Upload 3 files
Browse files- app.py +35 -0
- geochat_demo.py +706 -0
- pyproject.toml +39 -0
app.py
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import gradio as gr
|
3 |
+
from torchvision import transforms
|
4 |
+
from PIL import Image
|
5 |
+
|
6 |
+
# Load model
|
7 |
+
class MyModel(torch.nn.Module):
|
8 |
+
def __init__(self):
|
9 |
+
super().__init__()
|
10 |
+
# Define layers here
|
11 |
+
|
12 |
+
def forward(self, x):
|
13 |
+
# Forward pass
|
14 |
+
return x
|
15 |
+
|
16 |
+
model = MyModel()
|
17 |
+
model.load_state_dict(torch.load("model.pth"))
|
18 |
+
model.eval()
|
19 |
+
|
20 |
+
# Define image preprocessing
|
21 |
+
transform = transforms.Compose([
|
22 |
+
transforms.Resize((224, 224)),
|
23 |
+
transforms.ToTensor(),
|
24 |
+
])
|
25 |
+
|
26 |
+
# Define prediction function
|
27 |
+
def predict(image):
|
28 |
+
image = transform(image).unsqueeze(0) # Add batch dimension
|
29 |
+
with torch.no_grad():
|
30 |
+
output = model(image)
|
31 |
+
return output.numpy().tolist()
|
32 |
+
|
33 |
+
# Create Gradio interface
|
34 |
+
iface = gr.Interface(fn=predict, inputs=gr.Image(), outputs="json")
|
35 |
+
iface.launch()
|
geochat_demo.py
ADDED
@@ -0,0 +1,706 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
from collections import defaultdict
|
5 |
+
|
6 |
+
import cv2
|
7 |
+
import re
|
8 |
+
import math
|
9 |
+
import numpy as np
|
10 |
+
from PIL import Image
|
11 |
+
import torch
|
12 |
+
import html
|
13 |
+
import gradio as gr
|
14 |
+
|
15 |
+
import torchvision.transforms as T
|
16 |
+
import torch.backends.cudnn as cudnn
|
17 |
+
|
18 |
+
from geochat.conversation import conv_templates, Chat
|
19 |
+
from geochat.model.builder import load_pretrained_model
|
20 |
+
from geochat.mm_utils import get_model_name_from_path
|
21 |
+
|
22 |
+
|
23 |
+
def parse_args():
|
24 |
+
parser = argparse.ArgumentParser(description="Demo")
|
25 |
+
# parser = argparse.ArgumentParser()
|
26 |
+
parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
|
27 |
+
parser.add_argument("--model-base", type=str, default=None)
|
28 |
+
parser.add_argument("--gpu-id", type=str,default=0)
|
29 |
+
parser.add_argument("--device", type=str, default="cuda")
|
30 |
+
parser.add_argument("--conv-mode", type=str, default=None)
|
31 |
+
parser.add_argument("--max-new-tokens", type=int, default=300)
|
32 |
+
parser.add_argument("--load-8bit", action="store_true")
|
33 |
+
parser.add_argument("--load-4bit", action="store_true")
|
34 |
+
parser.add_argument("--debug", action="store_true")
|
35 |
+
parser.add_argument("--image-aspect-ratio", type=str, default='pad')
|
36 |
+
# args = parser.parse_args()
|
37 |
+
args = parser.parse_args()
|
38 |
+
return args
|
39 |
+
|
40 |
+
|
41 |
+
random.seed(42)
|
42 |
+
np.random.seed(42)
|
43 |
+
torch.manual_seed(42)
|
44 |
+
|
45 |
+
cudnn.benchmark = False
|
46 |
+
cudnn.deterministic = True
|
47 |
+
|
48 |
+
print('Initializing Chat')
|
49 |
+
args = parse_args()
|
50 |
+
# cfg = Config(args)
|
51 |
+
|
52 |
+
model_name = get_model_name_from_path(args.model_path)
|
53 |
+
tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name, args.load_8bit, args.load_4bit, device=args.device)
|
54 |
+
|
55 |
+
device = 'cuda:{}'.format(args.gpu_id)
|
56 |
+
|
57 |
+
# model_config = cfg.model_cfg
|
58 |
+
# model_config.device_8bit = args.gpu_id
|
59 |
+
# model_cls = registry.get_model_class(model_config.arch)
|
60 |
+
# model = model_cls.from_config(model_config).to(device)
|
61 |
+
bounding_box_size = 100
|
62 |
+
|
63 |
+
# vis_processor_cfg = cfg.datasets_cfg.cc_sbu_align.vis_processor.train
|
64 |
+
# vis_processor = registry.get_processor_class(vis_processor_cfg.name).from_config(vis_processor_cfg)
|
