def main():
# move up and down
train_data_sub_translation = []
for i in range(0, 3000):
train_data_sub_translation.extend(map(lambda x: img_translation(trainset_import.train_data[i], x), np.random.randint(1,4,10)))
# add noise
train_data_sub_noise = []
for i in range(0, 3000):
train_data_sub_noise.extend(map(lambda x: add_noise(trainset_import.train_data[i], x), range(0, 10)))
# rotate
train_data_sub_rotate = []
for i in range(0, 3000):
train_data_sub_rotate.extend(rotate_random(trainset_import.train_data[i]))
train_data_sub = train_data_sub_rotate + train_data_sub_translation + train_data_sub_noise
# label array
train_labels_sub = []
for i in range(0, 3000):
train_labels_sub.extend([trainset_import.train_labels[i]] * 10)
train_labels = train_labels_sub + train_labels_sub + train_labels_sub
train_labels = torch.from_numpy(np.array(train_labels))
transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
trainset_new = subMNIST(root='./data/', train=True, download=True, transform=transform, k=90000)
trainset_new.train_data = train_data_sub
trainset_new.train_labels = train_labels
pickle.dump(trainset_new, open("./data/train_labeled_allmethod.p", "wb" ))
def loadDataForLabeling():
train_loader2 = torch.utils.data.DataLoader(
trainset_new_unl, batch_size=1,
shuffle=False) #Careful must not be shuffled !!!WORKS LOSS IS REDUCED
#Add unlabeled test set
testAndLabel(1, trainedModel, train_loader2)
#Train on full newly labeled set
print('Beginning training loops phase 2')
#Load new Labels
train_labels_sub_unl = torch.from_numpy(
np.load("trainset_np_unlLabels.npy"))
train_labels_sub_unl = train_labels_sub_unl.type(torch.long)
trainset_new_unl.train_labels = train_labels_sub_unl
trainset_full = subMNIST(root='./data',
train=True,
transform=transform,
k=50000)
trainset_full.train_data = torch.cat(
(trainset_new_unl.train_data, trainset_new.train_data), 0)
trainset_full.train_labels = torch.cat(
(trainset_new_unl.train_labels, trainset_new.train_labels), 0)
train_loader3 = torch.utils.data.DataLoader(trainset_full,
batch_size=64,
shuffle=False)
print('Trainset train_data ' + str(trainset_full.train_data.size()))
return train_loader3
def join_MNIST_tensors(in1, in2):
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))])
joint_data = torch.cat((in1.train_data, in2.train_data), 0)
joint_labels = torch.cat((in1.train_labels, in2.train_labels), 0)
joint_total = subMNIST(root='./data',
train=True,
download=True,
transform=transform,
k=in1.train_data.size()[0] +
in2.train_data.size()[0])
joint_total.train_data = joint_data.clone()
joint_total.train_labels = joint_labels.clone()
return joint_total
da = DataAug()
augmented_data, augmented_label = da.dataaug(train_data_sub[0],
train_labels_sub[0])
for i in range(1, train_data_sub.shape[0]):
tdata, tlabel = da.dataaug(train_data_sub[i], train_labels_sub[i])
augmented_data = np.append(augmented_data, tdata, axis=0)
augmented_label = np.append(augmented_label, tlabel)
train_data_sub = np.append(train_data_sub, augmented_data, axis=0)
train_labels_sub = np.append(train_labels_sub, augmented_label, axis=0)
augdata = train_data_sub
auglabel = train_labels_sub
print(augdata.shape)
print(auglabel.shape)
train_data_sub = torch.from_numpy(augdata)
train_labels_sub = torch.from_numpy(auglabel)
print(train_labels_sub.size())
print(train_data_sub.size())
trainset_new = subMNIST(root='./data',
train=True,
download=True,
transform=transform,
k=18000)
trainset_new.train_data = train_data_sub.clone()
trainset_new.train_labels = train_labels_sub.clone()
pickle.dump(trainset_new, open("data/train_labeled_aug.p", "wb"))
def dataLoader():
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))])
trainset_original = datasets.MNIST('../data',
train=True,
download=True,
transform=transform)
train_label_index = []
valid_label_index = []
for i in range(10):
train_label_list = trainset_original.train_labels.numpy()
label_index = np.where(train_label_list == i)[0]
label_subindex = list(label_index[:300])
valid_subindex = list(label_index[300:1000 + 300])
train_label_index += label_subindex
valid_label_index += valid_subindex
#Train Set
trainset_np = trainset_original.train_data.numpy()
trainset_label_np = trainset_original.train_labels.numpy()
train_data_sub = torch.from_numpy(trainset_np[train_label_index])
train_labels_sub = torch.from_numpy(trainset_label_np[train_label_index])
trainset_new = subMNIST(root='./data',
train=True,
download=True,
transform=transform,
k=3000)
trainset_new.train_data = train_data_sub.clone()
trainset_new.train_labels = train_labels_sub.clone()
pickle.dump(trainset_new, open("train_labeled.p", "wb"))
#### Augmenting training set
## initialize trainset as usual
trainset_aug = subMNIST(root='./data',
train=True,
download=True,
transform=transform,
k=30000)
## turns out you can just repeat a tensor, cool
## http://pytorch.org/docs/tensors.html#torch.Tensor.repeat
trainset_aug.train_data = train_data_sub.clone().repeat(
10, 1, 1) ## 4 in the first axis
print(train_data_sub.size())
print(trainset_aug.train_data.size())
trainset_aug.train_labels = train_labels_sub.clone().repeat(
10) ## only one axi
print(train_labels_sub.size())
print(trainset_aug.train_labels.size())
