1.1.1.19124

Author: LatentLinker

Generate_image/augmentor.py

@@ -0,0 +1,16 @@
+import Augmentor  # 图像增强 图像预处理  https://github.com/mdbloice/Augmentor     https://augmentor.readthedocs.io/en/master/
+import torchvision
+
+p = Augmentor.Pipeline("/home/tian/Desktop/spiders/design/design/spiders/image_test/dog")
+
+p.rotate90(probability=0.5)
+p.rotate270(probability=0.5)
+p.flip_left_right(probability=0.8)
+p.flip_top_bottom(probability=0.3)
+p.crop_random(probability=1, percentage_area=0.5)
+p.resize(probability=1.0, width=224, height=224)
+
+transforms = torchvision.transforms.Compose([
+    p.torch_transform(),
+    torchvision.transforms.ToTensor(),
+])

Image_recognition/config.py

@@ -7,32 +7,32 @@
 class DefaultConfig(object):
     env = 'opalus_recognltion'  # visdom 环境
     vis_port = 8097  # visdom 端口
+    image_size = 224
     model = 'ResNet152'  # 使用的模型，名字必须与models/__init__.py中的名字一致
 
     data_root = "/home/tian/Desktop/spiders/design/design/spiders/image"  # 数据集存放路径
-    load_model_path = None  # 加载训练的模型的路径，为None代表不加载
+    # load_model_path = None  # 加载训练的模型的路径，为None代表不加载
     load_model_path = 'checkpoint/ResNet152_0124_11-57-28.pth.tar'
 
     batch_size = 16  # 每批训练数据的个数,显存不足,适当减少
     use_gpu = True  # user GPU or not
     num_workers = 4  # how many workers for loading data
     print_freq = 2  # print info every N batch
-    vis = True  # 是否使用visdom可视化
+    vis = False  # 是否使用visdom可视化
 
     cate_classes = get_classes(data_root)['class2num']  # 分类列表
     num_classes = len(cate_classes)  # 分类个数
-    # pretrained = False  # 不加载预训练
-    pretrained = True  # 加载预训练模型
-    result_file = 'result.csv'
+    pretrained = False  # 不加载预训练
+    # pretrained = True  # 加载预训练模型
 
     max_epoch = 10  # 学习次数
     lr = 0.001  # initial learning rate
     lr_decay = 0.5  # when val_loss increase, lr = lr*lr_decay
     weight_decay = 0e-5  # 损失函数
     # url = 'https://ss3.bdstatic.com/70cFv8Sh_Q1YnxGkpoWK1HF6hhy/it/u=614134999,3540271868&fm=27&gp=0.jpg'  # 识别图片地址
     # url = 'https://ss1.bdstatic.com/70cFuXSh_Q1YnxGkpoWK1HF6hhy/it/u=688429408,3192272581&fm=27&gp=0.jpg'
-    # url = 'https://ss1.bdstatic.com/70cFvXSh_Q1YnxGkpoWK1HF6hhy/it/u=1515206672,3808938099&fm=27&gp=0.jpg'
-    url = 'https://ss0.bdstatic.com/70cFuHSh_Q1YnxGkpoWK1HF6hhy/it/u=3211343338,3677737612&fm=27&gp=0.jpg'
+    url = 'https://ss1.bdstatic.com/70cFvXSh_Q1YnxGkpoWK1HF6hhy/it/u=1515206672,3808938099&fm=27&gp=0.jpg'
+    # url = 'https://ss0.bdstatic.com/70cFuHSh_Q1YnxGkpoWK1HF6hhy/it/u=3211343338,3677737612&fm=27&gp=0.jpg'
     # url = 'https://ss0.bdstatic.com/70cFuHSh_Q1YnxGkpoWK1HF6hhy/it/u=1173573129,2720567755&fm=27&gp=0.jpg'
 
     def _parse(self, kwargs):

Image_recognition/data/dataset.py

@@ -3,6 +3,7 @@
 from torch.utils import data
 from torchvision import transforms as T
 from utils.imagefolder_splitter import ImageFolderSplitter
+from config import opt
 import cv2
 