65 |
+
|
66 |
+
model = model.eval()
|
67 |
+
|
68 |
+
CONV_VISION = conv_templates['llava_v1'].copy()
|
69 |
+
|
70 |
+
def bbox_and_angle_to_polygon(x1, y1, x2, y2, a):
|
71 |
+
# Calculate center coordinates
|
72 |
+
x_ctr = (x1 + x2) / 2
|
73 |
+
y_ctr = (y1 + y2) / 2
|
74 |
+
|
75 |
+
# Calculate width and height
|
76 |
+
w = abs(x2 - x1)
|
77 |
+
h = abs(y2 - y1)
|
78 |
+
|
79 |
+
# Calculate the angle in radians
|
80 |
+
angle_rad = math.radians(a)
|
81 |
+
|
82 |
+
# Calculate coordinates of the four corners of the rotated bounding box
|
83 |
+
cos_a = math.cos(angle_rad)
|
84 |
+
sin_a = math.sin(angle_rad)
|
85 |
+
|
86 |
+
x1_rot = cos_a * (-w / 2) - sin_a * (-h / 2) + x_ctr
|
87 |
+
y1_rot = sin_a * (-w / 2) + cos_a * (-h / 2) + y_ctr
|
88 |
+
|
89 |
+
x2_rot = cos_a * (w / 2) - sin_a * (-h / 2) + x_ctr
|
90 |
+
y2_rot = sin_a * (w / 2) + cos_a * (-h / 2) + y_ctr
|
91 |
+
|
92 |
+
x3_rot = cos_a * (w / 2) - sin_a * (h / 2) + x_ctr
|
93 |
+
y3_rot = sin_a * (w / 2) + cos_a * (h / 2) + y_ctr
|
94 |
+
|
95 |
+
x4_rot = cos_a * (-w / 2) - sin_a * (h / 2) + x_ctr
|
96 |
+
y4_rot = sin_a * (-w / 2) + cos_a * (h / 2) + y_ctr
|
97 |
+
|
98 |
+
# Return the polygon coordinates
|
99 |
+
polygon_coords = np.array((x1_rot, y1_rot, x2_rot, y2_rot, x3_rot, y3_rot, x4_rot, y4_rot))
|
100 |
+
|
101 |
+
return polygon_coords
|
102 |
+
|
103 |
+
def rotate_bbox(top_right, bottom_left, angle_degrees):
|
104 |
+
# Convert angle to radians
|
105 |
+
angle_radians = np.radians(angle_degrees)
|
106 |
+
|
107 |
+
# Calculate the center of the rectangle
|
108 |
+
center = ((top_right[0] + bottom_left[0]) / 2, (top_right[1] + bottom_left[1]) / 2)
|
109 |
+
|
110 |
+
# Calculate the width and height of the rectangle
|
111 |
+
width = top_right[0] - bottom_left[0]
|
112 |
+
height = top_right[1] - bottom_left[1]
|
113 |
+
|
114 |
+
# Create a rotation matrix
|
115 |
+
rotation_matrix = cv2.getRotationMatrix2D(center, angle_degrees, 1)
|
116 |
+
|
117 |
+
# Create an array of the rectangle corners
|
118 |
+
rectangle_points = np.array([[bottom_left[0], bottom_left[1]],
|
119 |
+
[top_right[0], bottom_left[1]],
|
120 |
+
[top_right[0], top_right[1]],
|
121 |
+
[bottom_left[0], top_right[1]]], dtype=np.float32)
|
122 |
+
|
123 |
+
# Rotate the rectangle points
|
124 |
+
rotated_rectangle = cv2.transform(np.array([rectangle_points]), rotation_matrix)[0]
|
125 |
+
|
126 |
+
return rotated_rectangle
|
127 |
+
def extract_substrings(string):
|
128 |
+
# first check if there is no-finished bracket
|
129 |
+
index = string.rfind('}')
|
130 |
+
if index != -1:
|
131 |
+
string = string[:index + 1]
|
132 |
+
|
133 |
+
pattern = r'<p>(.*?)\}(?!<)'
|
134 |
+
matches = re.findall(pattern, string)
|
135 |
+
substrings = [match for match in matches]
|
136 |
+
|
137 |
+
return substrings
|
138 |
+
|
139 |
+
|
140 |
+
def is_overlapping(rect1, rect2):
|
141 |
+
x1, y1, x2, y2 = rect1
|
142 |
+
x3, y3, x4, y4 = rect2
|
143 |
+
return not (x2 < x3 or x1 > x4 or y2 < y3 or y1 > y4)
|
144 |
+
|
145 |
+
|
146 |
+
def computeIoU(bbox1, bbox2):
|
147 |
+
x1, y1, x2, y2 = bbox1
|
148 |
+
x3, y3, x4, y4 = bbox2
|
149 |
+
intersection_x1 = max(x1, x3)
|
150 |
+
intersection_y1 = max(y1, y3)
|
151 |
+
intersection_x2 = min(x2, x4)
|
152 |
+
intersection_y2 = min(y2, y4)
|
153 |
+
intersection_area = max(0, intersection_x2 - intersection_x1 + 1) * max(0, intersection_y2 - intersection_y1 + 1)
|
154 |
+
bbox1_area = (x2 - x1 + 1) * (y2 - y1 + 1)
|
155 |
+
bbox2_area = (x4 - x3 + 1) * (y4 - y3 + 1)
|
156 |
+
union_area = bbox1_area + bbox2_area - intersection_area
|
157 |
+
iou = intersection_area / union_area
|
158 |
+
return iou
|
159 |
+
|
160 |
+
|
161 |
+
