Spaces:
Sleeping
Sleeping
AAAAAAyq
commited on
Commit
•
bd6726a
1
Parent(s):
7c3570f
try to fix CUDA bug
Browse files- app.py +176 -238
- app_copy.py +196 -0
- requirements.txt +1 -1
app.py
CHANGED
@@ -1,239 +1,177 @@
|
|
1 |
-
from ultralytics import YOLO
|
2 |
-
import numpy as np
|
3 |
-
import matplotlib.pyplot as plt
|
4 |
-
import gradio as gr
|
5 |
-
import cv2
|
6 |
-
import torch
|
7 |
-
# import queue
|
8 |
-
# import threading
|
9 |
-
# from PIL import Image
|
10 |
-
|
11 |
-
|
12 |
-
model = YOLO('checkpoints/FastSAM.pt') # load a custom model
|
13 |
-
|
14 |
-
|
15 |
-
def fast_process(annotations, image, high_quality, device):
|
16 |
-
if isinstance(annotations[0],dict):
|
17 |
-
annotations = [annotation['segmentation'] for annotation in annotations]
|
18 |
-
|
19 |
-
original_h = image.height
|
20 |
-
original_w = image.width
|
21 |
-
fig = plt.figure(figsize=(10, 10))
|
22 |
-
plt.imshow(image)
|
23 |
-
if high_quality == True:
|
24 |
-
if isinstance(annotations[0],torch.Tensor):
|
25 |
-
annotations = np.array(annotations.cpu())
|
26 |
-
for i, mask in enumerate(annotations):
|
27 |
-
mask = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_CLOSE, np.ones((3, 3), np.uint8))
|
28 |
-
annotations[i] = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_OPEN, np.ones((8, 8), np.uint8))
|
29 |
-
if device == 'cpu':
|
30 |
-
annotations = np.array(annotations)
|
31 |
-
fast_show_mask(annotations,
|
32 |
-
plt.gca(),
|
33 |
-
bbox=None,
|
34 |
-
points=None,
|
35 |
-
pointlabel=None,
|
36 |
-
retinamask=True,
|
37 |
-
target_height=original_h,
|
38 |
-
target_width=original_w)
|
39 |
-
else:
|
40 |
-
if isinstance(annotations[0],np.ndarray):
|
41 |
-
annotations = torch.from_numpy(annotations)
|
42 |
-
fast_show_mask_gpu(annotations,
|
43 |
-
plt.gca(),
|
44 |
-
bbox=None,
|
45 |
-
points=None,
|
46 |
-
pointlabel=None)
|
47 |
-
if isinstance(annotations, torch.Tensor):
|
48 |
-
annotations = annotations.cpu().numpy()
|
49 |
-
if high_quality == True:
|
50 |
-
contour_all = []
|
51 |
-
temp = np.zeros((original_h, original_w,1))
|
52 |
-
for i, mask in enumerate(annotations):
|
53 |
-
if type(mask) == dict:
|
54 |
-
mask = mask['segmentation']
|
55 |
-
annotation = mask.astype(np.uint8)
|
56 |
-
contours, _ = cv2.findContours(annotation, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
57 |
-
for contour in contours:
|
58 |
-
contour_all.append(contour)
|
59 |
-
cv2.drawContours(temp, contour_all, -1, (255, 255, 255), 2)
|
60 |
-
color = np.array([0 / 255, 0 / 255, 255 / 255, 0.8])
|
61 |
-
contour_mask = temp / 225 * color.reshape(1, 1, -1)
|
62 |
-
plt.imshow(contour_mask)
|
63 |
-
|
64 |
-
plt.axis('off')
|
65 |
-
plt.tight_layout()
|
66 |
-
return fig
|
67 |
-
|
68 |
-
|
69 |
-
# CPU post process
|
70 |
-
def fast_show_mask(annotation, ax, bbox=None,
|
71 |
-
points=None, pointlabel=None,
|
72 |
-
retinamask=True, target_height=960,
|
73 |
-
target_width=960):
|
74 |
-
msak_sum = annotation.shape[0]
|
75 |
-
height = annotation.shape[1]
|
76 |
-
weight = annotation.shape[2]
|
77 |
-
# 将annotation 按照面积 排序
|
78 |
-
areas = np.sum(annotation, axis=(1, 2))
|
79 |
-
sorted_indices = np.argsort(areas)[::1]
|
80 |
-
annotation = annotation[sorted_indices]
|
81 |
-
|
