class MaskedLMConfig(cfg.TaskConfig):
"""The model config."""
network: bert.BertPretrainerConfig = bert.BertPretrainerConfig(cls_heads=[
bert.ClsHeadConfig(
inner_dim=768, num_classes=2, dropout_rate=0.1, name='next_sentence')
])
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class QuestionAnsweringConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
network: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class DualEncoderConfig(cfg.TaskConfig): """The model config.""" # At most one of `init_checkpoint` and `hub_module_url` can # be specified. init_checkpoint: str = '' hub_module_url: str = '' # Defines the concrete model config at instantiation time. model: ModelConfig = ModelConfig() train_data: cfg.DataConfig = cfg.DataConfig() validation_data: cfg.DataConfig = cfg.DataConfig()
class ElectraPretrainConfig(cfg.TaskConfig):
"""The model config."""
model: electra.ElectraPretrainerConfig = electra.ElectraPretrainerConfig(
cls_heads=[
bert.ClsHeadConfig(inner_dim=768,
num_classes=2,
dropout_rate=0.1,
name='next_sentence')
])
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class QuestionAnsweringConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
n_best_size: int = 20
max_answer_length: int = 30
null_score_diff_threshold: float = 0.0
model: ModelConfig = ModelConfig()
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class SentencePredictionConfig(cfg.TaskConfig): """The model config.""" # At most one of `init_checkpoint` and `hub_module_url` can # be specified. init_checkpoint: str = '' init_cls_pooler: bool = False hub_module_url: str = '' metric_type: str = 'accuracy' # Defines the concrete model config at instantiation time. model: ModelConfig = ModelConfig() train_data: cfg.DataConfig = cfg.DataConfig() validation_data: cfg.DataConfig = cfg.DataConfig()
class TaggingConfig(cfg.TaskConfig): """The model config.""" # At most one of `init_checkpoint` and `hub_module_url` can be specified. init_checkpoint: str = '' hub_module_url: str = '' model: ModelConfig = ModelConfig() # The real class names, the order of which should match real label id. # Note that a word may be tokenized into multiple word_pieces tokens, and # we asssume the real label id (non-negative) is assigned to the first token # of the word, and a negative label id is assigned to the remaining tokens. # The negative label id will not contribute to loss and metrics. class_names: Optional[List[str]] = None train_data: cfg.DataConfig = cfg.DataConfig() validation_data: cfg.DataConfig = cfg.DataConfig()
class TaggingConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
model: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
num_classes: int = 0
# The ignored label id will not contribute to loss.
# A word may be tokenized into multiple word_pieces tokens, and we usually
# assign the real label id for the first token of the word, and
# `ignore_label_id` for the remaining tokens.
ignore_label_id: int = 0
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class SentencePredictionConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `pretrain_checkpoint_dir` and `hub_module_url` can
# be specified.
pretrain_checkpoint_dir: str = ''
hub_module_url: str = ''
network: bert.BertPretrainerConfig = bert.BertPretrainerConfig(
num_masked_tokens=0,
cls_heads=[
bert.ClsHeadConfig(inner_dim=768,
num_classes=3,
dropout_rate=0.1,
name='sentence_prediction')
])
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
def test_parser(self, output_size, dtype, is_training):
params = cfg.DataConfig(input_path='imagenet-2012-tfrecord/train*',
global_batch_size=2,
is_training=True,
examples_consume=4)
decoder = classification_input.Decoder()
parser = classification_input.Parser(output_size=output_size[:2],
num_classes=1001,
aug_rand_hflip=False,
dtype=dtype)
reader = input_reader.InputReader(params,
dataset_fn=tf.data.TFRecordDataset,
decoder_fn=decoder.decode,
parser_fn=parser.parse_fn(
params.is_training))
dataset = reader.read()
images, labels = next(iter(dataset))
self.assertAllEqual(images.numpy().shape,
[params.global_batch_size] + output_size)
self.assertAllEqual(labels.numpy().shape, [params.global_batch_size])
if dtype == 'float32':
self.assertAllEqual(images.dtype, tf.float32)
elif dtype == 'float16':
self.assertAllEqual(images.dtype, tf.float16)
elif dtype == 'bfloat16':
self.assertAllEqual(images.dtype, tf.bfloat16)
class TaggingConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
model: encoders.TransformerEncoderConfig = (
encoders.TransformerEncoderConfig())
# The number of real labels. Note that a word may be tokenized into
# multiple word_pieces tokens, and we asssume the real label id (non-negative)
# is assigned to the first token of the word, and a negative label id is
# assigned to the remaining tokens. The negative label id will not contribute
# to loss and metrics.
num_classes: int = 0
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
class SentencePredictionConfig(cfg.TaskConfig):
"""The model config."""
