output_dir: the output folder for saved model.
tensorrt: If not None, must be {'FP32', 'FP16', 'INT8'}.
tflite: Type for post-training quantization.
file_pattern: Glob for tfrecords, e.g. coco/val-*.tfrecord.
num_calibration_steps: Number of post-training quantization calibration
steps to run.
"""
export_model, input_spec = self._get_model_and_spec(tflite)
image_size = utils.parse_image_size(self.params['image_size'])
if output_dir:
tf.saved_model.save(
export_model,
output_dir,
signatures=export_model.__call__.get_concrete_function(
input_spec))
logging.info('Model saved at %s', output_dir)
# also save freeze pb file.
graphdef = self.freeze(
export_model.__call__.get_concrete_function(input_spec))
proto_path = tf.io.write_graph(graphdef,
output_dir,
self.model_name + '_frozen.pb',
as_text=False)
logging.info('Frozen graph saved at %s', proto_path)
if tflite:
shape = (self.batch_size, *image_size, 3)
input_spec = tf.TensorSpec(shape=shape,
dtype=input_spec.dtype,
name=input_spec.name)
# from_saved_model supports advanced converter features like op fusing.
converter = tf.lite.TFLiteConverter.from_saved_model(output_dir)
if tflite == 'FP32':
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.float32]
elif tflite == 'FP16':
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.float16]
elif tflite == 'INT8':
# Enables MLIR-based post-training quantization.
converter.experimental_new_quantizer = True
if file_pattern:
config = hparams_config.get_efficientdet_config(
self.model_name)
config.override(self.params)
ds = dataloader.InputReader(file_pattern,
is_training=False,
max_instances_per_image=config.
max_instances_per_image)(
config,
batch_size=self.batch_size)
def representative_dataset_gen():
for image, _ in ds.take(num_calibration_steps):
yield [image]
else: # Used for debugging, can remove later.
logging.warn(
'Use real representative dataset instead of fake ones.'
)
num_calibration_steps = 10
def representative_dataset_gen(
): # rewrite this for real data.
for _ in range(num_calibration_steps):
yield [tf.ones(shape, dtype=input_spec.dtype)]
converter.representative_dataset = representative_dataset_gen
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.inference_input_type = tf.uint8
# TFLite's custom NMS op isn't supported by post-training quant,
# so we add TFLITE_BUILTINS as well.
supported_ops = [
tf.lite.OpsSet.TFLITE_BUILTINS_INT8,
tf.lite.OpsSet.TFLITE_BUILTINS
]
converter.target_spec.supported_ops = supported_ops
else:
raise ValueError(
f'Invalid tflite {tflite}: must be FP32, FP16, INT8.')
tflite_path = os.path.join(output_dir, tflite.lower() + '.tflite')
tflite_model = converter.convert()
tf.io.gfile.GFile(tflite_path, 'wb').write(tflite_model)
logging.info('TFLite is saved at %s', tflite_path)
if tensorrt:
trt_path = os.path.join(output_dir, 'tensorrt_' + tensorrt.lower())
conversion_params = tf.experimental.tensorrt.ConversionParams(
max_workspace_size_bytes=(2 << 20),
maximum_cached_engines=1,
precision_mode=tensorrt.upper())
converter = tf.experimental.tensorrt.Converter(
output_dir, conversion_params=conversion_params)
converter.convert()
converter.save(trt_path)
logging.info('TensorRT model is saved at %s', trt_path)