def create_ellipses_dataloader(num_images_train=100,
num_images_test=30,
resolution=(64, 64)):
"""
Creating dataloders of ellipses from dataset
param: num_images_train: int: number of ellipses in train loader
param: num_images_test: int: number of ellipses in test loader
param: resolution: tuple(int,int): image resolution
return: tuple of dataloaders for train and test
"""
ToTensor_pair = Lambda(lambda image_mask_pair: (torch.Tensor(
image_mask_pair[0]), torch.Tensor(image_mask_pair[1])))
Transpose_pair = Lambda(lambda image_mask_pair: ((image_mask_pair[
0]).transpose(0, 2).transpose(1, 2), torch.Tensor(image_mask_pair[1])))
e_train = Ellipses(
num_images=100,
image_shape=(64, 64),
transform=Compose([RandomD4_for_pair(), ToTensor_pair,
Transpose_pair]))
e_test = Ellipses(num_images=30,
image_shape=(64, 64),
transform=Compose([ToTensor_pair, Transpose_pair]))
ellipses_train_loader = DataLoader(e_train, batch_size=2, shuffle=False)
ellipses_test_loader = DataLoader(e_test, batch_size=2, shuffle=False)
return ellipses_train_loader, ellipses_test_loader
def __getitem__(self, index):
transforms = list()
transforms.append(
Lambda(self.__to_numpy
)) # First convert PIL.Image.Image to numpy.ndarray. HxWxC
transforms.append(Lambda(self.__per_pixel_subtraction_normalization)
) # Subtract per-pixel mean
if self.train:
if random.random() > 0.5:
transforms.append(Lambda(self.__horizontal_flip)
) # Flip horizontally with 50:50 chance
transforms.append(Lambda(self.__pad_and_random_crop)
) # Pad 0 along H and W dims and randomly crop
transforms.append(
ToTensor()
) # convert numpy.ndarray to torch.tensor (notice that HxWxC -> CxHxW)
transforms = Compose(transforms)
if self.train:
return transforms(
self.cifar10_train[index][0]), self.cifar10_train[index][1]
else:
return transforms(
self.cifar10_test[index][0]), self.cifar10_test[index][1]
def get_transforms(kmeans):
clusters = kmeans.kmeans.n_clusters
_transform_colorisation = Compose([
Resize((32, 32)),
ToNumpy(),
ConvertChannel(),
QuantizeAB(kmeans),
OneHotEncoding(clusters),
ToTensor()
])
transform_colorisation = Compose([
Lambda(lambda batch: torch.stack(
[_transform_colorisation(im) for im in batch]))
])
_transform_training = Compose([
Resize((256, 256)),
Grayscale(),
ToTensor(),
Normalize(mean=config.IMAGE_MEAN, std=config.IMAGE_STD)
])
transform_training = Compose([
Lambda(lambda batch: torch.stack(
[_transform_training(im) for im in batch]))
])
_transform_testing = Compose(
[Resize((256, 256)),
ToNumpy(), ConvertChannel()])
transform_testing = Compose(
[Lambda(lambda batch: [_transform_testing(im) for im in batch])])
return [transform_training, transform_colorisation, transform_testing]
def valid_lr_transform(crop_size, normalize=True):
return Compose([
ToPILImage(),
Resize(crop_size, interpolation=Image.BICUBIC),
ToTensor(),
Lambda(imagenet_normalise) if normalize else Lambda(lambda img: img),
])
def get_params(brightness, contrast, saturation, hue):
"""Get a randomized transform to be applied on image.
Arguments are same as that of __init__.
Returns:
Transform which randomly adjusts brightness, contrast and
saturation in a random order.
