def __init__(self, config):
width = height = config.DATA.SCALE # 300/512
self.train_transform = A.Compose( # Yolo
[
# A.RandomSizedCrop(min_max_height=(800, 1024), height=1024, width=1024, p=0.5),
# A.RandomScale(scale_limit=0.3, p=1.0), # 这个有问题
C.RandomResize(scale_limit=0.3, p=1.0), # 调节长宽比 [1/1.3, 1.3]
A.OneOf(
[
A.Sequential(
[
A.SmallestMaxSize(min(height, width), p=1.0),
A.RandomCrop(
height, width,
p=1.0) # 先resize到短边544，再crop成544×544
],
p=0.4),
A.LongestMaxSize(max(height, width),
p=0.6), # resize到长边544
],
p=1.0),
# A.LongestMaxSize(max(height, width), p=1.0),
A.OneOf([
A.HueSaturationValue(hue_shift_limit=0.4,
sat_shift_limit=0.4,
val_shift_limit=0.4,
p=0.9),
A.RandomBrightnessContrast(
brightness_limit=0.3, contrast_limit=0.3, p=0.9),
],
p=0.9),
# A.PadIfNeeded(min_height=height, min_width=width, border_mode=0, value=(0.5,0.5,0.5), p=1.0),
C.RandomPad(min_height=height,
min_width=width,
border_mode=0,
value=(123 / 255, 117 / 255, 104 / 255),
p=1.0),
A.HorizontalFlip(p=0.5),
ToTensorV2(p=1.0),
],
p=1.0,
bbox_params=A.BboxParams(format='pascal_voc',
min_area=0,
min_visibility=0,
label_fields=['labels']),
)
self.val_transform = A.Compose([
A.Resize(height=height, width=width, p=1.0),
ToTensorV2(p=1.0),
],
p=1.0,
bbox_params=A.BboxParams(
format='pascal_voc',
min_area=0,
min_visibility=0,
label_fields=['labels']))
def make_coco_transforms(image_set='train', image_size=(640, 640)):
normalize = A.Sequential([
A.Normalize(), #imagenet norms default
ToTensorV2()
])
if image_set == 'train':
transforms = A.Compose([
A.PadIfNeeded(*image_size, border_mode=0), #pad with zeros
A.RandomResizedCrop(*image_size),
A.HorizontalFlip(),
normalize
])
elif image_set == 'val':
transforms = A.Compose([
A.Resize(*image_size),
normalize
])
else:
raise ValueError(f'{image_set} not recognized!')
return transforms
def train_albu_augment(record):
verbose = record.get('verbose', False)
image_size = record['image_size']
image = record['image']
mask = record['mask']
if verbose:
pipeline = albu.ReplayCompose
else:
pipeline = albu.Compose
aug = pipeline([
albu.Sequential([
albu.GaussNoise(0.05, p=1),
albu.augmentations.transforms.Blur(
blur_limit=5, always_apply=False, p=1),
],
p=0.001),
albu.OneOf([
albu.ElasticTransform(p=0.5),
albu.IAAPiecewiseAffine(p=0.5),
albu.OpticalDistortion(p=1)
],
p=0.001),
albu.OneOf([
albu.RandomBrightnessContrast(brightness_limit=0.2,
contrast_limit=0.2,
brightness_by_max=True,
always_apply=False,
p=1),
albu.RandomBrightnessContrast(brightness_limit=(-0.2, 0.6),
contrast_limit=.2,
brightness_by_max=True,
always_apply=False,
p=1),
albu.augmentations.transforms.ColorJitter(brightness=0.2,
contrast=0.2,
saturation=0.1,
hue=0.1,
always_apply=False,
p=1),
albu.RandomGamma(p=0.5)
],
p=0.5),
albu.OneOf([
albu.GaussNoise(0.02, p=.5),
albu.IAAAffine(p=.5),
],
p=.25),
albu.OneOf([
albu.augmentations.transforms.Blur(
blur_limit=15, always_apply=False, p=0.25),
albu.augmentations.transforms.Blur(
blur_limit=3, always_apply=False, p=0.5)
],
p=0.5),
albu.RandomRotate90(p=.5),
albu.HorizontalFlip(p=.5),
albu.VerticalFlip(p=.5),
albu.RandomCrop(width=image_size, height=image_size),
])
data = aug(image=image, mask=mask)