82 |
-
index = (annotation != 0).argmax(axis=0)
|
83 |
-
color = np.random.random((msak_sum,1,1,3))
|
84 |
-
transparency = np.ones((msak_sum,1,1,1)) * 0.6
|
85 |
-
visual = np.concatenate([color,transparency],axis=-1)
|
86 |
-
mask_image = np.expand_dims(annotation,-1) * visual
|
87 |
-
|
88 |
-
show = np.zeros((height,weight,4))
|
89 |
-
|
90 |
-
h_indices, w_indices = np.meshgrid(np.arange(height), np.arange(weight), indexing='ij')
|
91 |
-
indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))
|
92 |
-
# 使用向量化索引更新show的值
|
93 |
-
show[h_indices, w_indices, :] = mask_image[indices]
|
94 |
-
if bbox is not None:
|
95 |
-
x1, y1, x2, y2 = bbox
|
96 |
-
ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor='b', linewidth=1))
|
97 |
-
# draw point
|
98 |
-
if points is not None:
|
99 |
-
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==1], [point[1] for i, point in enumerate(points) if pointlabel[i]==1], s=20, c='y')
|
100 |
-
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==0], [point[1] for i, point in enumerate(points) if pointlabel[i]==0], s=20, c='m')
|
101 |
-
|
102 |
-
if retinamask==False:
|
103 |
-
show = cv2.resize(show,(target_width,target_height),interpolation=cv2.INTER_NEAREST)
|
104 |
-
ax.imshow(show)
|
105 |
-
|
106 |
-
|
107 |
-
def fast_show_mask_gpu(annotation, ax,
|
108 |
-
bbox=None, points=None,
|
109 |
-
pointlabel=None):
|
110 |
-
msak_sum = annotation.shape[0]
|
111 |
-
height = annotation.shape[1]
|
112 |
-
weight = annotation.shape[2]
|
113 |
-
areas = torch.sum(annotation, dim=(1, 2))
|
114 |
-
sorted_indices = torch.argsort(areas, descending=False)
|
115 |
-
annotation = annotation[sorted_indices]
|
116 |
-
# 找每个位置第一个非零值下标
|
117 |
-
index = (annotation != 0).to(torch.long).argmax(dim=0)
|
118 |
-
color = torch.rand((msak_sum,1,1,3)).to(annotation.device)
|
119 |
-
transparency = torch.ones((msak_sum,1,1,1)).to(annotation.device) * 0.6
|
120 |
-
visual = torch.cat([color,transparency],dim=-1)
|
121 |
-
mask_image = torch.unsqueeze(annotation,-1) * visual
|
122 |
-
# 按index取数,index指每个位置选哪个batch的数,把mask_image转成一个batch的形式
|
123 |
-
show = torch.zeros((height,weight,4)).to(annotation.device)
|
124 |
-
h_indices, w_indices = torch.meshgrid(torch.arange(height), torch.arange(weight))
|
125 |
-
indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))
|
126 |
-
# 使用向量化索引更新show的值
|
127 |
-
show[h_indices, w_indices, :] = mask_image[indices]
|
128 |
-
show_cpu = show.cpu().numpy()
|
129 |
-
if bbox is not None:
|
130 |
-
x1, y1, x2, y2 = bbox
|
131 |
-
ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor='b', linewidth=1))
|
132 |
-
# draw point
|
133 |
-
if points is not None:
|
134 |
-
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==1], [point[1] for i, point in enumerate(points) if pointlabel[i]==1], s=20, c='y')
|
135 |
-
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==0], [point[1] for i, point in enumerate(points) if pointlabel[i]==0], s=20, c='m')
|
136 |
-
ax.imshow(show_cpu)
|
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 |
-
# print('4')
|
178 |
-
# # 将结果放回队列
|
179 |
-
# prediction_queue.put(fig)
|
180 |
-
|
181 |
-
# # 在一个新的线程中启动工作函数
|
182 |
-
# threading.Thread(target=worker).start()
|
183 |
-
|
184 |
-
