File size: 4,595 Bytes
3d9fba4
 
 
 
 
 
 
 
 
e5c9ee0
3d9fba4
 
e5c9ee0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6144294
3d9fba4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6144294
3d9fba4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e5c9ee0
 
 
 
6144294
 
 
e5c9ee0
 
 
 
3d9fba4
e5c9ee0
3d9fba4
e5c9ee0
 
 
 
3d9fba4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e5c9ee0
3d9fba4
 
 
 
 
 
e5c9ee0
 
3d9fba4
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import json
import os
import random

import cv2
import torch
import torch.nn.functional as F
from transformers import CLIPImageProcessor

from model.llava import conversation as conversation_lib
from model.segment_anything.utils.transforms import ResizeLongestSide

from .utils import DEFAULT_IMAGE_TOKEN


def preprocess_multimodal(source, mm_use_im_start_end):
    for sentence in source:
        if DEFAULT_IMAGE_TOKEN in sentence["value"]:
            sentence["value"] = (
                sentence["value"].replace(DEFAULT_IMAGE_TOKEN, "").strip()
            )
            sentence["value"] = DEFAULT_IMAGE_TOKEN + "\n" + sentence["value"]
            sentence["value"] = sentence["value"].strip()
            if "mmtag" in conversation_lib.default_conversation.version:
                sentence["value"] = sentence["value"].replace(
                    DEFAULT_IMAGE_TOKEN, "<Image>" + DEFAULT_IMAGE_TOKEN + "</Image>"
                )
    return source


class VQADataset(torch.utils.data.Dataset):
    pixel_mean = torch.Tensor([123.675, 116.28, 103.53]).view(-1, 1, 1)
    pixel_std = torch.Tensor([58.395, 57.12, 57.375]).view(-1, 1, 1)
    img_size = 1024
    ignore_label = 255

    def __init__(
        self,
        base_image_dir,
        tokenizer,
        vision_tower,
        samples_per_epoch=500 * 8 * 2 * 10,
        precision: str = "fp32",
        image_size: int = 224,
        num_classes_per_sample: int = 3,
        exclude_val=False,
        vqa_data="llava_instruct_150k",
    ):
        self.exclude_val = exclude_val
        self.samples_per_epoch = samples_per_epoch
        self.num_classes_per_sample = num_classes_per_sample

        self.base_image_dir = base_image_dir
        self.image_size = image_size
        self.tokenizer = tokenizer
        self.precision = precision
        self.transform = ResizeLongestSide(image_size)
        self.clip_image_processor = CLIPImageProcessor.from_pretrained(vision_tower)

        DATA_DIR = os.path.join(base_image_dir, "llava_dataset")
        self.vqa_image_root = os.path.join(base_image_dir, "coco/train2017")
        with open(os.path.join(DATA_DIR, "{}.json".format(vqa_data))) as f:
            vqa_data = json.load(f)
        self.vqa_data = vqa_data

        print("vqa_data: ", len(self.vqa_data))

    def __len__(self):
        return self.samples_per_epoch

    def preprocess(self, x: torch.Tensor) -> torch.Tensor:
        """Normalize pixel values and pad to a square input."""
        # Normalize colors
        x = (x - self.pixel_mean) / self.pixel_std

        # Pad
        h, w = x.shape[-2:]
        padh = self.img_size - h
        padw = self.img_size - w
        x = F.pad(x, (0, padw, 0, padh))
        return x

    def __getitem__(self, idx):
        idx = random.randint(0, len(self.vqa_data) - 1)
        item = self.vqa_data[idx]
        image_path = os.path.join(self.vqa_image_root, item["image"])
        image = cv2.imread(image_path)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        ori_size = image.shape[:2]
        image_clip = self.clip_image_processor.preprocess(image, return_tensors="pt")[
            "pixel_values"
        ][
            0
        ]  # preprocess image for clip

        image = self.transform.apply_image(image)  # preprocess image for sam
        resize = image.shape[:2]

        conv = conversation_lib.default_conversation.copy()
        source = item["conversations"]
        source = preprocess_multimodal(
            source,
            mm_use_im_start_end=conv.sep_style == conversation_lib.SeparatorStyle.TWO,
        )
        roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
        conversations = []
        if roles[source[0]["from"]] != conv.roles[0]:
            # Skip the first one if it is not from human
            source = source[1:]
        conv.messages = []
        for j, sentence in enumerate(source):
            role = roles[sentence["from"]]
            assert role == conv.roles[j % 2], f"{i}"
            conv.append_message(role, sentence["value"])
        conversations.append(conv.get_prompt())

        questions = conversations
        sampled_classes = conversations

        image = self.preprocess(torch.from_numpy(image).permute(2, 0, 1).contiguous())

        masks = torch.rand(0, *ori_size)
        label = torch.ones(ori_size) * self.ignore_label

        return (
            image_path,
            image,
            image_clip,
            conversations,
            masks,
            label,
            resize,
            questions,
            sampled_classes,
        )