Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,28 +1,29 @@
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
-
from torchvision import transforms
|
5 |
import pickle
|
6 |
from resnest.torch import resnest50
|
|
|
|
|
|
|
7 |
|
|
|
8 |
with open('class_names.pkl', 'rb') as f:
|
9 |
class_names = pickle.load(f)
|
10 |
|
11 |
-
#
|
12 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
13 |
-
|
14 |
-
model = resnest50(pretrained=None)
|
15 |
model.fc = nn.Sequential(
|
16 |
nn.Dropout(0.2),
|
17 |
nn.Linear(model.fc.in_features, len(class_names))
|
18 |
)
|
19 |
-
|
20 |
-
# 加载模型权重
|
21 |
-
model.load_state_dict(torch.load('best_model.pth', map_location=device, weights_only=True))
|
22 |
model = model.to(device)
|
23 |
model.eval()
|
24 |
|
25 |
-
#
|
26 |
preprocess = transforms.Compose([
|
27 |
transforms.Resize((100, 100)),
|
28 |
transforms.ToTensor(),
|
@@ -30,23 +31,42 @@ preprocess = transforms.Compose([
|
|
30 |
])
|
31 |
|
32 |
|
33 |
-
def
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
-
# 应用预处理
|
37 |
-
input_tensor = preprocess(img)
|
38 |
|
39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
input_batch = input_tensor.unsqueeze(0).to(device)
|
41 |
|
42 |
# 预测
|
43 |
with torch.no_grad():
|
44 |
output = model(input_batch)
|
45 |
|
46 |
-
# 计算概率
|
47 |
probabilities = torch.nn.functional.softmax(output[0], dim=0)
|
48 |
-
|
49 |
-
# 获取前3个预测结果
|
50 |
top3_probs, top3_indices = torch.topk(probabilities, 3)
|
51 |
|
52 |
results = {
|
@@ -54,38 +74,66 @@ def predict_image(img):
|
|
54 |
for p, i in zip(top3_probs, top3_indices)
|
55 |
}
|
56 |
|
57 |
-
#
|
58 |
best_class = class_names[top3_indices[0]]
|
59 |
best_conf = top3_probs[0].item() * 100
|
60 |
-
|
61 |
-
return best_class, best_conf, results
|
62 |
|
63 |
-
|
|
|
|
|
64 |
def create_interface():
|
65 |
examples = [
|
66 |
"r0_0_100.jpg",
|
67 |
-
"r0_18_100.jpg"
|
|
|
|
|
|
|
68 |
]
|
69 |
|
70 |
with gr.Blocks(title="Fruit Classification", theme=gr.themes.Soft()) as demo:
|
71 |
-
gr.Markdown("
|
|
|
|
|
72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
with gr.Row():
|
74 |
with gr.Column():
|
75 |
-
|
76 |
-
gr.Examples(examples=examples, inputs=
|
77 |
-
submit_btn = gr.Button("
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
with gr.Column():
|
80 |
-
|
81 |
-
|
82 |
-
|
|
|
83 |
|
84 |
-
# ‘分类’按钮点击事件
|
85 |
submit_btn.click(
|
86 |
fn=predict_image,
|
87 |
-
inputs=
|
88 |
-
outputs=[best_pred, confidence, full_results]
|
89 |
)
|
90 |
|
91 |
return demo
|
@@ -93,4 +141,4 @@ def create_interface():
|
|
93 |
|
94 |
if __name__ == "__main__":
|
95 |
interface = create_interface()
|
96 |
-
interface.launch(share=
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
+
from torchvision import transforms
|
5 |
import pickle
|
6 |
from resnest.torch import resnest50
|
7 |
+
from rembg import remove
|
8 |
+
from PIL import Image
|
9 |
+
import io
|
10 |
|
11 |
+
# 加载类别名称
|
12 |
with open('class_names.pkl', 'rb') as f:
|
13 |
class_names = pickle.load(f)
|
14 |
|
15 |
+
# 初始化模型
|
16 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
17 |
+
model = resnest50(pretrained=False)
|
|
|
18 |
model.fc = nn.Sequential(
|
19 |
nn.Dropout(0.2),
|
20 |
nn.Linear(model.fc.in_features, len(class_names))
|
21 |
)
|
22 |
+
model.load_state_dict(torch.load('best_model.pth', map_location=device))
|
|
|
|
|
23 |
model = model.to(device)
|
24 |
model.eval()
|
25 |
|
26 |
+
# 预处理流程
|
27 |
preprocess = transforms.Compose([
|
28 |