# # 将耗时的函数包装在另一个函数中,用于控制队列和线程同步
|
185 |
-
# def process_request():
|
186 |
-
# while True:
|
187 |
-
# if not request_queue.empty():
|
188 |
-
# # 如果请求队列不为空,则处理该请求
|
189 |
-
# try:
|
190 |
-
# lock.put(1) # 加锁,防止同时处理多个请求
|
191 |
-
# input, input_size, high_visual_quality = request_queue.get()
|
192 |
-
# fig = predict(input, input_size, high_visual_quality)
|
193 |
-
# request_queue.task_done() # 请求处理结束,移除请求
|
194 |
-
# lock.get() # 解锁
|
195 |
-
# yield fig # 返回预测结果
|
196 |
-
# except:
|
197 |
-
# lock.get() # 出错时也需要解锁
|
198 |
-
# else:
|
199 |
-
# # 如果请求队列为空,则等待新的请求到达
|
200 |
-
# time.sleep(1)
|
201 |
-
|
202 |
-
|
203 |
-
# input_size=1024
|
204 |
-
# high_quality_visual=True
|
205 |
-
# inp = 'assets/sa_192.jpg'
|
206 |
-
# input = Image.open(inp)
|
207 |
-
# device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
208 |
-
# input_size = int(input_size) # 确保 imgsz 是整数
|
209 |
-
# results = model(input, device=device, retina_masks=True, iou=0.7, conf=0.25, imgsz=input_size)
|
210 |
-
# pil_image = fast_process(annotations=results[0].masks.data,
|
211 |
-
# image=input, high_quality=high_quality_visual, device=device)
|
212 |
-
app_interface = gr.Interface(fn=predict,
|
213 |
-
inputs=[gr.components.Image(type='pil'),
|
214 |
-
gr.components.Slider(minimum=512, maximum=1024, value=1024, step=64, label='input_size'),
|
215 |
-
gr.components.Checkbox(value=False, label='high_visual_quality')],
|
216 |
-
outputs=['plot'],
|
217 |
-
examples=[["assets/sa_8776.jpg", 1024, True]],
|
218 |
-
# # ["assets/sa_1309.jpg", 1024]],
|
219 |
-
# examples=[["assets/sa_192.jpg"], ["assets/sa_414.jpg"],
|
220 |
-
# ["assets/sa_561.jpg"], ["assets/sa_862.jpg"],
|
221 |
-
# ["assets/sa_1309.jpg"], ["assets/sa_8776.jpg"],
|
222 |
-
# ["assets/sa_10039.jpg"], ["assets/sa_11025.jpg"],],
|
223 |
-
cache_examples=True,
|
224 |
-
title="Fast Segment Anything (Everything mode)"
|
225 |
-
)
|
226 |
-
|
227 |
-
# # 定义一个请求处理函数��将请求添加到队列中
|
228 |
-
# def handle_request(value):
|
229 |
-
# try:
|
230 |
-
# request_queue.put_nowait(value) # 添加请求到队列
|
231 |
-
# except:
|
232 |
-
# return "当前队列已满,请稍后再试!"
|
233 |
-
# return None
|
234 |
-
|
235 |
-
# # 添加请求处理函数到应用程序界面
|
236 |
-
# app_interface.call_function()
|
237 |
-
|
238 |
-
app_interface.queue(concurrency_count=1, max_size=20)
|
239 |
app_interface.launch()
|
|
|
1 |
+
from ultralytics import YOLO
|
2 |
+
import numpy as np
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
import gradio as gr
|
5 |
+
import cv2
|
6 |
+
import torch
|
7 |
+
# import queue
|
8 |
+
# import threading
|
9 |
+
# from PIL import Image
|
10 |
+
|
11 |
+
|
12 |
+
model = YOLO('checkpoints/FastSAM.pt') # load a custom model
|
13 |
+
|
14 |
+
|
15 |
+
def fast_process(annotations, image, high_quality, device):
|
16 |
+
if isinstance(annotations[0],dict):
|
17 |
+
annotations = [annotation['segmentation'] for annotation in annotations]
|
18 |
+
|
19 |
+
original_h = image.height
|
20 |
+
original_w = image.width
|
21 |
+
fig = plt.figure(figsize=(10, 10))
|
22 |
+
plt.imshow(image)
|
23 |
+
if high_quality == True:
|
24 |
+
if isinstance(annotations[0],torch.Tensor):
|
25 |
+
annotations = np.array(annotations.cpu())
|
26 |
+
for i, mask in enumerate(annotations):
|
27 |
+