"""
transforms = []
brightness_factor = random.uniform(max(0, brightness[0]),
brightness[1])
transforms.append(
Lambda(lambda img: F.adjust_brightness(img, brightness_factor)))
contrast_factor = random.uniform(max(0, contrast[0]), contrast[1])
transforms.append(
Lambda(lambda img: F.adjust_contrast(img, contrast_factor)))
saturation_factor = random.uniform(max(0, saturation[0]),
saturation[1])
transforms.append(
Lambda(lambda img: F.adjust_saturation(img, saturation_factor)))
hue_factor = random.uniform(-hue[0], hue[1])
transforms.append(Lambda(lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
transform = Compose(transforms)
return transform
def get_dataloader(batch_size):
train_data = KgForestDataset(
split='train-37479',
transform=Compose([
Lambda(lambda x: randomShiftScaleRotate(
x, u=0.75, shift_limit=6, scale_limit=6, rotate_limit=45)),
Lambda(lambda x: randomFlip(x)),
Lambda(lambda x: randomTranspose(x)),
Lambda(lambda x: toTensor(x)),
Normalize(mean=mean, std=std)
]),
height=256,
width=256)
train_data_loader = DataLoader(batch_size=batch_size,
dataset=train_data,
shuffle=True)
validation = KgForestDataset(
split='validation-3000',
transform=Compose([
# Lambda(lambda x: randomShiftScaleRotate(x, u=0.75, shift_limit=6, scale_limit=6, rotate_limit=45)),
# Lambda(lambda x: randomFlip(x)),
# Lambda(lambda x: randomTranspose(x)),
Lambda(lambda x: toTensor(x)),
Normalize(mean=mean, std=std)
]),
height=256,
width=256)
valid_dataloader = DataLoader(dataset=validation,
shuffle=False,
batch_size=batch_size)
return train_data_loader, valid_dataloader
def load_image(filename, root="", ensemble=False):
"""Load an image.
:param filename: path to image
:param root: can be specified, if filename is a relative path
:param ensemble: if `True`, perform ten crops and return that
instead.
:returns: an image of dimension (1,3,224,224) if `ensemble`
is `False`, otherwise (10,3,224,224).
"""
transform_list = []
if ensemble:
norm = Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
transform_list.append(Resize((256, 256)))
transform_list.append(TenCrop(224))
transform_list.append(
Lambda(lambda crops: torch.stack(
[ToTensor()(crop) for crop in crops])))
transform_list.append(
Lambda(lambda crops: torch.stack([norm(crop) for crop in crops])))
else:
transform_list.append(Resize((224, 224)))
transform_list.append(ToTensor())
transform_list.append(Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)))
transformer = Compose(transform_list)
filepath = "%s/%s" % (root, filename)
img = Image.open(filepath)
img = img.convert("RGB")
img = transformer(img)
if ensemble:
return img
else:
return img.unsqueeze(0)
def get_params(brightness, contrast, saturation):
"""Get a randomized transform to be applied on image.
Arguments are same as that of __init__.
Returns:
Transform which randomly adjusts brightness, contrast and
saturation in a random order.
"""
transforms = []
if brightness is not None:
brightness_factor = random.gauss(brightness[0], brightness[1])
transforms.append(
Lambda(lambda img: adjust_brightness(img, brightness_factor)))
if contrast is not None:
contrast_factor = random.gauss(contrast[0], contrast[1])
transforms.append(
Lambda(lambda img: adjust_contrast(img, contrast_factor)))
if saturation is not None:
saturation_factor = random.gauss(saturation[0], saturation[1])
transforms.append(
Lambda(lambda img: adjust_saturation(img, saturation_factor)))
random.shuffle(transforms)
transform = Compose(transforms)
print(brightness_factor)
return transform
def get_params(brightness, contrast, saturation, hue):
"""Get a randomized transform to be applied on image.
Arguments are same as that of __init__.
Returns:
Transform which randomly adjusts brightness, contrast and
saturation in a random order.
"""
transforms = []
if brightness > 0:
brightness_factor = random.uniform(max(0, 1 - brightness),
1 + brightness)
transforms.append(
Lambda(lambda img: adjust_brightness(img, brightness_factor)))
if contrast > 0:
contrast_factor = random.uniform(max(0, 1 - contrast),
1 + contrast)
transforms.append(
Lambda(lambda img: adjust_contrast(img, contrast_factor)))
if saturation > 0:
saturation_factor = random.uniform(max(0, 1 - saturation),
1 + saturation)
transforms.append(
Lambda(lambda img: adjust_saturation(img, saturation_factor)))
if hue > 0:
hue_factor = random.uniform(-hue, hue)
transforms.append(Lambda(lambda img: adjust_hue(img, hue_factor)))
random.shuffle(transforms)
transform = Compose(transforms)
return transform
def test_transforms_groups_constructor_error(self):
original_dataset = MNIST('./data/mnist', download=True)
with self.assertRaises(Exception):
# Test tuple has only one element
dataset = AvalancheDataset(original_dataset,
transform_groups=dict(
train=(ToTensor(), None),
eval=(Lambda(lambda t: float(t)))))
with self.assertRaises(Exception):
# Test is not a tuple has only one element
dataset = AvalancheDataset(
original_dataset,
transform_groups=dict(train=(ToTensor(), None),
eval=[None,
Lambda(lambda t: float(t))]))
with self.assertRaises(Exception):
# Train is None
dataset = AvalancheDataset(
original_dataset,
transform_groups=dict(train=None,
eval=(None, Lambda(lambda t: float(t)))))
with self.assertRaises(Exception):
# transform_groups is not a dictionary
dataset = AvalancheDataset(original_dataset,
transform_groups='Hello world!')
