def get_presize_combine_transforms_V4():
transforms_presize = A.Compose([
transforms.PadIfNeeded(600, 800),
geometric.Perspective(
scale=[0, .1],
pad_mode=cv2.BORDER_REFLECT,
interpolation=cv2.INTER_AREA, p = .3),
transforms.Flip(),
geometric.ShiftScaleRotate(interpolation=cv2.INTER_LANCZOS4, p = 0.95, scale_limit=0.0),
crops.RandomResizedCrop(
TARGET_SIZE, TARGET_SIZE,
scale=(config['rrc_scale_min'], config['rrc_scale_max']),
ratio=(.70, 1.4),
interpolation=cv2.INTER_CUBIC,
p=1.0),
transforms.Transpose()
#rotate.Rotate(interpolation=cv2.INTER_LANCZOS4, p = 0.99),
])
transforms_postsize = A.Compose([
#imgaug.IAAPiecewiseAffine(),
transforms.CoarseDropout(),
transforms.CLAHE(p=.1),
transforms.RandomToneCurve(scale=.1, p=0.2),
transforms.RandomBrightnessContrast(
brightness_limit=.1,
contrast_limit=0.4,
p=.8),
transforms.HueSaturationValue(
hue_shift_limit=20,
sat_shift_limit=50,
val_shift_limit=0,
p=0.5),
transforms.Equalize(p=0.05),
transforms.FancyPCA(p=0.05),
transforms.RandomGridShuffle(p=0.1),
A.OneOf([
transforms.MotionBlur(blur_limit=(3, 9)),
transforms.GaussianBlur(),
transforms.MedianBlur()
], p=0.1),
transforms.ISONoise(p=.2),
transforms.GaussNoise(var_limit=127., p=.3),
A.OneOf([
transforms.GridDistortion(interpolation=cv2.INTER_AREA, distort_limit=[0.7, 0.7], p=0.5),
transforms.OpticalDistortion(interpolation=cv2.INTER_AREA, p=.3),
], p=.3),
geometric.ElasticTransform(alpha=4, sigma=4, alpha_affine=4, interpolation=cv2.INTER_AREA, p=0.3),
transforms.CoarseDropout(),
transforms.Normalize(),
ToTensorV2()
])
return transforms_presize, transforms_postsize
def get_train_transforms():
return A.Compose([
transforms.PadIfNeeded(600, 800),
geometric.ShiftScaleRotate(interpolation=cv2.INTER_LANCZOS4, p = 0.99, scale_limit=0.8),
geometric.Perspective(pad_mode=cv2.BORDER_REFLECT,interpolation=cv2.INTER_AREA),
crops.RandomResizedCrop(
TARGET_SIZE, TARGET_SIZE,
scale=(config['rrc_scale_min'], config['rrc_scale_max']),
interpolation=cv2.INTER_CUBIC,
p=1.0),
transforms.Transpose(),
transforms.Flip(),
transforms.CoarseDropout(),
transforms.CLAHE(p=.1),
transforms.RandomToneCurve(scale=.1),
transforms.RandomBrightnessContrast(
brightness_limit=.1,
contrast_limit=0.3,
p=.7),
transforms.HueSaturationValue(
hue_shift_limit=20,
sat_shift_limit=60,
val_shift_limit=0,
p=0.6),
transforms.RandomGridShuffle(p=0.1),
A.OneOf([
transforms.MotionBlur(blur_limit=(3, 9)),
transforms.GaussianBlur(),
transforms.MedianBlur()
], p=0.2),
transforms.ISONoise(p=.3),
transforms.GaussNoise(var_limit=255., p=.3),
A.OneOf([
transforms.GridDistortion(interpolation=cv2.INTER_AREA, distort_limit=[0.7, 0.7], p=0.5),
transforms.OpticalDistortion(interpolation=cv2.INTER_AREA, p=.3),
], p=.3),
geometric.ElasticTransform(alpha=4, sigma=100, alpha_affine=100, interpolation=cv2.INTER_AREA, p=0.3),
transforms.CoarseDropout(),
transforms.Normalize(),
ToTensorV2()
])
def __init__(self, im_paths=None, labels=None, phase=None, resize=False):
"""
Args:
im_paths (numpy): image_data
y (numpy): label data
transform: pytorch transforms for transforms and tensor conversion
"""
self.im_paths = im_paths
self.labels = labels
self.resize = resize
self.albumentations_transform = {
'train':
Compose([
ab_transforms.HorizontalFlip(p=0.2),
ab_transforms.VerticalFlip(p=0.2),
ab_transforms.Rotate(limit=180, p=0.2),
ab_transforms.HueSaturationValue(p=0.1),
ab_transforms.RandomContrast(p=0.1),
ab_transforms.GaussianBlur(blur_limit=3, p=0.2),
ab_transforms.GaussNoise(p=0.05),
ab_transforms.CLAHE(p=0.2),
ab_transforms.Normalize(mean=[0.5944, 0.4343, 0.5652],
std=[0.2593, 0.2293, 0.2377]),
ToTensor()
]),
'val':
Compose([
ab_transforms.Normalize(mean=[0.5944, 0.4343, 0.5652],
std=[0.2593, 0.2293, 0.2377]),
ToTensor()
]),
}
if phase == 'train':
self.transform = self.albumentations_transform['train']
else:
self.transform = self.albumentations_transform['val']
