torchsat.transforms package¶

Submodules¶

torchsat.transforms.functional module¶

torchsat.transforms.functional.adjust_brightness(img, value=0)¶

torchsat.transforms.functional.adjust_contrast(img, factor)¶

torchsat.transforms.functional.adjust_hue()¶

torchsat.transforms.functional.adjust_saturation()¶

torchsat.transforms.functional.bbox_crop(bboxes, top, left, height, width)¶

crop bbox

Arguments:: img {ndarray} – image to be croped top {int} – top size left {int} – left size height {int} – croped height width {int} – croped width

torchsat.transforms.functional.bbox_hflip(bboxes, img_width)¶

horizontal flip the bboxes: ^

. . . . . . . . . . . . ………….

^

Args:: bbox (ndarray): bbox ndarray [box_nums, 4] flip_code (int, optional): [description]. Defaults to 0.

torchsat.transforms.functional.bbox_pad(bboxes, padding)¶

torchsat.transforms.functional.bbox_resize(bboxes, img_size, target_size)¶

resize the bbox

Args:: bboxes (ndarray): bbox ndarray [box_nums, 4] img_size (tuple): the image height and width target_size (int, or tuple): the target bbox size.

Int or Tuple, if tuple the shape should be (height, width)

torchsat.transforms.functional.bbox_shift(bboxes, top, left)¶

shift the bbox

Arguments:: bboxes {ndarray} – input bboxes, should be num x 4 top {int} – shit to top value, positive mean move down, negative mean move up. left {int} – shit to left value, positive mean move right, negative mean move left.
Returns:: [ndarray] – shifted bboxes

torchsat.transforms.functional.bbox_vflip(bboxes, img_height)¶

vertical flip the bboxes: . . . .
>………..<: . . . . ……….. Args:

bbox (ndarray): bbox ndarray [box_nums, 4] flip_code (int, optional): [description]. Defaults to 0.

torchsat.transforms.functional.center_crop(img, output_size)¶

crop image

Arguments:: img {ndarray} – input image output_size {number or sequence} – the output image size. if sequence, should be [h, w]
Raises:: ValueError – the input image is large than original image.
Returns:: ndarray image – return croped ndarray image.

torchsat.transforms.functional.crop(img, top, left, height, width)¶

crop image

Arguments:: img {ndarray} – image to be croped top {int} – top size left {int} – left size height {int} – croped height width {int} – croped width

torchsat.transforms.functional.elastic_transform(image, alpha, sigma, alpha_affine, interpolation=1, border_mode=4, random_state=None, approximate=False)¶: Elastic deformation of images as described in [Simard2003] (with modifications). Based on https://gist.github.com/erniejunior/601cdf56d2b424757de5 .. [Simard2003] Simard, Steinkraus and Platt, “Best Practices for

Convolutional Neural Networks applied to Visual Document Analysis”, in Proc. of the International Conference on Document Analysis and Recognition, 2003.

torchsat.transforms.functional.flip(img, flip_code)¶

torchsat.transforms.functional.gaussian_blur(img, kernel_size)¶

torchsat.transforms.functional.hflip(img)¶

torchsat.transforms.functional.noise(img, mode='gaussain', percent=0.02)¶: TODO: Not good for uint16 data

torchsat.transforms.functional.normalize(tensor, mean, std, inplace=False)¶

Normalize a tensor image with mean and standard deviation.

Note

This transform acts out of place by default, i.e., it does not mutates the input tensor.

See Normalize for more details.

