Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,332 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch.nn as nn
|
3 |
+
import math
|
4 |
+
import torch.utils.model_zoo as model_zoo
|
5 |
+
import torch
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
__all__ = ['Res2Net', 'res2net50_v1b', 'res2net101_v1b']
|
9 |
+
|
10 |
+
model_urls = {
|
11 |
+
'res2net50_v1b_26w_4s': 'https://shanghuagao.oss-cn-beijing.aliyuncs.com/res2net/res2net50_v1b_26w_4s-3cf99910.pth',
|
12 |
+
'res2net101_v1b_26w_4s': 'https://shanghuagao.oss-cn-beijing.aliyuncs.com/res2net/res2net101_v1b_26w_4s-0812c246.pth',
|
13 |
+
}
|
14 |
+
|
15 |
+
|
16 |
+
class Bottle2neck(nn.Module):
|
17 |
+
expansion = 4
|
18 |
+
|
19 |
+
def __init__(self, inplanes, planes, stride=1, downsample=None, baseWidth=26, scale=4, stype='normal'):
|
20 |
+
""" Constructor
|
21 |
+
Args:
|
22 |
+
inplanes: input channel dimensionality
|
23 |
+
planes: output channel dimensionality
|
24 |
+
stride: conv stride. Replaces pooling layer.
|
25 |
+
downsample: None when stride = 1
|
26 |
+
baseWidth: basic width of conv3x3
|
27 |
+
scale: number of scale.
|
28 |
+
type: 'normal': normal set. 'stage': first block of a new stage.
|
29 |
+
"""
|
30 |
+
super(Bottle2neck, self).__init__()
|
31 |
+
|
32 |
+
width = int(math.floor(planes * (baseWidth / 64.0)))
|
33 |
+
self.conv1 = nn.Conv2d(inplanes, width * scale, kernel_size=1, bias=False)
|
34 |
+
self.bn1 = nn.BatchNorm2d(width * scale)
|
35 |
+
|
36 |
+
if scale == 1:
|
37 |
+
self.nums = 1
|
38 |
+
else:
|
39 |
+
self.nums = scale - 1
|
40 |
+
if stype == 'stage':
|
41 |
+
self.pool = nn.AvgPool2d(kernel_size=3, stride=stride, padding=1)
|
42 |
+
convs = []
|
43 |
+
bns = []
|
44 |
+
for i in range(self.nums):
|
45 |
+
convs.append(nn.Conv2d(width, width, kernel_size=3, stride=stride, padding=1, bias=False))
|
46 |
+
bns.append(nn.BatchNorm2d(width))
|
47 |
+
self.convs = nn.ModuleList(convs)
|
48 |
+
self.bns = nn.ModuleList(bns)
|
49 |
+
|
50 |
+
self.conv3 = nn.Conv2d(width * scale, planes * self.expansion, kernel_size=1, bias=False)
|
51 |
+
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
|
52 |
+
|
53 |
+
self.relu = nn.ReLU(inplace=True)
|
54 |
+
self.downsample = downsample
|
55 |
+
self.stype = stype
|
56 |
+
self.scale = scale
|
57 |
+
self.width = width
|
58 |
+
|
59 |
+
def forward(self, x):
|
60 |
+
residual = x
|
61 |
+
|
62 |
+
out = self.conv1(x)
|
63 |
+
out = self.bn1(out)
|
64 |
+
out = self.relu(out)
|
65 |
+
|
66 |
+
spx = torch.split(out, self.width, 1)
|
67 |
+
for i in range(self.nums):
|
68 |
+
if i == 0 or self.stype == 'stage':
|
69 |
+
sp = spx[i]
|
70 |
+
else:
|
71 |
+
sp = sp + spx[i]
|
72 |
+
sp = self.convs[i](sp)
|
73 |
+
sp = self.relu(self.bns[i](sp))
|
74 |
+
if i == 0:
|
75 |
+
out = sp
|
76 |
+
else:
|
77 |
+
out = torch.cat((out, sp), 1)
|
78 |
+
if self.scale != 1 and self.stype == 'normal':
|
79 |
+
out = torch.cat((out, spx[self.nums]), 1)
|
80 |
+
elif self.scale != 1 and self.stype == 'stage':
|
81 |
+
out = torch.cat((out, self.pool(spx[self.nums])), 1)
|
82 |
+
|
83 |
+
out = self.conv3(out)
|
84 |
