Add fragmented code

Author: hmellor

pytorch-superpixels/__init__.py

@@ -0,0 +1,104 @@
+'''For superpixel validation'''
+
+
+def mask_accuracy(target, mask):
+    target_s = torch.zeros_like(target)
+    superpixels = mask.unique().numel()
+    for superpixel in range(superpixels):
+        # Define mask for cluster idx
+        segment_mask = mask == superpixel
+        # Take slices to select image, apply mask, mode for majority class
+        target_s[segment_mask] = target[segment_mask].view(-1).mode()[0]
+    accuracy = torch.mean((target == target_s).float())
+    return accuracy
+
+
+def dataset_accuracy(superpixels):
+    # Generate image list
+    if superpixels is not None:
+        image_list = get_image_list('trainval_super')
+    else:
+        image_list = get_image_list()
+
+    mask_acc = 0
+    mask_dir = "SegmentationClass/{}_sp".format(superpixels)
+    target_dir = "SegmentationClass/pre_encoded"
+    for image_number in tqdm(image_list):
+        mask_path = join(root, mask_dir, image_number + ".pt")
+        target_path = join(root, target_dir, image_number + ".png")
+        mask = torch.load(mask_path)
+        target = io.imread(target_path)
+        target = torch.from_numpy(target)
+        mask_acc += mask_accuracy(target, mask)
+    dataset_acc = mask_acc / len(image_list)
+    return dataset_acc
+
+
+def find_smallest_object():
+    # Generate image list
+    image_list = get_image_list()
+    smallest_object = 1e6
+    for image_number in tqdm(image_list):
+        target_name = image_number + ".png"
+        target_path = join(root, "SegmentationClass/pre_encoded", target_name)
+        target = io.imread(target_path)
+        target = torch.from_numpy(target)
+        object_size = torch.ne(target, 0).sum()
+        if object_size < smallest_object:
+            smallest_object = object_size
+            print(smallest_object, image_number)
+    return smallest_object
+
+
+def find_usable_images(split, superpixels):
+    # Generate image list
+    image_list = get_image_list(split)
+    usable = []
+    target_dir = join(
+        root,
+        "SegmentationClass/pre_encoded_{}_sp".format(superpixels)
+    )
+    for image_number in image_list:
+        target_name = image_number + ".pt"
+        target_path = join(target_dir, target_name)
+        target = torch.load(target_path)
+        if target.nonzero().numel() > 0:
+            usable.append(image_number)
+    return usable
+
+
+def fix_broken_images(superpixels):
+    for split in ["train", "val", "trainval"]:
+        usable = find_usable_images(split=split, superpixels=superpixels)
+        super_path = join(root, "ImageSets/Segmentation", split + "_super.txt")
+        if exists(super_path):
+            remove(super_path)
+        with open(super_path, "w+") as file:
+            for image_number in usable:
+                file.write(image_number + "\n")
+
+
+def find_size_variance(superpixels):
+    # Generate image list
+    if superpixels is not None:
+        image_list = get_image_list('trainval_super')
+    else:
+        image_list = get_image_list()
+    mask_dir = "SegmentationClass/{}_sp".format(superpixels)
+    dataset_variance = 0
+    for image_number in tqdm(image_list):
+        mask_path = join(root, mask_dir, image_number + ".pt")
+        mask = torch.load(mask_path)
+        # Initialise number of superpixels tensors
+        Q = mask.unique().numel()
+        size = torch.zeros(Q)
+        counter = torch.ones_like(mask)
+        # Calculate the size of each superpixel
+        size.put_(mask, counter.float(), True)
+        # Calculate the mean and standard deviation of the sizes
+        std = size.std()
+        mean = size.mean()
+        # Add to the variance of the total datasets
+        dataset_variance += std / mean
+    dataset_variance /= len(image_list)
+    return dataset_variance

