def train_bowl():
global config
torch.backends.cudnn.benchmark = True
cleanup_mac_hidden_files(config.dataset_path)
sample_count = len(os.listdir(config.dataset_path))
idx = list(range(sample_count))
random.seed(1)
random.shuffle(idx)
split = 0.95
train_idx, val_idx = idx[:int(split *
sample_count)], idx[int(split *
sample_count):]
im_type = BorderImageType if not config.sigmoid else SigmoidBorderImageType
im_val_type = PaddedImageType if not config.sigmoid else PaddedSigmoidImageType
ds = CachingImageProvider(im_type, paths, fn_mapping)
val_ds = CachingImageProvider(im_val_type, paths, fn_mapping)
fold = args.fold
train(ds,
val_ds,
fold,
train_idx,
val_idx,
config,
num_workers=num_workers,
transforms=aug_victor(.97))
def train_roads():
ds = ReadingImageProvider(RawImageType, paths, fn_mapping, image_suffix=image_suffix)
folds = get_csv_folds('folds4.csv', ds.im_names)
num_workers = 0 if os.name == 'nt' else 2
for fold, (train_idx, val_idx) in enumerate(folds):
if args.fold is not None and int(args.fold) != fold:
continue
train(ds, fold, train_idx, val_idx, config, num_workers=num_workers, transforms=get_flips_colors_augmentation())
def train_roads():
ds = ReadingImageProvider(RawImageType, paths, fn_mapping, image_suffix=image_suffix)
folds = get_csv_folds('folds4.csv', ds.im_names)
num_workers = 0 if os.name == 'nt' else 2
for fold, (train_idx, val_idx) in enumerate(folds):
if args.fold is not None and int(args.fold) != fold:
continue
train(ds, fold, train_idx, val_idx, config, num_workers=num_workers, transforms=get_flips_colors_augmentation())
def train_stage0():
"""
heat up weights for 5 epochs
"""
ds = ReadingImageProvider(TiffImageType,
paths,
fn_mapping,
image_suffix='RGB')
folds = get_folds(ds, 5)
num_workers = 0 if os.name == 'nt' else 8
train(ds,
folds,
config,
num_workers=num_workers,
transforms=augment_flips_color)
def train_bowl():
torch.backends.cudnn.benchmark = True
im_type = BorderImageType if not config.sigmoid else SigmoidBorderImageType
im_val_type = PaddedImageType if not config.sigmoid else PaddedSigmoidImageType
ds = CachingImageProvider(im_type, paths, fn_mapping)
val_ds = CachingImageProvider(im_val_type, paths, fn_mapping)
folds = get_csv_folds(ds, os.path.join(config.dataset_path, 'folds.csv'))
for fold, (train_idx, val_idx) in enumerate(folds):
if args.fold is not None and int(args.fold) != fold:
continue
train(ds,
val_ds,
fold,
train_idx,
val_idx,
config,
num_workers=num_workers,
transforms=aug_victor(.97))
def train_roads(config,
paths,
fn_mapping,
image_suffix,
folds_file_loc,
save_path,
log_path,
num_channels=3):
#t0 = time.time()
ds = ReadingImageProvider(RawImageType,
paths,
fn_mapping,
image_suffix=image_suffix,
num_channels=num_channels)
print("len ds:", len(ds))
print("folds_file_loc:", folds_file_loc)
print("save_path:", save_path)
folds = get_csv_folds(folds_file_loc, ds.im_names)
#folds = get_csv_folds('folds4.csv', ds.im_names)
num_workers = 0 if os.name == 'nt' else 2
for fold, (train_idx, val_idx) in enumerate(folds):
if args.fold is not None and int(args.fold) != fold:
continue
print("num workers:", num_workers)
print("fold:", fold)
#print ("(train_idx, val_idx):", (train_idx, val_idx))
print("len(train_idx):", len(train_idx))
print("len(val_idx):", len(val_idx))
if config.num_channels == 3:
transforms = get_flips_colors_augmentation()
else:
# can't do hsv rescaling with multiband imagery, so skip this part
transforms = get_flips_shifts_augmentation()
train(ds,
fold,
train_idx,
val_idx,
config,
save_path,
log_path,
num_workers=num_workers,
transforms=transforms)
def train_stage2(sal_map: bool, three=False):
"""
train with other loss function
three = True ===> use only RGB for training
"""
im_type = TiffDemImageType
if sal_map:
im_type = SalImageType
if three:
im_type = TiffImageType
ds = ReadingImageProvider(im_type, paths, fn_mapping, image_suffix='RGB')
folds = get_folds(ds, 5)
num_workers = 0 if os.name == 'nt' else 8
train(ds,
folds,
config,
num_workers=num_workers,
transforms=augment_flips_color)
def train_stage1(sal_map: bool, three=False):
"""
main training stage with dtm/dsm data
three = True ===> use only RGB for training
updates channels from warm start with only RGB to final number of channels in config.num_channels
"""
im_type = TiffDemImageType
if sal_map:
im_type = SalImageType
if three:
im_type = TiffImageType
ds = ReadingImageProvider(im_type, paths, fn_mapping, image_suffix='RGB')
folds = get_folds(ds, 5)
num_workers = 0 if os.name == 'nt' else 8
train(ds,
folds,
config,
num_workers=num_workers,
transforms=augment_flips_color,
num_channels_changed=not three)