+
out = self.bn3(out)
|
85 |
+
|
86 |
+
if self.downsample is not None:
|
87 |
+
residual = self.downsample(x)
|
88 |
+
|
89 |
+
out += residual
|
90 |
+
out = self.relu(out)
|
91 |
+
|
92 |
+
return out
|
93 |
+
|
94 |
+
|
95 |
+
class Res2Net(nn.Module):
|
96 |
+
|
97 |
+
def __init__(self, block, layers, baseWidth=26, scale=4, num_classes=1000):
|
98 |
+
self.inplanes = 64
|
99 |
+
super(Res2Net, self).__init__()
|
100 |
+
self.baseWidth = baseWidth
|
101 |
+
self.scale = scale
|
102 |
+
self.conv1 = nn.Sequential(
|
103 |
+
nn.Conv2d(3, 32, 3, 2, 1, bias=False),
|
104 |
+
nn.BatchNorm2d(32),
|
105 |
+
nn.ReLU(inplace=True),
|
106 |
+
nn.Conv2d(32, 32, 3, 1, 1, bias=False),
|
107 |
+
nn.BatchNorm2d(32),
|
108 |
+
nn.ReLU(inplace=True),
|
109 |
+
nn.Conv2d(32, 64, 3, 1, 1, bias=False)
|
110 |
+
)
|
111 |
+
self.bn1 = nn.BatchNorm2d(64)
|
112 |
+
self.relu = nn.ReLU()
|
113 |
+
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
|
114 |
+
self.layer1 = self._make_layer(block, 64, layers[0])
|
115 |
+
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
|
116 |
+
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
|
117 |
+
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
|
118 |
+
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
119 |
+
self.fc = nn.Linear(512 * block.expansion, num_classes)
|
120 |
+
|
121 |
+
for m in self.modules():
|
122 |
+
if isinstance(m, nn.Conv2d):
|
123 |
+
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
|
124 |
+
elif isinstance(m, nn.BatchNorm2d):
|
125 |
+
nn.init.constant_(m.weight, 1)
|
126 |
+
nn.init.constant_(m.bias, 0)
|
127 |
+
|
128 |
+
def _make_layer(self, block, planes, blocks, stride=1):
|
129 |
+
downsample = None
|
130 |
+
if stride != 1 or self.inplanes != planes * block.expansion:
|
131 |
+
downsample = nn.Sequential(
|
132 |
+
nn.AvgPool2d(kernel_size=stride, stride=stride,
|
133 |
+
ceil_mode=True, count_include_pad=False),
|
134 |
+
nn.Conv2d(self.inplanes, planes * block.expansion,
|
135 |
+
kernel_size=1, stride=1, bias=False),
|
136 |
+
nn.BatchNorm2d(planes * block.expansion),
|
137 |
+
)
|
138 |
+
|
139 |
+
layers = []
|
140 |
+
layers.append(block(self.inplanes, planes, stride, downsample=downsample,
|
141 |
+
stype='stage', baseWidth=self.baseWidth, scale=self.scale))
|
142 |
+
self.inplanes = planes * block.expansion
|
143 |
+
for i in range(1, blocks):
|
144 |
+
layers.append(block(self.inplanes, planes, baseWidth=self.baseWidth, scale=self.scale))
|
145 |
+
|
146 |
+
return nn.Sequential(*layers)
|
147 |
+
|
148 |
+
def forward(self, x):
|
149 |
+
x = self.conv1(x)
|
150 |
+
x = self.bn1(x)
|
151 |
+
x = self.relu(x)
|
152 |
+
x = self.maxpool(x)
|
153 |
+
|
154 |
+
x = self.layer1(x)
|
155 |
+
x = self.layer2(x)
|
156 |
+
x = self.layer3(x)
|
157 |
+
x = self.layer4(x)
|
158 |
+
|
159 |
+
x = self.avgpool(x)
|
160 |
+
x = x.view(x.size(0), -1)
|
161 |
+
x = self.fc(x)
|
162 |
+
|
163 |
+
return x
|
164 |
+
|
165 |
+
|
166 |
+
def res2net50_v1b(pretrained=False, **kwargs):
|
167 |
+
"""Constructs a Res2Net-50_v1b model.
|
168 |
+
Res2Net-50 refers to the Res2Net-50_v1b_26w_4s.