def CLAHE(clip_limit=8.0, tile_grid_size=(4, 4), always_apply=False, p=0.5):
return transforms.CLAHE(clip_limit=clip_limit,
tile_grid_size=tile_grid_size,
always_apply=always_apply,
p=p)
def main():
config = vars(parse_args())
if config['name'] is None:
config['name'] = '%s_%s' % (config['arch'],
datetime.now().strftime('%m%d%H'))
config['num_filters'] = [int(n) for n in config['num_filters'].split(',')]
if not os.path.exists('models/detection/%s' % config['name']):
os.makedirs('models/detection/%s' % config['name'])
if config['resume']:
with open('models/detection/%s/config.yml' % config['name'], 'r') as f:
config = yaml.load(f, Loader=yaml.FullLoader)
config['resume'] = True
with open('models/detection/%s/config.yml' % config['name'], 'w') as f:
yaml.dump(config, f)
print('-' * 20)
for key in config.keys():
print('- %s: %s' % (key, str(config[key])))
print('-' * 20)
cudnn.benchmark = True
df = pd.read_csv('inputs/train.csv')
img_paths = np.array('inputs/train_images/' + df['ImageId'].values +
'.jpg')
mask_paths = np.array('inputs/train_masks/' + df['ImageId'].values +
'.jpg')
labels = np.array(
[convert_str_to_labels(s) for s in df['PredictionString']])
test_img_paths = None
test_mask_paths = None
test_outputs = None
if config['pseudo_label'] is not None:
test_df = pd.read_csv('inputs/sample_submission.csv')
test_img_paths = np.array('inputs/test_images/' +
test_df['ImageId'].values + '.jpg')
test_mask_paths = np.array('inputs/test_masks/' +
test_df['ImageId'].values + '.jpg')
ext = os.path.splitext(config['pseudo_label'])[1]
if ext == '.pth':
test_outputs = torch.load('outputs/raw/test/%s' %
config['pseudo_label'])
elif ext == '.csv':
test_labels = pd.read_csv('outputs/submissions/test/%s' %
config['pseudo_label'])
null_idx = test_labels.isnull().any(axis=1)
test_img_paths = test_img_paths[~null_idx]
test_mask_paths = test_mask_paths[~null_idx]
test_labels = test_labels.dropna()
test_labels = np.array([
convert_str_to_labels(
s, names=['pitch', 'yaw', 'roll', 'x', 'y', 'z', 'score'])
for s in test_labels['PredictionString']
])
print(test_labels)
else:
raise NotImplementedError
if config['resume']:
checkpoint = torch.load('models/detection/%s/checkpoint.pth.tar' %
config['name'])
heads = OrderedDict([
('hm', 1),
('reg', 2),
('depth', 1),
])
if config['rot'] == 'eular':
heads['eular'] = 3
elif config['rot'] == 'trig':
heads['trig'] = 6
elif config['rot'] == 'quat':
heads['quat'] = 4
else:
raise NotImplementedError
if config['wh']:
heads['wh'] = 2
criterion = OrderedDict()
for head in heads.keys():
criterion[head] = losses.__dict__[config[head + '_loss']]().cuda()
train_transform = Compose([
transforms.ShiftScaleRotate(shift_limit=config['shift_limit'],
scale_limit=0,
rotate_limit=0,
border_mode=cv2.BORDER_CONSTANT,
value=0,
p=config['shift_p'])
if config['shift'] else NoOp(),
OneOf([
transforms.HueSaturationValue(hue_shift_limit=config['hue_limit'],
sat_shift_limit=config['sat_limit'],
val_shift_limit=config['val_limit'],
p=config['hsv_p'])
if config['hsv'] else NoOp(),
transforms.RandomBrightness(
limit=config['brightness_limit'],
p=config['brightness_p'],
) if config['brightness'] else NoOp(),
transforms.RandomContrast(
limit=config['contrast_limit'],
p=config['contrast_p'],
) if config['contrast'] else NoOp(),
],
p=1),
transforms.ISONoise(p=config['iso_noise_p'], )
if config['iso_noise'] else NoOp(),
transforms.CLAHE(p=config['clahe_p'], ) if config['clahe'] else NoOp(),
],
keypoint_params=KeypointParams(
format='xy', remove_invisible=False))
val_transform = None
folds = []
best_losses = []
# best_scores = []
kf = KFold(n_splits=config['n_splits'], shuffle=True, random_state=41)
for fold, (train_idx, val_idx) in enumerate(kf.split(img_paths)):
print('Fold [%d/%d]' % (fold + 1, config['n_splits']))
if (config['resume'] and fold < checkpoint['fold'] - 1) or (
not config['resume']
and os.path.exists('models/%s/model_%d.pth' %