def save_tmp_img(visual_img):
|
162 |
+
file_name = "".join([str(random.randint(0, 9)) for _ in range(5)]) + ".jpg"
|
163 |
+
file_path = "/tmp/gradio" + file_name
|
164 |
+
visual_img.save(file_path)
|
165 |
+
return file_path
|
166 |
+
|
167 |
+
|
168 |
+
def mask2bbox(mask):
|
169 |
+
if mask is None:
|
170 |
+
return ''
|
171 |
+
mask = mask.resize([100, 100], resample=Image.NEAREST)
|
172 |
+
mask = np.array(mask)[:, :, 0]
|
173 |
+
|
174 |
+
rows = np.any(mask, axis=1)
|
175 |
+
cols = np.any(mask, axis=0)
|
176 |
+
|
177 |
+
if rows.sum():
|
178 |
+
# Get the top, bottom, left, and right boundaries
|
179 |
+
rmin, rmax = np.where(rows)[0][[0, -1]]
|
180 |
+
cmin, cmax = np.where(cols)[0][[0, -1]]
|
181 |
+
bbox = '{{<{}><{}><{}><{}>}}'.format(cmin, rmin, cmax, rmax)
|
182 |
+
else:
|
183 |
+
bbox = ''
|
184 |
+
|
185 |
+
return bbox
|
186 |
+
|
187 |
+
|
188 |
+
def escape_markdown(text):
|
189 |
+
# List of Markdown special characters that need to be escaped
|
190 |
+
md_chars = ['<', '>']
|
191 |
+
|
192 |
+
# Escape each special character
|
193 |
+
for char in md_chars:
|
194 |
+
text = text.replace(char, '\\' + char)
|
195 |
+
|
196 |
+
return text
|
197 |
+
|
198 |
+
|
199 |
+
def reverse_escape(text):
|
200 |
+
md_chars = ['\\<', '\\>']
|
201 |
+
|
202 |
+
for char in md_chars:
|
203 |
+
text = text.replace(char, char[1:])
|
204 |
+
|
205 |
+
return text
|
206 |
+
|
207 |
+
|
208 |
+
colors = [
|
209 |
+
(255, 0, 0),
|
210 |
+
(0, 255, 0),
|
211 |
+
(0, 0, 255),
|
212 |
+
(210, 210, 0),
|
213 |
+
(255, 0, 255),
|
214 |
+
(0, 255, 255),
|
215 |
+
(114, 128, 250),
|
216 |
+
(0, 165, 255),
|
217 |
+
(0, 128, 0),
|
218 |
+
(144, 238, 144),
|
219 |
+
(238, 238, 175),
|
220 |
+
(255, 191, 0),
|
221 |
+
(0, 128, 0),
|
222 |
+
(226, 43, 138),
|
223 |
+
(255, 0, 255),
|
224 |
+
(0, 215, 255),
|
225 |
+
]
|
226 |
+
|
227 |
+
color_map = {
|
228 |
+
f"{color_id}": f"#{hex(color[2])[2:].zfill(2)}{hex(color[1])[2:].zfill(2)}{hex(color[0])[2:].zfill(2)}" for
|
229 |
+
color_id, color in enumerate(colors)
|
230 |
+
}
|
231 |
+
|
232 |
+
used_colors = colors
|
233 |
+
|
234 |
+
|
235 |
+
def visualize_all_bbox_together(image, generation):
|
236 |
+
if image is None:
|
237 |
+
return None, ''
|
238 |
+
|
239 |
+
generation = html.unescape(generation)
|
240 |
+
|
241 |
+
image_width, image_height = image.size
|
242 |
+
image = image.resize([500, int(500 / image_width * image_height)])
|
243 |
+
image_width, image_height = image.size
|
244 |
+
|
245 |
+
string_list = extract_substrings(generation)
|
246 |
+
if string_list: # it is grounding or detection
|
247 |
+
mode = 'all'
|
248 |
+
entities = defaultdict(list)
|
249 |
+
i = 0
|
250 |
+
j = 0
|
251 |
+
for string in string_list:
|
252 |
+
try:
|
253 |
+
obj, string = string.split('</p>')
|
254 |
+
except ValueError:
|
255 |
+
print('wrong string: ', string)
|
256 |
+
continue
|
257 |
+
if "}{" in string:
|
258 |
+
string=string.replace("}{","}<delim>{")
|
259 |
+
bbox_list = string.split('<delim>')
|
260 |
+
flag = False
|
261 |
+
for bbox_string in bbox_list:
|
262 |
+
integers = re.findall(r'-?\d+', bbox_string)
|
263 |
+
if len(integers)==4:
|
264 |
+
angle=0
|
265 |
+
else:
|
266 |
+
angle=integers[4]
|
267 |
+
integers=integers[:-1]
|
268 |
+
|
269 |
+
if len(integers) == 4:
|
270 |
+
x0, y0, x1, y1 = int(integers[0]), int(integers[1]), int(integers[2]), int(integers[3])
|
271 |
+
left = x0 / bounding_box_size * image_width
|
272 |
+
bottom = y0 / bounding_box_size * image_height
|
273 |
+
right = x1 / bounding_box_size * image_width
|
274 |
+
top = y1 / bounding_box_size * image_height
|
275 |
+
|
276 |
+
entities[obj].append([left, bottom, right, top,angle])
|
277 |
+
|
278 |
+
j += 1
|
