Upload 6 files

app.py

@@ -0,0 +1,62 @@
+import gradio as gr
+import torch
+from PIL import Image
+import torchvision.transforms as transforms
+
+
+def img2label(left, right):
+    left_img = Image.open(left).convert('RGB')
+    right_img = Image.open(right).convert('RGB')
+    # 将右眼底镜像反转
+    r2l = transforms.RandomHorizontalFlip(p=1)
+    right_img = r2l(right_img)
+
+    # 调整图片
+    left_img = my_transforms(left_img).to(device)
+    right_img = my_transforms(right_img).to(device)
+
+    # 读取模型
+    model = torch.load('densenet_FD_e4_l5e-4_b32.pkl', map_location='cpu').to(device)
+
+    with torch.no_grad():
+        output = model(left=left_img.unsqueeze(0), right=right_img.unsqueeze(0))
+
+    output = torch.sigmoid(output.squeeze(0))
+    pred = output.cpu().numpy().tolist()
+
+    return {LABELS[i]: pred[i] for i in range(len(pred))}
+
+
+if __name__ == '__main__':
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    # 标题
+    title = "基于眼底图像的智能健康诊断分析系统"
+    # 标题下的描述，支持md格式
+    description = "上传并输入左右眼底图像后，点击 submit 按钮，可根据双目眼底图像智能分析出可能有的疾病！" \
+                  "包含的疾病种类有：糖尿病、青光眼、白内障、年龄性黄斑变性、高血压、病理性近视、其他疾病以及正常共计8类"
+
+    # transforms设置
+    norm_mean = [0.485, 0.456, 0.406]
+    norm_std = [0.229, 0.224, 0.225]
+    my_transforms = transforms.Compose([
+        transforms.Resize((224, 224)),
+        transforms.ToTensor(),
+        transforms.Normalize(norm_mean, norm_std)
+    ])
+
+    LABELS = {0: '正常',
+              1: '糖尿病',
+              2: '青光眼',
+              3: '白内障',
+              4: '年龄性黄斑变性',
+              5: '高血压',
+              6: '病理性近视',
+              7: '其他疾病'}
+
+    left_img_dir = 'left.jpg'
+    right_img_dir = 'right.jpg'
+    r = img2label(left_img_dir, right_img_dir)
+    demo = gr.Interface(fn=img2label, inputs=[gr.inputs.Image(), gr.inputs.Image()], outputs='label',
+                        title=title, description=description)
+    demo.launch(share=True)

config/finetune_config.py

@@ -0,0 +1,33 @@
+import argparse
+
+
+def set_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--save_name', default='densenet_ce_e4_b32_lr1e-4.pkl', type=str,
+                        help='保存模型的名字，默认路径在 ./model_parameters/f')
+    parser.add_argument('--model_selection', default='MX', type=str,
+                        help='模型选择，M 表示模改过的，没有则表示原来的')
+    parser.add_argument('--pt', default='FD', type=str,
+                        help='有 FD 和 IN 两种，FD 表示预训练是眼底图像，IN 表示预训练是 imagenet')
+    parser.add_argument('--finetune_path', default='model_parameters/p/resnext_ce_e4_b32_lr1e-4.pkl', type=str,
+                        help='所选用预训练模型的路径')
+    parser.add_argument('--feature_module', default='cat', type=str,
+                        help='特征融合的方式，有 cat、mul、sum 三种方式，只在原始模型有效')
+    parser.add_argument('--opt', default='adamw', type=str, help='优化器')
+    parser.add_argument('--warmup_select', default='linear', type=str,
+                        help='')
+    parser.add_argument('--MAX_EPOCH', default=4, type=int,
+                        help='')
+    parser.add_argument('--BATCH_SIZE', default=32, type=int,
+                        help='')
+    parser.add_argument('--start_epoch', default=0, type=int,
+                        help='')
+    parser.add_argument('--LR', default=5e-4, type=float,
+                        help='')
+    parser.add_argument('--WD', default=1e-2, type=float,
+                        help='')
+    parser.add_argument('--adam_epsilon', default=1e-8, type=float,
+                        help='')
+    parser.add_argument('--warmup_proportion', default=0.1, type=float,
+                        help='')
+    return parser.parse_args()

densenet_FD_e4_l5e-4_b32.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:746019148ee61f72424b1f866b28ec220815413f3c5ca52436c6de24d9054b9a
+size 732694486

models/modified_dual_densenet.py

@@ -0,0 +1,78 @@
+import torch
+from torch import nn
+from torchvision.models import densenet169
+from config.finetune_config import set_args
+
+args = set_args()
+
+
+class Classifier(nn.Module):
+    def __init__(self, num_classes):
+        super(Classifier, self).__init__()
+        self.GDConv1 = nn.Conv2d(1664 * 2, 1024, kernel_size=4, padding=0, dilation=2)
+        self.GDConv2 = nn.Conv2d(1664 * 2, 1024, kernel_size=5, padding=1, dilation=2)
+        self.GDConv3 = nn.Conv2d(1664 * 2, 1024, kernel_size=3, padding=0, dilation=3)
+        self.LN1 = nn.LayerNorm([1024, 1, 1])
+        self.LN2 = nn.LayerNorm([1024, 1, 1])
+        self.LN3 = nn.LayerNorm([1024, 1, 1])
+        self.gelu = nn.GELU()
+        self.fc_dropout = nn.Dropout(0.2)
+        self.fc = nn.Linear(1024 * 3, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        x1 = self.GDConv1(x)
+        x1 = self.LN1(x1)
+        x1 = x1.view(x1.size(0), -1)
+
+        x2 = self.GDConv2(x)
+        x2 = self.LN2(x2)
+        x2 = x2.view(x2.size(0), -1)
+
+        x3 = self.GDConv3(x)
+        x3 = self.LN3(x3)
+        x3 = x3.view(x3.size(0), -1)
+
+        X = torch.cat((x1, x2, x3), 1)
+        X = self.gelu(X)
+        output = self.fc(self.fc_dropout(X))
+
+        return output
+
+
+class M_DenseNet(nn.Module):
+    def __init__(self, pretrain='IN', num_classes=8):
+        super(M_DenseNet, self).__init__()
+        # feature layer
+        if pretrain == 'IN':
+            model = densenet169(pretrained=True)  # 此处的model参数是已经加载了预训练参数的模型
+            self.feature = nn.Sequential(*list(model.children())[:-1])
+        else:
+            model = torch.load(args.finetune_path)
+            self.feature = nn.Sequential(*list(model.children())[:-2])
+        self.classifier = Classifier(num_classes)
+
+    def forward(self, left, right):
+        left = self.feature(left)
+        right = self.feature(right)
+        x = torch.cat((left, right), 1)
+
+        X = self.classifier(x)
+
+        return X
+
+
+if __name__ == '__main__':
+    model = M_DenseNet()
+    input1 = torch.normal(0, 1, size=(4, 3, 224, 224))
+    input2 = torch.normal(0, 1, size=(4, 3, 224, 224))
+    output = model(input1, input2)
+    print(output)