(config['name'], fold + 1))):
log = pd.read_csv('models/detection/%s/log_%d.csv' %
(config['name'], fold + 1))
best_loss = log.loc[log['val_loss'].values.argmin(), 'val_loss']
# best_loss, best_score = log.loc[log['val_loss'].values.argmin(), ['val_loss', 'val_score']].values
folds.append(str(fold + 1))
best_losses.append(best_loss)
# best_scores.append(best_score)
continue
train_img_paths, val_img_paths = img_paths[train_idx], img_paths[
val_idx]
train_mask_paths, val_mask_paths = mask_paths[train_idx], mask_paths[
val_idx]
train_labels, val_labels = labels[train_idx], labels[val_idx]
if config['pseudo_label'] is not None:
train_img_paths = np.hstack((train_img_paths, test_img_paths))
train_mask_paths = np.hstack((train_mask_paths, test_mask_paths))
train_labels = np.hstack((train_labels, test_labels))
# train
train_set = Dataset(
train_img_paths,
train_mask_paths,
train_labels,
input_w=config['input_w'],
input_h=config['input_h'],
transform=train_transform,
lhalf=config['lhalf'],
hflip=config['hflip_p'] if config['hflip'] else 0,
scale=config['scale_p'] if config['scale'] else 0,
scale_limit=config['scale_limit'],
# test_img_paths=test_img_paths,
# test_mask_paths=test_mask_paths,
# test_outputs=test_outputs,
)
train_loader = torch.utils.data.DataLoader(
train_set,
batch_size=config['batch_size'],
shuffle=True,
num_workers=config['num_workers'],
# pin_memory=True,
)
val_set = Dataset(val_img_paths,
val_mask_paths,
val_labels,
input_w=config['input_w'],
input_h=config['input_h'],
transform=val_transform,
lhalf=config['lhalf'])
val_loader = torch.utils.data.DataLoader(
val_set,
batch_size=config['batch_size'],
shuffle=False,
num_workers=config['num_workers'],
# pin_memory=True,
)
# create model
model = get_model(config['arch'],
heads=heads,
head_conv=config['head_conv'],
num_filters=config['num_filters'],
dcn=config['dcn'],
gn=config['gn'],
ws=config['ws'],
freeze_bn=config['freeze_bn'])
model = model.cuda()
if config['load_model'] is not None:
model.load_state_dict(
torch.load('models/detection/%s/model_%d.pth' %
(config['load_model'], fold + 1)))
params = filter(lambda p: p.requires_grad, model.parameters())
if config['optimizer'] == 'Adam':
optimizer = optim.Adam(params,
lr=config['lr'],
weight_decay=config['weight_decay'])
elif config['optimizer'] == 'AdamW':
optimizer = optim.AdamW(params,
lr=config['lr'],
weight_decay=config['weight_decay'])
elif config['optimizer'] == 'RAdam':
optimizer = RAdam(params,
lr=config['lr'],
weight_decay=config['weight_decay'])
elif config['optimizer'] == 'SGD':
optimizer = optim.SGD(params,
lr=config['lr'],
momentum=config['momentum'],
nesterov=config['nesterov'],
weight_decay=config['weight_decay'])
else:
raise NotImplementedError
if config['apex']:
amp.initialize(model, optimizer, opt_level='O1')
if config['scheduler'] == 'CosineAnnealingLR':
scheduler = lr_scheduler.CosineAnnealingLR(
optimizer, T_max=config['epochs'], eta_min=config['min_lr'])
elif config['scheduler'] == 'ReduceLROnPlateau':
scheduler = lr_scheduler.ReduceLROnPlateau(
optimizer,
factor=config['factor'],
patience=config['patience'],
verbose=1,
min_lr=config['min_lr'])
elif config['scheduler'] == 'MultiStepLR':
scheduler = lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(e) for e in config['milestones'].split(',')],
gamma=config['gamma'])
else:
raise NotImplementedError
log = {
'epoch': [],
'loss': [],
# 'score': [],
'val_loss': [],
# 'val_score': [],
}
best_loss = float('inf')
# best_score = float('inf')
start_epoch = 0
if config['resume'] and fold == checkpoint['fold'] - 1:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
start_epoch = checkpoint['epoch']
log = pd.read_csv(
'models/detection/%s/log_%d.csv' %
(config['name'], fold + 1)).to_dict(orient='list')
best_loss = checkpoint['best_loss']
for epoch in range(start_epoch, config['epochs']):
