def Proc(record):
"""Parses a serialized tf.Example record."""
features = [
('uttid', tf.VarLenFeature(tf.string)),
('transcript', tf.VarLenFeature(tf.string)),
('frames', tf.VarLenFeature(tf.float32)),
]
example = tf.parse_single_example(record, dict(features))
fval = {k: v.values for k, v in six.iteritems(example)}
# Reshape the flattened vector into its original time-major
# representation.
fval['frames'] = tf.reshape(fval['frames'],
shape=[-1, self.params.frame_size])
# Input duration determines the bucket.
bucket_key = tf.cast(tf.shape(fval['frames'])[0], tf.int32)
if self.params.append_eos_frame:
bucket_key += 1
tgt_ids, tgt_labels, tgt_paddings = self.StringsToIds(
fval['transcript'])
src_paddings = tf.zeros([tf.shape(fval['frames'])[0]],
dtype=tf.float32)
return [
fval['uttid'], tgt_ids, tgt_labels, tgt_paddings,
fval['frames'], src_paddings
], bucket_key
def Proc(record):
"""Parses a serialized tf.Example record."""
outputs = [
('inputs', tf.VarLenFeature(tf.int64)),
('targets', tf.VarLenFeature(tf.int64)),
]
features = tf.parse_single_example(record, dict(outputs))
for k, v in six.iteritems(features):
features[k] = v.values
src_ids = features['inputs']
tgt_labels = features['targets']
# Derive src_paddings, tgt_ids, tgt_paddings.
# tgt_ids is tgt_labels shifted right by one, with a SOS ID prepended.
tgt_ids = tf.concat([[p.sos_id], tgt_labels[:-1]], axis=0)
src_paddings = tf.zeros(tf.shape(src_ids), dtype=tf.float32)
tgt_paddings = tf.zeros(tf.shape(tgt_ids), dtype=tf.float32)
tgt_weights = tf.ones(tf.shape(tgt_ids), dtype=tf.float32)
bucket_key = tf.cast(
tf.maximum(
tf.reduce_sum(1.0 - src_paddings),
tf.reduce_sum(1.0 - tgt_paddings)), tf.int32)
return [
src_ids, src_paddings, tgt_ids, tgt_paddings, tgt_labels, tgt_weights
], bucket_key
def FeatureMap(self):
"""Return a dictionary from tf.Example feature names to Features."""
p = self.params
feature_map = {}
for laser in p.cbr_laser_names + p.gbr_laser_names:
feature_map['%s_beam_inclinations' %
laser] = (tf.VarLenFeature(dtype=tf.float32))
feature_map['%s_beam_inclination_min' %
laser] = (tf.VarLenFeature(dtype=tf.float32))
feature_map['%s_beam_inclination_max' %
laser] = (tf.VarLenFeature(dtype=tf.float32))
feature_map['%s_extrinsics' %
laser] = tf.VarLenFeature(dtype=tf.float32)
if laser in p.gbr_laser_names:
feature_map['%s_pose' %
laser] = tf.VarLenFeature(dtype=tf.float32)
for returns in p.returns:
feature_map['%s_%s' % (laser, returns)] = tf.VarLenFeature(
dtype=tf.float32)
feature_map['%s_%s_shape' %
(laser, returns)] = tf.VarLenFeature(
dtype=tf.int64)
feature_map['pose'] = tf.VarLenFeature(dtype=tf.float32)
return feature_map
def FeatureMap(self):
"""Return a dictionary from tf.Example feature names to Features."""
