class WG_IFN_Dataset(data.Dataset):
def __init__(self, cf, train=True, transform=None, data_idx_WG = np.arange(1),
data_idx_IFN = np.arange(1), complement_idx=False):
self.train = train # training set or test set
if len(data_idx_WG)==1:
self.datasetWG = WashingtonDataset(cf, train=self.train, transform = transform)
else:
self.datasetWG = WashingtonDataset(cf, train=self.train, transform = transform,
data_idx = data_idx_WG, complement_idx = True)
if len(data_idx_IFN)==1:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform)
else:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform,
data_idx = data_idx_IFN, complement_idx = True)
self.data_idx_WG = self.datasetWG.data_idx # this is needed, to be passed from one set to another
self.data_idx_IFN = self.datasetIFN.data_idx # this is needed, to be passed from one set to another
def add_weights_of_words(self): # weights to balance the loss, if the data is unbalanced
self.datasetWG.add_weights_of_words()
self.datasetIFN.add_weights_of_words()
def num_classes(self):
return self.datasetIFN.num_classes() #IFN and WG have the same phoc size
def __getitem__(self, index):
if index < len(self.datasetWG):
return self.datasetWG[index]
else:
return self.datasetIFN[index - len(self.datasetWG)] # check: are we skipping a sample here?
def __len__(self):
return len(self.datasetWG) + len(self.datasetIFN)
def __init__(self, cf, train=True, transform=None, data_idx_WG = np.arange(1),
data_idx_IFN = np.arange(1), complement_idx=False):
self.train = train # training set or test set
if len(data_idx_WG)==1:
self.datasetWG = WashingtonDataset(cf, train=self.train, transform = transform)
else:
self.datasetWG = WashingtonDataset(cf, train=self.train, transform = transform,
data_idx = data_idx_WG, complement_idx = True)
if len(data_idx_IFN)==1:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform)
else:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform,
data_idx = data_idx_IFN, complement_idx = True)
self.data_idx_WG = self.datasetWG.data_idx # this is needed, to be passed from one set to another
self.data_idx_IFN = self.datasetIFN.data_idx # this is needed, to be passed from one set to another
def __init__(self, cf, train=True, transform=None):
# cf.train_split = False # this should always be false, as we are keeping one folder for testing
self.train = train # training set or test set
trn_folder = cf.folders_to_use.replace(
cf.IFN_test[-1], '') # removing the test set from train folders
# backing up the original paths
dataset_path = cf.dataset_path_IFN
gt_path = cf.gt_path_IFN
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[0])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[0])
self.datasetIFN_1 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[1])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[1])
self.datasetIFN_2 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[2])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[2])
self.datasetIFN_3 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[3])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[3])
self.datasetIFN_4 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
self.IFN_1_len = len(self.datasetIFN_1)
self.IFN_2_len = len(self.datasetIFN_2)
self.IFN_3_len = len(self.datasetIFN_3)
self.IFN_4_len = len(self.datasetIFN_4)
cf.dataset_path_IFN = dataset_path
cf.gt_path_IFN = gt_path # this needs to be used in loading the test set
class IAM_IFN_Dataset(data.Dataset):
def __init__(self, cf, train=True, mode = 'train', transform=None,
data_idx_IAM = np.arange(1),
data_idx_IFN = np.arange(1),
complement_idx=False):
self.train = train # training set or test set
self.mode = mode
if len(data_idx_IFN)==1:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform)
else:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform,
data_idx = data_idx_IFN, complement_idx = True)
if len(data_idx_IAM)==1:
if mode == 'train':
self.datasetIAM = iam_train_valid_combined_dataset(cf, train=True, transform = transform) # mode is one of train, test, or validate
else:
assert(mode == 'test')
self.datasetIAM = IAM_words(cf, mode = self.mode, transform = transform)
else:
# this is deprecated for IAM dataset as we are splitting based on train, validate, and test folders
# self.datasetIAM = IAM_words(cf, mode = self.mode, transform = transform)
print('Deprecated by Rawi, as the split is based on train, validate and test')
self.data_idx_IFN = self.datasetIFN.data_idx # this is needed, to be passed from one set to another
# self.data_idx_IAM = self.datasetIAM.data_idx # this is needed, to be passed from one set to another
def add_weights_of_words(self): # weights to balance the loss, if the data is unbalanced
self.datasetIFN.add_weights_of_words()
self.datasetIAM.add_weights_of_words()
def num_classes(self):
return self.datasetIAM.num_classes() #IFN and WG have the same phoc size
def __getitem__(self, index):
if index < len(self.datasetIFN):
return self.datasetIFN[index]
else:
return self.datasetIAM[index - len(self.datasetIFN)] # check: are we skipping a sample here?
def __len__(self):
return len(self.datasetIFN) + len(self.datasetIAM)
def get_ifn(cf, image_transform):
print('...................Loading IFN dataset...................')