279 |
+
flag = True
|
280 |
+
if flag:
|
281 |
+
i += 1
|
282 |
+
else:
|
283 |
+
integers = re.findall(r'-?\d+', generation)
|
284 |
+
# if len(integers)==4:
|
285 |
+
angle=0
|
286 |
+
# else:
|
287 |
+
# angle=integers[4]
|
288 |
+
integers=integers[:-1]
|
289 |
+
if len(integers) == 4: # it is refer
|
290 |
+
mode = 'single'
|
291 |
+
|
292 |
+
entities = list()
|
293 |
+
x0, y0, x1, y1 = int(integers[0]), int(integers[1]), int(integers[2]), int(integers[3])
|
294 |
+
left = x0 / bounding_box_size * image_width
|
295 |
+
bottom = y0 / bounding_box_size * image_height
|
296 |
+
right = x1 / bounding_box_size * image_width
|
297 |
+
top = y1 / bounding_box_size * image_height
|
298 |
+
entities.append([left, bottom, right, top,angle])
|
299 |
+
else:
|
300 |
+
# don't detect any valid bbox to visualize
|
301 |
+
return None, ''
|
302 |
+
|
303 |
+
if len(entities) == 0:
|
304 |
+
return None, ''
|
305 |
+
|
306 |
+
if isinstance(image, Image.Image):
|
307 |
+
image_h = image.height
|
308 |
+
image_w = image.width
|
309 |
+
image = np.array(image)
|
310 |
+
|
311 |
+
elif isinstance(image, str):
|
312 |
+
if os.path.exists(image):
|
313 |
+
pil_img = Image.open(image).convert("RGB")
|
314 |
+
image = np.array(pil_img)[:, :, [2, 1, 0]]
|
315 |
+
image_h = pil_img.height
|
316 |
+
image_w = pil_img.width
|
317 |
+
else:
|
318 |
+
raise ValueError(f"invaild image path, {image}")
|
319 |
+
elif isinstance(image, torch.Tensor):
|
320 |
+
|
321 |
+
image_tensor = image.cpu()
|
322 |
+
reverse_norm_mean = torch.tensor([0.48145466, 0.4578275, 0.40821073])[:, None, None]
|
323 |
+
reverse_norm_std = torch.tensor([0.26862954, 0.26130258, 0.27577711])[:, None, None]
|
324 |
+
image_tensor = image_tensor * reverse_norm_std + reverse_norm_mean
|
325 |
+
pil_img = T.ToPILImage()(image_tensor)
|
326 |
+
image_h = pil_img.height
|
327 |
+
image_w = pil_img.width
|
328 |
+
image = np.array(pil_img)[:, :, [2, 1, 0]]
|
329 |
+
else:
|
330 |
+
raise ValueError(f"invalid image format, {type(image)} for {image}")
|
331 |
+
|
332 |
+
indices = list(range(len(entities)))
|
333 |
+
|
334 |
+
new_image = image.copy()
|
335 |
+
|
336 |
+
previous_bboxes = []
|
337 |
+
# size of text
|
338 |
+
text_size = 0.4
|
339 |
+
# thickness of text
|
340 |
+
text_line = 1 # int(max(1 * min(image_h, image_w) / 512, 1))
|
341 |
+
box_line = 2
|
342 |
+
(c_width, text_height), _ = cv2.getTextSize("F", cv2.FONT_HERSHEY_COMPLEX, text_size, text_line)
|
343 |
+
base_height = int(text_height * 0.675)
|
344 |
+
text_offset_original = text_height - base_height
|
345 |
+
text_spaces = 2
|
346 |
+
|
347 |
+
# num_bboxes = sum(len(x[-1]) for x in entities)
|
348 |
+
used_colors = colors # random.sample(colors, k=num_bboxes)
|
349 |
+
|
350 |
+
color_id = -1
|
351 |
+
for entity_idx, entity_name in enumerate(entities):
|
352 |
+
if mode == 'single' or mode == 'identify':
|
353 |
+
bboxes = entity_name
|
354 |
+
bboxes = [bboxes]
|
355 |
+
else:
|
356 |
+
bboxes = entities[entity_name]
|
357 |
+
color_id += 1
|
358 |
+
for bbox_id, (x1_norm, y1_norm, x2_norm, y2_norm,angle) in enumerate(bboxes):
|
359 |
+
skip_flag = False
|
360 |
+
orig_x1, orig_y1, orig_x2, orig_y2,angle = int(x1_norm), int(y1_norm), int(x2_norm), int(y2_norm), int(angle)
|
361 |
+
|
362 |
+
color = used_colors[entity_idx % len(used_colors)] # tuple(np.random.randint(0, 255, size=3).tolist())
|
363 |
+
top_right=(orig_x1,orig_y1)
|
364 |
+
bottom_left=(orig_x2,orig_y2)
|
365 |
+
angle=angle
|
366 |
+
rotated_bbox = rotate_bbox(top_right, bottom_left, angle)
|
367 |
+
new_image=cv2.polylines(new_image, [rotated_bbox.astype(np.int32)], isClosed=True,thickness=2, color=color)
|
368 |
+
|
369 |
+
