fix errors

antodima · May 13, 2021 · e512490 · e512490
1 parent b928471
commit e512490
Show file tree

Hide file tree

Showing 18 changed files with 34 additions and 23 deletions.
diff --git a/LAB3_1/Assignment1/README.txt b/LAB3_1/Assignment1/README.txt
@@ -1,2 +1,2 @@
-IDNN MSEs: training error=0.00476, validation error=0.00473, test error=0.00442
-RNN MSEs: training error=0.00499, validation error=0.00556, test error=0.00490
+IDNN MSEs: training error=0.00478, validation error=0.00484, test error=0.00451
+RNN MSEs: training error=0.00257, validation error=0.00280, test error=0.00459
diff --git a/LAB3_1/Assignment1/idnn.m b/LAB3_1/Assignment1/idnn.m
@@ -37,6 +37,7 @@
     infos = [];
     infoFieldnames = [];
     valErrors = [];
+    trErrors = [];
 
     for g=1:size(grid,1)
         params = grid(g,:);
@@ -65,39 +66,43 @@
         err = immse(cell2mat(y_train),cell2mat(y_pred));
         val_err = immse(cell2mat(y_val),cell2mat(y_val_pred));
         valErrors(end+1) = val_err;
+        trErrors(end+1) = err;
 
         fprintf('#%d/%d: delays=%d, hiddenSize=%d, epochs=%d, lr=%0.3f, lambda=%0.3f, tr.err=%0.5f, val.err=%0.5f \n',g,size(grid,1),delay,hiddenSize,epochs,lr,lambda,err,val_err);
     end
-    [val, idx] = min(valErrors);
+    [val_err, idx] = min(valErrors);
+    tr_err = trErrors(idx);
     infoCell = infos(:,idx);
     info = cell2struct(infoCell, infoFieldnames);
-    fprintf('min.val.err.=%0.5f, idx=%d, params={delays=%d, hiddenSize=%d, epochs=%d, lr=%0.3f, lambda=%0.3f} \n',val,idx,grid(idx,:));
+    fprintf('Best params: tr.err=%0.5f, val.err=%0.5f, idx=%d, params={delays=%d, hiddenSize=%d, epochs=%d, lr=%0.3f, lambda=%0.3f} \n',tr_err,val_err,idx,grid(idx,:));
     plotperform(info);
 else
     disp('>> Model assessment...');
-    % min.val.err.=0.00477, idx=391, params={delays=2, hiddenSize=10, epochs=90, lr=0.100, lambda=0.100} 
+    % Best params: tr.err=0.00478, val.err=0.00484, idx=391, params={delays=2, hiddenSize=10, epochs=90, lr=0.100, lambda=0.100} 
     net = timedelaynet(0:2,10,'traingdx');
-    net.performParam.regularization = 0.100;
+    net.performParam.regularization = 0.1;
     net.trainParam.epochs = 90;
     net.divideFcn = 'dividetrain';
     net.trainParam.lr = 0.1;
     net.trainParam.showWindow = false;
     [Xs,Xi,Ai,Ts] = preparets(net,X_design,y_design);
     [net,info] = train(net,Xs,Ts,'useParallel','yes');
-    plotperform(info);
+    %plotperform(info);
 
+    ds_pred = net(X_design);
     tr_pred = net(X_train);
     vl_pred = net(X_val);
     ts_pred = net(X_test);
 
-    tr_err = immse( cell2mat(y_train), cell2mat(tr_pred) );
-    vl_err = immse( cell2mat(y_val)  , cell2mat(vl_pred) );
-    ts_err = immse( cell2mat(y_test) , cell2mat(ts_pred) );
+    ds_err = immse( cell2mat(y_design), cell2mat(ds_pred) );
+    tr_err = immse( cell2mat(y_train), cell2mat(tr_pred)  );
+    vl_err = immse( cell2mat(y_val)  , cell2mat(vl_pred)  );
+    ts_err = immse( cell2mat(y_test) , cell2mat(ts_pred)  );
 
     fig = figure;
-    scatter((1:size(y_train,2)),cell2mat(y_train));
+    scatter((1:size(y_design,2)),cell2mat(y_design));
     hold on;
-    scatter((1:size(tr_pred,2)),cell2mat(tr_pred));
+    scatter((1:size(ds_pred,2)),cell2mat(ds_pred));
     xlabel('time')
     ylabel('target');
     title('Training targets and predictions');

diff --git a/LAB3_1/Assignment1/idnn_tr.mat b/LAB3_1/Assignment1/idnn_tr.mat
diff --git a/LAB3_1/Assignment1/images/idnn_learning_curve.png b/LAB3_1/Assignment1/images/idnn_learning_curve.png
diff --git a/LAB3_1/Assignment1/images/idnn_test_targets.png b/LAB3_1/Assignment1/images/idnn_test_targets.png
diff --git a/LAB3_1/Assignment1/images/idnn_training_targets.png b/LAB3_1/Assignment1/images/idnn_training_targets.png
diff --git a/LAB3_1/Assignment1/images/rnn_learning_curve.png b/LAB3_1/Assignment1/images/rnn_learning_curve.png
diff --git a/LAB3_1/Assignment1/images/rnn_test_targets.png b/LAB3_1/Assignment1/images/rnn_test_targets.png
diff --git a/LAB3_1/Assignment1/images/rnn_training_targets.png b/LAB3_1/Assignment1/images/rnn_training_targets.png
diff --git a/LAB3_1/Assignment1/rnn.m b/LAB3_1/Assignment1/rnn.m
@@ -37,6 +37,7 @@
     infos = [];
     infoFieldnames = [];
     valErrors = [];
+    trErrors = [];
 