Args:: tensor (Tensor): Tensor image of size (C, H, W) to be normalized. mean (sequence): Sequence of means for each channel. std (sequence): Sequence of standard deviations for each channel.
Returns:: Tensor: Normalized Tensor image.

torchsat.transforms.functional.pad(img, padding, fill=0, padding_mode='constant')¶

torchsat.transforms.functional.preserve_channel_dim(func)¶: Preserve dummy channel dim.

torchsat.transforms.functional.resize(img, size, interpolation=2)¶

resize the image TODO: opencv resize 之后图像就成了0~1了 Arguments:

img {ndarray} – the input ndarray image size {int, iterable} – the target size, if size is intger, width and height will be resized to same otherwise, the size should be tuple (height, width) or list [height, width]

Keyword Arguments:: interpolation {Image} – the interpolation method (default: {Image.BILINEAR})
Raises:: TypeError – img should be ndarray ValueError – size should be intger or iterable vaiable and length should be 2.
Returns:: img – resize ndarray image

torchsat.transforms.functional.resized_crop(img, top, left, height, width, size, interpolation=2)¶

torchsat.transforms.functional.rotate(img, angle, center=None, scale=1.0)¶

torchsat.transforms.functional.rotate_box(bboxes, angle, center_x, center_y)¶

rotate_box rotate the bboxes

Args:: bboxes (ndarray): the original bboxes, N(numbers) x 4 angle (float): rotate angle should be degree center_x (int): rotate center x center_y (int): rotate center y
Returns:: ndarray: the rotated bboxes

torchsat.transforms.functional.shift(img, top, left)¶

torchsat.transforms.functional.to_grayscale(img, output_channels=1)¶

convert input ndarray image to gray sacle image.

Arguments:: img {ndarray} – the input ndarray image
Keyword Arguments:: output_channels {int} – output gray image channel (default: {1})
Returns:: ndarray – gray scale ndarray image

torchsat.transforms.functional.to_pil_image(tensor)¶

torchsat.transforms.functional.to_tensor(img)¶

convert numpy.ndarray to torch tensor.

if the image is uint8 , it will be divided by 255;

if the image is uint16 , it will be divided by 65535;

if the image is float , it will not be divided, we suppose your image range should between [0~1] ;

Arguments:: img {numpy.ndarray} – image to be converted to tensor. torch.float32

torchsat.transforms.functional.to_tiff_image(tensor)¶

torchsat.transforms.functional.vflip(img)¶

torchsat.transforms.transforms_cls module¶

class torchsat.transforms.transforms_cls.Compose(transforms)¶

Bases: object

Composes serveral classification transform together.

Args:

transforms (list of transform objects): list of classification transforms to compose.

Example:

>>> transforms_cls.Compose([
>>>     transforms_cls.Resize(300),
>>>     transforms_cls.ToTensor()
>>>     ])

class torchsat.transforms.transforms_cls.Lambda(lambd)¶

Bases: object

Apply a user-defined lambda as function.

Args:: lambd (function): Lambda/function to be used for transform.

class torchsat.transforms.transforms_cls.ToTensor¶

Bases: object

onvert numpy.ndarray to torch tensor.

if the image is uint8 , it will be divided by 255; if the image is uint16 , it will be divided by 65535; if the image is float , it will not be divided, we suppose your image range should between [0~1] ;

Args:: img {numpy.ndarray} – image to be converted to tensor.

class torchsat.transforms.transforms_cls.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], inplace=False)¶

Bases: object

Normalize a tensor image with mean and standard deviation.

Given mean: (M1,...,Mn) and std: (S1,..,Sn) for n channels, this transform will normalize each channel of the input torch.*Tensor i.e. input[channel] = (input[channel] - mean[channel]) / std[channel] .. note:

This transform acts out of place, i.e., it does not mutates the input tensor.

Args:: tensor (tensor): input torch tensor data. mean (sequence): Sequence of means for each channel. std (sequence): Sequence of standard deviations for each channel. inplace (boolean): inplace apply the transform or not. (default: False)

class torchsat.transforms.transforms_cls.ToGray(output_channels=1)¶

Bases: object

Convert the image to grayscale

Args:: output_channels (int): number of channels desired for output image. (default: 1)
Returns:: [ndarray]: the graysacle version of input - If output_channels=1 : returned single channel image (height, width) - If output_channels>1 : returned multi-channels ndarray image (height, width, channels)

class torchsat.transforms.transforms_cls.GaussianBlur(kernel_size=3)¶

Bases: object

Convert the input ndarray image to blurred image by gaussian method.