# At most one of `init_checkpoint` and `hub_module_url` can
# be specified.
init_checkpoint: str = ''
hub_module_url: str = ''
metric_type: str = 'accuracy'
network: bert.BertPretrainerConfig = bert.BertPretrainerConfig(
num_masked_tokens=0, # No masked language modeling head.
cls_heads=[
bert.ClsHeadConfig(inner_dim=768,
num_classes=3,
dropout_rate=0.1,
name='sentence_prediction')
])
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
from official.nlp.configs import bert
from official.nlp.data import data_loader_factory
@dataclasses.dataclass
class MaskedLMConfig(cfg.TaskConfig):
"""The model config."""
<<<<<<< HEAD
init_checkpoint: str = ''
=======
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
model: bert.BertPretrainerConfig = bert.BertPretrainerConfig(cls_heads=[
bert.ClsHeadConfig(
inner_dim=768, num_classes=2, dropout_rate=0.1, name='next_sentence')
])
train_data: cfg.DataConfig = cfg.DataConfig()
validation_data: cfg.DataConfig = cfg.DataConfig()
<<<<<<< HEAD
@base_task.register_task_cls(MaskedLMConfig)
=======
@task_factory.register_task_cls(MaskedLMConfig)
>>>>>>> a811a3b7e640722318ad868c99feddf3f3063e36
class MaskedLMTask(base_task.Task):
"""Mock task object for testing."""
def build_model(self, params=None):
params = params or self.task_config.model
return bert.instantiate_pretrainer_from_cfg(params)
def testRetinanetInputReader(self, output_size, skip_crowd_during_training,
use_autoaugment, is_training):
batch_size = 2
min_level = 3
max_level = 7
num_scales = 3
aspect_ratios = [0.5, 1.0, 2.0]
anchor_size = 3
max_num_instances = 100
params = cfg.DataConfig(
input_path='/placer/prod/home/snaggletooth/test/data/coco/val*',
global_batch_size=batch_size,
is_training=is_training)
decoder = tf_example_decoder.TfExampleDecoder()
parser = retinanet_input.Parser(
output_size=output_size,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size,
skip_crowd_during_training=skip_crowd_during_training,
use_autoaugment=use_autoaugment,
max_num_instances=max_num_instances,
dtype='bfloat16')
reader = input_reader.InputReader(params,
dataset_fn=tf.data.TFRecordDataset,
decoder_fn=decoder.decode,
parser_fn=parser.parse_fn(
params.is_training))
dataset = reader.read()
iterator = iter(dataset)
image, labels = next(iterator)
np_image = image.numpy()
np_labels = tf.nest.map_structure(lambda x: x.numpy(), labels)
# Checks image shape.
self.assertEqual(list(np_image.shape),
[batch_size, output_size[0], output_size[1], 3])
# Checks keys in labels.
if is_training:
self.assertCountEqual(np_labels.keys(), [
'cls_targets', 'box_targets', 'anchor_boxes', 'cls_weights',
'box_weights', 'image_info'
])
else:
self.assertCountEqual(np_labels.keys(), [
'cls_targets', 'box_targets', 'anchor_boxes', 'cls_weights',
'box_weights', 'groundtruths', 'image_info'
])
# Checks shapes of `image_info` and `anchor_boxes`.
self.assertEqual(np_labels['image_info'].shape, (batch_size, 4, 2))
n_anchors = 0
for level in range(min_level, max_level + 1):
stride = 2**level
output_size_l = [output_size[0] / stride, output_size[1] / stride]
anchors_per_location = num_scales * len(aspect_ratios)
self.assertEqual(list(np_labels['anchor_boxes'][level].shape), [
batch_size, output_size_l[0], output_size_l[1],
4 * anchors_per_location
])
