def default_configs(phi, batch_size=4, image_size=512, num_classes=3):
h = Config()
h.dtype = "float32"
# backbone
h.model = dict(
model="EfficientNetB%d" % phi,
convolution="depthwise_conv2d",
dropblock=None,
# dropblock=dict(keep_prob=None,
# block_size=None)
normalization=dict(normalization="batch_norm",
momentum=0.99,
epsilon=1e-3,
axis=-1,
trainable=True),
activation=dict(activation="swish"),
strides=[2, 1, 2, 2, 2, 1, 2, 1],
dilation_rates=[1, 1, 1, 1, 1, 1, 1, 1],
output_indices=[
-1,
],
frozen_stages=[
-1,
],
num_classes=num_classes)
# loss
h.use_sigmoid = False
h.loss = dict(loss="CrossEntropy",
weight=1.,
from_logits=True,
reduction="none")
h.weight_decay = 4e-5
# dataset
h.num_classes = num_classes
h.train = dict(dataset=dict(
dataset="ShuaidaoDataset",
batch_size=batch_size,
dataset_dir=
"/home/bail/Data/data1/Dataset/RongChuangActions/shuaidao/shuaidaoThree/train",
training=True,
augmentations=[
dict(Resize=dict(input_size=image_size)),
# dict(RandomDistortColor=dict()),
dict(RandAugment=dict(num_layers=2,
magnitude=10.,
cutout_const=40.,
translate_const=100.))
]))
h.val = dict(dataset=dict(
dataset="ShuaidaoDataset",
batch_size=batch_size,
dataset_dir=
"/home/bail/Data/data1/Dataset/RongChuangActions/shuaidao/shuaidaoThree/val",
training=False,
augmentations=[dict(Resize=dict(input_size=image_size))]))
# train
h.pretrained_weights_path = "/data/bail/pretrained_weights/efficientnet-b%d.h5" % phi
h.optimizer = dict(optimizer="SGD", momentum=0.9)
h.lookahead = None
h.learning_rate_scheduler = dict(scheduler="CosineDecay",
initial_learning_rate=0.01,
warmup_steps=1200,
warmup_learning_rate=0.001,
train_steps=32001)
h.checkpoint_dir = "checkpoints/efficientnet_b%d_shuaidao" % phi
h.summary_dir = "logs/efficientnet_b%d_shuaidao" % phi
h.gradient_clip_norm = 0.
h.log_every_n_steps = 100
h.save_ckpt_steps = 1000
h.val_every_n_steps = 1000
return h
def default_configs(name, batch_size=4, image_size=512):
h = Config()
h.dtype = "float32"
# backbone
h.model = dict(
model=name,
convolution="conv2d",
dropblock=None,
# dropblock=dict(keep_prob=None,
# block_size=None)
normalization=dict(normalization="batch_norm",
momentum=0.99,
epsilon=1e-3,
axis=-1,
trainable=True),
activation=dict(activation="relu"),
strides=[2, 1, 2, 2, 2, 1, 2, 1],
dilation_rates=[1, 1, 1, 1, 1, 1, 1, 1],
output_indices=[
-1,
],
frozen_stages=[
-1,
],
num_classes=1)
# loss
h.use_sigmoid = True
h.loss = dict(loss="BinaryCrossEntropy",
weight=1.,
from_logits=True,
reduction="none")
h.weight_decay = 4e-5
# dataset
h.num_classes = 1
h.train = dict(dataset=dict(
dataset="SmokingDataset",
batch_size=batch_size,
dataset_dir="/data/bail/smoking/train",
training=True,
augmentations=[
dict(Resize=dict(input_size=image_size)),
# dict(RandAugment=dict(num_layers=2,
# magnitude=10.,
# cutout_const=40.,
# translate_const=100.))
]))
h.val = dict(
dataset=dict(dataset="SmokingDataset",
batch_size=batch_size,
dataset_dir="/data/bail/smoking/val",
training=False,
augmentations=[dict(Resize=dict(input_size=image_size))]))
# train
h.pretrained_weights_path = "/data/bail/pretrained_weights/efficientnet-b%d.h5" % phi
h.optimizer = dict(optimizer="SGD", momentum=0.9)
h.lookahead = None
h.learning_rate_scheduler = dict(scheduler="CosineDecay",
initial_learning_rate=0.016,
warmup_steps=800,
warmup_learning_rate=0.001,
train_steps=40001)
h.checkpoint_dir = "checkpoints/%s" % name
h.summary_dir = "logs/%s" % name
h.gradient_clip_norm = 0.
h.log_every_n_steps = 100
h.save_ckpt_steps = 2000
h.val_every_n_steps = 2000
return h