Args:: kernel_size (int): kernel size of gaussian blur method. (default: 3)
Returns:: ndarray: the blurred image.

class torchsat.transforms.transforms_cls.RandomNoise(mode='gaussian', percent=0.02)¶

Bases: object

Add noise to the input ndarray image. Args:

mode (str): the noise mode, should be one of gaussian, salt, pepper, s&p, (default: gaussian). percent (float): noise percent, only work for salt, pepper, s&p mode. (default: 0.02)

Returns:: ndarray: noised ndarray image.

class torchsat.transforms.transforms_cls.RandomBrightness(max_value=0)¶: Bases: object

class torchsat.transforms.transforms_cls.RandomContrast(max_factor=0)¶: Bases: object

class torchsat.transforms.transforms_cls.RandomShift(max_percent=0.4)¶

Bases: object

random shift the ndarray with value or some percent.

Args:: max_percent (float): shift percent of the image.
Returns:: ndarray: return the shifted ndarray image.

class torchsat.transforms.transforms_cls.RandomRotation(degrees, center=None)¶

Bases: object

random rotate the ndarray image with the degrees.

Args:

degrees (number or sequence): the rotate degree.: If single number, it must be positive. if squeence, it’s length must 2 and first number should small than the second one.

Raises:

ValueError: If degrees is a single number, it must be positive. ValueError: If degrees is a sequence, it must be of len 2.

Returns:

ndarray: return rotated ndarray image.

class torchsat.transforms.transforms_cls.Resize(size, interpolation=2)¶

Bases: object

resize the image Args:

img {ndarray} : the input ndarray image size {int, iterable} : the target size, if size is intger, width and height will be resized to same otherwise, the size should be tuple (height, width) or list [height, width]

Keyword Arguments:: interpolation {Image} : the interpolation method (default: {Image.BILINEAR})
Raises:: TypeError : img should be ndarray ValueError : size should be intger or iterable vaiable and length should be 2.
Returns:: img (ndarray) : resize ndarray image

class torchsat.transforms.transforms_cls.Pad(padding, fill=0, padding_mode='constant')¶

Bases: object

Pad the given ndarray image with padding width. Args:

padding : {int, sequence}, padding width

If int, each border same. If sequence length is 2, this is the padding for left/right and top/bottom. If sequence length is 4, this is the padding for left, top, right, bottom.

fill: {int, sequence}: Pixel padding_mode: str or function. contain{‘constant’,‘edge’,‘linear_ramp’,‘maximum’,‘mean’

, ‘median’, ‘minimum’, ‘reflect’,‘symmetric’,‘wrap’} (default: constant)

Examples:

>>> transformed_img = Pad(img, 20, mode='reflect')
>>> transformed_img = Pad(img, (10,20), mode='edge')
>>> transformed_img = Pad(img, (10,20,30,40), mode='reflect')

class torchsat.transforms.transforms_cls.CenterCrop(out_size)¶

Bases: object

crop image

Args:: img {ndarray}: input image output_size {number or sequence}: the output image size. if sequence, should be [height, width]
Raises:: ValueError: the input image is large than original image.
Returns:: ndarray: return croped ndarray image.

class torchsat.transforms.transforms_cls.RandomCrop(size)¶

Bases: object

random crop the input ndarray image

Args:: size (int, sequence): th output image size, if sequeue size should be [height, width]
Returns:: ndarray: return random croped ndarray image.

class torchsat.transforms.transforms_cls.RandomHorizontalFlip(p=0.5)¶

Bases: object

Flip the input image on central horizon line.

Args:: p (float): probability apply the horizon flip.(default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_cls.RandomVerticalFlip(p=0.5)¶

Bases: object

Flip the input image on central vertical line.

Args:: p (float): probability apply the vertical flip. (default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_cls.RandomFlip(p=0.5)¶

Bases: object

Flip the input image vertical or horizon.