print('Epoch [%d/%d]' % (epoch + 1, config['epochs']))
# train for one epoch
train_loss = train(config, heads, train_loader, model, criterion,
optimizer, epoch)
# evaluate on validation set
val_loss = validate(config, heads, val_loader, model, criterion)
if config['scheduler'] == 'CosineAnnealingLR':
scheduler.step()
elif config['scheduler'] == 'ReduceLROnPlateau':
scheduler.step(val_loss)
print('loss %.4f - val_loss %.4f' % (train_loss, val_loss))
# print('loss %.4f - score %.4f - val_loss %.4f - val_score %.4f'
# % (train_loss, train_score, val_loss, val_score))
log['epoch'].append(epoch)
log['loss'].append(train_loss)
# log['score'].append(train_score)
log['val_loss'].append(val_loss)
# log['val_score'].append(val_score)
pd.DataFrame(log).to_csv('models/detection/%s/log_%d.csv' %
(config['name'], fold + 1),
index=False)
if val_loss < best_loss:
torch.save(
model.state_dict(), 'models/detection/%s/model_%d.pth' %
(config['name'], fold + 1))
best_loss = val_loss
# best_score = val_score
print("=> saved best model")
state = {
'fold': fold + 1,
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
}
torch.save(
state,
'models/detection/%s/checkpoint.pth.tar' % config['name'])
print('val_loss: %f' % best_loss)
# print('val_score: %f' % best_score)
folds.append(str(fold + 1))
best_losses.append(best_loss)
# best_scores.append(best_score)
results = pd.DataFrame({
'fold':
folds + ['mean'],
'best_loss':
best_losses + [np.mean(best_losses)],
# 'best_score': best_scores + [np.mean(best_scores)],
})
print(results)
results.to_csv('models/detection/%s/results.csv' % config['name'],
index=False)
del model
torch.cuda.empty_cache()
del train_set, train_loader
del val_set, val_loader
gc.collect()
if not config['cv']:
break
def main():
config = vars(parse_args())
if config['name'] is None:
config['name'] = '%s_%s' % (config['arch'], datetime.now().strftime('%m%d%H'))
if not os.path.exists('models/pose/%s' % config['name']):
os.makedirs('models/pose/%s' % config['name'])
if config['resume']:
with open('models/pose/%s/config.yml' % config['name'], 'r') as f:
config = yaml.load(f, Loader=yaml.FullLoader)
config['resume'] = True
with open('models/pose/%s/config.yml' % config['name'], 'w') as f:
yaml.dump(config, f)
print('-'*20)
for key in config.keys():
print('- %s: %s' % (key, str(config[key])))
print('-'*20)
cudnn.benchmark = True
df = pd.read_csv('inputs/train.csv')
img_ids = df['ImageId'].values
pose_df = pd.read_csv('processed/pose_train.csv')
pose_df['img_path'] = 'processed/pose_images/train/' + pose_df['img_path']
if config['resume']:
checkpoint = torch.load('models/pose/%s/checkpoint.pth.tar' % config['name'])
if config['rot'] == 'eular':
num_outputs = 3
elif config['rot'] == 'trig':
num_outputs = 6
elif config['rot'] == 'quat':
num_outputs = 4
else:
raise NotImplementedError
if config['loss'] == 'L1Loss':
criterion = nn.L1Loss().cuda()
elif config['loss'] == 'MSELoss':
criterion = nn.MSELoss().cuda()
else:
raise NotImplementedError
train_transform = Compose([
transforms.ShiftScaleRotate(
shift_limit=config['shift_limit'],
scale_limit=0,
rotate_limit=0,
border_mode=cv2.BORDER_CONSTANT,
value=0,
p=config['shift_p']
) if config['shift'] else NoOp(),
OneOf([
transforms.HueSaturationValue(
hue_shift_limit=config['hue_limit'],
sat_shift_limit=config['sat_limit'],
val_shift_limit=config['val_limit'],
p=config['hsv_p']
) if config['hsv'] else NoOp(),
transforms.RandomBrightness(
limit=config['brightness_limit'],
p=config['brightness_p'],
) if config['brightness'] else NoOp(),
transforms.RandomContrast(
limit=config['contrast_limit'],
p=config['contrast_p'],
) if config['contrast'] else NoOp(),
], p=1),
transforms.ISONoise(
p=config['iso_noise_p'],
) if config['iso_noise'] else NoOp(),
transforms.CLAHE(
p=config['clahe_p'],
) if config['clahe'] else NoOp(),
transforms.Resize(config['input_w'], config['input_h']),
transforms.Normalize(),