n_anchors += output_size_l[0] * output_size_l[
1] * anchors_per_location
# Checks shapes of training objectives.
self.assertEqual(np_labels['cls_weights'].shape,
(batch_size, n_anchors))
for level in range(min_level, max_level + 1):
stride = 2**level
output_size_l = [output_size[0] / stride, output_size[1] / stride]
anchors_per_location = num_scales * len(aspect_ratios)
self.assertEqual(list(np_labels['cls_targets'][level].shape), [
batch_size, output_size_l[0], output_size_l[1],
anchors_per_location
])
self.assertEqual(list(np_labels['box_targets'][level].shape), [
batch_size, output_size_l[0], output_size_l[1],
4 * anchors_per_location
])
# Checks shape of groundtruths for eval.
if not is_training:
self.assertEqual(np_labels['groundtruths']['source_id'].shape,
(batch_size, ))
self.assertEqual(np_labels['groundtruths']['classes'].shape,
(batch_size, max_num_instances))
self.assertEqual(np_labels['groundtruths']['boxes'].shape,
(batch_size, max_num_instances, 4))
self.assertEqual(np_labels['groundtruths']['areas'].shape,
(batch_size, max_num_instances))
self.assertEqual(np_labels['groundtruths']['is_crowds'].shape,
(batch_size, max_num_instances))
def testMaskRCNNInputReader(self, output_size, skip_crowd_during_training,
include_mask, is_training):
min_level = 3
max_level = 7
num_scales = 3
aspect_ratios = [1.0, 2.0, 0.5]
max_num_instances = 100
batch_size = 2
mask_crop_size = 112
anchor_size = 4.0
params = cfg.DataConfig(
input_path='/placer/prod/home/snaggletooth/test/data/coco/val*',
global_batch_size=batch_size,
is_training=is_training)
parser = maskrcnn_input.Parser(
output_size=output_size,
min_level=min_level,
max_level=max_level,
num_scales=num_scales,
aspect_ratios=aspect_ratios,
anchor_size=anchor_size,
rpn_match_threshold=0.7,
rpn_unmatched_threshold=0.3,
rpn_batch_size_per_im=256,
rpn_fg_fraction=0.5,
aug_rand_hflip=True,
aug_scale_min=0.8,
aug_scale_max=1.2,
skip_crowd_during_training=skip_crowd_during_training,
max_num_instances=max_num_instances,
include_mask=include_mask,
mask_crop_size=mask_crop_size,
dtype='bfloat16')
decoder = tf_example_decoder.TfExampleDecoder(
include_mask=include_mask)
reader = input_reader.InputReader(params,
dataset_fn=tf.data.TFRecordDataset,
decoder_fn=decoder.decode,
parser_fn=parser.parse_fn(
params.is_training))
dataset = reader.read()
iterator = iter(dataset)
images, labels = next(iterator)
np_images = images.numpy()
np_labels = tf.nest.map_structure(lambda x: x.numpy(), labels)
if is_training:
self.assertAllEqual(
np_images.shape,
[batch_size, output_size[0], output_size[1], 3])
self.assertAllEqual(np_labels['image_info'].shape,
[batch_size, 4, 2])
self.assertAllEqual(np_labels['gt_boxes'].shape,
[batch_size, max_num_instances, 4])
self.assertAllEqual(np_labels['gt_classes'].shape,
[batch_size, max_num_instances])
if include_mask:
self.assertAllEqual(np_labels['gt_masks'].shape, [
batch_size, max_num_instances, mask_crop_size,
mask_crop_size
])
for level in range(min_level, max_level + 1):
stride = 2**level
output_size_l = [
output_size[0] / stride, output_size[1] / stride
]
anchors_per_location = num_scales * len(aspect_ratios)
self.assertAllEqual(
np_labels['rpn_score_targets'][level].shape, [
batch_size, output_size_l[0], output_size_l[1],
anchors_per_location
])
self.assertAllEqual(
np_labels['rpn_box_targets'][level].shape, [
batch_size, output_size_l[0], output_size_l[1],
4 * anchors_per_location
])
self.assertAllEqual(np_labels['anchor_boxes'][level].shape, [
batch_size, output_size_l[0], output_size_l[1],
4 * anchors_per_location
])
else:
self.assertAllEqual(
np_images.shape,
[batch_size, output_size[0], output_size[1], 3])
self.assertAllEqual(np_labels['image_info'].shape,
[batch_size, 4, 2])