Args:: p (float): probability apply flip. (default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_cls.RandomResizedCrop(crop_size, target_size, interpolation=2)¶

Bases: object

[summary]

Args:: object ([type]): [description]
Returns:: [type]: [description]

class torchsat.transforms.transforms_cls.ElasticTransform(alpha=1, sigma=50, alpha_affine=50, interpolation=1, border_mode=4, random_state=None, approximate=False)¶

Bases: object

code modify from https://github.com/albu/albumentations. Elastic deformation of images as described in [Simard2003] (with modifications). Based on https://gist.github.com/erniejunior/601cdf56d2b424757de5 .. [Simard2003] Simard, Steinkraus and Platt, “Best Practices for

Convolutional Neural Networks applied to Visual Document Analysis”, in Proc. of the International Conference on Document Analysis and Recognition, 2003.

Args:

approximate (boolean): Whether to smooth displacement map with fixed kernel size.: Enabling this option gives ~2X speedup on large images.

Image types:

uint8, uint16 float32

torchsat.transforms.transforms_det module¶

class torchsat.transforms.transforms_det.Compose(transforms)¶

Bases: object

Composes serveral classification transform together.

Args:

transforms (list of transform objects): list of classification transforms to compose.

Example:

>>> transforms_cls.Compose([
>>>     transforms_cls.Resize(300),
>>>     transforms_cls.ToTensor()
>>>     ])

class torchsat.transforms.transforms_det.Lambda(lambd)¶

Bases: object

Apply a user-defined lambda as function.

Args:: lambd (function): Lambda/function to be used for transform.

class torchsat.transforms.transforms_det.ToTensor¶

Bases: object

onvert numpy.ndarray to torch tensor.

if the image is uint8 , it will be divided by 255; if the image is uint16 , it will be divided by 65535; if the image is float , it will not be divided, we suppose your image range should between [0~1] ;

Args:: img {numpy.ndarray} – image to be converted to tensor. bboxes {numpy.ndarray} – target bbox to be converted to tensor. the input should be [box_nums, 4] labels {numpy.ndarray} – target labels to be converted to tensor. the input shape shold be [box_nums]

class torchsat.transforms.transforms_det.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], inplace=False)¶

Bases: object

Normalize a tensor image with mean and standard deviation.

Given mean: (M1,...,Mn) and std: (S1,..,Sn) for n channels, this transform will normalize each channel of the input torch.*Tensor i.e. input[channel] = (input[channel] - mean[channel]) / std[channel] .. note:

This transform acts out of place, i.e., it does not mutates the input tensor.

Args:: tensor (tensor): input torch tensor data. mean (sequence): Sequence of means for each channel. std (sequence): Sequence of standard deviations for each channel. inplace (boolean): inplace apply the transform or not. (default: False)

class torchsat.transforms.transforms_det.ToGray(output_channels=1)¶

Bases: object

Convert the image to grayscale

Args:: output_channels (int): number of channels desired for output image. (default: 1)
Returns:: [ndarray]: the graysacle version of input - If output_channels=1 : returned single channel image (height, width) - If output_channels>1 : returned multi-channels ndarray image (height, width, channels)

class torchsat.transforms.transforms_det.GaussianBlur(kernel_size=3)¶

Bases: object

Convert the input ndarray image to blurred image by gaussian method.

Args:: kernel_size (int): kernel size of gaussian blur method. (default: 3)
Returns:: ndarray: the blurred image.

class torchsat.transforms.transforms_det.RandomNoise(mode='gaussian', percent=0.02)¶

Bases: object

Add noise to the input ndarray image. Args:

mode (str): the noise mode, should be one of gaussian, salt, pepper, s&p, (default: gaussian). percent (float): noise percent, only work for salt, pepper, s&p mode. (default: 0.02)

Returns:: ndarray: noised ndarray image.

class torchsat.transforms.transforms_det.RandomBrightness(max_value=0)¶: Bases: object

class torchsat.transforms.transforms_det.RandomContrast(max_factor=0)¶: Bases: object

class torchsat.transforms.transforms_det.Resize(size, interpolation=2)¶

Bases: object

resize the image Args:

img {ndarray} : the input ndarray image size {int, iterable} : the target size, if size is intger, width and height will be resized to same otherwise, the size should be tuple (height, width) or list [height, width]