if not (cf.IFN_based_on_folds_experiment):
''' randomly split training and testing according to split percentage
the folder left for tesing is the cf.IFN_test '''
train_set = IfnEnitDataset(
cf, train=True, transform=image_transform['image_transform_hdr'])
test_set = IfnEnitDataset(
cf,
train=False,
transform=image_transform['image_transform_hdr'],
data_idx=train_set.data_idx,
complement_idx=True)
else:
''' leave one folder out of 'abcde' folders '''
train_set = IFN_XVAL_Dataset(
cf, train=True, transform=image_transform['image_transform_hdr'])
test_set = IfnEnitDataset(
cf, train=False, transform=image_transform['image_transform_hdr'])
return train_set, test_set
def test_dataload(cf):
logger = logging.getLogger('test_dataloader_wg')
# Image transformations
if cf.pad_images:
pad_image = PadImage(
(globals.MAX_IMAGE_WIDTH, globals.MAX_IMAGE_HEIGHT))
if cf.resize_images:
if cf.pad_images:
image_transfrom = transforms.Compose([
pad_image,
transforms.ToPILImage(),
transforms.Scale((cf.input_size[0], cf.input_size[1])),
transforms.ToTensor()
])
else:
image_transfrom = transforms.Compose([
transforms.ToPILImage(),
transforms.Scale((cf.input_size[0], cf.input_size[1])),
transforms.ToTensor()
])
else:
if cf.pad_images:
image_transfrom = transforms.Compose(
[pad_image, transforms.ToTensor()])
else:
image_transfrom = transforms.ToTensor()
if cf.dataset_name == 'WG':
input_dataset = WashingtonDataset(cf, transform=image_transfrom)
elif cf.dataset_name == 'IFN':
input_dataset = IfnEnitDataset(cf, transform=image_transfrom)
else:
logger.fatal('The dataset \'%s\' is unknown. Use: [WG, IFN]',
cf.dataset_name)
sys.exit(0)
# dataloader = DataLoader(input_dataset, batch_size=cf.batch_size,
# shuffle=cf.shuffle, num_workers=cf.num_workers)
for i in range(len(input_dataset)):
plt.figure(i)
plt.xticks([])
plt.yticks([])
data, target = input_dataset[i]
plt.imshow(data.numpy()[0, :, :], 'gray')
plt.show()
if i == 102: break
def __init__(self, cf, train=True, mode = 'train', transform=None,
data_idx_IAM = np.arange(1),
data_idx_IFN = np.arange(1),
complement_idx=False):
self.train = train # training set or test set
self.mode = mode
if len(data_idx_IFN)==1:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform)
else:
self.datasetIFN = IfnEnitDataset(cf, train=self.train, transform = transform,
data_idx = data_idx_IFN, complement_idx = True)
if len(data_idx_IAM)==1:
if mode == 'train':
self.datasetIAM = iam_train_valid_combined_dataset(cf, train=True, transform = transform) # mode is one of train, test, or validate
else:
assert(mode == 'test')
self.datasetIAM = IAM_words(cf, mode = self.mode, transform = transform)
else:
# this is deprecated for IAM dataset as we are splitting based on train, validate, and test folders
# self.datasetIAM = IAM_words(cf, mode = self.mode, transform = transform)
print('Deprecated by Rawi, as the split is based on train, validate and test')
self.data_idx_IFN = self.datasetIFN.data_idx # this is needed, to be passed from one set to another
class IFN_XVAL_Dataset(data.Dataset):
def __init__(self, cf, train=True, transform=None):
# cf.train_split = False # this should always be false, as we are keeping one folder for testing
self.train = train # training set or test set
trn_folder = cf.folders_to_use.replace(
cf.IFN_test[-1], '') # removing the test set from train folders
# backing up the original paths
dataset_path = cf.dataset_path_IFN
gt_path = cf.gt_path_IFN
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[0])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[0])
self.datasetIFN_1 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[1])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[1])
self.datasetIFN_2 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[2])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[2])
self.datasetIFN_3 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
cf.dataset_path_IFN = dataset_path.replace(cf.IFN_test,
'set_' + trn_folder[3])
cf.gt_path_IFN = gt_path.replace(cf.IFN_test, 'set_' + trn_folder[3])
self.datasetIFN_4 = IfnEnitDataset(cf,
train=self.train,
transform=transform)
self.IFN_1_len = len(self.datasetIFN_1)
self.IFN_2_len = len(self.datasetIFN_2)
self.IFN_3_len = len(self.datasetIFN_3)
self.IFN_4_len = len(self.datasetIFN_4)
cf.dataset_path_IFN = dataset_path
cf.gt_path_IFN = gt_path # this needs to be used in loading the test set
def __getitem__(self, index):
if index < self.IFN_1_len:
return self.datasetIFN_1[index]
elif index < (self.IFN_1_len + self.IFN_2_len):
index = index - (self.IFN_1_len)
return self.datasetIFN_2[
index] # check: are we skipping a sample here?
elif index < (self.IFN_1_len + self.IFN_2_len + self.IFN_3_len):
index = index - (self.IFN_1_len + self.IFN_2_len)
return self.datasetIFN_3[index]
else: # This is IFN_4
index = index - (self.IFN_1_len + self.IFN_2_len + self.IFN_3_len)
return self.datasetIFN_4[index]
def __len__(self):
return self.IFN_1_len + self.IFN_2_len + self.IFN_3_len + self.IFN_4_len
def add_weights_of_words(
self): # weights to balance the loss, if the data is unbalanced
self.datasetIFN_1.add_weights_of_words()
def num_classes(self):
return self.datasetIFN_1.num_classes(
) # Does not matter which one as they all have the same phoc length