 
@@ -38,16 +39,16 @@ def __init__(self, root, transforms=None, train=True, test=False):
             if self.test or not self.train:
                 # 训练 测试
                 self.transforms = T.Compose([
-                    T.Resize(224),  # #缩放图片（Image）,保持长宽比不变，最短边为224像素
-                    T.CenterCrop(224),  # 在图片的中间区域进行裁剪
+                    T.Resize(opt.image_size),  # #缩放图片（Image）,保持长宽比不变，最短边为224像素
+                    T.CenterCrop(opt.image_size),  # 在图片的中间区域进行裁剪
                     T.ToTensor(),  # 转tensor
                     normalize  # 归一化
                 ])
             else:
                 # 验证
                 self.transforms = T.Compose([
                     T.Resize(256),  # #缩放图片（Image）,保持长宽比不变，最短边为224像素
-                    T.RandomResizedCrop(224),  # 在一个随机的位置进行裁剪
+                    T.RandomResizedCrop(opt.image_size),  # 在一个随机的位置进行裁剪
                     T.RandomHorizontalFlip(),
                     T.ToTensor(),
                     normalize

Image_recognition/main.py

@@ -55,7 +55,7 @@ def recognition(**kwargs):
             model.load_state_dict(checkpoint["state_dict"])  # 预加载模型
         model.to(opt.device)
         model.eval()
-        image = image.view(1, 3, 224, 224).to(opt.device)  # 转换image
+        image = image.view(1, 3, opt.image_size, opt.image_size).to(opt.device)  # 转换image
         outputs = model(image)
         result = {}
         for i in range(opt.num_classes):  # 计算各分类比重
@@ -119,17 +119,15 @@ def train(**kwargs):
             target = label.to(opt.device)
 
             score = model(input)
-            loss = criterion(score, target)  # 计算损失
-
+            # loss = criterion(score, target)  # 计算损失
+            loss = criterion(score[0], target)  # 计算损失   Inception3网络
             optimizer.zero_grad()  # 参数梯度设成0
             loss.backward()  # 反向传播
             optimizer.step()  # 更新参数
             # meters update and visualize
-            precision1_train, precision2_train = accuracy(score, target, topk=(1, 2))
+            # precision1_train, precision2_train = accuracy(score, target, topk=(1, 2))
+            precision1_train, precision2_train = accuracy(score[0], target, topk=(1, 2))  # Inception3网络
             train_losses.update(loss.item(), input.size(0))
-            a = precision1_train[0]
-            b = input.size(0)
-            c = precision1_train[0].item()
             train_top1.update(precision1_train[0].item(), input.size(0))
             train_progressor.current_loss = train_losses.avg
             train_progressor.current_top1 = train_top1.avg

Image_recognition/models/alexnet.py

@@ -4,7 +4,7 @@
 import torch
 
 
-def alexnet(pretrained=False, **kwargs):
+def alexnet(pretrained=False, **kwargs): # 224*224
     if pretrained:
         model = AlexNet(**kwargs)
         model.load_state_dict(torch.load('./checkpoint/inception_v3_google-1a9a5a14.pth'))
@@ -21,6 +21,12 @@ def __init__(self):
     def forward(self, x):
         return self.model(x)
 
+    def get_optimizer(self, lr, weight_decay):
+        if not opt.pretrained:
+            return super(AlexNet1, self).get_optimizer(lr, weight_decay)
+        else:
+            return torch.optim.Adam(self.model.fc.parameters(), lr=lr, weight_decay=weight_decay)
+
 
 if __name__ == '__main__':
     a = AlexNet1()

Image_recognition/models/denseNet201.py

@@ -33,6 +33,12 @@ def __init__(self):
     def forward(self, x):
         return self.model(x)
 
+    def get_optimizer(self, lr, weight_decay):
+        if not opt.pretrained:
+            return super(DenseNet201, self).get_optimizer(lr, weight_decay)
+        else:
+            return torch.optim.Adam(self.model.fc.parameters(), lr=lr, weight_decay=weight_decay)
+
 
 if __name__ == '__main__':
     a = DenseNet201()

