-
Notifications
You must be signed in to change notification settings - Fork 0
/
xbd_inf_gen.py
304 lines (252 loc) · 10.3 KB
/
xbd_inf_gen.py
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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
import argparse
import os
import sys
import glob
import cv2
import numpy as np
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, BitsAndBytesConfig, CLIPImageProcessor
from model.LISA import LISAForCausalLM
from model.llava import conversation as conversation_lib
from model.llava.mm_utils import tokenizer_image_token
from model.segment_anything.utils.transforms import ResizeLongestSide
from utils.utils import (DEFAULT_IM_END_TOKEN, DEFAULT_IM_START_TOKEN,
DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_INDEX)
def parse_args(args):
parser = argparse.ArgumentParser(description="LISA chat")
parser.add_argument("--version", default="xinlai/LISA-13B-llama2-v1")
parser.add_argument("--vis_save_path", default="./vis_output/xbd_test", type=str)
parser.add_argument("--data_dir", default="./data2/xbd/test")
parser.add_argument(
"--precision",
default="bf16",
type=str,
choices=["fp32", "bf16", "fp16"],
help="precision for inference",
)
parser.add_argument("--image_size", default=1024, type=int, help="image size")
parser.add_argument("--model_max_length", default=512, type=int)
parser.add_argument("--lora_r", default=8, type=int)
parser.add_argument(
"--vision-tower", default="openai/clip-vit-large-patch14", type=str
)
parser.add_argument("--local-rank", default=0, type=int, help="node rank")
parser.add_argument("--load_in_8bit", action="store_true", default=False)
parser.add_argument("--load_in_4bit", action="store_true", default=False)
parser.add_argument("--use_mm_start_end", action="store_true", default=True)
parser.add_argument(
"--conv_type",
default="llava_v1",
type=str,
choices=["llava_v1", "llava_llama_2"],
)
return parser.parse_args(args)
def load_imgs(dataset_root):
images_root = dataset_root + "/images/"
in_filenames = glob.glob(images_root + "*.png")
return in_filenames
def color_mask(mask):
uniques = np.unique(mask)
colored_mask = np.zeros((mask.shape[0], mask.shape[1], 3))
unique_to_color = {0: [0, 0, 0], 1: [255, 0, 0], 2: [0, 255, 0], 3: [0, 0, 255], 4: [255, 255, 255]}
if uniques[-1] == 1:
unique_to_color = {0: [0, 0, 0], 1: [255, 255, 255]}
for u in unique_to_color.keys():
colored_mask[mask == u] = unique_to_color[u]
return colored_mask
def preprocess(
x,
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,
) -> torch.Tensor:
"""Normalize pixel values and pad to a square input."""
# Normalize colors
x = (x - pixel_mean) / pixel_std
# Pad
h, w = x.shape[-2:]
padh = img_size - h
padw = img_size - w
x = F.pad(x, (0, padw, 0, padh))
return x
def main(args):
args = parse_args(args)
os.makedirs(args.vis_save_path, exist_ok=True)
# Create model
tokenizer = AutoTokenizer.from_pretrained(
args.version,
cache_dir=None,
model_max_length=args.model_max_length,
padding_side="right",
use_fast=False,
)
tokenizer.pad_token = tokenizer.unk_token
args.seg_token_idx = tokenizer("[SEG]", add_special_tokens=False).input_ids[0]
torch_dtype = torch.float32
if args.precision == "bf16":
torch_dtype = torch.bfloat16
elif args.precision == "fp16":
torch_dtype = torch.half
kwargs = {"torch_dtype": torch_dtype}
if args.load_in_4bit:
kwargs.update(
{
"torch_dtype": torch.half,
"load_in_4bit": True,
"quantization_config": BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
llm_int8_skip_modules=["visual_model"],
),
}
)
elif args.load_in_8bit:
kwargs.update(
{
"torch_dtype": torch.half,
"quantization_config": BitsAndBytesConfig(
llm_int8_skip_modules=["visual_model"],
load_in_8bit=True,
),
}
)
model = LISAForCausalLM.from_pretrained(
args.version, low_cpu_mem_usage=True, vision_tower=args.vision_tower, seg_token_idx=args.seg_token_idx, **kwargs
)
model.config.eos_token_id = tokenizer.eos_token_id
model.config.bos_token_id = tokenizer.bos_token_id
model.config.pad_token_id = tokenizer.pad_token_id
model.get_model().initialize_vision_modules(model.get_model().config)
vision_tower = model.get_model().get_vision_tower()
vision_tower.to(dtype=torch_dtype)
if args.precision == "bf16":
model = model.bfloat16().cuda()
elif (
args.precision == "fp16" and (not args.load_in_4bit) and (not args.load_in_8bit)
):
vision_tower = model.get_model().get_vision_tower()
model.model.vision_tower = None
import deepspeed
model_engine = deepspeed.init_inference(
model=model,
dtype=torch.half,
replace_with_kernel_inject=True,
replace_method="auto",
)
model = model_engine.module
model.model.vision_tower = vision_tower.half().cuda()
elif args.precision == "fp32":
model = model.float().cuda()
vision_tower = model.get_model().get_vision_tower()
vision_tower.to(device=args.local_rank)
clip_image_processor = CLIPImageProcessor.from_pretrained(model.config.vision_tower)
transform = ResizeLongestSide(args.image_size)
model.eval()
pre_prompts = [
"can you segment region with undamaged building from this image?"