|
169 |
+
Args:
|
170 |
+
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
171 |
+
"""
|
172 |
+
model = Res2Net(Bottle2neck, [3, 4, 6, 3], baseWidth=26, scale=4, **kwargs)
|
173 |
+
if pretrained:
|
174 |
+
model.load_state_dict(model_zoo.load_url(model_urls['res2net50_v1b_26w_4s']))
|
175 |
+
return model
|
176 |
+
|
177 |
+
|
178 |
+
def res2net101_v1b(pretrained=False, **kwargs):
|
179 |
+
"""Constructs a Res2Net-50_v1b_26w_4s model.
|
180 |
+
Args:
|
181 |
+
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
182 |
+
"""
|
183 |
+
model = Res2Net(Bottle2neck, [3, 4, 23, 3], baseWidth=26, scale=4, **kwargs)
|
184 |
+
if pretrained:
|
185 |
+
model.load_state_dict(model_zoo.load_url(model_urls['res2net101_v1b_26w_4s']))
|
186 |
+
return model
|
187 |
+
|
188 |
+
|
189 |
+
def res2net50_v1b_26w_4s(pretrained=False, **kwargs):
|
190 |
+
"""Constructs a Res2Net-50_v1b_26w_4s model.
|
191 |
+
Args:
|
192 |
+
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
193 |
+
"""
|
194 |
+
model = Res2Net(Bottle2neck, [3, 4, 6, 3], baseWidth=26, scale=4, **kwargs)
|
195 |
+
if pretrained:
|
196 |
+
model.load_state_dict(torch.load(pthfile, map_location='cpu')) # load model
|
197 |
+
return model
|
198 |
+
|
199 |
+
|
200 |
+
def res2net101_v1b_26w_4s(pretrained=False, **kwargs):
|
201 |
+
"""Constructs a Res2Net-50_v1b_26w_4s model.
|
202 |
+
Args:
|
203 |
+
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
204 |
+
"""
|
205 |
+
model = Res2Net(Bottle2neck, [3, 4, 23, 3], baseWidth=26, scale=4, **kwargs)
|
206 |
+
if pretrained:
|
207 |
+
model.load_state_dict(model_zoo.load_url(model_urls['res2net101_v1b_26w_4s']))
|
208 |
+
return model
|
209 |
+
|
210 |
+
|
211 |
+
def res2net152_v1b_26w_4s(pretrained=False, **kwargs):
|
212 |
+
"""Constructs a Res2Net-50_v1b_26w_4s model.
|
213 |
+
Args:
|
214 |
+
pretrained (bool): If True, returns a model pre-trained on ImageNet
|
215 |
+
"""
|
216 |
+
model = Res2Net(Bottle2neck, [3, 8, 36, 3], baseWidth=26, scale=4, **kwargs)
|
217 |
+
if pretrained:
|
218 |
+
model.load_state_dict(model_zoo.load_url(model_urls['res2net152_v1b_26w_4s']))
|
219 |
+
return model
|
220 |
+
|
221 |
+
|
222 |
+
class mutil_model(nn.Module):
|
223 |
+
|
224 |
+
def __init__(self, category_num=10):
|
225 |
+
super(mutil_model, self).__init__()
|
226 |
+
self.model1 = res2net50_v1b_26w_4s(pretrained=False)
|
227 |
+
self.model1.fc = nn.Sequential(
|
228 |
+
nn.Linear(in_features=2048, out_features=category_num, bias=True),
|
229 |
+
)
|
230 |
+
self.model2 = torch.load('./enet_b2_8' + '.pt', map_location=torch.device('cpu'))
|
231 |
+
self.model2.classifier = nn.Sequential(
|
232 |
+
nn.Linear(in_features=1408, out_features=category_num, bias=True),
|
233 |
+
)
|
234 |
+
self.fc = nn.Linear(in_features=category_num * 2, out_features=category_num, bias=True)
|
235 |
+
|
236 |
+
def forward(self, x):
|
237 |
+
x1 = self.model1(x)
|
238 |
+
x2 = self.model2(x)
|
239 |
+
x = torch.cat((x1, x2), 1)
|
240 |
+
x = self.fc(x)
|
241 |
+
return x
|
242 |
+
|
243 |
+
|
244 |
+
pth_path = './image_loader10new_model.pt'
|
245 |
+
category_num = 10
|
246 |
+
|
247 |
+
# "cuda" only when GPUs are available.
|
248 |
+
#device = "cuda" if torch.cuda.is_available() else "cpu"
|
249 |
+
device = "cpu"
|
250 |
+
#Initialize a model, and put it on the device specified.