def get_test_data(data_path):
with open(data_path, 'rb') as f:
train_test_paths_labels = pickle.load(f)
test_paths = train_test_paths_labels[2]
test_labels = train_test_paths_labels[5]
test_num_each = train_test_paths_labels[8]
print('test_paths : {:6d}'.format(len(test_paths)))
print('test_labels : {:6d}'.format(len(test_labels)))
test_labels = np.asarray(test_labels, dtype=np.int64)
test_transforms = None
if crop_type == 0:
test_transforms = transforms.Compose([
transforms.RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.4310, 0.2971, 0.3126], [0.2405, 0.1863, 0.1935])
])
elif crop_type == 1:
test_transforms = transforms.Compose([
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.4310, 0.2971, 0.3126], [0.2405, 0.1863, 0.1935])
])
elif crop_type == 5:
test_transforms = transforms.Compose([
transforms.FiveCrop(224),
Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
Lambda(
lambda crops: torch.stack(
[transforms.Normalize([0.4310, 0.2971, 0.3126], [0.2405, 0.1863, 0.1935])(crop) for crop in crops]))
])
elif crop_type == 10:
test_transforms = transforms.Compose([
transforms.TenCrop(224),
Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
Lambda(
lambda crops: torch.stack(
[transforms.Normalize([0.4310, 0.2971, 0.3126], [0.2405, 0.1863, 0.1935])(crop) for crop in crops]))
])
elif crop_type == 2:
test_transforms = transforms.Compose([
transforms.Resize((250, 250)),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.4310, 0.2971, 0.3126], [0.2405, 0.1863, 0.1935])
])
elif crop_type == 3:
test_transforms = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize([0.4310, 0.2971, 0.3126], [0.2405, 0.1863, 0.1935])
])
test_dataset = CholecDataset(test_paths, test_labels, test_transforms)
return test_dataset, test_num_each
def __init__(self):
self.transform = Compose([
Lambda(lambda x: x / 255),
Normalize(-mean, std),
Lambda(lambda x: x.clamp(0, 1)),
Lambda(lambda x: x.flip(0)),
Lambda(lambda x: x.cpu()),
ToPILImage()
])
def forward(self, batch):
# 1.取数据，模型输入的特征维度应该等于self.feature_dim
x = batch['X']
# 2.根据输入数据计算模型的输出结果，模型输出的特征维度应该等于self.output_dim
k = 1
dgc_mode = 'hybrid'
obs_timesteps = 10
pred_timesteps = 3
nb_nodes = 208
encoder = torch.nn.RNN(DGCRNNCell(k, dgc_mode=dgc_mode),
return_state=True)
decoder = torch.nn.RNN(DGCRNNCell(k, dgc_mode=dgc_mode),
return_sequences=True,
return_state=True)
unstack_k = Lambda(unstack)
choice = Scheduled()
input_obs = torch.randn(size=(obs_timesteps, nb_nodes, 1))
input_gt = torch.randn(size=(pred_timesteps, nb_nodes,
1)) # (None, T, N, 1)
encoder_inputs = Lambda(lambda x: torch.squeeze(x, dim=-1))(
input_obs) # (None, T, N)
encoder_outputs, state_h = encoder(encoder_inputs)
unstacked = unstack_k(input_gt) # [(None, N, 1) x T] list
initial = unstacked[0] # (None, N, 1)
decoder_inputs = Lambda(lambda x: torch.transpose(x, 1, 2))(
initial) # (None, 1, N)
decoder_outputs_new, state_h_new = decoder(decoder_inputs,
initial_state=state_h)
state_h = state_h_new
