# 'offset': (8, 20, 20),
'out_channels': 2,
# 'in_memory': True # Uncomment to avoid disk I/O (if you have enough host memory for the data)
}
norm_mean = (155.291411, )
norm_std = (42.599973, )
valid_transform = transforms.Normalize(mean=norm_mean,
std=norm_std,
inplace=True)
print('Loading dataset...')
valid_dataset = PatchCreator(
input_sources=input_h5data,
target_sources=target_h5data,
train=False,
epoch_size=40, # How many samples to use for each validation run
warp_prob=0,
warp_kwargs={'sample_aniso': aniso_factor != 1},
transform=valid_transform,
**common_data_kwargs)
valid_loader = DataLoader(valid_dataset, num_workers=4, pin_memory=True)
# Validation metrics
valid_metrics = { # mean metrics
'val_accuracy_mean': metrics.Accuracy(),
'val_precision_mean': metrics.Precision(),
'val_recall_mean': metrics.Recall(),
'val_DSC_mean': metrics.DSC(),
'val_IoU_mean': metrics.IoU(),
}
if valid_dataset.out_channels > 2:
aniso_factor = 2 # Anisotropy in z dimension. E.g. 2 means half resolution in z dimension.
common_data_kwargs = { # Common options for training and valid sets.
'aniso_factor': aniso_factor,
'patch_shape': (44, 88, 88),
# 'offset': (8, 20, 20),
# 'in_memory': True # Uncomment to avoid disk I/O (if you have enough host memory for the data)
}
train_dataset = PatchCreator(input_sources=[
input_h5data[i] for i in range(len(input_h5data)) if i not in valid_indices
],
target_sources=[
target_h5data[i]
for i in range(len(input_h5data))
if i not in valid_indices
],
train=True,
epoch_size=args.epoch_size,
warp_prob=0.2,
warp_kwargs={
'sample_aniso': aniso_factor != 1,
'perspective': True,
'warp_amount': 1.0,
},
transform=train_transform,
**common_data_kwargs)
valid_dataset = None if not valid_indices else PatchCreator(
input_sources=[
input_h5data[i] for i in range(len(input_h5data)) if i in valid_indices
],
target_sources=[
target_h5data[i] for i in range(len(input_h5data))
if i in valid_indices
from elektronn3.data.knossos import KnossosRawData
train_dataset = KnossosRawData(
conf_path=
'/wholebrain/songbird/j0126/areaxfs_v5/knossosdatasets/mag1/knossos.conf',
patch_shape=common_data_kwargs['patch_shape'],
transform=train_transform,
epoch_size=args.epoch_size,
mode='caching',
)
else:
train_dataset = PatchCreator(input_sources=[
input_h5data[i] for i in range(len(input_h5data))
if i not in valid_indices
],
train=True,
epoch_size=args.epoch_size,
warp_prob=0,
transform=train_transform,
**common_data_kwargs)
valid_dataset = PatchCreator(input_sources=[
input_h5data[i] for i in range(len(input_h5data)) if i in valid_indices
],
train=False,
epoch_size=args.epoch_size,
warp_prob=0,
transform=valid_transform,
**common_data_kwargs)
# Use first validation cube for previews. Can be set to any other data source.
'num_classes': 6,
# 'offset': (20, 46, 46),
'target_discrete_ix': [3, 4, 5]
}
type_args = list(range(len(input_h5data)))
train_dataset = PatchCreator(input_h5data=[
input_h5data[i] for i in type_args if i not in valid_indices
],
target_h5data=[
target_h5data[i] for i in type_args
if i not in valid_indices
],
train=True,
epoch_size=args.epoch_size,
cube_prios=None,
warp_prob=0.2,
in_memory=True,
warp_kwargs={
'sample_aniso': aniso_factor != 1,
'perspective': True,
'warp_amount': 0.05,
},
transform=train_transform,
**common_data_kwargs)
valid_dataset = None if (valid_indices is None) or (
len(valid_indices) == 0) else PatchCreator(
input_h5data=[
input_h5data[i] for i in range(len(input_h5data))
if i in valid_indices
],
])
valid_transform = transforms.Compose(common_transforms + [])
# Specify data set
common_data_kwargs = { # Common options for training and valid sets.
'aniso_factor': 2,
'patch_shape': (48, 96, 96),
'classes': [0, 1],
}
train_dataset = PatchCreator(
input_h5data=input_h5data[:2],
target_h5data=target_h5data[:2],
train=True,
epoch_size=args.epoch_size,
warp=0.5,
warp_kwargs={
'sample_aniso': True,
'perspective': True,
},
transform=train_transform,**common_data_kwargs
)
valid_dataset = PatchCreator(
input_h5data=[input_h5data[2]],
target_h5data=[target_h5data[2]],
train=False,
epoch_size=10, # How many samples to use for each validation run
preview_shape=(64, 144, 144),
warp=0,
warp_kwargs={
'sample_aniso': True,