pytorch-superpixels/metrics.py

@@ -0,0 +1,98 @@
+'''For pre-processing'''
+
+
+def create_masks(numSegments=100, limOverseg=None):
+    # Generate image list
+    image_list = get_image_list()
+    for image_number in tqdm(image_list):
+        # Load image/target pair
+        image_name = image_number + ".jpg"
+        target_name = image_number + ".png"
+        image_path = join(root, "JPEGImages", image_name)
+        target_path = join(root, "SegmentationClass/pre_encoded", target_name)
+        image = img_as_float(io.imread(image_path))
+        target = io.imread(target_path)
+        target = torch.from_numpy(target)
+        # Create mask for image/target pair
+        mask, target_s = create_mask(
+            image=image,
+            target=target,
+            numSegments=numSegments,
+            limOverseg=limOverseg
+        )
+
+        # Save for later
+        image_save_dir = join(
+            root,
+            "SegmentationClass/{}_sp".format(numSegments)
+        )
+        target_s_save_dir = join(
+            root,
+            "SegmentationClass/pre_encoded_{}_sp".format(numSegments)
+        )
+        if not exists(image_save_dir):
+            mkdir(image_save_dir)
+        if not exists(target_s_save_dir):
+            mkdir(target_s_save_dir)
+        save_name = image_number + ".pt"
+        image_save_path = join(image_save_dir, save_name)
+        target_s_save_path = join(target_s_save_dir, save_name)
+        torch.save(mask, image_save_path)
+        torch.save(target_s, target_s_save_path)
+
+
+def create_mask(image, target, numSegments, limOverseg):
+    # Perform SLIC segmentation
+    mask = slic(image, n_segments=numSegments, slic_zero=True)
+    mask = torch.from_numpy(mask)
+
+    if limOverseg is not None:
+        # Oversegmentation step
+        superpixels = mask.unique().numel()
+        overseg = superpixels
+        for superpixel in range(superpixels):
+            overseg -= 1
+            # Define mask for superpixel
+            segment_mask = mask == superpixel
+            # Classes in this superpixel
+            classes = target[segment_mask].unique(sorted=True)
+            # Check if superpixel is on target boundary
+            on_boundary = classes.numel() > 1
+            # If current superpixel is on a gt boundary
+            if on_boundary:
+                # Find how many of each class is in superpixel
+                class_hist = torch.bincount(target[segment_mask])
+                # Remove zero elements
+                class_hist = class_hist[class_hist.nonzero()].float()
+                # Find minority class in superpixel
+                min_class = min(class_hist)
+                # Is the minority class large enough for oversegmentation
+                above_threshold = min_class > class_hist.sum() * limOverseg
+                if above_threshold:
+                    # Leaving one class in supperpixel be
+                    for c in classes[1:]:
+                        # Adding to the oversegmentation offset
+                        overseg += 1
+                        # Add offset to class c in the mask
+                        mask[segment_mask] += (target[segment_mask]
+                                               == c).long() * overseg
+
+    # (Re)define how many superpixels there are and create target_s
+    superpixels = mask.unique().numel()
+    target_s = torch.zeros(superpixels, dtype=torch.long)
+    for superpixel in range(superpixels):
+        # Define mask for superpixel
+        segment_mask = mask == superpixel
+        # Apply mask, the mode for majority class
+        target_s[superpixel] = target[segment_mask].view(-1).mode()[0]
+    return mask, target_s
+
+
+def get_image_list(split=None):
+    if split is None:
+        image_list_path = join(root, "ImageSets/Segmentation/trainval.txt")
+    else:
+        image_list_path = join(root, "ImageSets/Segmentation/", split + ".txt")
+    image_list = tuple(open(image_list_path, "r"))
+    image_list = [id_.rstrip() for id_ in image_list]
+    return image_list

pytorch-superpixels/preprocess.py

@@ -0,0 +1,74 @@
+import torch
+from os.path import join
+from os.path import exists
+from os.path import dirname
+from os.path import abspath
+from os import mkdir
+from os import remove
+from os import listdir
+from tqdm import tqdm
+from skimage import io
+from skimage.util import img_as_float
+from skimage.segmentation import slic
+
+# Define absolute path for accessing dataset files
+package_dir = dirname(abspath(__file__))
+dataset_dir = "../../datasets/VOCdevkit/VOC2011"
+root = join(package_dir, dataset_dir)
+'''For use during runtime'''
+
+
+def convert_to_superpixels(input, target, mask):
+    # Extract size data from input and target
+    images, c, h, w = input.size()
+    if images > 1:
+        raise RuntimeError("Not implemented for batch sizes greater than 1")
+    # Initialise vairables to use
+    Q = mask.unique().numel()
+    output = torch.zeros((Q, c), device=input.device)
+    size = torch.zeros(Q, device=input.device)
+    counter = torch.ones(mask.size(), device=input.device)
+    # Calculate the size of each superpixel
+    size.put_(mask, counter, True)
+    # Calculate the mean value of each superpixel
+    input = input.view(c, -1)
+    mask = mask.view(1, -1).repeat(c, 1)
+    arange = torch.arange(start=1, end=c, device=input.device)
+    mask[arange, :] += Q * arange.view(-1, 1)
+    output = output.put_(mask, input, True).view(c, Q).t()
+    output = (output.t() / size).t()
+    return output, target.view(-1), size
+
+
+def convert_to_pixels(input, output, mask):
+    n, c, h, w = output.size()
+    for k in range(c):
+        output[0, k, :, :] = torch.gather(
+            input[:, k], 0, mask.view(-1)).view(h, w)
+    return output
+
+
+def to_super_to_pixels(input, mask):
+    target = torch.tensor([])
+    input_s, _, _ = convert_to_superpixels(input, target, mask)
+    output = convert_to_pixels(input_s, input, mask)
+    return output
+
+
+def setup_superpixels(superpixels):
+    image_save_dir = join(
+        root,
+        "SegmentationClass/{}_sp".format(superpixels)
+    )
+    target_s_save_dir = join(
+        root,
+        "SegmentationClass/pre_encoded_{}_sp".format(superpixels)
+    )
+    dirs = [image_save_dir, target_s_save_dir]
+    dataset_len = len(get_image_list())
+    if not any(exists(x) and len(listdir(x)) == dataset_len for x in dirs):
+            print("Superpixel dataset of scale {} superpixels either doesn't exist or is incomplete".format(superpixels))
+            print("Generating superpixel dataset now...")
+            create_masks(superpixels)
+
+    fix_broken_images(superpixels)