def get_yolov4_config(input_size=None,
num_classes=80,
num_anchors=3,
depth_multiplier=1.,
width_multiplier=1.,
label_assignment="iou",
name="YOLOv4"):
h = Config()
if input_size is None:
input_size = default_input_sizes[name]
h.detector = name
h.dtype = "float16"
h.data_format = "channels_last"
h.num_classes = num_classes
h.depth_multiplier = depth_multiplier_dict[name]
h.width_multiplier = width_multiplier_dict[name]
if name not in model:
raise ValueError(name + " not in ", list(model.keys()))
h.model = model[name]
h.min_level = min_level_dict[name]
h.max_level = max_level_dict[name]
h.strides = strides_dict[name]
h.anchors = anchors_dict[name]
h.num_anchors = len(anchors_dict[name][0]) // 2
h.input_size = input_size if isinstance(input_size, (tuple, list)) else (input_size, input_size)
h.label_assignment = label_assignment
h.anchor_threshold = 0.2
h.gr = 1.
h.bbox_loss = dict(loss="CIoULoss", weight=1., reduction="none")
h.label_loss = dict(loss="BinaryCrossEntropy", weight=1., from_logits=True, reduction="none") # .631 if finetuning else weight = 1.0
h.conf_loss = dict(loss="BinaryCrossEntropy", weight=1., from_logits=True, reduction="none") # 0.911 if finetuning else weight = 1.
h.balance = [1., 1., 1.] # [4.0, 1.0, 0.4] # if num_level == 3 else [4.0, 1.0, 0.4, 0.1]
h.box_weight = 0.05 # 0.0296 if finetune else 0.05
h.label_weight = .5 # 0.243 if finetune else 0.5
h.conf_weight = 1.0 # 0.301 if finetune else 1.0
h.weight_decay = 0.0005
h.excluding_weight_names = ["predicted"]
h.train=dict(dataset=dict(dataset="COCODataset",
batch_size=8,
dataset_dir="/data/bail/COCO",
training=True,
augmentations=[
dict(augmentation="FlipLeftToRight", probability=0.5),
dict(augmentation="RandomDistortColor"),
dict(augmentation="Resize", img_scale=(0.2, 2), keep_ratio=True),
dict(augmentation="Pad", size_divisor=32)
],
# mixup=dict(alpha=8.0, prob=0.5),
mosaic=dict(size=input_size, min_image_scale=0.25, prob=1.),
num_samples=118287),
pretrained_weights_path="/data/bail/pretrained_weights/darknet53-notop/darknet53.ckpt",
optimizer=dict(optimizer="SGD", momentum=0.937),
mixed_precision=dict(loss_scale=None), # The loss scale in mixed precision training. If None, use dynamic.
gradient_clip_norm=.0,
scheduler=dict(train_epochs=480,
# learning_rate_scheduler=dict(scheduler="PiecewiseConstantDecay",
# boundaries=[24, 32],
# values=[0.012, 0.0012, 0.00012]),
learning_rate_scheduler=dict(scheduler="CosineDecay", initial_learning_rate=0.012),
warmup=dict(warmup_learning_rate=0.0012, steps=12000)),
checkpoint_dir="checkpoints/%s" % name,
summary_dir="logs/%s" % name,
log_every_n_steps=100,
save_ckpt_steps=10000)
h.val=dict(dataset=dict(dataset="COCODataset",
batch_size=8,
dataset_dir="/data/bail/COCO",
training=False,
augmentations=[
dict(Resize=dict(size=(input_size, input_size), strides=32, min_scale=1., max_scale=1.0))
# dict(ResizeV2=dict(short_side=800, long_side=1333, strides=64, min_scale=1.0, max_scale=1.0))
]),
samples=5000)
# h.test=dict(nms="NonMaxSuppressionWithQuality",
# pre_nms_size=5000,
# post_nms_size=100,
# iou_threshold=0.6,
# score_threshold=0.5,
# sigma=0.5,
# nms_type="nms")
h.test=dict(nms="CombinedNonMaxSuppression",
pre_nms_size=2000,
post_nms_size=100,
iou_threshold=0.6,
score_threshold=0.35)
return h
def default_detection_configs(phi,
min_level=3,
max_level=7,
fpn_filters=64,
neck_repeats=3,
head_repeats=3,
anchor_scale=4,
num_scales=3,
