diff --git a/planet_models/resnext.py b/planet_models/resnext.py
index 8b35566..bcc9462 100644
--- a/planet_models/resnext.py
+++ b/planet_models/resnext.py
@@ -9,8 +9,26 @@
 
 class Bottleneck(Module):
     """Type C in the paper"""
-    def __init__(self, width, planes, cardinality, downsample=None, activation_fn=ELU):
+    def __init__(self, inplanes, planes, cardinality, stride, downsample=None, activation_fn=ELU):
+        """
+        Parameters:
+            inplanes: # of input channels
+            planes: # of output channels
+            cardinality: # of convolution groups
+            stride: convolution stride
+            downsample: convolution operation to increase the width of the output
+            activation_fn: activation function
+        """
         super(Bottleneck, self).__init__()
+        depth = planes/2
+        self.conv1 = Conv2d(inplanes, depth, kernel_size=1, stride=1, padding=0, bias=False)
+        self.bn1 = BatchNorm2d(planes/2)
+        # group convolution
+        self.conv2 = Conv2d(depth, depth, kernel_size=3, groups=cardinality, stride=stride, padding=1, bias=False)
+        self.bn2 = BatchNorm2d(planes/2)
+        # increase depth
+        self.conv3 = Conv2d(planes/2, planes, kernel_size=1, stride=1, padding=0, bias=False)
+        self.bn3 = BatchNorm2d(planes)
 
     def forward(self, x):
         pass
diff --git a/test.py b/test.py
index b3832b5..09db84b 100644
--- a/test.py
+++ b/test.py
@@ -12,7 +12,7 @@
 from planet_models.resnet_planet import *
 from trainers.train_simplenet import evaluate
 
-MODEL='models/resnet-101.pth'
+MODEL='models/resnet-34.pth'
 
 
 def test(model_dir, transform):
@@ -25,7 +25,7 @@ def test(model_dir, transform):
     ))
 
     if 'resnet' in model_dir:
-        model = nn.DataParallel(resnet101_planet())
+        model = nn.DataParallel(resnet34_planet())
     else:
         model = MultiLabelCNN(17)
     model.load_state_dict(torch.load(model_dir))
@@ -39,7 +39,7 @@ def test(model_dir, transform):
         result = F.sigmoid(result)
         result = result.data.cpu().numpy()
         for r, id in zip(result, im_ids):
-            r = np.where(r >= 0.15)[0]
+            r = np.where(r >= 0.24)[0]
             labels = [idx_to_label[index] for index in r]
             imid_to_label[id] = sorted(labels)
         print('Batch Index {}'.format(batch_idx))
diff --git a/trainers/threshold.py b/trainers/threshold.py
new file mode 100644
index 0000000..68b036c
--- /dev/null
+++ b/trainers/threshold.py
@@ -0,0 +1,5 @@
+from util import optimize_threshold
+from planet_models.resnet_planet import resnet34_planet
+
+model = resnet34_planet()
+optimize_threshold(model, '../models/resnet-34.pth')
diff --git a/trainers/resnet_forest.py b/trainers/train_resnet.py
similarity index 100%
rename from trainers/resnet_forest.py
rename to trainers/train_resnet.py
index 5a86444..6e29bd0 100644
--- a/trainers/resnet_forest.py
+++ b/trainers/train_resnet.py
@@ -32,6 +32,20 @@ def train_resnet_forest(epoch=50):
     best_loss = np.inf
     patience = 0
     for i in range(epoch):
+        # training
+        for batch_index, (target_x, target_y) in enumerate(train_data_set):
+            if is_cuda_availible:
+                target_x, target_y = target_x.cuda(), target_y.cuda()
+            resnet.train()
+            target_x, target_y = Variable(target_x), Variable(target_y)
+            optimizer.zero_grad()
+            output = resnet(target_x)
+            loss = criterion(output, target_y)
+            loss.backward()
+            optimizer.step()
+
+        print('Finished epoch {}'.format(i))
+
         # evaluating
         val_loss = 0.0
         f2_scores = 0.0
@@ -62,20 +76,6 @@ def train_resnet_forest(epoch=50):
         logger.add_record('train_loss', loss.data[0])
         logger.add_record('evaluation_loss', val_loss.data[0]/batch_index)
         logger.add_record('f2_score', f2_scores/batch_index)
-
-        # training
-        for batch_index, (target_x, target_y) in enumerate(train_data_set):
-            if is_cuda_availible:
-                target_x, target_y = target_x.cuda(), target_y.cuda()
-            resnet.train()
-            target_x, target_y = Variable(target_x), Variable(target_y)
-            optimizer.zero_grad()
-            output = resnet(target_x)
-            loss = criterion(output, target_y)
-            loss.backward()
-            optimizer.step()
-
-        print('Finished epoch {}'.format(i))
     logger.save()
     logger.save_plot()
 
diff --git a/util.py b/util.py
index d94f326..1f9da13 100644
--- a/util.py
+++ b/util.py
@@ -75,8 +75,11 @@ def optimize_threshold(model, mode_dir, resolution=10000):
     """
     model = nn.DataParallel(model)
     model.load_state_dict(torch.load(mode_dir))
-    model.cuda(0)
-    data = validation_jpg_loader(256, transform=input_transform(227))
+    model.cuda()
+    data = validation_jpg_loader(256, transform=Compose([
+        Scale(224),
+        ToTensor()
+    ]))
     num_class = 17
     pred = []
     targets = []
@@ -130,9 +133,9 @@ def save_plot(self):
         plt.plot(np.arange(len(eval_loss)), eval_loss, color='blue', label='eval_loss')
         plt.legend(loc='best')
 
-        plt.savefig('log/%s_losses.jpg' % self.name)
+        plt.savefig('../log/%s_losses.jpg' % self.name)
 
         plt.figure()
         plt.plot(np.arange(len(f2_scores)), f2_scores)
-        plt.savefig('log/%s_fcscore.jpg' % self.name)
+        plt.savefig('../log/%s_fcscore.jpg' % self.name)