From 4ee24b919fe74b89d57808d753bdfa2270081686 Mon Sep 17 00:00:00 2001
From: jxu7 <jxu7@bsu.edu>
Date: Wed, 10 May 2017 16:09:17 -0400
Subject: [PATCH] add resnext

---
 planet_models/resnetxt.py |  9 -------
 planet_models/resnext.py  | 39 +++++++++++++++++++++++++++
 trainers/resnet_forest.py | 56 ++++++++++++++++++++++++---------------
 3 files changed, 73 insertions(+), 31 deletions(-)
 delete mode 100644 planet_models/resnetxt.py
 create mode 100644 planet_models/resnext.py

diff --git a/planet_models/resnetxt.py b/planet_models/resnetxt.py
deleted file mode 100644
index 2be6e4d..0000000
--- a/planet_models/resnetxt.py
+++ /dev/null
@@ -1,9 +0,0 @@
-import torch
-from torch.nn import Module
-from torch.nn import functional as F
-
-
-class Block(Module):
-    def __init__(self):
-        super(Block, self).__init__()
-
diff --git a/planet_models/resnext.py b/planet_models/resnext.py
new file mode 100644
index 0000000..8b35566
--- /dev/null
+++ b/planet_models/resnext.py
@@ -0,0 +1,39 @@
+"""
+    A re-implementation of ResNeXT. All the blocks are of bottleneck type.
+    The code follows the style of resnet.py in pytorch vision model.
+"""
+import torch
+from torch.nn import *
+from torch.nn import functional as F
+
+
+class Bottleneck(Module):
+    """Type C in the paper"""
+    def __init__(self, width, planes, cardinality, downsample=None, activation_fn=ELU):
+        super(Bottleneck, self).__init__()
+
+    def forward(self, x):
+        pass
+
+
+class ResNeXT(Module):
+    def __init__(self, block, depths, num_classes, cardinality=32, activation_fn=ELU):
+        super(ResNeXT, self).__init__()
+        self.inplanes = 64
+        self.cardinality = cardinality
+        self.conv1 = Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = BatchNorm2d(64)
+        self.activation = activation_fn(inplace=True)
+        self.maxpool = MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.stage1 = self._make_layers(block, self.inplanes, depths[0])
+        self.stage2 = self._make_layers(block, self.inplanes, depths[1], stride=2)
+        self.stage3 = self._make_layers(block, self.inplanes, depths[2], stride=2)
+        self.stage4 = self._make_layers(block, self.inplanes, depths[3], stride=2)
+        self.avgpool = AvgPool2d(7)
+        self.fc = Linear(block.expansion*512, num_classes)
+
+    def _make_layers(self, block, planes, blocks, stride=1):
+        pass
+
+    def forward(self, x):
+        pass
diff --git a/trainers/resnet_forest.py b/trainers/resnet_forest.py
index 07af050..5a86444 100644
--- a/trainers/resnet_forest.py
+++ b/trainers/resnet_forest.py
@@ -29,10 +29,41 @@ def train_resnet_forest(epoch=50):
             ToTensor()
          ]
     ))
-
+    best_loss = np.inf
+    patience = 0
     for i in range(epoch):
-        # save the model for every epoch
-        torch.save(resnet.state_dict(), '../models/resnet-34.pth')
+        # evaluating
+        val_loss = 0.0
+        f2_scores = 0.0
+        resnet.eval()
+        for batch_index, (val_x, val_y) in enumerate(validation_data_set):
+            if is_cuda_availible:
+                val_y = val_y.cuda()
+            val_y = Variable(val_y, volatile=True)
+            val_output = evaluate(resnet, val_x)
+            val_loss += criterion(val_output, val_y)
+            binary_y = threshold_labels(val_output.data.cpu().numpy())
+            f2 = f2_score(val_y.data.cpu().numpy(), binary_y)
+            f2_scores += f2
+        if best_loss > val_loss:
+            best_loss = val_loss
+            torch.save(resnet.state_dict(), '../models/resnet-34.pth')
+        else:
+            print('Reload previous model')
+            patience += 1
+            resnet.load_state_dict(torch.load('../models/resnet-34.pth'))
+
+        if patience >= 5:
+            print('Early stopping!')
+            break
+
+        print('Evaluation loss is {}, Training loss is {}'.format(val_loss.data[0]/batch_index, loss.data[0]))
+        print('F2 Score is %s' % (f2_scores/batch_index))
+        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()
@@ -44,25 +75,6 @@ def train_resnet_forest(epoch=50):
             loss.backward()
             optimizer.step()
 
-            if batch_index % 50 == 0:
-                val_loss = 0.0
-                f2_scores = 0.0
-                resnet.eval()
-                for batch_index, (val_x, val_y) in enumerate(validation_data_set):
-                    if is_cuda_availible:
-                        val_y = val_y.cuda()
-                    val_y = Variable(val_y, volatile=True)
-                    val_output = evaluate(resnet, val_x)
-                    val_loss += criterion(val_output, val_y)
-                    binary_y = threshold_labels(val_output.data.cpu().numpy())
-                    f2 = f2_score(val_y.data.cpu().numpy(), binary_y)
-                    f2_scores += f2
-                print('Evaluation loss is {}, Training loss is {}'.format(val_loss.data[0]/batch_index,
-                                                                          loss.data[0]))
-                print('F2 Score is %s' % (f2_scores/batch_index))
-                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)
         print('Finished epoch {}'.format(i))
     logger.save()
     logger.save_plot()