# prediction part
prediction = []
decoded_results = decoder_outputs_new
prediction.append(decoded_results)
if pred_timesteps > 1:
for i in range(1, pred_timesteps):
decoder_inputs = choice([prediction[-1],
unstacked[i]]) # (None, 208, 1)
decoder_inputs = Lambda(lambda x: torch.transpose(x, 1, 2))(
decoder_inputs)
decoder_outputs_new, state_h_new = decoder(
decoder_inputs, initial_state=state_h)
state_h = state_h_new
decoded_results = decoder_outputs_new
prediction.append(decoded_results)
outputs = Lambda(stack)(prediction)
return outputs
def __init__(self, preserve_color, device):
self.normalize = Compose([
Lambda(lambda x: x.flip(0)),
Normalize(mean, std),
Lambda(lambda x: x * 255),
Lambda(lambda x: x.unsqueeze(0))
])
self.device = device
self.preserve_color = preserve_color
def get_data(data_path):
with open(data_path, 'rb') as f:
train_test_paths_labels = pickle.load(f)
train_paths = train_test_paths_labels[0]
val_paths = train_test_paths_labels[1]
test_paths = train_test_paths_labels[2]
train_labels = train_test_paths_labels[3]
val_labels = train_test_paths_labels[4]
test_labels = train_test_paths_labels[5]
train_num_each = train_test_paths_labels[6]
val_num_each = train_test_paths_labels[7]
test_num_each = train_test_paths_labels[8]
print('train_paths : {:6d}'.format(len(train_paths)))
print('train_labels : {:6d}'.format(len(train_labels)))
print('valid_paths : {:6d}'.format(len(val_paths)))
print('valid_labels : {:6d}'.format(len(val_labels)))
print('test_paths : {:6d}'.format(len(test_paths)))
print('test_labels : {:6d}'.format(len(test_labels)))
train_labels = np.asarray(train_labels, dtype=np.int64)
val_labels = np.asarray(val_labels, dtype=np.int64)
test_labels = np.asarray(test_labels, dtype=np.int64)
# train_transforms = transforms.Compose([
# transforms.RandomCrop(224),
# transforms.ToTensor(),
# transforms.Normalize([0.3456, 0.2281, 0.2233], [0.2528, 0.2135, 0.2104])
# ])
train_transforms = transforms.Compose([
transforms.TenCrop(224),
Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
Lambda(
lambda crops: torch.stack(
[transforms.Normalize([0.3456, 0.2281, 0.2233], [0.2528, 0.2135, 0.2104])(crop) for crop in crops]))
])
val_transforms = transforms.Compose([
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.3456, 0.2281, 0.2233], [0.2528, 0.2135, 0.2104])
])
test_transforms = transforms.Compose([
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.3456, 0.2281, 0.2233], [0.2528, 0.2135, 0.2104])
])
train_dataset = CholecDataset(train_paths, train_labels, train_transforms)
val_dataset = CholecDataset(val_paths, val_labels, val_transforms)
test_dataset = CholecDataset(test_paths, test_labels, test_transforms)
return train_dataset, train_num_each, val_dataset, val_num_each, test_dataset, test_num_each
def sr_output_train_transform(convert_luma):
def transform(pred, target):
return pred_transform(pred), target_transform(target)
transforms = [Lambda(lambda img: normalize_range(img, (-1., 1.)))]
