A Neural network on CIFAR-10 dataset adopted from Keras/example (unfinished)

readme.md

CIFAR10-CNN

A Neural network on CIFAR-10 dataset adopted from Keras' example.

cifar10_keras.ml

open Owl
open Owl_neural
open Algodiff.S
open Owl_neural_neuron

(*
let test_cnn nn x y =
  for i = 0 to 9 do
    let u = Dense.Ndarray.S.slice [[i]] x in
    Dense.Ndarray.S.reshape u [|3;32;32|]
    |> Dense.Matrix.S.of_ndarray
    |> Dataset.print_mnist_image;
    let p = Graph.run (Arr u) nn |> Algodiff.S.unpack_mat in
    Owl_dense_matrix_generic.print p;
    Printf.printf "prediction: %i\n" (let _, _, j = Owl_dense_matrix_generic.max_i p in j)
  done
  
 
let mnist_model () = 
  let open Owl_neural_graph in
  let nn = input [|28;28;1|]
    |> conv2d [|3;3;1;32|] [|1;1|] ~act_typ:Activation.Relu ~padding:Owl_dense_ndarray_generic.VALID
    |> conv2d [|3;3;32;64|] [|1;1|] ~act_typ:Activation.Relu ~padding:Owl_dense_ndarray_generic.VALID
    |> max_pool2d [|2;2|] [|2;2|]
    |> dropout 0.25
    |> fully_connected 1024 ~act_typ:Activation.Relu
    |> linear 10 ~act_typ:Activation.Softmax
    |> get_network
  in print nn; 
  nn
*)


let model () = 
  let open Owl_neural_graph in
  let nn = input [|32;32;3|]
    |> conv2d [|3;3;3;32|] [|1;1|] ~act_typ:Activation.Relu
    |> conv2d [|3;3;32;32|] [|1;1|] ~act_typ:Activation.Relu ~padding:Owl_dense_ndarray_generic.VALID
    |> max_pool2d [|2;2|] [|2;2|] ~padding:Owl_dense_ndarray_generic.VALID
    |> dropout 0.25
    |> conv2d [|3;3;32;64|] [|1;1|] ~act_typ:Activation.Relu
    |> conv2d [|3;3;64;64|] [|1;1|] ~act_typ:Activation.Relu ~padding:Owl_dense_ndarray_generic.VALID
    |> max_pool2d [|2;2|] [|2;2|] ~padding:Owl_dense_ndarray_generic.VALID
    |> dropout 0.25
    |> fully_connected 512 ~act_typ:Activation.Relu
    |> linear 10 ~act_typ:Activation.Softmax
    |> get_network
  in 
  print nn;
  nn

let train_cifar10_keras_graph () =

  let x, y = Dataset.load_cifar_train_data 1 in
  let m = Dense.Matrix.S.row_num x in (* 10000 *)
  let x = Dense.Matrix.S.to_ndarray x in
  let x = Dense.Ndarray.S.reshape x [|m;3;32;32|] in
  let x = Dense.Ndarray.S.transpose ~axis:[|0;2;3;1|] x in 
  let classes = 10 in 
  let y' = Dense.Matrix.S.zeros m classes in 
  for i = 0 to m - 1 do
    Dense.Matrix.S.set y' i (int_of_float y.{i,0}) 1.
  done;

  let nn = model () in 
  let params = Params.config
    ~batch:(Batch.Mini 32) ~learning_rate:(Learning_Rate.RMSprop (0.0001, 1e-6)) ~checkpoint:0.1 0.2 in
  Graph.train_cnn ~params nn x y' (* |> ignore;

  let a, b = Owl_neural_optimise.Utils.draw_samples (Arr x) (Mat y) 10 in 
  test_cnn nn (Algodiff.S.unpack_arr a) (Algodiff.S.unpack_mat b) *)
  
let _ = 
  train_cifar10_keras_graph ()