minor changes

SimpleNet.m

@@ -232,11 +232,6 @@
             for e=1:params.epochs,
                 idx = randsample(all_idx, batch_size, false);
                 Xtr = X(idx,:);
-                if (1 == 1)
-                    noise_scales = 0.05 * std(Xtr);
-                    noise = bsxfun(@times, randn(size(Xtr)), noise_scales);
-                    Xtr = Xtr + noise;
-                end
                 Ytr = Y(idx,:);
                 for r=1:params.batch_rounds,
                     % Run backprop to compute gradients for this training batch

test_mnist.m

@@ -2,10 +2,11 @@
 % Load and "preprocess" MNIST digit data %
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 load('mnist_data.mat');
-X = X_mnist;
-Y = Y_mnist;
-X = X((Y < 5),:);
-Y = Y((Y < 5),:);
+idx = randsample(size(X_mnist,1),30000);
+X = X_mnist(idx,:);
+Y = Y_mnist(idx,:);
+% X = X((Y < 5),:);
+% Y = Y((Y < 5),:);
 X = ZMUV(double(X));
 Y_vals = unique(Y);
 Ym = zeros(size(X,1),numel(Y_vals));
@@ -23,7 +24,7 @@
 obs_dim = size(X,2);
 out_dim = size(Y,2);
 obs_count = size(X,1);
-train_count = round(obs_count * 0.75);
+train_count = round(obs_count * 0.8);
 
 % Split data into training and testing portions
 tr_idx = randsample(obs_count,train_count,false);
@@ -38,22 +39,23 @@
 act_func = ActFunc(2);
 out_func = ActFunc(1);
 loss_func = LossFunc(3);
-layer_sizes = [obs_dim 2*obs_dim 2*obs_dim out_dim];
+layer_sizes = [obs_dim 1500 1500 out_dim];
 
 % Generate a SimpleNet instance
 net = SimpleNet(layer_sizes, act_func, out_func, loss_func);
 net.init_weights(0.1);
+net.drop_stride = 5;
 
 % Set up parameter struct for updates
 params = struct();
-params.epochs = 1000;
+params.epochs = 10000;
 params.start_rate = 1.0;
 params.decay_rate = 0.1^(1 / params.epochs);
 params.momentum = 0.5;
-params.weight_bound = 10;
+params.weight_bound = 20;
 params.batch_size = 100;
 params.batch_rounds = 1;
-params.dr_obs = 0.0;
+params.dr_obs = 0.1;
 params.dr_node = 0.5;
 params.do_validate = 1;
 params.X_v = X_te;