tokencut_dataset.py

import json
from torch.utils import data
from torchvision.datasets import ImageFolder
import torch
import os
from PIL import Image
import numpy as np
import argparse
from tqdm import tqdm
from munkres import Munkres
import multiprocessing
from multiprocessing import Process, Manager
import collections
import torchvision.transforms as transforms
import torchvision.transforms.functional as TF
import random
import torchvision
import cv2

torch.manual_seed(0)

SegItem = collections.namedtuple('SegItem', ('image_name', 'tag'))
normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5],
                                 std=[0.5, 0.5, 0.5])

TRANSFORM_TRAIN = transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(),
])

TRANSFORM_EVAL = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
])

IMAGE_TRANSFORMS = transforms.Compose([
    transforms.ToTensor(),
    normalize
])

MERGED_TAGS = {'n04356056', 'n04355933',
               'n04493381', 'n02808440',
               'n03642806', 'n03832673',
               'n04008634', 'n03773504',
               'n03887697', 'n15075141'}

TRAIN_PARTITION = "train"
VAL_PARTITION = "val"
LEGAL_PARTITIONS = {TRAIN_PARTITION, VAL_PARTITION}


# TRAIN_CLASSES = 500

class SegmentationDataset(ImageFolder):
    def __init__(self, seg_path, imagenet_path, partition=TRAIN_PARTITION, num_samples=2, train_classes=500
                 , imagenet_classes_path='imagenet_classes.json'):
        assert partition in LEGAL_PARTITIONS
        self._partition = partition
        self._seg_path = seg_path
        self._imagenet_path = imagenet_path
        with open(imagenet_classes_path, 'r') as f:
            self._imagenet_classes = json.load(f)
        self._tag_list = [tag for tag in os.listdir(self._seg_path) if tag not in MERGED_TAGS]
        if partition == TRAIN_PARTITION:
            # Skip merged tags as those cause a headache
            self._tag_list = self._tag_list[:train_classes]
        elif partition == VAL_PARTITION:
            # Skip merged tags as those cause a headache
            self._tag_list = self._tag_list[train_classes:]
        for tag in self._tag_list:
            assert tag in self._imagenet_classes
        self._all_segementations = []
        for tag in self._tag_list:
            base_dir = os.path.join(self._seg_path, tag)
            curr_num_samples = 0
            for i, seg in enumerate(os.listdir(base_dir)):
                seg_name = seg.split('.')[0]
                if 'bfs' not in seg_name:
                    continue
                seg_path = os.path.join(self._seg_path, tag, seg)
                seg_map = torch.load(seg_path)
                seg_map = torch.from_numpy(seg_map.astype(np.float32))
                if torch.sum(seg_map) < 520:
                    continue
                if curr_num_samples >= num_samples:
                    break
                self._all_segementations.append(SegItem(seg_name, tag))
                curr_num_samples += 1

    def __getitem__(self, item):
        seg_item = self._all_segementations[item]

        seg_path = os.path.join(self._seg_path, seg_item.tag, seg_item.image_name + ".pt")

        image_path = os.path.join(self._imagenet_path, seg_item.image_name.split('_tokencut_bfs')[0] + ".JPEG")
        image = Image.open(image_path)
        image = image.convert('RGB')

        seg_map = torch.load(seg_path)
        seg_map = torch.from_numpy(seg_map.astype(np.float32))

        # transforms - start
        seg_map = seg_map.reshape(1, seg_map.shape[-2], seg_map.shape[-1])

        resize = transforms.Resize(224)
        image = resize(image)

        if self._partition == VAL_PARTITION:
            image = TRANSFORM_EVAL(image)
            seg_map = TRANSFORM_EVAL(seg_map)

        elif self._partition == TRAIN_PARTITION:
            # Resize
            resize = transforms.Resize(size=(256, 256))
            image = resize(image)
            seg_map = resize(seg_map)

            # Random crop
            i, j, h, w = transforms.RandomCrop.get_params(
                image, output_size=(224, 224))
            image = TF.crop(image, i, j, h, w)
            seg_map = TF.crop(seg_map, i, j, h, w)

            # RandomHorizontalFlip
            if random.random() > 0.5:
                image = TF.hflip(image)
                seg_map = TF.hflip(seg_map)

        else:
            raise Exception(f"Unsupported partition type {self._partition}")
        image_ten = IMAGE_TRANSFORMS(image)
        # transforms - end

        class_name = int(self._imagenet_classes[seg_item.tag])

        return seg_map, image_ten, class_name

    def __len__(self):
        return len(self._all_segementations)