BatchMaker.py

from tifffile import imread, imsave
import numpy as np
import random
import os
import torch
import torch.utils.data
import ImageTools
import LearnTools
import math
from PIL import Image
import matplotlib.pyplot as plt

perms = [[1, 2, 3], [2, 1, 3], [3, 1, 2]]  # permutations for a 4d array.
perms_3d = np.array(perms) + 1 # permutations for a 5d array.
crop_size = 4  # crop pixels in each dimension when choosing train slices
# LOW_L_3D = 45  # length of low resolution 3d
HIGH_L_3D = 64  # length of high resolution 3d

if os.getcwd().endswith('code'):
    os.chdir('..')  # current directory from /SuperRes/code to SuperRes/
NMC_PATH = 'data/NMC.tif'
SOFC_CATHODE_PATH = 'netl-sofc-cathode-segmented.tif'


class BatchMaker:
    """
    Makes and saves training and test batch images.
    """

    def __init__(self, device, to_low_idx=None, path=NMC_PATH, sf=4, dims=3,
                 stack=True, down_sample=False, separator=False,
                 low_res=False, rot_and_mir=True, squash=False,
                 super_sample=False):
        """
        :param device: the device that the image is on.
        :param to_low_idx: the indices of the phases to be down-sampled.
        :param path: the path of the tif file
        TODO make it more general than just tif..
        :param sf: the scale factor between low and high res.
        :param dims: number of dimensions for the batches (2 or 3)
        :param stack: if the data is a stack of 2D images
        :param down_sample: whether to down-sample the data or not.
        :param separator: whether the material is a separator.
        :param rot_and_mir: if True, the stack of 2D images will rotate and
        mirror for another 8 configurations
        :param squash: whether to squash all phases (other than pore) to one phase.
        :param super_sample: whether when down sampling to make it with
        super-sampling and not with blurring.
        """
        # TODO currently training images are on cpu, maybe better to move
        #  them to gpu.
        # down-sample parameters:
        self.down_sample, self.to_low_idx, self.squash = down_sample, \
                                                         to_low_idx, squash
        self.scale_factor = sf
        self.path = path  # the path of the tif file
        self.dims = dims  # if G is 3D to 3D or 2D to 2D
        self.device = device
        self.stack = stack  # if the data is a stack of 2D images
        self.im = imread(path)
        self.high_l = HIGH_L_3D
        if stack and not low_res:  # it is the high-res training data
            self.hr_metrics = ImageTools.vf_sa_metrics(self.im)
            self.high_l -= crop_size*2
        if rot_and_mir:
            self.rotate_and_mirror()
        self.dim_im = len(self.im.shape)  # the dimension of the image
        self.phases = np.unique(self.im)  # the unique values in image
        self.im = ImageTools.one_hot_encoding(self.im, self.phases)
        if low_res:
            self.high_l = int(HIGH_L_3D/self.scale_factor)
        if self.dims == 2:
            self.high_l = self.high_l*2
        if self.down_sample:
            self.down_sample_object = LearnTools.\
                DownSample(self.squash, self.dims, self.to_low_idx,
                           self.scale_factor, device, super_sample,
                           separator).to(self.device)
            self.im = np.array(self.down_sample_im(self.im).detach().cpu())
            self.phases = [self.phases[0]] + list(np.array(self.phases)[
                           np.array(self.to_low_idx.detach().cpu())])
            self.high_l = int(HIGH_L_3D / self.scale_factor)

    def down_sample_im(self, image):
        """
        :return: a down-sample image of the high resolution image for the input
        of G.
        """
        torch_im = torch.FloatTensor(image).unsqueeze(0).to(self.device)
        return self.down_sample_object(torch_im, low_res_input=True).squeeze(0)

    def rotate_and_mirror(self):
        """
        Given a stack of 2D images, in the form of num_images X width X heigth
        return a num_images*8 X width X height stack, with all 8 different
        90deg rotations and mirrors of the images.
        """
        num_ims = self.im.shape[0]
        flip_im = np.flip(self.im, -1)
        res = np.zeros((num_ims*8, *self.im.shape[1:]), dtype=self.im.dtype)
        for k in np.arange(4):  # for each 90 deg rotation
            first_i, second_i = 2*k*num_ims, (2*k+1)*num_ims
            # rotation images of original image:
            res[first_i:second_i, ...] = np.rot90(self.im, k, [-2, -1])
            # rotation images of flipped image:
            res[second_i:second_i + num_ims, ...] = np.rot90(flip_im, k,
                                                             [-2, -1])
        self.im = res

    def random_batch_for_real(self, batch_size):
        return self.random_batch2d(batch_size)

    def random_batch_for_fake(self, batch_size, dim_chosen):
        if self.dims == 3:
            return self.random_batch3d(batch_size, dim_chosen)
        else:  # dims = 2
            return self.random_batch2d(batch_size)

    def random_batch3d(self, batch_size, dim_chosen):
        """
        :return: A batch of high resolution images,
        along the dimension chosen (0->x,1->y,2->z) in the 3d tif image.
        """
        res = np.zeros((batch_size, len(self.phases),
                       *self.high_l * np.ones(self.dims, dtype=int)),
                       dtype=self.im.dtype)
        for i in range(batch_size):
            res[i, ...] = self.generate_a_random_image3d(dim_chosen)
        # return a torch tensor:
        return torch.FloatTensor(res).to(device=self.device)

    def generate_a_random_image3d(self, dim_chosen):
        """
        :param dim_chosen: the dimension chosen for the slice
        :return: A random image of size res from the dimension chosen of the
        image.
        """
        h_r = self.high_l
        # starting voxels
        s_ind = np.random.randint(np.array(self.im.shape[1:]) - h_r)
        e_ind = s_ind + h_r  # the end indices
        res_image = self.im[:, s_ind[0]:e_ind[0], s_ind[1]:e_ind[1],
                                s_ind[2]:e_ind[2]]
        # for different view, change the cube around..
        return res_image.transpose(0, *perms[dim_chosen])

    def random_batch2d(self, batch_size):
        """
        :return: A batch of high resolution images,
        along the dimension chosen (0->x,1->y,2->z) in the 3d tif image.
        """
        res = np.zeros((batch_size, len(self.phases), self.high_l,
                        self.high_l), dtype=self.im.dtype)
        for i in range(batch_size):
            res[i, :, :, :] = self.generate_a_random_image2d()
        # return a torch tensor:
        return torch.FloatTensor(res).to(device=self.device)

    def generate_a_random_image2d(self):
        """
        :return: A random image of size res from the dimension chosen of the
        image.
        """
        # the starting pixels of the other dimensions:
        if self.stack:  # has to be, since this function is only called for
            # by the discriminator batch maker.
            s_ind = np.random.randint(np.array(self.im.shape[2:]) -
                                      self.high_l)
            e_ind = s_ind + self.high_l
            slice_chosen = np.random.randint(self.im.shape[1])
            return self.im[:, slice_chosen, s_ind[0]:e_ind[0],
                                    s_ind[1]:e_ind[1]]
        else:
            raise ValueError("2D images should be stacked as "
                             "n_stack X height X width")

    def all_image_batch(self):
        """
        :return: the 3d image ready to be fed into the G with dimensions
        1xCxDxHxW or 1xCxHxW
        """
        return torch.FloatTensor(self.im).to(self.device).unsqueeze(0)