]
post_prompts = [
"can you segment region with undamaged building from this image?",
"can you segment the building with minor damage from this image?",
"can you segment the building with major damage from this image?",
"can you segment completely destroyed building from in image?"
]
image_paths = load_imgs(args.data_dir)
for image_path in image_paths:
save_path = "{}/{}_mask_{}.jpg".format(
args.vis_save_path, image_path.split("/")[-1].split(".")[0], 0
)
if os.path.isfile(save_path):
continue
mask = None
lower_case_path = image_path.lower()
prompts = post_prompts
if "_pre_" in lower_case_path:
prompts = pre_prompts
for prompt_idx, prompt in enumerate(prompts):
conv = conversation_lib.conv_templates[args.conv_type].copy()
conv.messages = []
# prompt = input("Please input your prompt: ")
prompt = DEFAULT_IMAGE_TOKEN + "\n" + prompt
if args.use_mm_start_end:
replace_token = (
DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN
)
prompt = prompt.replace(DEFAULT_IMAGE_TOKEN, replace_token)
conv.append_message(conv.roles[0], prompt)
conv.append_message(conv.roles[1], "")
prompt = conv.get_prompt()
# image_path = input("Please input the image path: ")
if not os.path.exists(image_path):
print("File not found in {}".format(image_path))
continue
image_np = cv2.imread(image_path)
image_np = cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB)
original_size_list = [image_np.shape[:2]]
image_clip = (
clip_image_processor.preprocess(image_np, return_tensors="pt")[
"pixel_values"
][0]
.unsqueeze(0)
.cuda()
)
if args.precision == "bf16":
image_clip = image_clip.bfloat16()
elif args.precision == "fp16":
image_clip = image_clip.half()
else:
image_clip = image_clip.float()
image = transform.apply_image(image_np)
resize_list = [image.shape[:2]]
image = (
preprocess(torch.from_numpy(image).permute(2, 0, 1).contiguous())
.unsqueeze(0)
.cuda()
)
if args.precision == "bf16":
image = image.bfloat16()
elif args.precision == "fp16":
image = image.half()
else:
image = image.float()
print("prompt : ,",prompt)
input_ids = tokenizer_image_token(prompt, tokenizer, return_tensors="pt")
input_ids = input_ids.unsqueeze(0).cuda()
print("input_ids : ,",input_ids)
output_ids, pred_masks = model.evaluate(
image_clip,
image,
input_ids,
resize_list,
original_size_list,
max_new_tokens=512,
tokenizer=tokenizer,
)
output_ids = output_ids[0][output_ids[0] != IMAGE_TOKEN_INDEX]
text_output = tokenizer.decode(output_ids, skip_special_tokens=False)
text_output = text_output.replace("\n", "").replace(" ", " ")
print("text_output: ", text_output)
for i, pred_mask in enumerate(pred_masks):
if pred_mask.shape[0] == 0:
continue
pred_mask = pred_mask.detach().cpu().numpy()[0]
pred_mask = pred_mask > 0
if mask is None:
mask = pred_mask * (prompt_idx + 1)
else:
mask += pred_mask * (prompt_idx + 1)
save_path = "{}/{}_mask_{}.jpg".format(
args.vis_save_path, image_path.split("/")[-1].split(".")[0], i
)
save_path_colored = "{}/{}_colored_mask_{}.jpg".format(
args.vis_save_path, image_path.split("/")[-1].split(".")[0], i
)
cv2.imwrite(save_path, mask)
print("{} has been saved.".format(save_path))
cv2.imwrite(save_path_colored, color_mask(mask))
print("{} has been saved.".format(save_path_colored))
if __name__ == "__main__":
main(sys.argv[1:])