if convert_luma:
transforms.append(Lambda(lambda img: convert_to_luma(img)))
pred_transform = Compose(transforms)
target_transform = Compose(transforms)
return transform
def __call__(self, img1, img2):
if self.istrain:
img1 = F.resize(img1, self.resize, Image.BILINEAR)
img2 = F.resize(img2, self.resize, Image.BILINEAR)
# random crop
w1, h1 = img1.size
w2, h2 = img2.size
w = min(w1, w2)
h = min(h1, h2)
th, tw = self.output_size, self.output_size
i = random.randint(0, h - th)
j = random.randint(0, w - tw)
img1 = F.crop(img1, i, j, th, tw)
img2 = F.crop(img2, i, j, th, tw)
# random flip
if random.random() < 0.5:
img1 = F.hflip(img1)
img2 = F.hflip(img2)
if random.random() < 0.5:
img1 = F.vflip(img1)
img2 = F.vflip(img2)
# color jitter
brightness = 0.4
saturation = 0.4
hue = 0.4
brightness_factor = random.uniform(max(0, 1 - brightness),
1 + brightness)
saturation_factor = random.uniform(max(0, 1 - saturation),
1 + saturation)
hue_factor = random.uniform(-hue, hue)
transforms = []
transforms.append(
Lambda(
lambda img: F.adjust_brightness(img, brightness_factor)))
transforms.append(
Lambda(
lambda img: F.adjust_saturation(img, saturation_factor)))
transforms.append(
Lambda(lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
transform = Compose(transforms)
img1 = transform(img1)
img2 = transform(img2)
else:
img1 = F.resize(img1, (self.output_size, self.output_size),
Image.BILINEAR)
img2 = F.resize(img2, (self.output_size, self.output_size),
Image.BILINEAR)
img1 = F.to_tensor(img1)
img2 = F.to_tensor(img2)
img1 = F.normalize(img1, self.mean, self.std)
img2 = F.normalize(img2, self.mean, self.std)
return img1, img2
def _get_transforms(self, augmentation_config: Optional[CfgNode],
dataset_name: str,
is_ssl_encoder_module: bool) -> Tuple[Any, Any]:
# is_ssl_encoder_module will be True for ssl training, False for linear head training
train_transforms = ImageTransformationPipeline([Lambda(lambda x: x)]) # do nothing
val_transforms = ImageTransformationPipeline([Lambda(lambda x: x + 1)]) # add 1
if is_ssl_encoder_module:
train_transforms = DualViewTransformWrapper(train_transforms) # type: ignore
val_transforms = DualViewTransformWrapper(val_transforms) # type: ignore
return train_transforms, val_transforms
def test_transform(self):
train_transform = Lambda(lambda k: 1)
test_transform = Lambda(lambda k: 0)
dataset = ActiveLearningDataset(MyDataset(train_transform), make_unlabelled=lambda x: (x[0], -1),
pool_specifics={'transform': test_transform})
dataset.label(np.arange(10))
pool = dataset.pool
assert np.equal([i for i in pool], [(0, -1) for i in np.arange(10, 100)]).all()
assert np.equal([i for i in dataset], [(1, i) for i in np.arange(10)]).all()
with pytest.raises(ValueError) as e:
ActiveLearningDataset(MyDataset(train_transform), pool_specifics={'whatever': 123}).pool
def test_transform(self):
train_transform = Lambda(lambda k: 1)
test_transform = Lambda(lambda k: 0)
dataset = ActiveLearningDataset(MyDataset(train_transform),
test_transform,
make_unlabelled=lambda x: (x[0], -1))
dataset.label(np.arange(10))
pool = dataset.pool
assert np.equal([i for i in pool],
[(0, -1) for i in np.arange(10, 100)]).all()
assert np.equal([i for i in dataset],
[(1, i) for i in np.arange(10)]).all()
def __init__(self, config, data_feature):
# 1.初始化父类
super().__init__(config, data_feature)
# 2.从data_feature(TrafficStatePointDataset)中获取需要的信息
self._scaler = self.data_feature.get('scaler') # 归一化方法
self.adj_mx = self.data_feature.get('adj_mx') # 邻接矩阵
self.num_nodes = self.data_feature.get('adj_mx') # 点的个数
self.feature_dim = self.data_feature.get('feature_dim') # 输入数据的维度
self.output_dim = self.data_feature.get('output_dim') # 模型输出的维度
# 3.初始化log用于必要的输出
self._logger = getLogger()
# 4.初始化device
self.device = config.get('device', torch.device('cpu'))
# 5.Build network
self.encoder = torch.nn.RNN()
self.decoder = torch.nn.RNN()
self.input_obs = torch.randn()
self.input_gt = torch.randn()
self.encoder_inputs = Lambda(lambda x: torch.squeeze(x, dim=-1))(