ToTensor(),
])
val_transform = Compose([
transforms.Resize(config['input_w'], config['input_h']),
transforms.Normalize(),
ToTensor(),
])
folds = []
best_losses = []
kf = KFold(n_splits=config['n_splits'], shuffle=True, random_state=41)
for fold, (train_idx, val_idx) in enumerate(kf.split(img_ids)):
print('Fold [%d/%d]' %(fold + 1, config['n_splits']))
if (config['resume'] and fold < checkpoint['fold'] - 1) or (not config['resume'] and os.path.exists('pose_models/%s/model_%d.pth' % (config['name'], fold+1))):
log = pd.read_csv('models/pose/%s/log_%d.csv' %(config['name'], fold+1))
best_loss = log.loc[log['val_loss'].values.argmin(), 'val_loss']
# best_loss, best_score = log.loc[log['val_loss'].values.argmin(), ['val_loss', 'val_score']].values
folds.append(str(fold + 1))
best_losses.append(best_loss)
# best_scores.append(best_score)
continue
train_img_ids, val_img_ids = img_ids[train_idx], img_ids[val_idx]
train_img_paths = []
train_labels = []
for img_id in train_img_ids:
tmp = pose_df.loc[pose_df.ImageId == img_id]
img_path = tmp['img_path'].values
train_img_paths.append(img_path)
yaw = tmp['yaw'].values
pitch = tmp['pitch'].values
roll = tmp['roll'].values
roll = rotate(roll, np.pi)
if config['rot'] == 'eular':
label = np.array([
yaw,
pitch,
roll
]).T
elif config['rot'] == 'trig':
label = np.array([
np.cos(yaw),
np.sin(yaw),
np.cos(pitch),
np.sin(pitch),
np.cos(roll),
np.sin(roll),
]).T
elif config['rot'] == 'quat':
raise NotImplementedError
else:
raise NotImplementedError
train_labels.append(label)
train_img_paths = np.hstack(train_img_paths)
train_labels = np.vstack(train_labels)
val_img_paths = []
val_labels = []
for img_id in val_img_ids:
tmp = pose_df.loc[pose_df.ImageId == img_id]
img_path = tmp['img_path'].values
val_img_paths.append(img_path)
yaw = tmp['yaw'].values
pitch = tmp['pitch'].values
roll = tmp['roll'].values
roll = rotate(roll, np.pi)
if config['rot'] == 'eular':
label = np.array([
yaw,
pitch,
roll
]).T
elif config['rot'] == 'trig':
label = np.array([
np.cos(yaw),
np.sin(yaw),
np.cos(pitch),
np.sin(pitch),
np.cos(roll),
np.sin(roll),
]).T
elif config['rot'] == 'quat':
raise NotImplementedError
else:
raise NotImplementedError
val_labels.append(label)
val_img_paths = np.hstack(val_img_paths)
val_labels = np.vstack(val_labels)
# train
train_set = PoseDataset(
train_img_paths,
train_labels,
transform=train_transform,
)
train_loader = torch.utils.data.DataLoader(
train_set,
batch_size=config['batch_size'],
shuffle=True,
num_workers=config['num_workers'],
# pin_memory=True,
)
val_set = PoseDataset(
val_img_paths,
val_labels,
transform=val_transform,
)
val_loader = torch.utils.data.DataLoader(
val_set,
batch_size=config['batch_size'],
shuffle=False,
num_workers=config['num_workers'],
# pin_memory=True,
)
# create model
model = get_pose_model(config['arch'],
num_outputs=num_outputs,
freeze_bn=config['freeze_bn'])
model = model.cuda()
params = filter(lambda p: p.requires_grad, model.parameters())
if config['optimizer'] == 'Adam':
optimizer = optim.Adam(params, lr=config['lr'], weight_decay=config['weight_decay'])
elif config['optimizer'] == 'AdamW':
optimizer = optim.AdamW(params, lr=config['lr'], weight_decay=config['weight_decay'])
elif config['optimizer'] == 'RAdam':
optimizer = RAdam(params, lr=config['lr'], weight_decay=config['weight_decay'])
elif config['optimizer'] == 'SGD':
optimizer = optim.SGD(params, lr=config['lr'], momentum=config['momentum'],
nesterov=config['nesterov'], weight_decay=config['weight_decay'])
else:
raise NotImplementedError
if config['scheduler'] == 'CosineAnnealingLR':
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=config['epochs'], eta_min=config['min_lr'])
elif config['scheduler'] == 'ReduceLROnPlateau':
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, factor=config['factor'], patience=config['patience'],
verbose=1, min_lr=config['min_lr'])