# new_image = cv2.rectangle(new_image, (orig_x1, orig_y1), (orig_x2, orig_y2), color, box_line)
|
370 |
+
|
371 |
+
if mode == 'all':
|
372 |
+
l_o, r_o = box_line // 2 + box_line % 2, box_line // 2 + box_line % 2 + 1
|
373 |
+
|
374 |
+
x1 = orig_x1 - l_o
|
375 |
+
y1 = orig_y1 - l_o
|
376 |
+
|
377 |
+
if y1 < text_height + text_offset_original + 2 * text_spaces:
|
378 |
+
y1 = orig_y1 + r_o + text_height + text_offset_original + 2 * text_spaces
|
379 |
+
x1 = orig_x1 + r_o
|
380 |
+
|
381 |
+
# add text background
|
382 |
+
(text_width, text_height), _ = cv2.getTextSize(f" {entity_name}", cv2.FONT_HERSHEY_COMPLEX, text_size,
|
383 |
+
text_line)
|
384 |
+
text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2 = x1, y1 - (
|
385 |
+
text_height + text_offset_original + 2 * text_spaces), x1 + text_width, y1
|
386 |
+
|
387 |
+
for prev_bbox in previous_bboxes:
|
388 |
+
if computeIoU((text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2), prev_bbox['bbox']) > 0.95 and \
|
389 |
+
prev_bbox['phrase'] == entity_name:
|
390 |
+
skip_flag = True
|
391 |
+
break
|
392 |
+
while is_overlapping((text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2), prev_bbox['bbox']):
|
393 |
+
text_bg_y1 += (text_height + text_offset_original + 2 * text_spaces)
|
394 |
+
text_bg_y2 += (text_height + text_offset_original + 2 * text_spaces)
|
395 |
+
y1 += (text_height + text_offset_original + 2 * text_spaces)
|
396 |
+
|
397 |
+
if text_bg_y2 >= image_h:
|
398 |
+
text_bg_y1 = max(0, image_h - (text_height + text_offset_original + 2 * text_spaces))
|
399 |
+
text_bg_y2 = image_h
|
400 |
+
y1 = image_h
|
401 |
+
break
|
402 |
+
if not skip_flag:
|
403 |
+
alpha = 0.5
|
404 |
+
for i in range(text_bg_y1, text_bg_y2):
|
405 |
+
for j in range(text_bg_x1, text_bg_x2):
|
406 |
+
if i < image_h and j < image_w:
|
407 |
+
if j < text_bg_x1 + 1.35 * c_width:
|
408 |
+
# original color
|
409 |
+
bg_color = color
|
410 |
+
else:
|
411 |
+
# white
|
412 |
+
bg_color = [255, 255, 255]
|
413 |
+
new_image[i, j] = (alpha * new_image[i, j] + (1 - alpha) * np.array(bg_color)).astype(
|
414 |
+
np.uint8)
|
415 |
+
|
416 |
+
cv2.putText(
|
417 |
+
new_image, f" {entity_name}", (x1, y1 - text_offset_original - 1 * text_spaces),
|
418 |
+
cv2.FONT_HERSHEY_COMPLEX, text_size, (0, 0, 0), text_line, cv2.LINE_AA
|
419 |
+
)
|
420 |
+
|
421 |
+
previous_bboxes.append(
|
422 |
+
{'bbox': (text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2), 'phrase': entity_name})
|
423 |
+
|
424 |
+
if mode == 'all':
|
425 |
+
def color_iterator(colors):
|
426 |
+
while True:
|
427 |
+
for color in colors:
|
428 |
+
yield color
|
429 |
+
|
430 |
+
color_gen = color_iterator(colors)
|
431 |
+
|
432 |
+
# Add colors to phrases and remove <p></p>
|
433 |
+
def colored_phrases(match):
|
434 |
+
phrase = match.group(1)
|
435 |
+
color = next(color_gen)
|
436 |
+
return f'<span style="color:rgb{color}">{phrase}</span>'
|
437 |
+
|
438 |
+
generation = re.sub(r'{<\d+><\d+><\d+><\d+>}|<delim>', '', generation)
|
439 |
+
generation_colored = re.sub(r'<p>(.*?)</p>', colored_phrases, generation)
|
440 |
+
else:
|
441 |
+
generation_colored = ''
|
442 |
+
|
443 |
+
pil_image = Image.fromarray(new_image)
|
444 |
+
return pil_image, generation_colored
|
445 |
+
|
446 |
+
|
447 |
+
def gradio_reset(chat_state, img_list):
|
448 |
+
if chat_state is not None:
|
449 |
+
chat_state.messages = []
|
450 |
+
if img_list is not None:
|
451 |
+
img_list = []
|
452 |
+
return None, gr.update(value=None, interactive=True), gr.update(placeholder='Upload your image and chat',
|
453 |
+
interactive=True), chat_state, img_list
|
454 |
+
|
455 |
+
|
456 |
+
def image_upload_trigger(upload_flag, replace_flag, img_list):
