How can I prevent overfitting when the dataset is not to large. My dataset consists of 5 classes with a total dataset size about 15k images.
I have tried data augmentation but doesn’t help too much.
class Net(nn.Module):
def __init__(self, num_channels):
super(Net, self).__init__()
self.num_channels = num_channels
self.conv1 = nn.Conv2d(in_channels=1,
out_channels=self.num_channels,
kernel_size=3,
stride=2,
padding=0)
self.bn1 = nn.BatchNorm2d(self.num_channels)
self.conv2 = nn.Conv2d(in_channels=self.num_channels,
out_channels=self.num_channels * 2,
kernel_size=3,
stride=1,
padding=0)
self.bn2 = nn.BatchNorm2d(self.num_channels * 2)
self.conv3 = nn.Conv2d(in_channels=self.num_channels * 2,
out_channels=self.num_channels * 4,
kernel_size=3,
stride=1,
padding=0)
self.bn3 = nn.BatchNorm2d(self.num_channels * 4)
self.fc1 = nn.Linear(self.num_channels * 4 * 6 * 6, self.num_channels * 4)
self.fcbn1 = nn.BatchNorm1d(self.num_channels * 4)
self.fc2 = nn.Linear(self.num_channels * 4, 5)
def forward(self, x):
x = self.bn1(self.conv1(x)) # Salida: self.num_channels x 128 x 128
x = F.relu(F.max_pool2d(x, 2)) # salida: num_channels x 63 x 63
x = self.bn2(self.conv2(x)) # Salida: self.num_channels*2 x 31 x 31
x = F.relu(F.max_pool2d(x, 2)) # salida: num_channels*2 x 29 x 29
x = self.bn3(self.conv3(x)) # Salida: self.num_channels*4 x 14 x 14
x = F.relu(F.max_pool2d(x, 2)) # salida: num_channels*4 x 12 x 12
# Flatten
x = x.view(-1, self.num_channels * 4 * 6 * 6)
# fc
x = F.relu(self.fcbn1(self.fc1(x)))
x = F.dropout(x, p=0.5, training=True)
x = self.fc2(x)
return x
Here is the loss