## dims look correct!
## load scipy image tools and distributions for sampling
from scipy import ndimage
from scipy import stats
import random
random.seed(1337)
#from math import ceil, floor
## iterate through and augment
n = trainset_aug.train_data.size()[0] / 10
print(n)
iter_vals = range(0, n)
#iter_vals = range(0, 5)
'''
for i in iter_vals:
image_in = trainset_aug.train_data[i]
trainset_aug.train_data[i+n] = zoom_image(image_in)
trainset_aug.train_data[i+(n*2)] = translate_image(image_in)
trainset_aug.train_data[i+(n*3)] = rotate_image(image_in)
'''
for i in iter_vals:
image_in = trainset_aug.train_data[i]
trainset_aug.train_data[i + n] = zoom_image(image_in)
trainset_aug.train_data[i + (n * 2)] = zoom_image(image_in)
trainset_aug.train_data[i + (n * 3)] = zoom_image(image_in)
trainset_aug.train_data[i + (n * 4)] = translate_image(image_in)
trainset_aug.train_data[i + (n * 5)] = translate_image(image_in)
trainset_aug.train_data[i + (n * 6)] = translate_image(image_in)
trainset_aug.train_data[i + (n * 7)] = rotate_image(image_in)
trainset_aug.train_data[i + (n * 8)] = rotate_image(image_in)
trainset_aug.train_data[i + (n * 9)] = rotate_image(image_in)
## dump to pickle
pickle.dump(trainset_aug, open("train_labeled_aug.p", "wb"))
#Validation Set
validset_np = trainset_original.train_data.numpy()
validset_label_np = trainset_original.train_labels.numpy()
valid_data_sub = torch.from_numpy(validset_np[valid_label_index])
valid_labels_sub = torch.from_numpy(validset_label_np[valid_label_index])
validset = subMNIST(root='./data',
train=False,
download=True,
transform=transform,
k=10000)
validset.test_data = valid_data_sub.clone()
validset.test_labels = valid_labels_sub.clone()
pickle.dump(validset, open("validation.p", "wb"))
#Unlabeled Data
train_unlabel_index = []
for i in range(60000):
if i in train_label_index or i in valid_label_index:
pass
else:
train_unlabel_index.append(i)
trainset_np = trainset_original.train_data.numpy()
trainset_label_np = trainset_original.train_labels.numpy()
train_data_sub_unl = torch.from_numpy(trainset_np[train_unlabel_index])
#train_labels_sub_unl = torch.from_numpy(trainset_label_np[train_unlabel_index])
temp = np.empty(47000)
temp.fill(-1)
train_labels_sub_unl = torch.from_numpy(temp)
trainset_new_unl = subMNIST(root='./data',
train=True,
download=True,
transform=transform,
k=47000)
trainset_new_unl.train_data = train_data_sub_unl.clone()
trainset_new_unl.train_labels = train_labels_sub_unl.clone()
pickle.dump(trainset_new_unl, open("train_unlabeled.p", "wb"))
def split_dataset(trainset_original,
n_train_labels_pc=10,
n_train_unlabeled_pc=None,
n_validation_pc=1000):
"""
Parameters
----------
trainset_original : torch.utils.data.Dataset
A dataset object as defined by torch
n_train_labels_pc : int
Number of labeled samples per class to use for training.
n_train_unlabeled_pc : int
Number of unlabeled samples per class to use for training.
n_validation_pc : int
Number of labeled samples per class to use for validation.