elif config['scheduler'] == 'MultiStepLR':
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[int(e) for e in config['milestones'].split(',')], gamma=config['gamma'])
else:
raise NotImplementedError
log = {
'epoch': [],
'loss': [],
# 'score': [],
'val_loss': [],
# 'val_score': [],
}
best_loss = float('inf')
# best_score = float('inf')
start_epoch = 0
if config['resume'] and fold == checkpoint['fold'] - 1:
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
start_epoch = checkpoint['epoch']
log = pd.read_csv('models/pose/%s/log_%d.csv' % (config['name'], fold+1)).to_dict(orient='list')
best_loss = checkpoint['best_loss']
for epoch in range(start_epoch, config['epochs']):
print('Epoch [%d/%d]' % (epoch + 1, config['epochs']))
# train for one epoch
train_loss = train(config, train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
val_loss = validate(config, val_loader, model, criterion)
if config['scheduler'] == 'CosineAnnealingLR':
scheduler.step()
elif config['scheduler'] == 'ReduceLROnPlateau':
scheduler.step(val_loss)
print('loss %.4f - val_loss %.4f' % (train_loss, val_loss))
# print('loss %.4f - score %.4f - val_loss %.4f - val_score %.4f'
# % (train_loss, train_score, val_loss, val_score))
log['epoch'].append(epoch)
log['loss'].append(train_loss)
# log['score'].append(train_score)
log['val_loss'].append(val_loss)
# log['val_score'].append(val_score)
pd.DataFrame(log).to_csv('models/pose/%s/log_%d.csv' % (config['name'], fold+1), index=False)
if val_loss < best_loss:
torch.save(model.state_dict(), 'models/pose/%s/model_%d.pth' % (config['name'], fold+1))
best_loss = val_loss
# best_score = val_score
print("=> saved best model")
state = {
'fold': fold + 1,
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_loss': best_loss,
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
}
torch.save(state, 'models/pose/%s/checkpoint.pth.tar' % config['name'])
print('val_loss: %f' % best_loss)
# print('val_score: %f' % best_score)
folds.append(str(fold + 1))
best_losses.append(best_loss)
# best_scores.append(best_score)
results = pd.DataFrame({
'fold': folds + ['mean'],
'best_loss': best_losses + [np.mean(best_losses)],
# 'best_score': best_scores + [np.mean(best_scores)],
})
print(results)
results.to_csv('models/pose/%s/results.csv' % config['name'], index=False)
del model
torch.cuda.empty_cache()
del train_set, train_loader
del val_set, val_loader
gc.collect()
if not config['cv']:
break
def augment(im, params=None):
"""
Perform data augmentation on some image using the albumentations package.
Parameters
----------
im : Numpy array
params : dict or None
Contains the data augmentation parameters
Mandatory keys:
- h_flip ([0,1] float): probability of performing an horizontal left-right mirroring.
- v_flip ([0,1] float): probability of performing an vertical up-down mirroring.
- rot ([0,1] float): probability of performing a rotation to the image.
- rot_lim (int): max degrees of rotation.
- stretch ([0,1] float): probability of randomly stretching an image.
- crop ([0,1] float): randomly take an image crop.
- zoom ([0,1] float): random zoom applied to crop_size.
--> Therefore the effective crop size at each iteration will be a
random number between 1 and crop*(1-zoom). For example:
* crop=1, zoom=0: no crop of the image
* crop=1, zoom=0.1: random crop of random size between 100% image and 90% of the image
* crop=0.9, zoom=0.1: random crop of random size between 90% image and 80% of the image
* crop=0.9, zoom=0: random crop of always 90% of the image
Image size refers to the size of the shortest side.
- blur ([0,1] float): probability of randomly blurring an image.
- pixel_noise ([0,1] float): probability of randomly adding pixel noise to an image.
- pixel_sat ([0,1] float): probability of randomly using HueSaturationValue in the image.
- cutout ([0,1] float): probability of using cutout in the image.