Keyword Arguments:: interpolation {Image} : the interpolation method (default: {Image.BILINEAR})
Raises:: TypeError : img should be ndarray ValueError : size should be intger or iterable vaiable and length should be 2.
Returns:: img (ndarray) : resize ndarray image

class torchsat.transforms.transforms_det.Pad(padding, fill=0, padding_mode='constant')¶

Bases: object

Pad the given ndarray image with padding width. Args:

padding : {int, sequence}, padding width

If int, each border same. If sequence length is 2, this is the padding for left/right and top/bottom. If sequence length is 4, this is the padding for left, top, right, bottom.

fill: {int, sequence}: Pixel padding_mode: str or function. contain{‘constant’,‘edge’,‘linear_ramp’,‘maximum’,‘mean’

, ‘median’, ‘minimum’, ‘reflect’,‘symmetric’,‘wrap’} (default: constant)

Examples:

>>> transformed_img = Pad(img, 20, mode='reflect')
>>> transformed_img = Pad(img, (10,20), mode='edge')
>>> transformed_img = Pad(img, (10,20,30,40), mode='reflect')

class torchsat.transforms.transforms_det.CenterCrop(out_size)¶

Bases: object

crop image

Args:: img {ndarray}: input image output_size {number or sequence}: the output image size. if sequence, should be [height, width]
Raises:: ValueError: the input image is large than original image.
Returns:: ndarray: return croped ndarray image.

class torchsat.transforms.transforms_det.RandomCrop(size)¶

Bases: object

random crop the input ndarray image

Args:: size (int, sequence): th output image size, if sequeue size should be [height, width]
Returns:: ndarray: return random croped ndarray image.

class torchsat.transforms.transforms_det.RandomHorizontalFlip(p=0.5)¶

Bases: object

Flip the input image on central horizon line.

Args:: p (float): probability apply the horizon flip.(default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_det.RandomVerticalFlip(p=0.5)¶

Bases: object

Flip the input image on central vertical line.

Args:: p (float): probability apply the vertical flip. (default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_det.RandomFlip(p=0.5)¶

Bases: object

Flip the input image vertical or horizon.

Args:: p (float): probability apply flip. (default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_det.RandomResizedCrop(crop_size, target_size, interpolation=2)¶

Bases: object

[summary]

Args:: object ([type]): [description]
Returns:: [type]: [description]

torchsat.transforms.transforms_seg module¶

class torchsat.transforms.transforms_seg.Compose(transforms)¶: Bases: object

class torchsat.transforms.transforms_seg.Lambda(lambd)¶: Bases: object

class torchsat.transforms.transforms_seg.ToTensor¶: Bases: object

class torchsat.transforms.transforms_seg.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], inplace=False)¶: Bases: object

class torchsat.transforms.transforms_seg.ToGray(output_channels=1)¶

Bases: object

Convert the image to grayscale

Args:: output_channels (int): number of channels desired for output image. (default: 1)
Returns:: [ndarray]: the graysacle version of input - If output_channels=1 : returned single channel image (height, width) - If output_channels>1 : returned multi-channels ndarray image (height, width, channels)

class torchsat.transforms.transforms_seg.GaussianBlur(kernel_size=3)¶: Bases: object

class torchsat.transforms.transforms_seg.RandomNoise(mode='gaussian', percent=0.02)¶

Bases: object

Add noise to the input ndarray image. Args:

mode (str): the noise mode, should be one of gaussian, salt, pepper, s&p, (default: gaussian). percent (float): noise percent, only work for salt, pepper, s&p mode. (default: 0.02)

Returns:: ndarray: noised ndarray image.

class torchsat.transforms.transforms_seg.RandomBrightness(max_value=0)¶: Bases: object

class torchsat.transforms.transforms_seg.RandomContrast(max_factor=0)¶: Bases: object

class torchsat.transforms.transforms_seg.RandomShift(max_percent=0.4)¶

Bases: object

random shift the ndarray with value or some percent.

Args:: max_percent (float): shift percent of the image.
Returns:: ndarray: return the shifted ndarray image.

class torchsat.transforms.transforms_seg.RandomRotation(degrees, center=None)¶

Bases: object

random rotate the ndarray image with the degrees.

