finish code

README.md

@@ -1,16 +1,40 @@
 # reid_baseline
 reid baseline model for exploring softmax and triplet hard loss's influence.
 
-## Configuration
+## Architecture
+
+### softmax
+
+### triplet
+
+### softmax + triplet
+
 
 ### Classification
-resnet lr: 0.1
-classifier lr: 0.01
+<img src='https://ws3.sinaimg.cn/large/006tKfTcly1fs76ysvu3zj30kg0d60t5.jpg' height='200'>
+
+<img src='https://ws2.sinaimg.cn/large/006tKfTcly1fs76zbtfxcj30js0d674m.jpg' height='200'>
+
 
 ### Triplet Hard
-lr: 2e-4
+<img src='https://ws2.sinaimg.cn/large/006tNc79ly1fs3sxc54xjj30ka0d6dgd.jpg' height='200'>
+
+<img src='https://ws2.sinaimg.cn/large/006tNc79ly1fs3tpat6emj30k00d2t93.jpg' height=200>
+
 
 ### Classification + Triplet Hard
-lr: 2e-4
 
-exponetional decay at 150
+<img src='https://ws2.sinaimg.cn/large/006tKfTcly1fs79doog34j30ja0cudg5.jpg' height='200'>
+
+<img src='https://ws2.sinaimg.cn/large/006tNc79ly1fs3tpat6emj30k00d2t93.jpg' height=200>
+
+
+## Results
+
+| loss | rank1 | map |
+| --- | --| ---|
+| triplet hard | 89.9% | 76.8% | 
+| softmax | 87% | 65% |
+|triplet + softmax | 89.7% | 76.2% |
+
+

config.py

@@ -21,12 +21,12 @@ class DefaultConfig(object):
     width = 192
 
     # optimization options
-    optim = 'SGD'
-    max_epoch = 100
+    optim = 'Adam'
+    max_epoch = 60
     train_batch = 128
     test_batch = 128
     lr = 0.1
-    step_size = 60
+    step_size = 40
     gamma = 0.1
     weight_decay = 5e-4
     momentum = 0.9

main_reid.py

@@ -92,7 +92,8 @@ def train_classification(**kwargs):
 
     # get optimizer
     optimizer = torch.optim.SGD(
-        model.optim_policy(), lr=args.lr, weight_decay=args.weight_decay, momentum=args.momentum
+        model.optim_policy(), lr=args.lr, weight_decay=args.weight_decay,
+        momentum=args.momentum, nesterov=True
     )
 
     def adjust_lr(optimizer, ep, decay_ep, gamma):

models/resnet_reid.py

@@ -15,10 +15,10 @@ from torch import nn
 from .resnet import ResNet
 
 
-def weights_init(m):
+def weights_init_kaiming(m):
     classname = m.__class__.__name__
     if classname.find('Linear') != -1:
-        nn.init.normal_(m.weight, std=0.001)
+        nn.init.kaiming_normal_(m.weight, a=0, mode='fan_out')
         nn.init.constant_(m.bias, 0.0)
     elif classname.find('Conv') != -1:
         nn.init.kaiming_normal_(m.weight, a=0, mode='fan_in')
@@ -30,6 +30,13 @@ def weights_init(m):
             nn.init.constant_(m.bias, 0.0)
 
 
+def weights_init_classifier(m):
+    classname = m.__class__.__name__
+    if classname.find('Linear') != -1:
+        nn.init.normal_(m.weight, std=0.001)
+        nn.init.constant_(m.bias, 0.0)
+
+
 class ResNetBuilder(nn.Module):
     in_planes = 2048
 
@@ -37,20 +44,25 @@ class ResNetBuilder(nn.Module):
         super().__init__()
         self.base = ResNet(last_stride)
         self.base.load_param(model_path)
-        self.bottleneck = nn.Sequential(
-            nn.BatchNorm2d(self.in_planes),
-            nn.ReLU(True)
-        )
+
         self.num_classes = num_classes
         if num_classes is not None:
-            self.classifier = nn.Linear(self.in_planes, num_classes)
+            self.bottleneck = nn.Sequential(
+                nn.Linear(self.in_planes, 512),
+                nn.BatchNorm1d(512),
+                nn.LeakyReLU(0.1),
+                nn.Dropout(p=0.5)
+            )
+            self.bottleneck.apply(weights_init_kaiming)
+            self.classifier = nn.Linear(512, self.num_classes)
+            self.classifier.apply(weights_init_classifier)
 
     def forward(self, x):
         feat = self.base(x)
-        feat = self.bottleneck(feat)
         global_feat = F.avg_pool2d(feat, feat.shape[2:])  # (b, 2048, 1, 1)
         global_feat = global_feat.view(global_feat.shape[0], -1)
         if self.training and self.num_classes is not None:
+            global_feat = self.bottleneck(global_feat)
             cls_score = self.classifier(global_feat)
             return cls_score, global_feat
         else:
@@ -58,10 +70,12 @@ class ResNetBuilder(nn.Module):
 
     def optim_policy(self):
         base_param_group = self.base.parameters()
-        clf_param_group = self.classifier.parameters()
+        other_param_group = list()
+        other_param_group.extend(list(self.bottleneck.parameters()))
+        other_param_group.extend(list(self.classifier.parameters()))
         return [
             {'params': base_param_group, 'lr_multi': 0.1},
-            {'params': clf_param_group}
+            {'params': other_param_group}
         ]