mask = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_CLOSE, np.ones((3, 3), np.uint8))
|
28 |
+
annotations[i] = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_OPEN, np.ones((8, 8), np.uint8))
|
29 |
+
if device == 'cpu':
|
30 |
+
annotations = np.array(annotations)
|
31 |
+
fast_show_mask(annotations,
|
32 |
+
plt.gca(),
|
33 |
+
bbox=None,
|
34 |
+
points=None,
|
35 |
+
pointlabel=None,
|
36 |
+
retinamask=True,
|
37 |
+
target_height=original_h,
|
38 |
+
target_width=original_w)
|
39 |
+
else:
|
40 |
+
if isinstance(annotations[0],np.ndarray):
|
41 |
+
annotations = torch.from_numpy(annotations)
|
42 |
+
fast_show_mask_gpu(annotations,
|
43 |
+
plt.gca(),
|
44 |
+
bbox=None,
|
45 |
+
points=None,
|
46 |
+
pointlabel=None)
|
47 |
+
if isinstance(annotations, torch.Tensor):
|
48 |
+
annotations = annotations.cpu().numpy()
|
49 |
+
if high_quality == True:
|
50 |
+
contour_all = []
|
51 |
+
temp = np.zeros((original_h, original_w,1))
|
52 |
+
for i, mask in enumerate(annotations):
|
53 |
+
if type(mask) == dict:
|
54 |
+
mask = mask['segmentation']
|
55 |
+
annotation = mask.astype(np.uint8)
|
56 |
+
contours, _ = cv2.findContours(annotation, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
57 |
+
for contour in contours:
|
58 |
+
contour_all.append(contour)
|
59 |
+
cv2.drawContours(temp, contour_all, -1, (255, 255, 255), 2)
|
60 |
+
color = np.array([0 / 255, 0 / 255, 255 / 255, 0.8])
|
61 |
+
contour_mask = temp / 225 * color.reshape(1, 1, -1)
|
62 |
+
plt.imshow(contour_mask)
|
63 |
+
|
64 |
+
plt.axis('off')
|
65 |
+
plt.tight_layout()
|
66 |
+
return fig
|
67 |
+
|
68 |
+
|
69 |
+
# CPU post process
|
70 |
+
def fast_show_mask(annotation, ax, bbox=None,
|
71 |
+
points=None, pointlabel=None,
|
72 |
+
retinamask=True, target_height=960,
|
73 |
+
target_width=960):
|
74 |
+
msak_sum = annotation.shape[0]
|
75 |
+
height = annotation.shape[1]
|
76 |
+
weight = annotation.shape[2]
|
77 |
+
# 将annotation 按照面积 排序
|
78 |
+
areas = np.sum(annotation, axis=(1, 2))
|
79 |
+
sorted_indices = np.argsort(areas)[::1]
|
80 |
+
annotation = annotation[sorted_indices]
|
81 |
+
|
82 |
+
index = (annotation != 0).argmax(axis=0)
|
83 |
+
color = np.random.random((msak_sum,1,1,3))
|
84 |
+
transparency = np.ones((msak_sum,1,1,1)) * 0.6
|
85 |
+
visual = np.concatenate([color,transparency],axis=-1)
|
86 |
+
mask_image = np.expand_dims(annotation,-1) * visual
|
87 |
+
|
88 |
+
show = np.zeros((height,weight,4))
|
89 |
+
|
90 |
+
h_indices, w_indices = np.meshgrid(np.arange(height), np.arange(weight), indexing='ij')
|
91 |
+
indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))
|
92 |
+
# 使用向量化索引更新show的值
|
93 |
+
show[h_indices, w_indices, :] = mask_image[indices]
|
94 |
+
if bbox is not None:
|
95 |
+
x1, y1, x2, y2 = bbox
|
96 |
+
ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor='b', linewidth=1))
|
97 |
+
# draw point
|
98 |
+
if points is not None:
|
99 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==1], [point[1] for i, point in enumerate(points) if pointlabel[i]==1], s=20, c='y')
|
100 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==0], [point[1] for i, point in enumerate(points) if pointlabel[i]==0], s=20, c='m')
|
101 |
+
|
102 |
+
if retinamask==False:
|
103 |
+