p = self.params
features = {}
features['pose'] = tf.VarLenFeature(dtype=tf.float32)
for camera_name in p.camera_names:
features['image_%s' % camera_name] = tf.VarLenFeature(dtype=tf.string)
features['image_%s_shape' % camera_name] = (
tf.VarLenFeature(dtype=tf.int64))
features['camera_%s_intrinsics' %
camera_name] = tf.VarLenFeature(dtype=tf.float32)
features['camera_%s_extrinsics' %
camera_name] = tf.VarLenFeature(dtype=tf.float32)
features['camera_%s_rolling_shutter_direction' %
camera_name] = tf.FixedLenFeature(
dtype=tf.int64, shape=())
features['image_%s_pose' %
camera_name] = tf.VarLenFeature(dtype=tf.float32)
features['image_%s_velocity' %
camera_name] = tf.VarLenFeature(dtype=tf.float32)
for feat in [
'pose_timestamp', 'shutter', 'camera_trigger_time',
'camera_readout_done_time'
]:
features['image_%s_%s' % (camera_name, feat)] = tf.FixedLenFeature(
dtype=tf.float32, shape=())
return features
def FeatureMap(self):
"""Return a dictionary from tf.Example feature names to Features."""
p = self.params
features = {}
for lidar in p.lidar_names:
for ri in ['ri1', 'ri2']:
features['laser_%s_%s' %
(lidar, ri)] = tf.VarLenFeature(dtype=tf.float32)
return features
def FeatureMap(self):
"""Return a dictionary from tf.Example feature names to Features."""
feature_map = {}
feature_map['pose'] = tf.VarLenFeature(dtype=tf.float32)
feature_map['run_segment'] = tf.FixedLenFeature((), tf.string, '')
feature_map['run_start_offset'] = tf.FixedLenFeature((), tf.int64, 0)
feature_map['time_of_day'] = tf.FixedLenFeature((), tf.string, '')
feature_map['location'] = tf.FixedLenFeature((), tf.string, '')
feature_map['weather'] = tf.FixedLenFeature((), tf.string, '')
return feature_map
def _Proc(record):
"""Parses a serialized tf.Example record."""
outputs = [
('inputs', tf.VarLenFeature(tf.int64)),
('targets', tf.VarLenFeature(tf.int64)),
]
features = tf.parse_single_example(record, dict(outputs))
for k, v in six.iteritems(features):
features[k] = v.values
src_ids = features['inputs']
tgt_labels = features['targets']
src_paddings, tgt_ids, tgt_paddings, tgt_weights, bucket_key = _DerivePaddingsAndIds(
src_ids, tgt_labels)
return [
src_ids, src_paddings, tgt_ids, tgt_paddings, tgt_labels,
tgt_weights
], bucket_key
def FeatureMap(self):
"""Return a dictionary from tf.Example feature names to Features."""
feature_map = {}
feature_map['labels'] = tf.VarLenFeature(dtype=tf.int64)
feature_map['label_ids'] = tf.VarLenFeature(dtype=tf.string)
feature_map['detection_difficulties'] = tf.VarLenFeature(dtype=tf.int64)
feature_map['combined_detection_difficulties'] = tf.VarLenFeature(
dtype=tf.int64)
feature_map['tracking_difficulties'] = tf.VarLenFeature(dtype=tf.int64)
feature_map['bboxes_3d'] = tf.VarLenFeature(dtype=tf.float32)
feature_map['bboxes_3d_num_points'] = tf.VarLenFeature(dtype=tf.int64)
feature_map['label_metadata'] = tf.VarLenFeature(dtype=tf.float32)
return feature_map
def _Proc(record):
"""Parses a serialized tf.Example record."""
outputs = [
('inputs', tf.VarLenFeature(tf.int64)),
('targets', tf.VarLenFeature(tf.int64)),
]
features = tf.parse_single_example(record, dict(outputs))
for k, v in six.iteritems(features):
features[k] = v.values
src_ids = features['inputs']
tgt_labels = features['targets']
# Derive trivial segmentation for unpacked input.
src_paddings, tgt_ids, tgt_paddings, tgt_weights, bucket_key = _DerivePaddingsAndIds(
src_ids, tgt_labels)
src_len = tf.shape(src_ids)[0]
tgt_len = tf.shape(tgt_ids)[0]
src_pos = tf.range(src_len, dtype=tf.int32)
src_seg = tf.zeros_like(src_paddings)
tgt_pos = tf.range(tgt_len, dtype=tf.int32)
tgt_seg = tf.zeros_like(tgt_paddings)
return [
src_ids,
src_paddings,
tgt_ids,
tgt_paddings,
tgt_labels,
tgt_weights,
src_pos,
src_seg,
tgt_pos,
tgt_seg,
], bucket_key
def Proc(record):
"""Parses a serialized tf.Example record."""