|
457 |
+
# set the upload flag to true when receive a new image.
|
458 |
+
# if there is an old image (and old conversation), set the replace flag to true to reset the conv later.
|
459 |
+
upload_flag = 1
|
460 |
+
if img_list:
|
461 |
+
replace_flag = 1
|
462 |
+
return upload_flag, replace_flag
|
463 |
+
|
464 |
+
|
465 |
+
def example_trigger(text_input, image, upload_flag, replace_flag, img_list):
|
466 |
+
# set the upload flag to true when receive a new image.
|
467 |
+
# if there is an old image (and old conversation), set the replace flag to true to reset the conv later.
|
468 |
+
upload_flag = 1
|
469 |
+
if img_list or replace_flag == 1:
|
470 |
+
replace_flag = 1
|
471 |
+
|
472 |
+
return upload_flag, replace_flag
|
473 |
+
|
474 |
+
|
475 |
+
def gradio_ask(user_message, chatbot, chat_state, gr_img, img_list, upload_flag, replace_flag):
|
476 |
+
if len(user_message) == 0:
|
477 |
+
text_box_show = 'Input should not be empty!'
|
478 |
+
else:
|
479 |
+
text_box_show = ''
|
480 |
+
|
481 |
+
if isinstance(gr_img, dict):
|
482 |
+
gr_img, mask = gr_img['image'], gr_img['mask']
|
483 |
+
else:
|
484 |
+
mask = None
|
485 |
+
|
486 |
+
if '[identify]' in user_message:
|
487 |
+
# check if user provide bbox in the text input
|
488 |
+
integers = re.findall(r'-?\d+', user_message)
|
489 |
+
if len(integers) != 4: # no bbox in text
|
490 |
+
bbox = mask2bbox(mask)
|
491 |
+
user_message = user_message + bbox
|
492 |
+
|
493 |
+
if chat_state is None:
|
494 |
+
chat_state = CONV_VISION.copy()
|
495 |
+
|
496 |
+
if upload_flag:
|
497 |
+
if replace_flag:
|
498 |
+
chat_state = CONV_VISION.copy() # new image, reset everything
|
499 |
+
replace_flag = 0
|
500 |
+
chatbot = []
|
501 |
+
img_list = []
|
502 |
+
llm_message = chat.upload_img(gr_img, chat_state, img_list)
|
503 |
+
upload_flag = 0
|
504 |
+
|
505 |
+
chat.ask(user_message, chat_state)
|
506 |
+
|
507 |
+
chatbot = chatbot + [[user_message, None]]
|
508 |
+
|
509 |
+
if '[identify]' in user_message:
|
510 |
+
visual_img, _ = visualize_all_bbox_together(gr_img, user_message)
|
511 |
+
if visual_img is not None:
|
512 |
+
file_path = save_tmp_img(visual_img)
|
513 |
+
chatbot = chatbot + [[(file_path,), None]]
|
514 |
+
|
515 |
+
return text_box_show, chatbot, chat_state, img_list, upload_flag, replace_flag
|
516 |
+
|
517 |
+
|
518 |
+
# def gradio_answer(chatbot, chat_state, img_list, temperature):
|
519 |
+
# llm_message = chat.answer(conv=chat_state,
|
520 |
+
# img_list=img_list,
|
521 |
+
# temperature=temperature,
|
522 |
+
# max_new_tokens=500,
|
523 |
+
# max_length=2000)[0]
|
524 |
+
# chatbot[-1][1] = llm_message
|
525 |
+
# return chatbot, chat_state
|
526 |
+
|
527 |
+
|
528 |
+
def gradio_stream_answer(chatbot, chat_state, img_list, temperature):
|
529 |
+
if len(img_list) > 0:
|
530 |
+
if not isinstance(img_list[0], torch.Tensor):
|
531 |
+
chat.encode_img(img_list)
|
532 |
+
streamer = chat.stream_answer(conv=chat_state,
|
533 |
+
img_list=img_list,
|
534 |
+
temperature=temperature,
|
535 |
+
max_new_tokens=500,
|
536 |
+
max_length=2000)
|
537 |
+
# chatbot[-1][1] = output
|
538 |
+
# chat_state.messages[-1][1] = '</s>'
|
539 |
+
|
540 |
+
output = ''
|
541 |
+
for new_output in streamer:
|
542 |
+
# print(new_output)
|
543 |
+
output=output+new_output
|
544 |
+
print(output)
