DenseNet161 with SGD

planet_models/resnet_planet.py

@@ -1,26 +1,77 @@
-from torchvision.models.resnet import BasicBlock, ResNet, resnet34, resnet50, resnet101, resnet152, resnet18
+from torchvision.models.resnet import BasicBlock, ResNet, resnet34, resnet50, resnet101, resnet152, \
+    resnet18, model_urls, model_zoo
 import torch.nn as nn
 import math
 
 
-def resnet34_planet():
-    return resnet34(False, num_classes=17)
-
-
-def resnet101_planet():
-    return resnet101(pretrained=False, num_classes=17)
-
-
-def resnet50_planet():
-    return resnet50(pretrained=False, num_classes=17)
-
-
-def resnet152_planet():
-    return resnet152(pretrained=False, num_classes=17)
-
-
-def resnet18_planet():
-    return resnet18(pretrained=False, num_classes=17)
+def resnet34_planet(pretrained=False):
+    model = resnet34(False, num_classes=17)
+    if pretrained:
+        # load model dictionary
+        model_dict = model.state_dict()
+        # load pretrained model
+        pretrained_dict = model_zoo.load_url(model_urls['resnet34'])
+        # update model dictionary using pretrained model without classifier layer
+        model_dict.update({key: pretrained_dict[key] for key in pretrained_dict.keys() if 'fc' not in key})
+        model.load_state_dict(model_dict)
+
+    return model
+
+
+def resnet101_planet(pretrained=False):
+    model = resnet101(False, num_classes=17)
+    if pretrained:
+        # load model dictionary
+        model_dict = model.state_dict()
+        # load pretrained model
+        pretrained_dict = model_zoo.load_url(model_urls['resnet101'])
+        # update model dictionary using pretrained model without classifier layer
+        model_dict.update({key: pretrained_dict[key] for key in pretrained_dict.keys() if 'fc' not in key})
+        model.load_state_dict(model_dict)
+
+    return model
+
+
+def resnet50_planet(pretrained=False):
+    model = resnet50(False, num_classes=17)
+    if pretrained:
+        # load model dictionary
+        model_dict = model.state_dict()
+        # load pretrained model
+        pretrained_dict = model_zoo.load_url(model_urls['resnet50'])
+        # update model dictionary using pretrained model without classifier layer
+        model_dict.update({key: pretrained_dict[key] for key in pretrained_dict.keys() if 'fc' not in key})
+        model.load_state_dict(model_dict)
+
+    return model
+
+
+def resnet152_planet(pretrained=False):
+    model = resnet50(False, num_classes=17)
+    if pretrained:
+        # load model dictionary
+        model_dict = model.state_dict()
+        # load pretrained model
+        pretrained_dict = model_zoo.load_url(model_urls['resnet152'])
+        # update model dictionary using pretrained model without classifier layer
+        model_dict.update({key: pretrained_dict[key] for key in pretrained_dict.keys() if 'fc' not in key})
+        model.load_state_dict(model_dict)
+
+    return model
+
+
+def resnet18_planet(pretrained=False):
+    model = resnet18(False, num_classes=17)
+    if pretrained:
+        # load model dictionary
+        model_dict = model.state_dict()
+        # load pretrained model
+        pretrained_dict = model_zoo.load_url(model_urls['resnet18'])
+        # update model dictionary using pretrained model without classifier layer
+        model_dict.update({key: pretrained_dict[key] for key in pretrained_dict.keys() if 'fc' not in key})
+        model.load_state_dict(model_dict)
+
+    return model
 
 
 def resnet14_planet():

trainers/baseline_trainer.py

@@ -81,6 +81,27 @@ def get_dataloader(batch_size):
     return train_data_loader, valid_dataloader
 
 
+def get_optimizer(net, lr, resnet=False, pretrained=False):
+    if pretrained:
+        if resnet:
+            parameters = [
+                {'params': net.fc.parameters(), 'lr': lr*10},
+                {'params': net.layer1.parameters(), 'lr': lr},
+                {'params': net.layer2.parameters(), 'lr': lr},
+                {'params': net.layer3.parameters(), 'lr': lr},
+                {'params': net.layer4.parameters(), 'lr': lr}
+            ]
+        else:
+            parameters = [
+                {'params': net.features.parameters(), 'lr': lr},
+                {'params': net.classifier.parameters(), 'lr': lr * 10}
+            ]
+        optimizer = optim.SGD(params=parameters, weight_decay=5e-5, momentum=.9)
+    else:
+        optimizer = optim.SGD(params=net.parameters(), lr=lr, weight_decay=5e-5, momentum=.9)
+    return optimizer
+
+
 def train_baselines():
 
     train_data, val_data = get_dataloader(96)
@@ -103,7 +124,7 @@ def train_baselines():
 
         # optimizer
         # optimizer = optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=0.0005)
-        optimizer = optim.Adam(net.parameters(), lr=1e-4, weight_decay=5e-4)
+        # optimizer = optim.Adam(net.parameters(), lr=1e-4, weight_decay=5e-4)
 
         smooth_loss = 0.0
         train_loss = np.nan
@@ -114,10 +135,11 @@ def train_baselines():
         t = time.time()
 
         for epoch in range(num_epoches):  # loop over the dataset multiple times
+            optimizer = get_optimizer(net, lr=.01, pretrained=True, resnet=True if 'resnet' in name else False)
             # train loss averaged every epoch
             total_epoch_loss = 0.0
 
-            # lr_schedule(epoch, optimizer)
+            lr_schedule(epoch, optimizer, pretrained=True)
 
             rate = get_learning_rate(optimizer)[0]  # check
 

util.py

@@ -131,15 +131,25 @@ def get_learning_rate(optimizer):
     return lr
 
 
-def lr_schedule(epoch, optimizer):
-    if 0 <= epoch < 20:
-        lr = 1e-4
-    elif 20 <= epoch < 35:
-        lr = 9e-5
-    elif 35 <= epoch < 45:
-        lr = 5e-5
+def lr_schedule(epoch, optimizer, pretrained=False):
+    if pretrained:
+        if 0 <= epoch < 10:
+            lr = 1e-2
+        elif 10 <= epoch < 25:
+            lr = 5e-3
+        elif 25 <= epoch < 40:
+            lr = 1e-3
+        else:
+            lr = 1e-4
     else:
-        lr = 5e-5
+        if 0 <= epoch < 10:
+            lr = 1e-1
+        elif 10 <= epoch < 25:
+            lr = 5e-2
+        elif 25 <= epoch < 40:
+            lr = 1e-2
+        else:
+            lr = 1e-3
 
     for para_group in optimizer.param_groups:
         para_group['lr'] = lr