outputs = [
('source_id', tf.VarLenFeature(tf.int64)),
('source_padding', tf.VarLenFeature(tf.float32)),
('target_id', tf.VarLenFeature(tf.int64)),
('target_padding', tf.VarLenFeature(tf.float32)),
('target_label', tf.VarLenFeature(tf.int64)),
('target_weight', tf.VarLenFeature(tf.float32)),
]
features = tf.parse_single_example(record, dict(outputs))
for k, v in six.iteritems(features):
features[k] = v.values
bucket_key = tf.to_int32(
tf.maximum(tf.reduce_sum(1.0 - features['source_padding']),
tf.reduce_sum(1.0 - features['target_padding'])))
return [features[k] for k, _ in outputs] + [bucket_key]
def _ProcPacked(record):
"""TFExample -> Tensors for PackedInput."""
outputs = [
('inputs', tf.VarLenFeature(tf.int64)),
('targets', tf.VarLenFeature(tf.int64)),
('inputs_segmentation', tf.VarLenFeature(tf.int64)),
('inputs_position', tf.VarLenFeature(tf.int64)),
('targets_segmentation', tf.VarLenFeature(tf.int64)),
('targets_position', tf.VarLenFeature(tf.int64)),
]
features = tf.parse_single_example(record, dict(outputs))
for k, v in six.iteritems(features):
features[k] = v.values
src_ids = features['inputs']
tgt_labels = features['targets']
src_pos = features['inputs_position']
src_seg = features['inputs_segmentation']
tgt_pos = features['targets_position']
tgt_seg = features['targets_segmentation']
src_paddings, tgt_ids, tgt_paddings, tgt_weights, bucket_key = _DerivePaddingsAndIds(
src_ids, tgt_labels)
return [
src_ids,
src_paddings,
tgt_ids,
tgt_paddings,
tgt_labels,
tgt_weights,
src_pos,
src_seg,
tgt_pos,
tgt_seg,
], bucket_key
def FeatureMap(self):
feature_map = {
'pointcloud/xyz': tf.VarLenFeature(dtype=tf.float32),
'pointcloud/reflectance': tf.VarLenFeature(dtype=tf.float32),
}
return feature_map
def FeatureMap(self):
return {
'image/source_id':
tf.FixedLenFeature((), tf.string, ''),
'object/image/bbox/xmin':
tf.VarLenFeature(tf.float32),
'object/image/bbox/xmax':
tf.VarLenFeature(tf.float32),
'object/image/bbox/ymin':
tf.VarLenFeature(tf.float32),
'object/image/bbox/ymax':
tf.VarLenFeature(tf.float32),
'object/label':
tf.VarLenFeature(tf.string),
'object/has_3d_info':
tf.VarLenFeature(dtype=tf.int64),
'object/occlusion':
tf.VarLenFeature(dtype=tf.int64),
'object/truncation':
tf.VarLenFeature(dtype=tf.float32),
'object/velo/bbox/xyz':
tf.VarLenFeature(dtype=tf.float32),
'object/velo/bbox/dim_xyz':
tf.VarLenFeature(dtype=tf.float32),
'object/velo/bbox/phi':
tf.VarLenFeature(dtype=tf.float32),
'transform/velo_to_image_plane':
tf.FixedLenFeature(shape=(3, 4), dtype=tf.float32),
}