show = cv2.resize(show,(target_width,target_height),interpolation=cv2.INTER_NEAREST)
|
104 |
+
ax.imshow(show)
|
105 |
+
|
106 |
+
|
107 |
+
def fast_show_mask_gpu(annotation, ax,
|
108 |
+
bbox=None, points=None,
|
109 |
+
pointlabel=None):
|
110 |
+
msak_sum = annotation.shape[0]
|
111 |
+
height = annotation.shape[1]
|
112 |
+
weight = annotation.shape[2]
|
113 |
+
areas = torch.sum(annotation, dim=(1, 2))
|
114 |
+
sorted_indices = torch.argsort(areas, descending=False)
|
115 |
+
annotation = annotation[sorted_indices]
|
116 |
+
# 找每个位置第一个非零值下标
|
117 |
+
index = (annotation != 0).to(torch.long).argmax(dim=0)
|
118 |
+
color = torch.rand((msak_sum,1,1,3)).to(annotation.device)
|
119 |
+
transparency = torch.ones((msak_sum,1,1,1)).to(annotation.device) * 0.6
|
120 |
+
visual = torch.cat([color,transparency],dim=-1)
|
121 |
+
mask_image = torch.unsqueeze(annotation,-1) * visual
|
122 |
+
# 按index取数,index指每个位置选哪个batch的数,把mask_image转成一个batch的形式
|
123 |
+
show = torch.zeros((height,weight,4)).to(annotation.device)
|
124 |
+
h_indices, w_indices = torch.meshgrid(torch.arange(height), torch.arange(weight))
|
125 |
+
indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))
|
126 |
+
# 使用向量化索引更新show的值
|
127 |
+
show[h_indices, w_indices, :] = mask_image[indices]
|
128 |
+
show_cpu = show.cpu().numpy()
|
129 |
+
if bbox is not None:
|
130 |
+
x1, y1, x2, y2 = bbox
|
131 |
+
ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor='b', linewidth=1))
|
132 |
+
# draw point
|
133 |
+
if points is not None:
|
134 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==1], [point[1] for i, point in enumerate(points) if pointlabel[i]==1], s=20, c='y')
|
135 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==0], [point[1] for i, point in enumerate(points) if pointlabel[i]==0], s=20, c='m')
|
136 |
+
ax.imshow(show_cpu)
|
137 |
+
|
138 |
+
|
139 |
+
|
140 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
141 |
+
|
142 |
+
def predict(input, input_size=512, high_visual_quality=False):
|
143 |
+
input_size = int(input_size) # 确保 imgsz 是整数
|
144 |
+
results = model(input, device=device, retina_masks=True, iou=0.7, conf=0.25, imgsz=input_size)
|
145 |
+
fig = fast_process(annotations=results[0].masks.data,
|
146 |
+
image=input, high_quality=high_visual_quality, device=device)
|
147 |
+
return fig
|
148 |
+
|
149 |
+
|
150 |
+
|
151 |
+
# input_size=1024
|
152 |
+
# high_quality_visual=True
|
153 |
+
# inp = 'assets/sa_192.jpg'
|
154 |
+
# input = Image.open(inp)
|
155 |
+
# device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
156 |
+
# input_size = int(input_size) # 确保 imgsz 是整数
|
157 |
+
# results = model(input, device=device, retina_masks=True, iou=0.7, conf=0.25, imgsz=input_size)
|
158 |
+
# pil_image = fast_process(annotations=results[0].masks.data,
|
159 |
+
# image=input, high_quality=high_quality_visual, device=device)
|
160 |
+
app_interface = gr.Interface(fn=predict,
|
161 |
+
inputs=[gr.components.Image(type='pil'),
|
162 |
+
gr.components.Slider(minimum=512, maximum=1024, value=1024, step=64, label='input_size'),
|
163 |
+
gr.components.Checkbox(value=False, label='high_visual_quality')],
|
164 |
+
outputs=['plot'],
|
165 |
+
examples=[["assets/sa_8776.jpg", 1024, True]],
|
166 |
+