|
545 |
+
# if "{" in output:
|
546 |
+
# chatbot[-1][1]="Grounding and referring expression is still under work."
|
547 |
+
# else:
|
548 |
+
output = escape_markdown(output)
|
549 |
+
# output += escapped
|
550 |
+
chatbot[-1][1] = output
|
551 |
+
yield chatbot, chat_state
|
552 |
+
chat_state.messages[-1][1] = '</s>'
|
553 |
+
return chatbot, chat_state
|
554 |
+
|
555 |
+
|
556 |
+
def gradio_visualize(chatbot, gr_img):
|
557 |
+
if isinstance(gr_img, dict):
|
558 |
+
gr_img, mask = gr_img['image'], gr_img['mask']
|
559 |
+
|
560 |
+
unescaped = reverse_escape(chatbot[-1][1])
|
561 |
+
visual_img, generation_color = visualize_all_bbox_together(gr_img, unescaped)
|
562 |
+
if visual_img is not None:
|
563 |
+
if len(generation_color):
|
564 |
+
chatbot[-1][1] = generation_color
|
565 |
+
file_path = save_tmp_img(visual_img)
|
566 |
+
chatbot = chatbot + [[None, (file_path,)]]
|
567 |
+
|
568 |
+
return chatbot
|
569 |
+
|
570 |
+
|
571 |
+
def gradio_taskselect(idx):
|
572 |
+
prompt_list = [
|
573 |
+
'',
|
574 |
+
'Classify the image in the following classes: ',
|
575 |
+
'[identify] what is this ',
|
576 |
+
]
|
577 |
+
instruct_list = [
|
578 |
+
'**Hint:** Type in whatever you want',
|
579 |
+
'**Hint:** Type in the classes you want the model to classify in',
|
580 |
+
'**Hint:** Draw a bounding box on the uploaded image then send the command. Click the "clear" botton on the top right of the image before redraw',
|
581 |
+
]
|
582 |
+
return prompt_list[idx], instruct_list[idx]
|
583 |
+
|
584 |
+
|
585 |
+
|
586 |
+
|
587 |
+
chat = Chat(model, image_processor,tokenizer, device=device)
|
588 |
+
|
589 |
+
|
590 |
+
title = """<h1 align="center">GeoChat Demo</h1>"""
|
591 |
+
description = 'Welcome to Our GeoChat Chatbot Demo!'
|
592 |
+
article = """<div style="display: flex;"><p style="display: inline-block;"><a href='https://mbzuai-oryx.github.io/GeoChat'><img src='https://img.shields.io/badge/Project-Page-Green'></a></p><p style="display: inline-block;"><a href='https://arxiv.org/abs/2311.15826'><img src='https://img.shields.io/badge/Paper-PDF-red'></a></p><p style="display: inline-block;"><a href='https://github.com/mbzuai-oryx/GeoChat/tree/main'><img src='https://img.shields.io/badge/GitHub-Repo-blue'></a></p><p style="display: inline-block;"><a href='https://youtu.be/KOKtkkKpNDk?feature=shared'><img src='https://img.shields.io/badge/YouTube-Video-red'></a></p></div>"""
|
593 |
+
# article = """<p><a href='https://minigpt-v2.github.io'><img src='https://img.shields.io/badge/Project-Page-Green'></a></p>"""
|
594 |
+
|
595 |
+
introduction = '''
|
596 |
+
1. Identify: Draw the bounding box on the uploaded image window and CLICK **Send** to generate the bounding box. (CLICK "clear" button before re-drawing next time).
|
597 |
+
2. No Tag: Input whatever you want and CLICK **Send** without any tagging
|
598 |
+
|
599 |
+
You can also simply chat in free form!
|
600 |
+
'''
|
601 |
+
|
602 |
+
|
603 |
+
text_input = gr.Textbox(placeholder='Upload your image and chat', interactive=True, show_label=False, container=False,
|
604 |
+
scale=12)
|
605 |
+
with gr.Blocks() as demo:
|
606 |
+
gr.Markdown(title)
|
607 |
+
# gr.Markdown(description)
|
608 |
+
gr.Markdown(article)
|
609 |
+
|
610 |
+
with gr.Row():
|
611 |
+
with gr.Column(scale=0.5):
|
612 |
+
image = gr.Image(type="pil", tool='sketch', brush_radius=20)
|
613 |
+
|
614 |
+
temperature = gr.Slider(
|
615 |
+
minimum=0.1,
|
616 |
+
maximum=1.5,
|
617 |
+
value=0.6,
|
618 |
+
step=0.1,
|
619 |
+
interactive=True,
|
620 |
+
label="Temperature",
|
621 |
+
)
|
622 |
+
|
623 |
+
clear = gr.Button("Restart")
|
624 |
+
|
625 |
+
gr.Markdown(introduction)
|
626 |
+
|
627 |
+
with gr.Column():
|
628 |
+
chat_state = gr.State(value=None)
|
629 |
+
img_list = gr.State(value=[])
|
630 |
+
chatbot = gr.Chatbot(label='GeoChat')
|
631 |
+
|
632 |
+
dataset = gr.Dataset(
|
633 |
+
components=[gr.Textbox(visible=False)],
|
634 |
+
samples=[['No Tag'], ['Scene Classification'],['Identify']],
|
635 |
+
type="index",
|
636 |
+
label='Task Shortcuts',
|
637 |
+
)
|
638 |
+
task_inst = gr.Markdown('**Hint:** Upload your image and chat')
|
639 |
+
with gr.Row():
|
640 |
+
text_input.render()
|
641 |
+
send = gr.Button("Send", variant='primary', size='sm', scale=1)
