os.chdir(os.path.join(base_dir, 'src/training/'))
from sampling import DataLoader, CSVDataset
from sampling import transforms as tx
from models import create_unet_model3D
from keras import callbacks as cbks
# tx.Compose lets you string together multiple transforms
co_tx = tx.Compose([
tx.TypeCast('float32'),
tx.ExpandDims(axis=-1),
tx.RandomAffine(
rotation_range=(-15, 15), # rotate btwn -15 & 15 degrees
translation_range=(
0.1, 0.1), # translate btwn -10% and 10% horiz, -10% and 10% vert
shear_range=(-10, 10), # shear btwn -10 and 10 degrees
zoom_range=(0.85, 1.15), # between 15% zoom-in and 15% zoom-out
turn_off_frequency=5,
fill_value='min',
target_fill_mode='constant',
target_fill_value=0
) # how often to just turn off random affine transform (units=#samples)
])
input_tx = tx.MinMaxScaler((-1, 1)) # scale between -1 and 1
target_tx = tx.OneHot(
) # convert segmentation image to One-Hot representation for cross-entropy loss
# use a co-transform, meaning the same transform will be applied to input+target images at the same time
# this is necessary since Affine transforms have random parameter draws which need to be shared
base_dir = '/users/ncullen/desktop/projects/unet-ants/'
os.chdir(base_dir + 'code/')
# local imports
from sampling import DataLoader, CSVDataset
from sampling import transforms as tx
from models import create_unet_model2D
data_dir = base_dir + 'data_2D/'
results_dir = base_dir + 'results_2D/'
# tx.Compose lets you string together multiple transforms
co_tx = tx.Compose([
tx.MinMaxScaler((-1, 1)),
tx.ExpandDims(
axis=-1
) # expand from (128,128) to (128,128,1) -> RandomAffine and Keras expect that
])
# use a co-transform, meaning the same transform will be applied to input+target images at the same time
# this is necessary since Affine transforms have random parameter draws which need to be shared
dataset = CSVDataset(
filepath=data_dir + 'image_filemap.csv',
base_path=
data_dir, # this path will be appended to all of the filenames in the csv file
input_cols=[
'images'
], # column in dataframe corresponding to inputs (can be an integer also)
target_cols=[
'images'
], # column in dataframe corresponding to targets (can be an integer also)
from sampling import DataLoader, CSVDataset
from sampling import transforms as tx
from models import create_unet_model2D
data_dir = base_dir + 'data_2D/'
results_dir = base_dir+'results_2D/'
input_tx = tx.MinMaxScaler((0,1)) # scale input images between 0 and 1
target_tx = tx.BinaryMask(cutoff=0.5) # convert target segmentation to a binary mask
# tx.Compose lets you string together multiple transforms
co_tx = tx.Compose([tx.ExpandDims(axis=-1), # expand from (128,128) to (128,128,1) -> RandomAffine and Keras expect that
tx.RandomAffine(rotation_range=(-15,15), # rotate btwn -15 & 15 degrees
translation_range=(0.1,0.1), # translate btwn -10% and 10% horiz, -10% and 10% vert
shear_range=(-10,10), # shear btwn -10 and 10 degrees
zoom_range=(0.85,1.15), # between 15% zoom-in and 15% zoom-out
turn_off_frequency=5) # how often to just turn off random affine transform (units=#samples)
])
# use a co-transform, meaning the same transform will be applied to input+target images at the same time
# this is necessary since Affine transforms have random parameter draws which need to be shared
dataset = CSVDataset(filepath=data_dir+'image_filemap.csv',
base_path=data_dir, # this path will be appended to all of the filenames in the csv file
input_cols=['images'], # column in dataframe corresponding to inputs (can be an integer also)
target_cols=['masks'],# column in dataframe corresponding to targets (can be an integer also)
input_transform=input_tx, target_transform=target_tx, co_transform=co_tx)
# split into train and test set based on the `train-test` column in the csv file
# this splits alphabetically by values, and since 'test' comes before 'train' thus val_data is returned before train_data
data_dir = base_dir + 'data_3D/'
results_dir = base_dir+'results_3D/'
try:
os.mkdir(results_dir)
except:
pass
# tx.Compose lets you string together multiple transforms
co_tx = tx.Compose([tx.TypeCast('float32'),
tx.MinMaxScaler((-1,1)), # scale between -1 and 1
tx.ExpandDims(axis=-1), # expand from (128,128,128) to (128,128,128,1) -> RandomAffine and Keras expect that
tx.RandomAffine(rotation_range=(-15,15), # rotate btwn -15 & 15 degrees
translation_range=(0.1,0.1), # translate btwn -10% and 10% horiz, -10% and 10% vert
shear_range=(-10,10), # shear btwn -10 and 10 degrees
zoom_range=(0.85,1.15), # between 15% zoom-in and 15% zoom-out
turn_off_frequency=5,
fill_value='min',
target_fill_mode='constant',
target_fill_value='min') # how often to just turn off random affine transform (units=#samples)
])
# use a co-transform, meaning the same transform will be applied to input+target images at the same time
# this is necessary since Affine transforms have random parameter draws which need to be shared
dataset = CSVDataset(filepath=data_dir+'image_filemap.csv',
base_path=os.path.join(data_dir,'images'), # this path will be appended to all of the filenames in the csv file
input_cols=['Images'], # column in dataframe corresponding to inputs (can be an integer also)
target_cols=['Images'],# column in dataframe corresponding to targets (can be an integer also)
co_transform=co_tx)