Args:

degrees (number or sequence): the rotate degree.: If single number, it must be positive. if squeence, it’s length must 2 and first number should small than the second one.

Raises:

ValueError: If degrees is a single number, it must be positive. ValueError: If degrees is a sequence, it must be of len 2.

Returns:

ndarray: return rotated ndarray image.

class torchsat.transforms.transforms_seg.Resize(size, interpolation=2)¶

Bases: object

resize the image Args:

img {ndarray} : the input ndarray image size {int, iterable} : the target size, if size is intger, width and height will be resized to same otherwise, the size should be tuple (height, width) or list [height, width]

Keyword Arguments:: interpolation {Image} : the interpolation method (default: {Image.BILINEAR})
Raises:: TypeError : img should be ndarray ValueError : size should be intger or iterable vaiable and length should be 2.
Returns:: img (ndarray) : resize ndarray image

class torchsat.transforms.transforms_seg.Pad(padding, fill=0, padding_mode='constant')¶

Bases: object

Pad the given ndarray image with padding width. Args:

padding : {int, sequence}, padding width

If int, each border same. If sequence length is 2, this is the padding for left/right and top/bottom. If sequence length is 4, this is the padding for left, top, right, bottom.

fill: {int, sequence}: Pixel padding_mode: str or function. contain{‘constant’,‘edge’,‘linear_ramp’,‘maximum’,‘mean’

, ‘median’, ‘minimum’, ‘reflect’,‘symmetric’,‘wrap’} (default: constant)

Examples:

>>> transformed_img = Pad(img, 20, mode='reflect')
>>> transformed_img = Pad(img, (10,20), mode='edge')
>>> transformed_img = Pad(img, (10,20,30,40), mode='reflect')

class torchsat.transforms.transforms_seg.CenterCrop(out_size)¶

Bases: object

crop image

Args:: img {ndarray}: input image output_size {number or sequence}: the output image size. if sequence, should be [height, width]
Raises:: ValueError: the input image is large than original image.
Returns:: ndarray: return croped ndarray image.

class torchsat.transforms.transforms_seg.RandomCrop(size)¶

Bases: object

random crop the input ndarray image

Args:: size (int, sequence): th output image size, if sequeue size should be [height, width]
Returns:: ndarray: return random croped ndarray image.

class torchsat.transforms.transforms_seg.RandomHorizontalFlip(p=0.5)¶

Bases: object

Flip the input image on central horizon line.

Args:: p (float): probability apply the horizon flip.(default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_seg.RandomVerticalFlip(p=0.5)¶

Bases: object

Flip the input image on central vertical line.

Args:: p (float): probability apply the vertical flip. (default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_seg.RandomFlip(p=0.5)¶

Bases: object

Flip the input image vertical or horizon.

Args:: p (float): probability apply flip. (default: 0.5)
Returns:: ndarray: return the flipped image.

class torchsat.transforms.transforms_seg.RandomResizedCrop(crop_size, target_size, interpolation=2)¶

Bases: object

[summary]

Args:: object ([type]): [description]
Returns:: [type]: [description]

class torchsat.transforms.transforms_seg.ElasticTransform(alpha=1, sigma=50, alpha_affine=50, interpolation=1, border_mode=4, random_state=None, approximate=False)¶

Bases: object

code modify from https://github.com/albu/albumentations. Elastic deformation of images as described in [Simard2003] (with modifications). Based on https://gist.github.com/erniejunior/601cdf56d2b424757de5 .. [Simard2003] Simard, Steinkraus and Platt, “Best Practices for

Convolutional Neural Networks applied to Visual Document Analysis”, in Proc. of the International Conference on Document Analysis and Recognition, 2003.

Args:

approximate (boolean): Whether to smooth displacement map with fixed kernel size.: Enabling this option gives ~2X speedup on large images.

Image types:

uint8, uint16 float32

torchsat.transforms package¶

Submodules¶

torchsat.transforms.functional module¶

torchsat.transforms.transforms_cls module¶

torchsat.transforms.transforms_det module¶

torchsat.transforms.transforms_seg module¶

Module contents¶