def test_parser(self, output_size, dtype, is_training, aug_name,
is_multilabel, decode_jpeg_only, image_format):
serialized_example = tfexample_utils.create_classification_example(
output_size[0], output_size[1], image_format, is_multilabel)
if aug_name == 'randaug':
aug_type = common.Augmentation(
type=aug_name, randaug=common.RandAugment(magnitude=10))
elif aug_name == 'autoaug':
aug_type = common.Augmentation(
type=aug_name,
autoaug=common.AutoAugment(augmentation_name='test'))
else:
aug_type = None
decoder = classification_input.Decoder(image_field_key=IMAGE_FIELD_KEY,
label_field_key=LABEL_FIELD_KEY,
is_multilabel=is_multilabel)
parser = classification_input.Parser(output_size=output_size[:2],
num_classes=10,
image_field_key=IMAGE_FIELD_KEY,
label_field_key=LABEL_FIELD_KEY,
is_multilabel=is_multilabel,
decode_jpeg_only=decode_jpeg_only,
aug_rand_hflip=False,
aug_type=aug_type,
dtype=dtype)
decoded_tensors = decoder.decode(serialized_example)
image, label = parser.parse_fn(is_training)(decoded_tensors)
self.assertAllEqual(image.numpy().shape, output_size)
if not is_multilabel:
self.assertAllEqual(label, 0)
else:
self.assertAllEqual(label.numpy().shape, [10])
if dtype == 'float32':
self.assertAllEqual(image.dtype, tf.float32)
elif dtype == 'float16':
self.assertAllEqual(image.dtype, tf.float16)
elif dtype == 'bfloat16':
self.assertAllEqual(image.dtype, tf.bfloat16)
def image_classification_imagenet_resnetrs() -> cfg.ExperimentConfig:
"""Image classification on imagenet with resnet-rs."""
train_batch_size = 4096
eval_batch_size = 4096
steps_per_epoch = IMAGENET_TRAIN_EXAMPLES // train_batch_size
config = cfg.ExperimentConfig(
task=ImageClassificationTask(
model=ImageClassificationModel(
num_classes=1001,
input_size=[160, 160, 3],
backbone=backbones.Backbone(
type='resnet',
resnet=backbones.ResNet(
model_id=50,
stem_type='v1',
resnetd_shortcut=True,
replace_stem_max_pool=True,
se_ratio=0.25,
stochastic_depth_drop_rate=0.0)),
dropout_rate=0.25,
norm_activation=common.NormActivation(
norm_momentum=0.0,
norm_epsilon=1e-5,
use_sync_bn=False,
activation='swish')),
losses=Losses(l2_weight_decay=4e-5, label_smoothing=0.1),
train_data=DataConfig(
input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
is_training=True,
global_batch_size=train_batch_size,
aug_type=common.Augmentation(
type='randaug', randaug=common.RandAugment(magnitude=10))),
validation_data=DataConfig(
input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'valid*'),
is_training=False,
global_batch_size=eval_batch_size)),
trainer=cfg.TrainerConfig(
steps_per_loop=steps_per_epoch,
summary_interval=steps_per_epoch,
checkpoint_interval=steps_per_epoch,
train_steps=350 * steps_per_epoch,
validation_steps=IMAGENET_VAL_EXAMPLES // eval_batch_size,
validation_interval=steps_per_epoch,
optimizer_config=optimization.OptimizationConfig({
'optimizer': {
'type': 'sgd',
'sgd': {
'momentum': 0.9
}
},
'ema': {
'average_decay': 0.9999,
'trainable_weights_only': False,
},
'learning_rate': {
'type': 'cosine',
'cosine': {
'initial_learning_rate': 1.6,
'decay_steps': 350 * steps_per_epoch
}
},
'warmup': {
'type': 'linear',
'linear': {
'warmup_steps': 5 * steps_per_epoch,
'warmup_learning_rate': 0
}
}
})),
restrictions=[
'task.train_data.is_training != None',
'task.validation_data.is_training != None'
])
return config
def image_classification_imagenet_deit_pretrain() -> cfg.ExperimentConfig:
"""Image classification on imagenet with vision transformer."""
train_batch_size = 4096 # originally was 1024 but 4096 better for tpu v3-32
eval_batch_size = 4096 # originally was 1024 but 4096 better for tpu v3-32
num_classes = 1001
label_smoothing = 0.1
steps_per_epoch = IMAGENET_TRAIN_EXAMPLES // train_batch_size
config = cfg.ExperimentConfig(
task=ImageClassificationTask(
model=ImageClassificationModel(
num_classes=num_classes,
input_size=[224, 224, 3],
kernel_initializer='zeros',
backbone=backbones.Backbone(
type='vit',
vit=backbones.VisionTransformer(
model_name='vit-b16',
representation_size=768,
init_stochastic_depth_rate=0.1,
original_init=False,
transformer=backbones.Transformer(
dropout_rate=0.0, attention_dropout_rate=0.0)))),
losses=Losses(l2_weight_decay=0.0,
label_smoothing=label_smoothing,
one_hot=False,
soft_labels=True),
train_data=DataConfig(
input_path=os.path.join(IMAGENET_INPUT_PATH_BASE, 'train*'),
is_training=True,
global_batch_size=train_batch_size,
aug_type=common.Augmentation(type='randaug',
randaug=common.RandAugment(
magnitude=9,
exclude_ops=['Cutout'])),
mixup_and_cutmix=common.MixupAndCutmix(
label_smoothing=label_smoothing)),
validation_data=DataConfig(input_path=os.path.join(
IMAGENET_INPUT_PATH_BASE, 'valid*'),
is_training=False,
global_batch_size=eval_batch_size)),
trainer=cfg.TrainerConfig(
steps_per_loop=steps_per_epoch,
summary_interval=steps_per_epoch,
checkpoint_interval=steps_per_epoch,
train_steps=300 * steps_per_epoch,
validation_steps=IMAGENET_VAL_EXAMPLES // eval_batch_size,
validation_interval=steps_per_epoch,
optimizer_config=optimization.OptimizationConfig({
'optimizer': {
'type': 'adamw',
'adamw': {
'weight_decay_rate': 0.05,
'include_in_weight_decay': r'.*(kernel|weight):0$',
'gradient_clip_norm': 0.0
}
},
'learning_rate': {
'type': 'cosine',
'cosine': {
'initial_learning_rate':
0.0005 * train_batch_size / 512,
'decay_steps': 300 * steps_per_epoch,
}
},
'warmup': {
'type': 'linear',
'linear': {
'warmup_steps': 5 * steps_per_epoch,
'warmup_learning_rate': 0
}
}
})),
restrictions=[
'task.train_data.is_training != None',
'task.validation_data.is_training != None'
])
return config