     for g=1:size(grid,1)
         params = grid(g,:);
@@ -64,17 +65,19 @@
         err = immse(cell2mat(y_train),cell2mat(y_pred));
         val_err = immse(cell2mat(y_val),cell2mat(y_val_pred));
         valErrors(end+1) = val_err;
+        trErrors(end+1) = err;
 
         fprintf('#%d/%d: hiddenSize=%d, epochs=%d, lr=%0.3f, lambda=%0.3f, tr.err=%0.5f, val.err=%0.5f \n',g,size(grid,1),hiddenSize,epochs,lr,lambda,err,val_err);
     end
-    [val, idx] = min(valErrors);
+    [val_err, idx] = min(valErrors);
+    tr_err = trErrors(idx);
     infoCell = infos(:,idx);
     info = cell2struct(infoCell, infoFieldnames);
-    fprintf('min.val.err.=%0.5f, idx=%d, params={hiddenSize=%d, epochs=%d, lr=%0.3f, lambda=%0.3f} \n',val,idx,grid(idx,:));
+    fprintf('Best params: tr.err=%0.5f, val.err=%0.5f, idx=%d, params={hiddenSize=%d, epochs=%d, lr=%0.3f, lambda=%0.3f} \n',tr_err,val_err,idx,grid(idx,:));
     plotperform(info);
 else
     disp('>> Model assessment...');
-    % min.val.err.=0.00432, idx=42, params={hiddenSize=20, epochs=90, lr=0.100, lambda=0.001} 
+    % Best params: tr.err=0.00257, val.err=0.00280, idx=89, params={hiddenSize=40, epochs=90, lr=0.100, lambda=0.010} 
     net = layrecnet(1,20,'traingdx');
     net.performParam.regularization = 0.001;
     net.trainParam.epochs = 90;
@@ -83,20 +86,22 @@
     net.trainParam.showWindow = false;
     [Xs,Xi,Ai,Ts] = preparets(net,X_design,y_design);
     [net,info] = train(net,Xs,Ts,'useParallel','yes');
-    plotperform(info);
+    %plotperform(info);
 
+    ds_pred = net(X_design);
     tr_pred = net(X_train);
     vl_pred = net(X_val);
     ts_pred = net(X_test);
 
+    ds_err = immse( cell2mat(y_design), cell2mat(ds_pred) );
     tr_err = immse( cell2mat(y_train), cell2mat(tr_pred) );
     vl_err = immse( cell2mat(y_val)  , cell2mat(vl_pred) );
     ts_err = immse( cell2mat(y_test) , cell2mat(ts_pred) );
 
     fig = figure;
-    scatter((1:size(y_train,2)),cell2mat(y_train));
+    scatter((1:size(y_design,2)),cell2mat(y_design));
     hold on;
-    scatter((1:size(tr_pred,2)),cell2mat(tr_pred));
+    scatter((1:size(ds_pred,2)),cell2mat(ds_pred));
     xlabel('time')
     ylabel('target');
     title('Training targets and predictions');

diff --git a/LAB3_1/Assignment1/rnn_tr.mat b/LAB3_1/Assignment1/rnn_tr.mat
diff --git a/LAB3_2/Assignment1/README.txt b/LAB3_2/Assignment1/README.txt
@@ -8,6 +8,6 @@ Echo State Network best hyper-parameters:
     - Number of guesses = 10
 
  Mean Squared Errors:
-    - Training MSE = 0.00106
-    - Validation MSE = 0.00111
-    - Test MSE = 0.00097
+    - Training MSE = 0.00092
+    - Validation MSE = 0.00242
+    - Test MSE = 0.00118
diff --git a/LAB3_2/Assignment1/Win.mat b/LAB3_2/Assignment1/Win.mat
diff --git a/LAB3_2/Assignment1/Wout.mat b/LAB3_2/Assignment1/Wout.mat
diff --git a/LAB3_2/Assignment1/Wr.mat b/LAB3_2/Assignment1/Wr.mat
diff --git a/LAB3_2/Assignment1/esn.m b/LAB3_2/Assignment1/esn.m
@@ -127,6 +127,7 @@
     D = y_design(:,Nh+1:end);
     % train the readout
     V = D*H'*inv(H*H'+lambda*eye(Nh+1)); % solve the linear system (Nh+1 for the bias)
+    Y = V * H;
 
     % compute training error
     H_tr = [];
@@ -175,9 +176,9 @@
     fprintf('tr_err=%0.5f, vl_err=%0.5f, ts_err=%0.5f \n',tr_err,vl_err,ts_err);
 
     fig = figure;
-    scatter((1:size(y_train,2)),y_train);
+    scatter((1:size(y_design(:,Nh+1:end),2)),y_design(:,Nh+1:end));
     hold on;
-    scatter((1:size(Y_tr,2)),Y_tr);
+    scatter((1:size(Y,2)),Y);
     xlabel('time')
     ylabel('target');
     title('Training targets and predictions');

diff --git a/LAB3_2/Assignment1/images/esn_test_targets.png b/LAB3_2/Assignment1/images/esn_test_targets.png
diff --git a/LAB3_2/Assignment1/images/esn_training_targets.png b/LAB3_2/Assignment1/images/esn_training_targets.png