Returns
-------
Numpy array
"""
## 1) Crop the image
effective_zoom = np.random.rand() * params['zoom']
crop = params['crop'] - effective_zoom
ly, lx, channels = im.shape
crop_size = int(crop * min([ly, lx]))
rand_x = np.random.randint(low=0, high=lx - crop_size + 1)
rand_y = np.random.randint(low=0, high=ly - crop_size + 1)
crop = transforms.Crop(x_min=rand_x,
y_min=rand_y,
x_max=rand_x + crop_size,
y_max=rand_y + crop_size)
im = crop(image=im)['image']
## 2) Now add the transformations for augmenting the image pixels
transform_list = []
# Add random stretching
if params['stretch']:
transform_list.append(
imgaug_transforms.IAAPerspective(scale=0.1, p=params['stretch']))
# Add random rotation
if params['rot']:
transform_list.append(
transforms.Rotate(limit=params['rot_lim'], p=params['rot']))
# Add horizontal flip
if params['h_flip']:
transform_list.append(transforms.HorizontalFlip(p=params['h_flip']))
# Add vertical flip
if params['v_flip']:
transform_list.append(transforms.VerticalFlip(p=params['v_flip']))
# Add some blur to the image
if params['blur']:
transform_list.append(
albumentations.OneOf([
transforms.MotionBlur(blur_limit=7, p=1.),
transforms.MedianBlur(blur_limit=7, p=1.),
transforms.Blur(blur_limit=7, p=1.),
],
p=params['blur']))
# Add pixel noise
if params['pixel_noise']:
transform_list.append(
albumentations.OneOf(
[
transforms.CLAHE(clip_limit=2, p=1.),
imgaug_transforms.IAASharpen(p=1.),
imgaug_transforms.IAAEmboss(p=1.),
transforms.RandomBrightnessContrast(contrast_limit=0,
p=1.),
transforms.RandomBrightnessContrast(brightness_limit=0,
p=1.),
transforms.RGBShift(p=1.),
transforms.RandomGamma(p=1.) #,
# transforms.JpegCompression(),
# transforms.ChannelShuffle(),
# transforms.ToGray()
],
p=params['pixel_noise']))
# Add pixel saturation
if params['pixel_sat']:
transform_list.append(
transforms.HueSaturationValue(p=params['pixel_sat']))
# Remove randomly remove some regions from the image
if params['cutout']:
ly, lx, channels = im.shape
scale_low, scale_high = 0.05, 0.25 # min and max size of the squares wrt the full image
scale = np.random.uniform(scale_low, scale_high)
transform_list.append(
transforms.Cutout(num_holes=8,
max_h_size=int(scale * ly),
max_w_size=int(scale * lx),
p=params['cutout']))
# Compose all image transformations and augment the image
augmentation_fn = albumentations.Compose(transform_list)
im = augmentation_fn(image=im)['image']
return im
def augment(im, params=None):
"""
Perform data augmentation on some image using the albumentations package.
Parameters
----------
im : Numpy array
params : dict or None
Contains the data augmentation parameters
Mandatory keys:
- h_flip ([0,1] float): probability of performing an horizontal left-right mirroring.
- v_flip ([0,1] float): probability of performing an vertical up-down mirroring.
- rot ([0,1] float): probability of performing a rotation to the image.
- rot_lim (int): max degrees of rotation.
- stretch ([0,1] float): probability of randomly stretching an image.
- expand ([True, False] bool): whether to pad the image to a square shape with background color canvas.
- crop ([0,1] float): randomly take an image crop.
- invert_col ([0, 1] float): randomly invert the colors of the image. p=1 -> invert colors (VPR)
- zoom ([0,1] float): random zoom applied to crop_size.
--> Therefore the effective crop size at each iteration will be a
random number between 1 and crop*(1-zoom). For example:
* crop=1, zoom=0: no crop of the image
* crop=1, zoom=0.1: random crop of random size between 100% image and 90% of the image
* crop=0.9, zoom=0.1: random crop of random size between 90% image and 80% of the image
* crop=0.9, zoom=0: random crop of always 90% of the image
Image size refers to the size of the shortest side.
- blur ([0,1] float): probability of randomly blurring an image.
- pixel_noise ([0,1] float): probability of randomly adding pixel noise to an image.
- pixel_sat ([0,1] float): probability of randomly using HueSaturationValue in the image.
- cutout ([0,1] float): probability of using cutout in the image.