|
251 |
+
# 导入res2net预训练模型
|
252 |
+
pthfile = './res2net50_v1b.pth'
|
253 |
+
model = res2net50_v1b_26w_4s(pretrained=False)
|
254 |
+
# 修改全连接层,输出维度为预测 分类
|
255 |
+
num_ftrs = model.fc.in_features
|
256 |
+
model.fc = nn.Sequential(
|
257 |
+
nn.Linear(in_features=2048, out_features=1000, bias=True),
|
258 |
+
nn.Dropout(0.5),
|
259 |
+
nn.Linear(1000, out_features=category_num)
|
260 |
+
)
|
261 |
+
model.fc = nn.Sequential(
|
262 |
+
nn.Linear(in_features=2048, out_features=category_num, bias=True),
|
263 |
+
)
|
264 |
+
|
265 |
+
model = model.to(device)
|
266 |
+
model.device = device
|
267 |
+
model.load_state_dict(torch.load(pth_path,torch.device('cpu')))
|
268 |
+
model.eval()
|
269 |
+
|
270 |
+
|
271 |
+
# 增加人脸识别模型
|
272 |
+
#model = mutil_model(category_num=7)
|
273 |
+
#model_state = torch.load('./add_face_emotion_model_7.pt', map_location=torch.device('cpu')).state_dict()
|
274 |
+
#model.load_state_dict(model_state) # 加载模型参数
|
275 |
+
#model.eval()
|
276 |
+
|
277 |
+
labels = ['中国风', '古典', '电子', '摇滚', '乡村', '说唱', '民谣', '二次元', '轻音乐', '儿歌']
|
278 |
+
|
279 |
+
import requests
|
280 |
+
import torch
|
281 |
+
|
282 |
+
import gradio as gr
|
283 |
+
import torchvision.transforms as transforms
|
284 |
+
|
285 |
+
# import cv2
|
286 |
+
# from PIL import Image
|
287 |
+
# PIL
|
288 |
+
# from PIL import Image
|
289 |
+
# inception_net = tf.keras.applications.MobileNetV2() # load the model
|
290 |
+
|
291 |
+
# Download human-readable labels for ImageNet.
|
292 |
+
# response = requests.get("https://git.io/JJkYN")
|
293 |
+
# labels = response.text.split("\n")
|
294 |
+
print(len(labels))
|
295 |
+
|
296 |
+
|
297 |
+
def classify_image(inp):
|
298 |
+
# inp = inp.convert('RGB')
|
299 |
+
# inp = Image.fromarray(inp.astype('uint8'), 'RGB')
|
300 |
+
transform_test = transforms.Compose([
|
301 |
+
# transforms.ToPILImage(),
|
302 |
+
transforms.Resize((256, 256)),
|
303 |
+
transforms.ToTensor(),
|
304 |
+
transforms.Normalize((0.485, 0.456, 0.406),
|
305 |
+
(0.229, 0.224, 0.225)),
|
306 |
+
])
|
307 |
+
inp = transform_test(inp)
|
308 |
+
print(inp)
|
309 |
+
with torch.no_grad():
|
310 |
+
prediction = model(torch.unsqueeze(inp, 0)).flatten()
|
311 |
+
print(prediction)
|
312 |
+
prediction = torch.nn.Softmax(dim=0)(prediction)
|
313 |
+
print(prediction)
|
314 |
+
return {labels[i]: float(prediction[i].item()) for i in range(len(labels))}
|
315 |
+
|
316 |
+
|
317 |
+
# print(classify_image("/jj.jpg"))
|
318 |
+
# image = gr.inputs.Image(shape=(256, 256))
|
319 |
+
# image = gr.inputs.Image()
|
320 |
+
# print(image)
|
321 |
+
# label = gr.outputs.Label(num_top_classes=6)
|
322 |
+
|
323 |
+
gr.Interface(
|
324 |
+
classify_image,
|
325 |
+
# gr.inputs.Image(),
|
326 |
+
gr.inputs.Image(type='pil'),
|
327 |
+
outputs='label'
|
328 |
+
# inputs='image',
|
329 |
+
# outputs='label',
|
330 |
+
# examples=[["images/cheetah1.jpg"], ["images/lion.jpg"]],
|
331 |
+
).launch(share=True)
|
332 |
+
# share=True
|