batch_size=4,
image_size=512,
fusion_type="weighted_sum"):
h = Config()
# model name
h.detector = "efficientdet-d%d" % phi
h.min_level = min_level
h.max_level = max_level
h.dtype = "float16"
# backbone
h.backbone = dict(backbone="efficientnet-b%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=False),
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=[3, 4, 5],
frozen_stages=[-1])
# neck
h.neck = dict(neck="bifpn",
repeats=neck_repeats,
convolution="separable_conv2d",
dropblock=None,
# dropblock=dict(keep_prob=None,
# block_size=None)
feat_dims=fpn_filters,
normalization=dict(normalization="batch_norm",
momentum=0.99,
epsilon=1e-3,
axis=-1,
trainable=False),
activation=dict(activation="swish"),
add_extra_conv=False, # Add extra convolution for neck
fusion_type=fusion_type,
use_multiplication=False)
# head
h.head = dict(head="RetinaNetHead",
repeats=head_repeats,
convolution="separable_conv2d",
dropblock=None,
# dropblock=dict(keep_prob=None,
# block_size=None)
feat_dims=fpn_filters,
normalization=dict(normalization="batch_norm",
momentum=0.99,
epsilon=1e-3,
axis=-1,
trainable=False),
activation=dict(activation="swish"),
prior=0.01)
# anchors parameters
strides = [2 ** l for l in range(min_level, max_level + 1)]
h.anchor = dict(aspect_ratios=[[1., 0.5, 2.]] * (max_level - min_level + 1),
scales=[
[2 ** (i / num_scales) * s * anchor_scale
for i in range(num_scales)] for s in strides
],
num_anchors=9)
# assigner
h.assigner = dict(assigner="max_iou_assigner",
pos_iou_thresh=0.5,
neg_iou_thresh=0.5)
# sampler
h.sampler = dict(sampler="pseudo_sampler")
# loss
h.use_sigmoid = True
h.label_loss=dict(loss="focal_loss",
alpha=0.25,
gamma=1.5,
label_smoothing=0.,
weight=1.,
from_logits=True,
reduction="none")
h.bbox_loss=dict(loss="smooth_l1_loss",
weight=50., # 50.
delta=.1, # .1
reduction="none")
# h.box_loss=dict(loss="giou_loss",
# weight=10.,
# reduction="none")
h.weight_decay = 4e-5
h.bbox_mean = None # [0., 0., 0., 0.]
h.bbox_std = None # [0.1, 0.1, 0.2, 0.2]
# dataset
h.num_classes = 90
h.skip_crowd_during_training = True
h.dataset = "objects365"
h.batch_size = batch_size
h.input_size = [image_size, image_size]
h.train_dataset_dir = "/home/bail/Data/data1/Dataset/Objects365/train"
h.val_dataset_dir = "/home/bail/Data/data1/Dataset/Objects365/train"
h.augmentation = [
dict(ssd_crop=dict(patch_area_range=(0.3, 1.),
aspect_ratio_range=(0.5, 2.0),
min_overlaps=(0.1, 0.3, 0.5, 0.7, 0.9),
max_attempts=100,
probability=.5)),
# dict(data_anchor_sampling=dict(anchor_scales=(16, 32, 64, 128, 256, 512),
# overlap_threshold=0.7,
# max_attempts=50,
# probability=.5)),
dict(flip_left_to_right=dict(probability=0.5)),
dict(random_distort_color=dict(probability=1.))
]
# train
h.pretrained_weights_path = "/home/bail/Workspace/pretrained_weights/efficientdet-d%d" % phi
h.optimizer = dict(optimizer="sgd", momentum=0.9)
h.lookahead = None
h.train_steps = 240000
h.learning_rate_scheduler = dict(scheduler="cosine", initial_learning_rate=0.002)
h.warmup = dict(warmup_learning_rate = 0.00001, steps = 24000)
h.checkpoint_dir = "checkpoints/efficientdet_d%d" % phi
h.summary_dir = "logs/efficientdet_d%d" % phi
h.gradient_clip_norm = .0
h.log_every_n_steps = 500
h.save_ckpt_steps = 10000
h.val_every_n_steps = 4000
h.postprocess = dict(pre_nms_size=5000, # select top_k high confident detections for nms
post_nms_size=100,
iou_threshold=0.5,
score_threshold=0.2)
return h