Returns
-------
trainset_new, validset, trainset_new_unl: torch.utils.data.Dataset objects
"""
train_label_index = []
train_unlabel_index = []
valid_label_index = []
classes = np.unique(trainset_original.train_labels.numpy())
n_classes = len(classes)
for i in range(n_classes):
train_label_list = trainset_original.train_labels.numpy()
label_index = np.where(train_label_list == i)[0]
n_class_samples = len(label_index)
n_tv = n_train_labels_pc + n_validation_pc
if n_train_unlabeled_pc is not None:
n_tv += n_train_unlabeled_pc
if n_tv > n_class_samples:
raise ValueError('Class {} has not enough samples ({}) to split '
'in training labeled, training unlabeled and '
'validation set'.format(classes[i],
n_class_samples))
label_subindex = list(label_index[:n_train_labels_pc])
ind_end = n_train_labels_pc + n_validation_pc
valid_subindex = list(label_index[n_train_labels_pc:ind_end])
ind_start = ind_end
if n_train_unlabeled_pc is not None:
ind_end += n_train_labels_pc
else:
ind_end = n_class_samples
unlabel_subindex = list(label_index[ind_start:ind_end])
train_label_index += label_subindex
valid_label_index += valid_subindex
train_unlabel_index += unlabel_subindex
trainset_np = trainset_original.train_data.numpy()
trainset_label_np = trainset_original.train_labels.numpy()
train_data_sub = torch.from_numpy(trainset_np[train_label_index])
train_labels_sub = torch.from_numpy(trainset_label_np[train_label_index])
trainset_new = subMNIST(root='./../data',
train=True,
download=True,
transform=mnist_transform,
k=n_train_labels_pc * n_classes)
trainset_new.train_data = train_data_sub.clone()
trainset_new.train_labels = train_labels_sub.clone()
# pickle.dump(trainset_new, open("./../data/train_labeled.p", "wb"))
validset_np = trainset_original.train_data.numpy()
validset_label_np = trainset_original.train_labels.numpy()
valid_data_sub = torch.from_numpy(validset_np[valid_label_index])
valid_labels_sub = torch.from_numpy(validset_label_np[valid_label_index])
validset = subMNIST(root='./../data',
train=False,
download=True,
transform=mnist_transform,
k=n_validation_pc * n_classes)
validset.test_data = valid_data_sub.clone()
validset.test_labels = valid_labels_sub.clone()
# pickle.dump(validset, open("./../data/validation.p", "wb"))
n_unlabeled_set = len(train_unlabel_index)
trainset_np = trainset_original.train_data.numpy()
trainset_label_np = trainset_original.train_labels.numpy()
train_data_sub_unl = torch.from_numpy(trainset_np[train_unlabel_index])
# train_labels_sub_unl = torch.from_numpy(trainset_label_np[train_unlabel_index])
trainset_new_unl = subMNIST(root='./../data',
train=True,
download=True,
transform=mnist_transform,
k=n_unlabeled_set)
trainset_new_unl.train_data = train_data_sub_unl.clone()
trainset_new_unl.train_labels = None # Unlabeled
# pickle.dump(trainset_new_unl, open("./../data/train_unlabeled.p", "wb"))
return trainset_new, trainset_new_unl, validset
train_label_index = []
valid_label_index = []
for i in range(10):
train_label_list = trainset_original.train_labels.numpy()
label_index = np.where(train_label_list == i)[0]
label_subindex = list(label_index[:300])
valid_subindex = list(label_index[300: 1000 + 300])
train_label_index += label_subindex
valid_label_index += valid_subindex
trainset_np = trainset_original.train_data.numpy()
trainset_label_np = trainset_original.train_labels.numpy()
train_data_sub = torch.from_numpy(trainset_np[train_label_index])
train_labels_sub = torch.from_numpy(trainset_label_np[train_label_index])
trainset_new = subMNIST(root='data', train=True, download=True, transform=transform, k=3000)
trainset_new.data = train_data_sub.clone()
trainset_new.targets = train_labels_sub.clone()
pickle.dump(trainset_new, open("data/train_labeled.p", "wb"))
validset_np = trainset_original.train_data.numpy()
validset_label_np = trainset_original.train_labels.numpy()
valid_data_sub = torch.from_numpy(validset_np[valid_label_index])
valid_labels_sub = torch.from_numpy(validset_label_np[valid_label_index])
validset = subMNIST(root='data', train=False, download=True, transform=transform, k=10000)
validset.data = valid_data_sub.clone()
validset.targets = valid_labels_sub.clone()
def all_transformations(dataset_pickle,
data_type,
type_transformation=['rotation'],
value_rotation=45.0,
value_scale=0.8,
distance_translation=0.1,
direction_translation='right',
horizontal_translation=True,
elastic_alpha=34,
elastic_sigma=4):
'''
This function takes the pickle data (train and test) and pre-process with 3 different data-augmentation techniques.
We can change default values of each transformation.