Returns
-------
Numpy array
"""
## 1) Expand the image by padding it with bg-color canvas
if params["expand"]:
desired_size = max(im.shape)
# check bg
if np.argmax(im.shape) > 0:
bgcol = tuple(np.repeat(int(np.mean(im[[0, -1], :, :])), 3))
else:
bgcol = tuple(np.repeat(int(np.mean(im[:, [0, -1], :])), 3))
im = Image.fromarray(im)
old_size = im.size # old_size[0] is in (width, height) format
ratio = float(desired_size) / max(old_size)
new_size = tuple([int(x * ratio) for x in old_size])
im = im.resize(new_size, Image.ANTIALIAS)
# create a new image and paste the resized on it
new_im = Image.new("RGB", (desired_size, desired_size), color=bgcol)
new_im.paste(im, ((desired_size - new_size[0]) // 2,
(desired_size - new_size[1]) // 2))
im = np.array(new_im)
## 2) Crop the image
if params["crop"] and params["crop"] != 1:
effective_zoom = np.random.rand() * params['zoom']
crop = params['crop'] - effective_zoom
ly, lx, channels = im.shape
crop_size = int(crop * min([ly, lx]))
rand_x = np.random.randint(low=0, high=lx - crop_size + 1)
rand_y = np.random.randint(low=0, high=ly - crop_size + 1)
crop = transforms.Crop(x_min=rand_x,
y_min=rand_y,
x_max=rand_x + crop_size,
y_max=rand_y + crop_size)
im = crop(image=im)['image']
if params["enhance"]:
im = Image.fromarray(im)
enhancer = ImageEnhance.Contrast(im)
im = np.array(enhancer.enhance(params["enhance"]))
## 3) Now add the transformations for augmenting the image pixels
transform_list = []
if params['invert_col']:
transform_list.append(transforms.InvertImg(p=params['invert_col']))
# Add random stretching
if params['stretch']:
transform_list.append(
imgaug_transforms.IAAPerspective(scale=0.1, p=params['stretch']))
# Add random rotation
if params['rot']:
transform_list.append(
transforms.Rotate(limit=params['rot_lim'], p=params['rot']))
# Add horizontal flip
if params['h_flip']:
transform_list.append(transforms.HorizontalFlip(p=params['h_flip']))
# Add vertical flip
if params['v_flip']:
transform_list.append(transforms.VerticalFlip(p=params['v_flip']))
# Add some blur to the image
if params['blur']:
transform_list.append(
albumentations.OneOf([
transforms.MotionBlur(blur_limit=7, p=1.),
transforms.MedianBlur(blur_limit=7, p=1.),
transforms.Blur(blur_limit=7, p=1.),
],
p=params['blur']))
# Add pixel noise
if params['pixel_noise']:
transform_list.append(
albumentations.OneOf(
[
transforms.CLAHE(clip_limit=2, p=1.),
imgaug_transforms.IAASharpen(p=1.),
imgaug_transforms.IAAEmboss(p=1.),
transforms.RandomBrightnessContrast(contrast_limit=0,
p=1.),
transforms.RandomBrightnessContrast(brightness_limit=0,
p=1.),
transforms.RGBShift(p=1.),
transforms.RandomGamma(p=1.) #,
# transforms.JpegCompression(),
# transforms.ChannelShuffle(),
# transforms.ToGray()
],
p=params['pixel_noise']))
# Add pixel saturation
if params['pixel_sat']:
transform_list.append(
transforms.HueSaturationValue(p=params['pixel_sat']))
# Remove randomly remove some regions from the image
if params['cutout']:
ly, lx, channels = im.shape
scale_low, scale_high = 0.05, 0.25 # min and max size of the squares wrt the full image
scale = np.random.uniform(scale_low, scale_high)
transform_list.append(
transforms.Cutout(num_holes=8,
max_h_size=int(scale * ly),
max_w_size=int(scale * lx),
p=params['cutout']))
# Compose all image transformations and augment the image
augmentation_fn = albumentations.Compose(transform_list)
im = augmentation_fn(image=im)['image']
return im
import pandas as pd
import matplotlib.pyplot as plt
from tqdm.auto import tqdm
import pathlib
from dataset import TRAFFIC_LABELS_TO_NUM
alb_transforms = [
alb.IAAAdditiveGaussianNoise(p=1),
alb.GaussNoise(p=1),
alb.MotionBlur(p=1),
alb.MedianBlur(blur_limit=3, p=1),
alb.Blur(blur_limit=3, p=1),
alb.OpticalDistortion(p=1),
alb.GridDistortion(p=1),
alb.IAAPiecewiseAffine(p=1),
aat.CLAHE(clip_limit=2, p=1),
alb.IAASharpen(p=1),
alb.IAAEmboss(p=1),
aat.HueSaturationValue(p=0.3),
aat.HorizontalFlip(p=1),
aat.RGBShift(),
aat.RandomBrightnessContrast(),
aat.RandomGamma(p=1),
aat.Cutout(2, 10, 10, p=1),
aat.Equalize(mode='cv', p=1),
aat.FancyPCA(p=1),
aat.RandomFog(p=1),
aat.RandomRain(blur_value=3, p=1),
albumentations.IAAAffine(p=1),
albumentations.ShiftScaleRotate(rotate_limit=15, p=1)
]