Image_recognition/models/densenet161.py

@@ -33,6 +33,12 @@ def __init__(self):
     def forward(self, x):
         return self.model(x)
 
+    def get_optimizer(self, lr, weight_decay):
+        if not opt.pretrained:
+            return super(DenseNet161, self).get_optimizer(lr, weight_decay)
+        else:
+            return torch.optim.Adam(self.model.fc.parameters(), lr=lr, weight_decay=weight_decay)
+
 
 if __name__ == '__main__':
     a = DenseNet161()

Image_recognition/models/inceptionv3.py

@@ -4,13 +4,22 @@
 import torch
 
 
-def inception_v3(pretrained=False, **kwargs):
+def inception_v3(pretrained=False, **kwargs):  # 299*299
     if pretrained:
         if 'transform_input' not in kwargs:
             kwargs['transform_input'] = True
         model = Inception3(**kwargs)
-        model.load_state_dict(torch.load('./checkpoint/inception_v3_google-1a9a5a14.pth'))
-
+        pretrained_state_dict = torch.load(
+            './Authority/inception_v3_google-1a9a5a14.pth')  # load_url函数根据model_urls字典下载或导入相应的预训练模型
+        now_state_dict = model.state_dict()  # 返回model模块的字典
+        pretrained_state_dict.pop('AuxLogits.fc.weight')
+        pretrained_state_dict.pop('AuxLogits.fc.bias')
+        pretrained_state_dict.pop('fc.weight')
+        pretrained_state_dict.pop('fc.bias')
+        now_state_dict.update(pretrained_state_dict)
+        model.load_state_dict(
+            now_state_dict)  # 最后通过调用model的load_state_dict方法用预训练的模型参数来初始化你构建的网络结构，这个方法就是PyTorch中通用的用一个模型的参数初始化另一个模型的层的操作。load_state_dict方法还有一个重要的参数是strict，该参数默认是True，表示预训练模型的层和你的网络结构层严格对应相等（比如层名和维度）
+        return model
     return Inception3(**kwargs)
 
 
@@ -23,6 +32,15 @@ def __init__(self):
     def forward(self, x):
         return self.model(x)
 
+    def get_optimizer(self, lr, weight_decay):
+        if not opt.pretrained:
+            return super(InceptionV3, self).get_optimizer(lr, weight_decay)
+        else:
+            return torch.optim.Adam([
+                {'params': self.model.AuxLogits.fc.parameters()},
+                {'params': self.model.fc.parameters()}
+            ], lr=lr, weight_decay=weight_decay)
+
 
 if __name__ == '__main__':
     a = InceptionV3()

Image_recognition/models/resnet152.py

@@ -48,10 +48,8 @@ def forward(self, x):
 def resnet152(pretrained=False, **kwargs):
     if pretrained:
         model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
-        # model.load_state_dict(torch.load('./checkpoint/resnet152-b121ed2d.pth'))
-        # 网络结构不对等
         pretrained_state_dict = torch.load(
-            './checkpoint/resnet152-b121ed2d.pth')  # load_url函数根据model_urls字典下载或导入相应的预训练模型
+            './Authority/resnet152-b121ed2d.pth')  # load_url函数根据model_urls字典下载或导入相应的预训练模型
         now_state_dict = model.state_dict()  # 返回model模块的字典
         pretrained_state_dict.pop('fc.weight')
         pretrained_state_dict.pop('fc.bias')

Image_recognition/utils/image_loader.py

@@ -2,13 +2,14 @@
 from torchvision import transforms as T
 import requests
 from PIL import Image
+from config import opt
 
 normalize = T.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
 
 transforms = T.Compose([
-    T.Resize(224),
-    T.CenterCrop(224),
+    T.Resize(opt.image_size),
+    T.CenterCrop(opt.image_size),
     T.ToTensor(),
     normalize
 ])