transforms.Resize((100, 100)),
|
29 |
transforms.ToTensor(),
|
|
|
31 |
])
|
32 |
|
33 |
|
34 |
+
def remove_background(img):
|
35 |
+
"""使用rembg去除背景并添加白色背景"""
|
36 |
+
# 转换图像为字节流
|
37 |
+
img_byte_arr = io.BytesIO()
|
38 |
+
img.save(img_byte_arr, format='PNG')
|
39 |
+
img_bytes = img_byte_arr.getvalue()
|
40 |
+
|
41 |
+
# 去除背景
|
42 |
+
removed_bg_bytes = remove(img_bytes)
|
43 |
+
|
44 |
+
# 转换为PIL图像并处理透明度
|
45 |
+
removed_bg_img = Image.open(io.BytesIO(removed_bg_bytes)).convert('RGBA')
|
46 |
+
|
47 |
+
# 创建白色背景
|
48 |
+
white_bg = Image.new('RGBA', removed_bg_img.size, (255, 255, 255, 255))
|
49 |
+
combined = Image.alpha_composite(white_bg, removed_bg_img)
|
50 |
+
return combined.convert('RGB')
|
51 |
|
|
|
|
|
52 |
|
53 |
+
def predict_image(img, remove_bg=False):
|
54 |
+
"""分类预测主函数"""
|
55 |
+
# 根据选择处理图像
|
56 |
+
if remove_bg:
|
57 |
+
processed_img = remove_background(img)
|
58 |
+
else:
|
59 |
+
processed_img = img.convert('RGB') # 确保为RGB格式
|
60 |
+
|
61 |
+
# 预处理
|
62 |
+
input_tensor = preprocess(processed_img)
|
63 |
input_batch = input_tensor.unsqueeze(0).to(device)
|
64 |
|
65 |
# 预测
|
66 |
with torch.no_grad():
|
67 |
output = model(input_batch)
|
68 |
|
|
|
69 |
probabilities = torch.nn.functional.softmax(output[0], dim=0)
|
|
|
|
|
70 |
top3_probs, top3_indices = torch.topk(probabilities, 3)
|
71 |
|
72 |
results = {
|
|
|
74 |
for p, i in zip(top3_probs, top3_indices)
|
75 |
}
|
76 |
|
77 |
+
# 记录结果
|
78 |
best_class = class_names[top3_indices[0]]
|
79 |
best_conf = top3_probs[0].item() * 100
|
|
|
|
|
80 |
|
81 |
+
return processed_img, best_class, f"{best_conf:.2f}%", results
|
82 |
+
|
83 |
+
|
84 |
def create_interface():
|
85 |
examples = [
|
86 |
"r0_0_100.jpg",
|
87 |
+
"r0_18_100.jpg",
|
88 |
+
"9_100.jpg",
|
89 |
+
"127_100.jpg",
|
90 |
+
"r0_1_100.jpg",
|
91 |
]
|
92 |
|
93 |
with gr.Blocks(title="Fruit Classification", theme=gr.themes.Soft()) as demo:
|
94 |
+
gr.Markdown("""
|
95 |
+
# 🍎 智能水果识别系统
|
96 |
+
""")
|
97 |
|
98 |
+
# 新增:模式选择卡片(视觉强化)
|
99 |
+
with gr.Row():
|
100 |
+
with gr.Column(scale=3):
|
101 |
+
with gr.Group():
|
102 |
+
gr.Markdown("### ⚙️ 处理模式选择")
|
103 |
+
with gr.Row():
|
104 |
+
bg_removal = gr.Checkbox(
|
105 |
+
label="背景去除",
|
106 |
+
value=False,
|
107 |
+
interactive=True
|
108 |
+
)
|
109 |
+
|
110 |
+
# 主操作区域
|
111 |
with gr.Row():
|
112 |
with gr.Column():
|
113 |
+
original_image = gr.Image(label="📤 上传图片", type="pil")
|
114 |
+
gr.Examples(examples=examples, inputs=original_image)
|
115 |
+
submit_btn = gr.Button("🚀 开始识别", variant="primary")
|
116 |
+
|
117 |
+
# 添加模式说明提示
|
118 |
+
gr.Markdown("""
|
119 |
+
<div style="background: #f3f4f6; padding: 15px; border-radius: 8px; margin-top: 10px">
|
120 |
+
<b>💡 使用建议:</b><br>
|
121 |
+
• 上传图片:选择一张图片,点击'开始识别'按钮<br>
|
122 |
+
• 勾选背景去除:适合杂乱背景的图片(识别更准确)<br>
|
123 |
+
• 不勾选:适合纯色背景的图片(速度更快)
|
124 |
+
</div>
|
125 |
+
""")
|
126 |
|
127 |
with gr.Column():
|
128 |
+
processed_image = gr.Image(label="🖼️ 处理后图片", interactive=False)
|
129 |
+
best_pred = gr.Textbox(label="🔍 识别结果")
|
130 |
+
confidence = gr.Textbox(label="📊 置信度")
|
131 |
+
full_results = gr.Label(label="🏆 Top 3 可能结果", num_top_classes=3)
|
132 |
|
|
|
133 |
submit_btn.click(
|
134 |
fn=predict_image,
|
135 |
+
inputs=[original_image, bg_removal],
|
136 |
+
outputs=[processed_image, best_pred, confidence, full_results]
|
137 |
)
|
138 |
|
139 |
return demo
|
|
|
141 |
|
142 |
if __name__ == "__main__":
|
143 |
interface = create_interface()
|
144 |
+
interface.launch(share=True)
|