record['image'] = data['image']
record['mask'] = data['mask']
if verbose:
for transformation in data['replay']['transforms']:
if not isinstance(transformation, dict):
print('not a dict')
pass
elif transformation.get('applied', False):
print(30 * '-')
if 'OneOf' in transformation['__class_fullname__']:
print(30 * '=')
for _trans in transformation['transforms']:
if not _trans.get('applied', False): continue
_name = _trans['__class_fullname__']
if 'Flip' in _name: continue
print(_trans['__class_fullname__'])
for k, v in _trans.items():
if k in [
'__class_fullname__', 'applied',
'always_apply'
]:
continue
print(f"{k}: {v}")
else:
_name = transformation['__class_fullname__']
if 'Flip' in _name: continue
print(_name)
for k, v in transformation.items():
if k in [
'__class_fullname__', 'applied', 'always_apply'
]:
continue
print(f"{k}: {v}")
return record
def aug(source, images_output_path, size):
images_path = images_output_path + "/JPEGImages/"
os.makedirs(images_path, exist_ok=True)
xml_path = images_output_path + "/Annotations/"
os.makedirs(xml_path, exist_ok=True)
transform = A.Compose(
[
# A.CLAHE(),
# A.RandomScale(scale_limit=[0.5, 1]),
# A.RandomCrop(width=450, height=450),
A.OneOf([
A.Sequential(
[A.RandomCrop(width=800, height=600),
A.RandomRotate90()]),
# A.Sequential(
# [
# A.RandomSizedBBoxSafeCrop(width=800, height=600),
# A.RandomRotate90(),
# ]
# ),
A.Sequential([
A.RandomScale(scale_limit=0.2),
A.Flip(),
A.RandomRotate90(),
],
# p=0.3,
),
A.Sequential(
[
A.Rotate(),
],
p=0.3,
),
])
# A.Transpose(),
# A.Resize(0.9, 0.9),
# A.Blur(blur_limit=3),
# A.OpticalDistortion(),
# A.GridDistortion(),
# A.HueSaturationValue(),
],
bbox_params=A.BboxParams(format="pascal_voc",
min_visibility=0.5,
label_fields=["class_labels"]),
)
rows = []
random.seed(42)
images_index = 1
for name, group in source.groupby("filename"):
row = group.iloc[0]
print(row["filename"])
image = cv2.imread(row["filename"])
same = set()
bboxes = []
class_labels = []
aleady_box = {}
for _, vrow in group.iterrows():
bboxes.append(
[vrow["xmin"], vrow["ymin"], vrow["xmax"], vrow["ymax"]])
class_labels.append(vrow["class"])
aleady_box[vrow["class"]] = set()
all_count = 0
print(aleady_box)
while int(all_count) < size:
augmented = transform(
image=image,
bboxes=bboxes,
class_labels=class_labels,
)
file_name = f"{images_index}.jpg"
if len(augmented["bboxes"]) < 1:
continue
writer = Writer(file_name, augmented["image"].shape[1],
augmented["image"].shape[0])
findbox = False
for index, bbox in enumerate(augmented["bboxes"]):
x_min, y_min, x_max, y_max = map(lambda v: int(v), bbox[:4])
same.add(x_min)
rows.append({
"filename": f"{images_path}/{file_name}",
"width": augmented["image"].shape[1],
"height": augmented["image"].shape[0],
"class": augmented["class_labels"][index],
"xmin": x_min,
"ymin": y_min,
"xmax": x_max,
"ymax": y_max,
"imageindex": str(images_index),
})
writer.addObject(augmented["class_labels"][index], x_min,
y_min, x_max, y_max)
if len(aleady_box[augmented["class_labels"][index]]) >= size:
continue
aleady_box[augmented["class_labels"][index]].add(x_min)
findbox = True
if findbox:
