nChannels not according to paper

densenet.py

@@ -58,14 +58,14 @@ def forward(self, x):
 
 
 class DenseNet(nn.Module):
-    def __init__(self, growthRate, depth, reduction, nClasses, bottleneck):
+     def __init__(self, growthRate, depth, reduction, nClasses, bottleneck, compression):
         super(DenseNet, self).__init__()
-
+    
         nDenseBlocks = (depth-4) // 3
         if bottleneck:
             nDenseBlocks //= 2
 
-        nChannels = 2*growthRate
+        nChannels = 2*growthRate if compression and bottleneck else 16 # They only do this for BC type
         self.conv1 = nn.Conv2d(3, nChannels, kernel_size=3, padding=1,
                                bias=False)
         self.dense1 = self._make_dense(nChannels, growthRate, nDenseBlocks, bottleneck)
@@ -84,6 +84,7 @@ def __init__(self, growthRate, depth, reduction, nClasses, bottleneck):
         nChannels += nDenseBlocks*growthRate
 
         self.bn1 = nn.BatchNorm2d(nChannels)
+        self.avgpool = nn.AvgPool2d(kernel_size=8)
         self.fc = nn.Linear(nChannels, nClasses)
 
         for m in self.modules():
@@ -111,6 +112,49 @@ def forward(self, x):
         out = self.trans1(self.dense1(out))
         out = self.trans2(self.dense2(out))
         out = self.dense3(out)
-        out = torch.squeeze(F.avg_pool2d(F.relu(self.bn1(out)), 8))
-        out = F.log_softmax(self.fc(out))
+        #out = torch.squeeze(F.avg_pool2d(F.relu(self.bn1(out)), 8)) ## Why did you relu and bn again?
+        out = self.avgpool(8)
+        # out = F.log_softmax(self.fc(out))
+        out = self.fc(out)                      # You can define the softmax within the loss function right?
         return out
+
+
+def denseNet_40_12():
+    return DenseNet(12, 40, 1, 10, bottleneck=False, compression=False)
+
+def denseNet_100_12():
+    return DenseNet(12, 100, 1, 10, bottleneck=False, compression=False)
+
+def denseNet_100_24():
+    return DenseNet(24, 100, 1, 10, bottleneck=False, compression=False)
+
+def denseNetBC_100_12():
+    return DenseNet(12, 100, 0.5, 10, bottleneck=True, compression=True)
+
+def denseNetBC_250_24():
+    return DenseNet(24, 250, 0.5, 10, bottleneck=True, compression=True)
+
+def denseNetBC_190_40():
+    return DenseNet(40, 190, 0.5, 10, bottleneck=True, compression=True)
+
+
+''' Did this little check:
+DenseNets implemented on the paper <https://arxiv.org/pdf/1608.06993.pdf>
+
++-------------+-------------+-------+--------------+
+|    Model    | Growth Rate | Depth | M. of Params |
++-------------+-------------+-------+--------------+
+|  DenseNet   |     12      |  40   |     1.02     |
++-------------+-------------+-------+--------------+
+|  DenseNet   |     12      |  100  |     6.98     |
++-------------+-------------+-------+--------------+
+|  DenseNet   |     24      |  100  |    27.249    |
++-------------+-------------+-------+--------------+
+| DenseNet-BC |     12      |  100  |    0.769     |
++-------------+-------------+-------+--------------+
+| DenseNet-BC |     24      |  250  |    15.324    |
++-------------+-------------+-------+--------------+
+| DenseNet-BC |     40      |  190  |    25.624    |
++-------------+-------------+-------+--------------+
+
+''''