|
642 |
+
|
643 |
+
upload_flag = gr.State(value=0)
|
644 |
+
replace_flag = gr.State(value=0)
|
645 |
+
image.upload(image_upload_trigger, [upload_flag, replace_flag, img_list], [upload_flag, replace_flag])
|
646 |
+
|
647 |
+
with gr.Row():
|
648 |
+
with gr.Column():
|
649 |
+
gr.Examples(examples=[
|
650 |
+
["demo_images/train_2956_0001.png", "Where are the airplanes located and what is their type?", upload_flag, replace_flag,
|
651 |
+
img_list],
|
652 |
+
["demo_images/7292.JPG", "How many buildings are flooded?", upload_flag,
|
653 |
+
replace_flag, img_list],
|
654 |
+
], inputs=[image, text_input, upload_flag, replace_flag, img_list], fn=example_trigger,
|
655 |
+
outputs=[upload_flag, replace_flag])
|
656 |
+
with gr.Column():
|
657 |
+
gr.Examples(examples=[
|
658 |
+
["demo_images/church_183.png", "Classify the image in the following classes: Church, Beach, Dense Residential, Storage Tanks.",
|
659 |
+
upload_flag, replace_flag, img_list],
|
660 |
+
["demo_images/04444.png", "[identify] what is this {<8><26><22><37>}", upload_flag,
|
661 |
+
replace_flag, img_list],
|
662 |
+
], inputs=[image, text_input, upload_flag, replace_flag, img_list], fn=example_trigger,
|
663 |
+
outputs=[upload_flag, replace_flag])
|
664 |
+
|
665 |
+
dataset.click(
|
666 |
+
gradio_taskselect,
|
667 |
+
inputs=[dataset],
|
668 |
+
outputs=[text_input, task_inst],
|
669 |
+
show_progress="hidden",
|
670 |
+
postprocess=False,
|
671 |
+
queue=False,
|
672 |
+
)
|
673 |
+
|
674 |
+
text_input.submit(
|
675 |
+
gradio_ask,
|
676 |
+
[text_input, chatbot, chat_state, image, img_list, upload_flag, replace_flag],
|
677 |
+
[text_input, chatbot, chat_state, img_list, upload_flag, replace_flag], queue=False
|
678 |
+
).success(
|
679 |
+
gradio_stream_answer,
|
680 |
+
[chatbot, chat_state, img_list, temperature],
|
681 |
+
[chatbot, chat_state]
|
682 |
+
).success(
|
683 |
+
gradio_visualize,
|
684 |
+
[chatbot, image],
|
685 |
+
[chatbot],
|
686 |
+
queue=False,
|
687 |
+
)
|
688 |
+
|
689 |
+
send.click(
|
690 |
+
gradio_ask,
|
691 |
+
[text_input, chatbot, chat_state, image, img_list, upload_flag, replace_flag],
|
692 |
+
[text_input, chatbot, chat_state, img_list, upload_flag, replace_flag], queue=False
|
693 |
+
).success(
|
694 |
+
gradio_stream_answer,
|
695 |
+
[chatbot, chat_state, img_list, temperature],
|
696 |
+
[chatbot, chat_state]
|
697 |
+
).success(
|
698 |
+
gradio_visualize,
|
699 |
+
[chatbot, image],
|
700 |
+
[chatbot],
|
701 |
+
queue=False,
|
702 |
+
)
|
703 |
+
|
704 |
+
clear.click(gradio_reset, [chat_state, img_list], [chatbot, image, text_input, chat_state, img_list], queue=False)
|
705 |
+
|
706 |
+
demo.launch(share=True, enable_queue=True,server_name='0.0.0.0')
|
pyproject.toml
ADDED
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[build-system]
|
2 |
+
requires = ["setuptools>=61.0"]
|
3 |
+
build-backend = "setuptools.build_meta"
|
4 |
+
|
5 |
+
[project]
|
6 |
+
name = "geochat"
|
7 |
+
version = "1.1.1"
|
8 |
+
description = "Grounded VLM for Remote Sensing"
|
9 |
+
readme = "README.md"
|
10 |
+
requires-python = ">=3.8"
|
11 |
+
classifiers = [
|
12 |
+
"Programming Language :: Python :: 3",
|
13 |
+
"License :: OSI Approved :: Apache Software License",
|
14 |
+
]
|
15 |
+
dependencies = [
|
16 |
+
"einops", "fastapi", "gradio==3.35.2", "markdown2[all]", "numpy",
|
17 |
+
"requests", "sentencepiece", "tokenizers>=0.12.1",
|
18 |
+
"torch==2.0.1", "torchvision==0.15.2", "uvicorn", "wandb",
|
19 |
+
"shortuuid", "httpx==0.24.0",
|
20 |
+
"deepspeed==0.9.5",
|
21 |
+
"peft==0.4.0",
|
22 |
+
"transformers==4.31.0",
|
23 |
+
"accelerate==0.21.0",
|
24 |
+
"bitsandbytes==0.41.0",
|
25 |
+
"scikit-learn==1.2.2",
|
26 |
+
"sentencepiece==0.1.99",
|
27 |
+
"einops==0.6.1", "einops-exts==0.0.4", "timm==0.6.13",
|
28 |
+
"gradio_client==0.2.9"
|
29 |
+
]
|
30 |
+
|
31 |
+
[project.urls]
|
32 |
+
"Homepage" = "https://github.com/mbzuai-oryx/GeoChat"
|
33 |
+
"Bug Tracker" = "https://github.com/mbzuai-oryx/GeoChat/issues"
|
34 |
+
|
35 |
+
[tool.setuptools.packages.find]
|
36 |
+
exclude = ["assets*", "benchmark*", "docs", "dist*", "playground*", "scripts*", "tests*"]
|
37 |
+
|
38 |
+
[tool.wheel]
|
39 |
+
exclude = ["assets*", "benchmark*", "docs", "dist*", "playground*", "scripts*", "tests*"]
|