# # ["assets/sa_1309.jpg", 1024]],
|
167 |
+
# examples=[["assets/sa_192.jpg"], ["assets/sa_414.jpg"],
|
168 |
+
# ["assets/sa_561.jpg"], ["assets/sa_862.jpg"],
|
169 |
+
# ["assets/sa_1309.jpg"], ["assets/sa_8776.jpg"],
|
170 |
+
# ["assets/sa_10039.jpg"], ["assets/sa_11025.jpg"],],
|
171 |
+
cache_examples=True,
|
172 |
+
title="Fast Segment Anything (Everything mode)"
|
173 |
+
)
|
174 |
+
|
175 |
+
|
176 |
+
app_interface.queue(concurrency_count=1, max_size=20)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
app_interface.launch()
|
app_copy.py
ADDED
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from ultralytics import YOLO
|
2 |
+
import numpy as np
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
import gradio as gr
|
5 |
+
import cv2
|
6 |
+
import torch
|
7 |
+
# import queue
|
8 |
+
# import threading
|
9 |
+
from PIL import Image
|
10 |
+
|
11 |
+
|
12 |
+
model = YOLO('checkpoints/FastSAM.pt') # load a custom model
|
13 |
+
|
14 |
+
|
15 |
+
def fast_process(annotations, image, high_quality, device):
|
16 |
+
if isinstance(annotations[0],dict):
|
17 |
+
annotations = [annotation['segmentation'] for annotation in annotations]
|
18 |
+
|
19 |
+
original_h = image.height
|
20 |
+
original_w = image.width
|
21 |
+
# fig = plt.figure(figsize=(10, 10))
|
22 |
+
# plt.imshow(image)
|
23 |
+
if high_quality == True:
|
24 |
+
if isinstance(annotations[0],torch.Tensor):
|
25 |
+
annotations = np.array(annotations.cpu())
|
26 |
+
for i, mask in enumerate(annotations):
|
27 |
+
mask = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_CLOSE, np.ones((3, 3), np.uint8))
|
28 |
+
annotations[i] = cv2.morphologyEx(mask.astype(np.uint8), cv2.MORPH_OPEN, np.ones((8, 8), np.uint8))
|
29 |
+
if device == 'cpu':
|
30 |
+
annotations = np.array(annotations)
|
31 |
+
inner_mask = fast_show_mask(annotations,
|
32 |
+
plt.gca(),
|
33 |
+
bbox=None,
|
34 |
+
points=None,
|
35 |
+
pointlabel=None,
|
36 |
+
retinamask=True,
|
37 |
+
target_height=original_h,
|
38 |
+
target_width=original_w)
|
39 |
+
else:
|
40 |
+
if isinstance(annotations[0],np.ndarray):
|
41 |
+
annotations = torch.from_numpy(annotations)
|
42 |
+
inner_mask = fast_show_mask_gpu(annotations,
|
43 |
+
plt.gca(),
|
44 |
+
bbox=None,
|
45 |
+
points=None,
|
46 |
+
pointlabel=None)
|
47 |
+
if isinstance(annotations, torch.Tensor):
|
48 |
+
annotations = annotations.cpu().numpy()
|
49 |
+
if high_quality == True:
|
50 |
+
contour_all = []
|
51 |
+
temp = np.zeros((original_h, original_w,1))
|
52 |
+
for i, mask in enumerate(annotations):
|
53 |
+
if type(mask) == dict:
|
54 |
+
mask = mask['segmentation']
|
55 |
+
annotation = mask.astype(np.uint8)
|
56 |
+
contours, _ = cv2.findContours(annotation, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
|
57 |
+
for contour in contours:
|
58 |
+
contour_all.append(contour)
|
59 |
+
cv2.drawContours(temp, contour_all, -1, (255, 255, 255), 2)
|
60 |
+
color = np.array([0 / 255, 0 / 255, 255 / 255, 0.8])
|
61 |
+
contour_mask = temp / 255 * color.reshape(1, 1, -1)
|
62 |
+
# plt.imshow(contour_mask)
|
63 |
+
image = image.convert('RGBA')
|
64 |
+
overlay_inner = Image.fromarray((inner_mask * 255).astype(np.uint8), 'RGBA')
|
65 |
+
overlay_contour = Image.fromarray((contour_mask * 255).astype(np.uint8), 'RGBA')
