train densenet with dropout rate 0.2.

ajing · Jun 4, 2017 · 60793df · 60793df
1 parent d10fd53
commit 60793df
Show file tree

Hide file tree

Showing 5 changed files with 27 additions and 26 deletions.
diff --git a/planet_models/densenet_planet.py b/planet_models/densenet_planet.py
@@ -4,7 +4,8 @@
 
 
 def densenet121(num_classes=17, pretrained=False):
-    model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16), num_classes=num_classes)
+    model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16), drop_rate=0.2,
+                     num_classes=num_classes)
     if pretrained:
         # load model dictionary
         model_dict = model.state_dict()

diff --git a/test.py b/test.py
@@ -17,7 +17,7 @@
 from util import BEST_THRESHOLD
 
 
-MODEL='models/densenet121.pth'
+MODEL='models/pretrained_densenet121.pth'
 
 
 def test(model_dir, transform):

diff --git a/trainers/optimize_threshold.py b/trainers/optimize_threshold.py
@@ -23,9 +23,9 @@ def get_labels(prediction, t):
     large_net = nn.DataParallel(densenet121().cuda())
     large_net.load_state_dict(torch.load('../models/pretrained_densenet121.pth'))
     large_net.eval()
-    simple_v2 = nn.DataParallel(SimpleNetV3().cuda())
-    simple_v2.load_state_dict(torch.load('../models/simplenet_v3.1.pth'))
-    simple_v2.eval()
+    # simple_v2 = nn.DataParallel(SimpleNetV3().cuda())
+    # simple_v2.load_state_dict(torch.load('../models/simplenet_v3.1.pth'))
+    # simple_v2.eval()
 
     resnet_data = validation_jpg_loader(512, transform=Compose([
         Scale(224),
@@ -34,30 +34,30 @@ def get_labels(prediction, t):
         Normalize(mean, std)
     ]))
 
-    simplenet_data = validation_jpg_loader(
-        512, transform=Compose(
-            [
-                Scale(72),
-                RandomHorizontalFlip(),
-                ToTensor(),
-                Normalize(mean, std)
-            ]
-        )
-    )
+    # simplenet_data = validation_jpg_loader(
+    #     512, transform=Compose(
+    #         [
+    #             Scale(72),
+    #             RandomHorizontalFlip(),
+    #             ToTensor(),
+    #             Normalize(mean, std)
+    #         ]
+    #     )
+    # )
     pred = []
     targets = []
     # predict
-    for batch_index, ((resnet_img, resnet_target), (simplenet_img, simplenet_target)) \
-        in enumerate(zip(resnet_data, simplenet_data)):
-    #for batch_index, (simplenet_img, simplenet_target) in enumerate(simplenet_data):
-        resnet_output = evaluate(large_net, resnet_img)
+    # for batch_index, ((resnet_img, resnet_target), (simplenet_img, simplenet_target)) \
+    #     in enumerate(zip(resnet_data, simplenet_data)):
+    for batch_index, (simplenet_img, simplenet_target) in enumerate(resnet_data):
+        output = evaluate(large_net, simplenet_img)
         # resnet_output = F.sigmoid(resnet_output)
 
-        simplenet_output = evaluate(simple_v2, simplenet_img)
+        # simplenet_output = evaluate(simple_v2, simplenet_img)
         # simplenet_output = F.sigmoid(simplenet_output)
 
-        output = F.sigmoid((simplenet_output + resnet_output)/2)
-        # output = F.sigmoid(output)
+        # output = F.sigmoid((simplenet_output + resnet_output)/2)
+        output = F.sigmoid(output)
         pred.append(output.data.cpu().numpy())
         targets.append(simplenet_target.cpu().numpy())
 

diff --git a/trainers/train_densenet.py b/trainers/train_densenet.py
@@ -15,7 +15,7 @@ def __call__(self, img):
         return img
 
 
-def get_optimizer(model, lr=1e-4, weight_decay=1e-4):
+def get_optimizer(model, lr=5e-5, weight_decay=5e-4):
     params = [
         {'params': model.features.parameters(), 'lr': lr},
         {'params': model.classifier.parameters(), 'lr': lr * 10}
@@ -32,7 +32,7 @@ def train(epoch):
     net.cuda()
     net = torch.nn.DataParallel(net, device_ids=[0, 1])
     # resnet.load_state_dict(torch.load('../models/simplenet_v3.pth'))
-    train_data_set = train_jpg_loader(100, transform=Compose(
+    train_data_set = train_jpg_loader(90, transform=Compose(
         [
 
             Scale(256),
@@ -45,7 +45,7 @@ def train(epoch):
     ))
     validation_data_set = validation_jpg_loader(64, transform=Compose(
         [
-            Scale(256),
+            Scale(224),
             ToTensor(),
             Normalize(mean, std)
          ]

diff --git a/util.py b/util.py
@@ -12,7 +12,7 @@
 from planet_models.resnet_planet import resnet14_planet
 from planet_models.simplenet_v2 import SimpleNetV2
 
-BEST_THRESHOLD= [0.16, 0.093, 0.203, 0.203, 0.241, 0.175, 0.119, 0.225, 0.134, 0.074, 0.141, 0.22, 0.073, 0.184, 0.167, 0.049, 0.071]
+BEST_THRESHOLD= [0.166, 0.082, 0.237, 0.178, 0.242, 0.223, 0.168, 0.173, 0.197, 0.107, 0.14, 0.242, 0.103, 0.17, 0.273, 0.147, 0.09 ]
 
 
 def evaluate(model, image):