Caffe-Experiments

Collection of solvers, models of my experiments.

Usage

• Layer definition

  • Add the necessary paths to dataset.
  • Apply necessary transforms - Crop, mean, scale etc.
  • Fix Batchsize
  • Fix number of feature maps, kernel size, stride etc.
  • Fix activation function
  • Choose your loss function

• Solver

  • Add the relative path to layer definition
  • Add snapshot prefix
  • Add iterations, max iterations, learning rate, solver type, etc.

Then create a script similar to train_lent.sh in examples/mnist folder with the solver file as arguments and run it for training each layer.

Note : For training layer_(n), we must fix the weights till layer_(n-1) (copy weights from layer_(n-1) and set lr_param to zero). Do the necessary changes as above and edit the layer2/finetune_training.sh and fix the paths. Point the weights parameter to the snapshot of layer1 to copy weights and use the solver of layer2. You can extend the same procedure to multiple layers as you please.