|
66 |
+
# image = image.convert('RGBA')
|
67 |
+
# image = Image.alpha_composite(image, overlay_inner)
|
68 |
+
# image = Image.alpha_composite(image, overlay_contour)
|
69 |
+
image.paste(overlay_inner, (0, 0), overlay_inner)
|
70 |
+
image.paste(overlay_contour, (0, 0), overlay_contour)
|
71 |
+
|
72 |
+
return image
|
73 |
+
# plt.axis('off')
|
74 |
+
# plt.tight_layout()
|
75 |
+
# return fig
|
76 |
+
|
77 |
+
|
78 |
+
# CPU post process
|
79 |
+
def fast_show_mask(annotation, ax, bbox=None,
|
80 |
+
points=None, pointlabel=None,
|
81 |
+
retinamask=True, target_height=960,
|
82 |
+
target_width=960):
|
83 |
+
msak_sum = annotation.shape[0]
|
84 |
+
height = annotation.shape[1]
|
85 |
+
weight = annotation.shape[2]
|
86 |
+
# 将annotation 按照面积 排序
|
87 |
+
areas = np.sum(annotation, axis=(1, 2))
|
88 |
+
sorted_indices = np.argsort(areas)[::1]
|
89 |
+
annotation = annotation[sorted_indices]
|
90 |
+
|
91 |
+
index = (annotation != 0).argmax(axis=0)
|
92 |
+
color = np.random.random((msak_sum,1,1,3))
|
93 |
+
transparency = np.ones((msak_sum,1,1,1)) * 0.6
|
94 |
+
visual = np.concatenate([color,transparency],axis=-1)
|
95 |
+
mask_image = np.expand_dims(annotation,-1) * visual
|
96 |
+
|
97 |
+
mask = np.zeros((height,weight,4))
|
98 |
+
|
99 |
+
h_indices, w_indices = np.meshgrid(np.arange(height), np.arange(weight), indexing='ij')
|
100 |
+
indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))
|
101 |
+
# 使用向量化索引更新show的值
|
102 |
+
mask[h_indices, w_indices, :] = mask_image[indices]
|
103 |
+
if bbox is not None:
|
104 |
+
x1, y1, x2, y2 = bbox
|
105 |
+
ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor='b', linewidth=1))
|
106 |
+
# draw point
|
107 |
+
if points is not None:
|
108 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==1], [point[1] for i, point in enumerate(points) if pointlabel[i]==1], s=20, c='y')
|
109 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==0], [point[1] for i, point in enumerate(points) if pointlabel[i]==0], s=20, c='m')
|
110 |
+
|
111 |
+
if retinamask==False:
|
112 |
+
mask = cv2.resize(mask, (target_width, target_height), interpolation=cv2.INTER_NEAREST)
|
113 |
+
# ax.imshow(mask)
|
114 |
+
|
115 |
+
return mask
|
116 |
+
|
117 |
+
|
118 |
+
def fast_show_mask_gpu(annotation, ax,
|
119 |
+
bbox=None, points=None,
|
120 |
+
pointlabel=None):
|
121 |
+
msak_sum = annotation.shape[0]
|
122 |
+
height = annotation.shape[1]
|
123 |
+
weight = annotation.shape[2]
|
124 |
+
areas = torch.sum(annotation, dim=(1, 2))
|
125 |
+
sorted_indices = torch.argsort(areas, descending=False)
|
126 |
+
annotation = annotation[sorted_indices]
|
127 |
+
# 找每个位置第一个非零值下标
|
128 |
+
index = (annotation != 0).to(torch.long).argmax(dim=0)
|
129 |
+
color = torch.rand((msak_sum,1,1,3)).to(annotation.device)
|
130 |
+
transparency = torch.ones((msak_sum,1,1,1)).to(annotation.device) * 0.6
|
131 |
+
visual = torch.cat([color,transparency],dim=-1)
|
132 |
+
mask_image = torch.unsqueeze(annotation,-1) * visual
|
133 |
+
# 按index取数,index指每个位置选哪个batch的数,把mask_image转成一个batch的形式
|
134 |
+
mask = torch.zeros((height,weight,4)).to(annotation.device)
|
135 |
+
h_indices, w_indices = torch.meshgrid(torch.arange(height), torch.arange(weight))
