Python库 | imagebox-0.0.0.2.tar.gz

import math from random import randint from rasterio.enums import Resampling import numpy as np import imagebox.io as io import imagebox.processor as proc import imagebox.indices as indices from imagebox.config import FIRST, LAST, BAND_ORDERING, BANDS_FIRST # # CONSTANTS # INPUT_DTYPE=np.float TARGET_DTYPE=np.int64 DEFAULT_SIZE=256 DEFAULT_OVERLAP=0 TO_CATEGORICAL_ERROR='imagebox.handler: nb_categories required for to_categorical' INPUT_RESAMPLING=Resampling.bilinear TARGET_RESAMPLING=Resampling.mode # # InputTargetHandler # class InputTargetHandler(object): """ A handler for processing multiple input/target pairs in the same way. This is especially useful for constructing data loader/generators Args: input_bands<list|None>: list of bands to select from image means,stdevs<list|np.array>: - means/stdevs for all bands in input image (before selecting input_bands) - if means and not stdevs: center input around means - if means and stdevs normalize inputs band_indices<list>: - list of arguments for imagebox.indices.index - adds band_index bands to image indices_dict<dict>: - preset indices from which to select band_indices value_map<dict>: - target value map. - keys: new values - values: list of values to map to new value default_mapped_value<'image'|int|float>: - only used if value_name is not None - if 'image' keep all unmapped values - otherwise value for all unmapped values to_categorical<bool>: one-hot target image nb_categories<int>: - number of target categories - required for to_categorical augment<bool>: augment image flip_target/input<bool>: - flip target/input image before any additional augmentation - for use when target and input pairs have opposite y-axis orientation - set only one of these to be true target/input_resolution: rescale input or target with input/target resampling method input/target_resampling: resampling method for input or target padding|input/target_padding<int|None>: - amount to pad images|input/target-image after processing the image - will be ignored if there is cropping cropping<int|None>: - amount to crop both input and target image when reading the image input/target_cropping<int|None>: - amount to crop input/target image after processing the image float_cropping<int|None>: remove pixels from h/w of input and target starting at random i,j width<int|None>: image width (required if float cropping and not tiller) height<int|None>: image height (required if float cropping and not tiller) tiller<Tiller|True|None>: tile the image then read in a specific tile tiller_config<dict>: if tiller is True create tiller=Tiller(**tiller_config) input_dtype<str>: input data type target_dtype<str>: target data type """ def __init__(self, input_bands=None, means=None, stdevs=None, band_indices=None, indices_dict=None, value_map=None, default_mapped_value=proc.DEFAULT_VMAP_VALUE, to_categorical=False, nb_categories=None, augment=True, flip_target=False, flip_input=False, input_resolution=None, target_resolution=None, input_bounds=None, input_resampling=INPUT_RESAMPLING, target_resampling=TARGET_RESAMPLING, padding=None, input_padding=None, target_padding=None, input_padding_value=0, target_padding_value=0, float_cropping=None, cropping=None, input_cropping=None, target_cropping=None, window_index=None, size=None, width=None, height=None, example_path=None, tiller=None, tiller_config={}, target_expand_axis=None, input_preprocess=None, target_preprocess=None, target_squeeze=True, input_dtype=INPUT_DTYPE, target_dtype=TARGET_DTYPE ): if tiller is True: self.tiller=Tiller(**tiller_config) else: self.tiller=tiller self.input_bands=input_bands self.band_indices=band_indices self.indices_dict=indices_dict self.value_map=value_map self.default_mapped_value=default_mapped_value self.means=means self.stdevs=stdevs if to_categorical and (not nb_categories): raise ValueError(TO_CATEGORICAL_ERROR) self.to_categorical=to_categorical self.nb_categories=nb_categories self.augment=augment self.flip_target=flip_target self.flip_input=flip_input self.input_resolution=input_resolution self.target_resolution=target_resolution self.input_bounds=input_bounds self.input_resampling=input_resampling self.target_resampling=target_resampling self.set_augmentation() self.input_padding=input_padding or padding self.target_padding=target_padding or padding self.input_padding_value=input_padding_value self.target_padding_value=target_padding_value self.cropping=cropping or 0 self.input_cropping=input_cropping self.target_cropping=target_cropping self.float_cropping=float_cropping if size: self.width=size self.height=size else: self.width=width self.height=height self.set_window( window_index=window_index, example_path=example_path) self.target_expand_axis=target_expand_axis self.input_preprocess=input_preprocess self.target_preprocess=target_preprocess self.target_squeeze=target_squeeze self.input_dtype=input_dtype self.target_dtype=target_dtype def input(self,path,window=None,return_profile=False): self.input_path=path im,profile=self._read( path, self.input_resolution, self.target_resampling, window ) im=process_input( im, preprocess=self.input_preprocess, flip=self.flip_input, input_bands=self.input_bands, band_indices=self.band_indices, indices_dict=self.indices_dict, padding=self.input_padding, padding_value=self.input_padding_value, cropping=self.input_cropping, bounds=self.input_bounds, means=self.means, stdevs=self.stdevs, dtype=self.input_dtype ) return self._return_data( im, profile, return_profile ) def target(self,path,window=None,return_profile=False): self.target_path=path im,profile=self._read( path, self.target_resolution, self.target_resampling, window ) im=process_target( im, preprocess=self.target_preprocess, flip=self.flip_target, value_map=self.value_map, default_mapped_value=self.default_mapped_value, categorical=self.to_categorical, nb_categories=self.nb_categories, padding=self.target_padding, padding_value=self.target_padding_value, cropping=self.target_cropping, expand_axis=self.target_expand_axis, squeeze=self.target_squeeze, dtype=self.target_dtype )