Newbie here… I am working on a classification problem as a precursor for object detection work. The use case surrounds aerial imagery and in my dataset the classifiable area is very small relative to the image size. I have taken some measures with data augmentation to improve performance but am not seeing the improvement I would hope to see. Since my dataset is quite different from the ImageNet dataset, I have been trying to add an additional convolutional layer in place of the resnet18 fc layer. I believe I can do this with a custom_head. I have had success adding a head like the one below but am struggling to implement a convolution. Any suggestions would be very appreciated.
head= nn.Sequential( Flatten(), nn.ReLU(), nn.Dropout(0.5), nn.Linear(25088,256), nn.ReLU(), nn.BatchNorm1d(256), nn.Dropout(0.5), nn.Linear(256,2), )