|
136 |
+
indices = (index[h_indices, w_indices], h_indices, w_indices, slice(None))
|
137 |
+
# 使用向量化索引更新show的值
|
138 |
+
mask[h_indices, w_indices, :] = mask_image[indices]
|
139 |
+
mask_cpu = mask.cpu().numpy()
|
140 |
+
if bbox is not None:
|
141 |
+
x1, y1, x2, y2 = bbox
|
142 |
+
ax.add_patch(plt.Rectangle((x1, y1), x2 - x1, y2 - y1, fill=False, edgecolor='b', linewidth=1))
|
143 |
+
# draw point
|
144 |
+
if points is not None:
|
145 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==1], [point[1] for i, point in enumerate(points) if pointlabel[i]==1], s=20, c='y')
|
146 |
+
plt.scatter([point[0] for i, point in enumerate(points) if pointlabel[i]==0], [point[1] for i, point in enumerate(points) if pointlabel[i]==0], s=20, c='m')
|
147 |
+
# ax.imshow(mask_cpu)
|
148 |
+
return mask_cpu
|
149 |
+
|
150 |
+
|
151 |
+
# # 预测队列
|
152 |
+
# prediction_queue = queue.Queue(maxsize=5)
|
153 |
+
|
154 |
+
# # 线程锁
|
155 |
+
# lock = threading.Lock()
|
156 |
+
|
157 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
158 |
+
|
159 |
+
def predict(input, input_size=512, high_visual_quality=False):
|
160 |
+
input_size = int(input_size) # 确保 imgsz 是整数
|
161 |
+
results = model(input, device=device, retina_masks=True, iou=0.7, conf=0.25, imgsz=input_size)
|
162 |
+
fig = fast_process(annotations=results[0].masks.data,
|
163 |
+
image=input, high_quality=high_visual_quality, device=device)
|
164 |
+
return fig
|
165 |
+
|
166 |
+
|
167 |
+
|
168 |
+
# input_size=1024
|
169 |
+
# high_quality_visual=True
|
170 |
+
# inp = 'assets/sa_192.jpg'
|
171 |
+
# input = Image.open(inp)
|
172 |
+
# device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
173 |
+
# input_size = int(input_size) # 确保 imgsz 是整数
|
174 |
+
# results = model(input, device=device, retina_masks=True, iou=0.7, conf=0.25, imgsz=input_size)
|
175 |
+
# pil_image = fast_process(annotations=results[0].masks.data,
|
176 |
+
# image=input, high_quality=high_quality_visual, device=device)
|
177 |
+
|
178 |
+
app_interface = gr.Interface(fn=predict,
|
179 |
+
inputs=[gr.components.Image(type='pil'),
|
180 |
+
gr.components.Slider(minimum=512, maximum=1024, value=1024, step=64, label='input_size'),
|
181 |
+
gr.components.Checkbox(value=False, label='high_visual_quality')],
|
182 |
+
# outputs=['plot'],
|
183 |
+
outputs=gr.components.Image(type='pil'),
|
184 |
+
examples=[["assets/sa_8776.jpg", 1024, True]],
|
185 |
+
# # ["assets/sa_1309.jpg", 1024]],
|
186 |
+
# examples=[["assets/sa_192.jpg"], ["assets/sa_414.jpg"],
|
187 |
+
# ["assets/sa_561.jpg"], ["assets/sa_862.jpg"],
|
188 |
+
# ["assets/sa_1309.jpg"], ["assets/sa_8776.jpg"],
|
189 |
+
# ["assets/sa_10039.jpg"], ["assets/sa_11025.jpg"],],
|
190 |
+
cache_examples=True,
|
191 |
+
title="Fast Segment Anything (Everything mode)"
|
192 |
+
)
|
193 |
+
|
194 |
+
|
195 |
+
app_interface.queue(concurrency_count=1, max_size=20)
|
196 |
+
app_interface.launch()
|
requirements.txt
CHANGED
@@ -14,5 +14,5 @@ opencv-python
|
|
14 |
# seaborn>=0.11.0
|
15 |
|
16 |
# Ultralytics-----------------------------------
|
17 |
-
ultralytics==8.0.
|
18 |
|
|
|
14 |
# seaborn>=0.11.0
|
15 |
|
16 |
# Ultralytics-----------------------------------
|
17 |
+
ultralytics==8.0.121
|
18 |
|