cv2.imwrite(f"{images_path}/{file_name}", augmented["image"])
writer.save(f"{xml_path}/{images_index}.xml")
images_index += 1
print(aleady_box)
all_count = sum([min(len(v), size)
for k, v in aleady_box.items()]) / len(aleady_box)
df = pd.DataFrame(rows)
return df
bbox_params = A.BboxParams(format='pascal_voc', min_visibility=0.3)
train_transforms = A.Compose([
A.Resize(height=img_size[0], width=img_size[1], p=1.0),
A.HorizontalFlip(p=0.5),
# A.OneOf([
# A.Sequential([
# A.Resize(height=img_size[0], width=img_size[1], p=1.0),
# ], p=1.0),
# A.Sequential([
# A.RandomSizedBBoxSafeCrop(height=img_size[0], width=img_size[1], p=1.0),
# ], p=1.0)
# ], p=1.0),
A.OneOf([
A.Sequential([
A.GaussNoise(var_limit=(100, 150), p=0.5),
A.MotionBlur(blur_limit=17, p=0.5)
], p=1.0),
A.Sequential([
A.GaussNoise(var_limit=(100, 150), p=0.5),
A.MotionBlur(blur_limit=17, p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.5),
A.HueSaturationValue(hue_shift_limit=20, sat_shift_limit=30, val_shift_limit=0, p=0.5),
], p=1.0),
A.Sequential([
A.GaussNoise(var_limit=(100, 150), p=0.5),
A.MotionBlur(blur_limit=17, p=0.5),
A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.5),
A.HueSaturationValue(hue_shift_limit=20, sat_shift_limit=30, val_shift_limit=0, p=0.5),
A.ChannelShuffle(p=0.5),
# A.ShiftScaleRotate(shift_limit=0.1, scale_limit=(-0.15, 0.15), rotate_limit=30, p=0.5,
# border_mode=cv2.BORDER_CONSTANT, value=0),
def train_function(gpu, world_size, node_rank, gpus, fold_number, group_name):
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
torch.manual_seed(25)
np.random.seed(25)
rank = node_rank * gpus + gpu
dist.init_process_group(
backend='nccl',
init_method='env://',
world_size=world_size,
rank=rank
)
device = torch.device("cuda:{}".format(gpu) if torch.cuda.is_available() else "cpu")
batch_size = 64
width_size = 416
init_lr = 1e-4
end_lr = 1e-6
n_epochs = 20
emb_size = 512
margin = 0.5
dropout = 0.0
iters_to_accumulate = 1
if rank == 0:
wandb.init(project='shopee_effnet0', group=group_name, job_type=str(fold_number))
checkpoints_dir_name = 'effnet0_{}_{}_{}'.format(width_size, dropout, group_name)
os.makedirs(checkpoints_dir_name, exist_ok=True)
wandb.config.model_name = checkpoints_dir_name
wandb.config.batch_size = batch_size
wandb.config.width_size = width_size
wandb.config.init_lr = init_lr
wandb.config.n_epochs = n_epochs
wandb.config.emb_size = emb_size
wandb.config.dropout = dropout
wandb.config.iters_to_accumulate = iters_to_accumulate
wandb.config.optimizer = 'adam'
wandb.config.scheduler = 'ShopeeScheduler'
df = pd.read_csv('../../dataset/reliable_validation_tm.csv')
train_df = df[df['fold_group'] != fold_number]
train_transforms = alb.Compose([
alb.RandomResizedCrop(width_size, width_size),
alb.ShiftScaleRotate(shift_limit=0.1, rotate_limit=30),
alb.HorizontalFlip(),
alb.OneOf([
alb.Sequential([
alb.HueSaturationValue(hue_shift_limit=50),
alb.RandomBrightnessContrast(),
]),
alb.FancyPCA(),
alb.ChannelDropout(),
alb.ChannelShuffle(),
alb.RGBShift()
]),
alb.CoarseDropout(max_height=int(width_size*0.1), max_width=int(width_size*0.1)),
alb.OneOf([
alb.ElasticTransform(),
alb.OpticalDistortion(),
alb.GridDistortion()
