Spaces:
Paused
Paused
File size: 8,485 Bytes
3d9fba4 6144294 3d9fba4 da654fa 6144294 da654fa ba91034 da654fa ba91034 da654fa 3d9fba4 6144294 3d9fba4 620ddd7 3d9fba4 620ddd7 3d9fba4 620ddd7 3d9fba4 6144294 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 |
import glob
import json
import os
import random
import cv2
import numpy as np
import torch
import torch.nn.functional as F
from transformers import CLIPImageProcessor
from model.segment_anything.utils.transforms import ResizeLongestSide
from .conversation import get_default_conv_template
from .data_processing import get_mask_from_json
from .utils import (
ANSWER_LIST,
DEFAULT_IM_END_TOKEN,
DEFAULT_IM_START_TOKEN,
DEFAULT_IMAGE_PATCH_TOKEN,
DEFAULT_IMAGE_TOKEN,
EXPLANATORY_QUESTION_LIST,
LONG_QUESTION_LIST,
SHORT_QUESTION_LIST,
)
class ReasonSegDataset(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,
reason_seg_data="ReasonSeg|train",
explanatory=0.1,
):
self.exclude_val = exclude_val
self.reason_seg_data = reason_seg_data
self.samples_per_epoch = samples_per_epoch
self.explanatory = explanatory
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)
self.short_question_list = SHORT_QUESTION_LIST
self.long_question_list = LONG_QUESTION_LIST
self.answer_list = ANSWER_LIST
reason_seg_data, splits = reason_seg_data.split("|")
splits = splits.split("_")
images = []
for split in splits:
images_split = glob.glob(
os.path.join(
base_image_dir, "reason_seg", reason_seg_data, split, "*.jpg"
)
)
images.extend(images_split)
jsons = [path.replace(".jpg", ".json") for path in images]
self.reason_seg_data = (images, jsons)
if explanatory != -1:
self.explanatory_question_list = EXPLANATORY_QUESTION_LIST
self.img_to_explanation = {}
with open(
os.path.join(
base_image_dir,
"reason_seg",
reason_seg_data,
"explanatory",
"train.json",
)
) as f:
items = json.load(f)
for item in items:
img_name = item["image"]
self.img_to_explanation[img_name] = {
"query": item["query"],
"outputs": item["outputs"],
}
print("len(self.img_to_explanation): ", len(self.img_to_explanation))
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):
images, jsons = self.reason_seg_data
idx = random.randint(0, len(images) - 1)
image_path = images[idx]
json_path = jsons[idx]
img = cv2.imread(image_path)
images = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
ori_size = images.shape[:2]
# preprocess images for clip
images_clip = self.clip_image_processor.preprocess(images, return_tensors="pt")[
"pixel_values"
][0]
image_token_len = (images_clip.shape[1] // 14) * (
images_clip.shape[2] // 14
) # FIXME: 14 is hardcoded patch size
images = self.transform.apply_image(images) # preprocess images for sam
resize = images.shape[:2]
mask, sents, is_sentence = get_mask_from_json(json_path, img)
if len(sents) >= self.num_classes_per_sample:
sampled_inds = np.random.choice(
list(range(len(sents))), size=self.num_classes_per_sample, replace=False
)
else:
sampled_inds = list(range(len(sents)))
sampled_sents = np.vectorize(sents.__getitem__)(sampled_inds).tolist()
sampled_masks = [
(mask == 1).astype(np.float32) for _ in range(len(sampled_inds))
]
image_name = image_path.split("/")[-1]
if self.explanatory != -1 and image_name in self.img_to_explanation:
if random.random() < self.explanatory:
choice = 2
else:
choice = random.randint(0, 1)
questions = []
answers = []
for text in sampled_sents:
if is_sentence:
question_template = random.choice(self.long_question_list)
questions.append(question_template.format(sent=text))
else:
question_template = random.choice(self.short_question_list)
questions.append(question_template.format(class_name=text.lower()))
img_name = image_path.split("/")[-1]
if self.explanatory != -1 and img_name in self.img_to_explanation:
# choice = random.randint(0, 2)
if choice == 0: # [SEG] token
answers.append(random.choice(self.answer_list))
elif choice == 1: # [SEG] token + text answer
image_name = image_path.split("/")[-1]
answer = self.img_to_explanation[image_name]["outputs"]
answer = random.choice(self.answer_list) + " {}".format(answer)
questions[-1] = (
DEFAULT_IMAGE_TOKEN
+ " "
+ text
+ " {}".format(random.choice(self.explanatory_question_list))
)
answers.append(answer)
elif choice == 2: # vanilla text answer
image_name = image_path.split("/")[-1]
answer = self.img_to_explanation[image_name]["outputs"]
questions[-1] = DEFAULT_IMAGE_TOKEN + " " + text
answers.append(answer)
else:
raise ValueError("Not implemented yet.")
else:
answers.append(random.choice(self.answer_list))
conversations = []
conv = get_default_conv_template("vicuna").copy()
roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
i = 0
while i < len(questions):
conv.messages = []
conv.append_message(conv.roles[0], questions[i])
conv.append_message(conv.roles[1], answers[i])
conversations.append(conv.get_prompt())
i += 1
# replace <image> token
for i in range(len(conversations)):
replace_token = DEFAULT_IMAGE_PATCH_TOKEN * image_token_len
replace_token = (
DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
)
conversations[i] = conversations[i].replace(
DEFAULT_IMAGE_TOKEN, replace_token
)
images = self.preprocess(torch.from_numpy(images).permute(2, 0, 1).contiguous())
image_name = image_path.split("/")[-1]
if (
self.explanatory != -1
and image_name in self.img_to_explanation
and choice == 2
):
masks = torch.rand(0, *ori_size)
label = torch.ones(ori_size) * self.ignore_label
else:
masks = np.stack(sampled_masks, axis=0)
masks = torch.from_numpy(masks)
label = torch.ones(masks.shape[1], masks.shape[2]) * self.ignore_label
return (
image_path,
images,
images_clip,
conversations,
masks,
label,
resize,
questions,
sampled_sents,
)
|