INPUT: dataset_pickle (pickle pytorch),
data_type ='train' or 'train_unlabel' or 'test',
type_transformation=['rotation'], ['scale'], ['translation'], combinations (i.e. ['rotation','scale']) or ['all'],
RETURNS: Transformed data, ready to load with torch.utils.data.DataLoader
'''
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))])
if data_type == 'train':
dataset_loader_numpy = dataset_pickle.train_data.numpy()
dataset_label_numpy = dataset_pickle.train_labels.numpy()
dataset_import = subMNIST(root='./data',
train=True,
transform=transform,
download=True,
k=dataset_pickle.train_data.size()[0])
if data_type == 'train_unlabel':
dataset_loader_numpy = dataset_pickle.train_data.numpy()
dataset_label_numpy = dataset_pickle.train_labels.numpy()
dataset_import = subMNIST(root='./data',
train=True,
transform=transform,
download=True,
k=dataset_pickle.train_data.size()[0])
if type_transformation == ['rotation']:
dataset_loader_numpy_transformed = np.array(
[rotation(x, value_rotation) for x in dataset_loader_numpy])
if type_transformation == ['elastic']:
dataset_loader_numpy_transformed = np.array([
elastic_transform(x, elastic_alpha, elastic_sigma)
for x in dataset_loader_numpy
])
if type_transformation == ['scale']:
dataset_loader_numpy_transformed = np.array(
[scale(x, value_scale) for x in dataset_loader_numpy])
if type_transformation == ['translation']:
dataset_loader_numpy_transformed = np.array([
translation(x,
distance_translation,
direction=direction_translation,
horizontal=horizontal_translation)
for x in dataset_loader_numpy
])
if type_transformation == ['rotation', 'scale'] or type_transformation == [
'scale', 'rotation'
]:
dataset_loader_numpy_transformed = np.array(
[rotation(x, value_rotation) for x in dataset_loader_numpy])
dataset_loader_numpy_transformed = np.array(
[scale(x, value_scale) for x in dataset_loader_numpy_transformed])
if type_transformation == [
'rotation', 'translation'
] or type_transformation == ['translation', 'rotation']:
dataset_loader_numpy_transformed = np.array(
[rotation(x, value_rotation) for x in dataset_loader_numpy])
dataset_loader_numpy_transformed = np.array([
translation(x,
distance_translation,
direction=direction_translation,
horizontal=horizontal_translation)
for x in dataset_loader_numpy_transformed
])
if type_transformation == [
'scale', 'translation'
] or type_transformation == ['translation', 'scale']:
dataset_loader_numpy_transformed = np.array(
[scale(x, value_scale) for x in dataset_loader_numpy])
dataset_loader_numpy_transformed = np.array([
translation(x,
distance_translation,
direction=direction_translation,
horizontal=horizontal_translation)
for x in dataset_loader_numpy_transformed
])
if type_transformation == ['rotation', 'scale', 'translation', 'elastic']:
dataset_loader_numpy_transformed = np.array(
[rotation(x, value_rotation) for x in dataset_loader_numpy])
dataset_loader_numpy_transformed = np.array(
[scale(x, value_scale) for x in dataset_loader_numpy_transformed])
dataset_loader_numpy_transformed = np.array([
translation(x,
distance_translation,
direction=direction_translation,
horizontal=horizontal_translation)
for x in dataset_loader_numpy_transformed
])
dataset_loader_numpy_transformed = np.array([
elastic_transform(x, elastic_alpha, elastic_sigma)
for x in dataset_loader_numpy_transformed
])
if type_transformation == ['rotation', 'elastic']:
dataset_loader_numpy_transformed = np.array(
[rotation(x, value_rotation) for x in dataset_loader_numpy])
dataset_loader_numpy_transformed = np.array([
elastic_transform(x, elastic_alpha, elastic_sigma)
for x in dataset_loader_numpy_transformed
])
if type_transformation == ['all']:
dataset_loader_numpy_transformed = np.array(
[rotation(x, value_rotation) for x in dataset_loader_numpy])
dataset_loader_numpy_transformed = np.array(
[scale(x, value_scale) for x in dataset_loader_numpy_transformed])
dataset_loader_numpy_transformed = np.array([
translation(x,
distance_translation,
direction=direction_translation,
horizontal=horizontal_translation)
for x in dataset_loader_numpy_transformed
])
dataset_loader_preprocessed = torch.from_numpy(
dataset_loader_numpy_transformed)
dataset_loader_preprocessed2 = dataset_loader_preprocessed.type(
torch.ByteTensor)
if data_type == 'train':
print("TRAIN TYPE")
dataset_import.train_data = dataset_loader_preprocessed2.clone()
dataset_import.train_labels = torch.from_numpy(
dataset_label_numpy).clone()
if data_type == 'train_unlabel':
dataset_import.train_data = dataset_loader_preprocessed2.clone()
dataset_import.train_labels = torch.from_numpy(
np.repeat(-1,
dataset_import.train_data.size()[0])).clone()
return dataset_import