]),
alb.Resize(width_size, width_size),
alb.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2()
])
train_set = ImageDataset(train_df, train_df, '../../dataset/train_images', train_transforms)
sampler = DistributedSampler(train_set, num_replicas=world_size, rank=rank, shuffle=True)
train_dataloader = DataLoader(train_set, batch_size=batch_size // world_size, shuffle=False, num_workers=4,
sampler=sampler)
# valid_df = df[df['fold_strat'] == fold_number]
valid_transforms = alb.Compose([
alb.Resize(width_size, width_size),
alb.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2()
])
# valid_set = ImageDataset(train_df, valid_df, '../../dataset/train_images', valid_transforms)
# valid_dataloader = DataLoader(valid_set, batch_size=batch_size // world_size, shuffle=False, num_workers=4)
test_df = df[df['fold_group'] == fold_number]
test_set = ImageDataset(test_df, test_df, '../../dataset/train_images', valid_transforms)
test_dataloader = DataLoader(test_set, batch_size=batch_size // world_size, shuffle=False, num_workers=4)
model = EfficientNetArcFace(emb_size, train_df['label_group'].nunique(), device, dropout=dropout,
backbone='tf_efficientnet_b0_ns', pretrained=True, margin=margin, is_amp=True)
model = SyncBatchNorm.convert_sync_batchnorm(model)
model.to(device)
model = DistributedDataParallel(model, device_ids=[gpu])
scaler = GradScaler()
criterion = CrossEntropyLoss()
# criterion = LabelSmoothLoss(smoothing=0.1)
optimizer = optim.Adam(model.parameters(), lr=init_lr)
# scheduler = CosineAnnealingLR(optimizer, T_max=n_epochs, eta_min=end_lr,
# last_epoch=-1)
# scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=2000, T_mult=1,
# eta_min=end_lr, last_epoch=-1)
scheduler = ShopeeScheduler(optimizer, lr_start=init_lr,
lr_max=init_lr*batch_size, lr_min=end_lr)
for epoch in range(n_epochs):
train_loss, train_duration, train_f1 = train_one_epoch(
model, train_dataloader, optimizer, criterion, device, scaler,
scheduler=None, iters_to_accumulate=iters_to_accumulate)
scheduler.step()
if rank == 0:
# valid_loss, valid_duration, valid_f1 = evaluate(model, valid_dataloader, criterion, device)
embeddings = get_embeddings(model, test_dataloader, device)
embeddings_f1 = validate_embeddings_f1(embeddings, test_df)
wandb.log({'train_loss': train_loss, 'train_f1': train_f1,
'embeddings_f1': embeddings_f1, 'epoch': epoch})
filename = '{}_foldnum{}_epoch{}_train_loss{}_f1{}'.format(
checkpoints_dir_name, fold_number+1, epoch+1,
round(train_loss, 3), round(embeddings_f1, 3))
torch.save(model.module.state_dict(), os.path.join(checkpoints_dir_name, '{}.pth'.format(filename)))
# np.savez_compressed(os.path.join(checkpoints_dir_name, '{}.npz'.format(filename)), embeddings=embeddings)
print('FOLD NUMBER %d\tEPOCH %d:\t'
'TRAIN [duration %.3f sec, loss: %.3f, avg f1: %.3f]\t'
'VALID EMBEDDINGS [avg f1: %.3f]\tCurrent time %s' %
(fold_number + 1, epoch + 1, train_duration,
train_loss, train_f1, embeddings_f1,
str(datetime.now(timezone('Europe/Moscow')))))
if rank == 0:
wandb.finish()
def create_transform(self, input_dtype, p):
return A.Sequential([stage.create_transform(input_dtype) for stage in self.stages], p=p)