self.input_obs)
self.encoder_outputs, state_h = self.encoder(self.encoder_inputs)
self.unstacked = self.unstack_k(self.input_gt)
self.initial = self.unstacked[0] # (None, N, 1)
self.decoder_inputs = Lambda(lambda x: torch.transpose(x, 1, 2))(
self.initial)
self.decoder_outputs_new, self.state_h_new = self.decoder(
self.decoder_inputs, initial_state=state_h)
self.state_h = self.state_h_new
self.prediction = []
self.decoded_results = self.decoder_outputs_new
self.prediction.append(self.decoded_results)
self.pred_timesteps = 3
self.choice = Scheduled()
if self.pred_timesteps > 1:
for i in range(1, self.pred_timesteps):
self.decoder_inputs = self.choice(
[self.prediction[-1], self.unstacked[i]]) # (None, 208, 1)
self.decoder_inputs = Lambda(
lambda x: torch.transpose(x, 1, 2))(
self.decoder_inputs) # (None, 1, 208)
self.decoder_outputs_new, self.state_h_new = self.decoder(
self.decoder_inputs, initial_state=state_h)
self.state_h = self.state_h_new
self.decoded_results = self.decoder_outputs_new
self.prediction.append(self.decoded_results)
self.outputs = Lambda(stack)(self.prediction)
def __init__(self,
root,
preload_train=False,
preload_test=False,
im_size=28):
transforms = Compose([
Resize(im_size, Image.LANCZOS),
ToTensor(),
Lambda(lambda x: x.unsqueeze(0)), # used to add batch dimension
])
t_transforms = Lambda(lambda x: torch.tensor(x).unsqueeze(0))
self.cifar_train = CIFAR10(
root=root,
#background=True,
train=True,
download=True,
transform=transforms,
target_transform=t_transforms,
)
self.cifar_test = CIFAR10(
root=root,
#background=False,
train=False,
download=True,
transform=transforms,
target_transform=t_transforms,
)
# task_train contains tasks i.e. 963 sets of 20 characters
self.tasks_train = np.arange(len(self.cifar_train._character_images))
# chars_train contains the characters directly i.e. 964*20=19280 images-labels pairs
self.chars_train = np.arange(
len(self.cifar_train._flat_character_images))
# task_test contains tasks i.e. 659 sets of 20 characters
self.tasks_test = np.arange(len(self.cifar_test._character_images))
# chars_train contains the characters directly i.e. 600*20=13180 images-labels pairs
self.chars_test = np.arange(len(
self.cifar_test._flat_character_images))
if preload_train:
start = time()
print("Pre-loading CIFAR 10 train...")
_ = [img for img in self.cifar_train]
end = time()
print(f"{end - start:.1f}s :CIFAR 10 train pre-loaded.")
if preload_test:
start = time()
print("Pre-loading CIFAR 10 test...")
_ = [img for img in self.cifar_test]
end = time()
print(f"{end - start:.1f}s : CIFAR 10 test pre-loaded.")
def get_basic_input_transform(patch_size, mean, std):
cv2_scale = (lambda x: cv2.resize(
x, dsize=(patch_size, patch_size), interpolation=cv2.INTER_LINEAR)
if x.shape[0] != patch_size else x.copy())
np_reshape = lambda x: np.reshape(x, (patch_size, patch_size, 1))
tforms = [
Lambda(cv2_scale),
Lambda(np_reshape),
ToTensor(),
Normalize((mean, ), (std, )),
]
tforms = Compose(tforms)
return tforms
def create_encodings(self):
color_table = self.create_colors()
colors = []
images = []
X = []
instances = []
image_transform = Compose([
Lambda(lambda x: x.detach().cpu()),
ToPILImage(),
Resize(45),
Lambda(lambda x: np.asarray(x))
])
print("Retrieve visualization data...")
for idx, enc in enumerate(tqdm(self.encoding_dataset)):
# save metadata
instances.append({
"rooms": enc["anchor"]["meta"]["reference"],
"labels": enc["anchor"]["meta"]["label"],
"instances": enc["anchor"]["meta"]["reference"] + "_" + enc["anchor"]["meta"]["label"] + "_" + str(enc["anchor"]["meta"]["instance_id"])
})
# load image to be visualized
image = self.encoding_dataset.dataset[idx]["anchor"]["image"]
image = image_transform(image)
images.append(image)
# load color to be visualized
key = self.get_key(enc, idx)
colors.append(color_table[key])
# load encodings to be visualized. Always load encoding regardless of filter because otherwise the t-SNE and PCA would be different!!
anchor_encodings = enc["anchor"]["encodings"]
if self.dim == -1:
all_encs = torch.cat(anchor_encodings, dim=1)
X.append(all_encs.clone())
else:
X.append(anchor_encodings[self.dim].clone())
assert (len(X) == len(colors))
assert (len(colors) == len(images))
assert (len(images) == len(instances))
images = np.array(images)
X = torch.cat(X)
return X, colors, images, instances
def input_transform_augment(crop_size):
return Compose([
CenterCrop(crop_size),
Scale(256),
ToTensor(), # [0, 255] --> [ 0., 1.]
Lambda(lambda x: 2 * x - 1), # [0., 1.] --> [-1., 1.]
])
def __init__(self, phase, kwargs):
self.mode = Mode[kwargs['mode']]
self.image_size = kwargs['image_size']
self.hidden_size = kwargs['hidden_size']
self.debug_use_dataset = kwargs['debug_use_dataset']
self.debug_one_sentence = kwargs['debug_one_sentence']
self.__use_densenet = kwargs['__use_densenet']
self.sent_tokenizer = PunktSentenceTokenizer()
self.word_tokenizer = TweetTokenizer()
if phase == Phase.train:
jitter = [ColorJitter(brightness=0.5, contrast=0.5)]
else:
jitter = []
if self.__use_densenet:
self.transform = Compose((
[Lambda(lambda img: img.convert('RGB'))] +
[Resize((256, 256))] +
jitter +
[ToTensor()] +
[Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]
))
else:
self.transform = Compose((
[Resize((256, 256))] +
jitter +
[ToTensor()]
))
def test_transforms_groups_base_usage(self):
original_dataset = MNIST('./data/mnist', download=True)
dataset = AvalancheDataset(
original_dataset,
transform_groups=dict(train=(ToTensor(), None),
eval=(None, Lambda(lambda t: float(t)))))
x, y, _ = dataset[0]
self.assertIsInstance(x, Tensor)
self.assertIsInstance(y, int)
dataset_test = dataset.eval()
x2, y2, _ = dataset_test[0]
x3, y3, _ = dataset[0]
self.assertIsInstance(x2, Image)
self.assertIsInstance(y2, float)
self.assertIsInstance(x3, Tensor)
self.assertIsInstance(y3, int)
dataset_train = dataset.train()
dataset.transform = None
x4, y4, _ = dataset_train[0]
x5, y5, _ = dataset[0]
self.assertIsInstance(x4, Tensor)
self.assertIsInstance(y4, int)
self.assertIsInstance(x5, Image)
self.assertIsInstance(y5, int)
def preprocess(images):
'''
Performs preprocessing on a batch of images (bs, h, w, c) or on a single image (h, w, c).
It doesn't handle flickering!! (there is no flickering in breakout)
Use grayscale instead of luminance.
'''
size_preprocessed_image = 84
transformations = Compose([
Lambda(lambda image: image.permute(2, 0, 1)),
ToPILImage(),
Grayscale(),
Resize((size_preprocessed_image, size_preprocessed_image)),
ToTensor()
])
if len(images.shape) == 3:
images = images.unsqueeze(0)
assert len(images.shape) == 4
batch_size = images.shape[0]
preprocessed_images = []
for i in range(batch_size):
preprocessed_images.append(transformations(images[i]).squeeze(0))
preprocessed_images = torch.stack(preprocessed_images).permute(
1, 2, 0).squeeze()
if len(preprocessed_images.shape) == 3:
preprocessed_images, _ = torch.max(preprocessed_images,
dim=2,
keepdim=False)
return preprocessed_images
def test_if_applies_transforms(self, input_directory):
transform = Lambda(lambda x: np.pad(x, (2, 2)))
dataset = datasets.NiftiFolder.from_dir(input_directory, transform)
for idx in range(len(dataset)):
assert np.all(dataset[idx].shape == np.array(INPUT_IMAGE_SHAPE) +
4)
def valid_lr_transform(crop_size):
return Compose([
ToPILImage(),
Resize(crop_size, interpolation=Image.BICUBIC),
ToTensor(),
Lambda(imagenet_normalise),
])