% Author: Scott Sanner% Email:  ssanner@cs.stanford.edu% Course: CS223B, Winter% Desc:   Trains a neural net given a training set with target values%         and randomly selects training and validation data, training%         either for the max number of networks, network convergence,%         or increase in validation error.  Returns network and%         performance. % % [NET,PERF, ERR] = trainnn(NET,IMVECTOR,TVECTOR,percent_val,iter)function [NET,PERF, ERR] = trainnn(NET,IMVECTOR,TVECTOR,percent_val,iter)% Setup validation and test setsN_IMS = size(IMVECTOR,2);CHOICE  = rand(1,N_IMS) > percent_val;V_TRAIN = find(CHOICE);V_VALID = find(1-CHOICE);IM_TRAIN_V = IMVECTOR(:,V_TRAIN);IM_TRAIN_R = TVECTOR (:,V_TRAIN);IM_VALID_V = IMVECTOR(:,V_VALID);IM_VALID_R = TVECTOR (:,V_VALID);fprintf(1,'Training set quantity:   #%d\n', size(IM_TRAIN_V,2));fprintf(1,'Validation set quantity: #%d\n', size(IM_VALID_V,2));% Setup net parametersNET.trainParam.epochs = 1;NET.trainParam.goal   = 0.0001;TRAIN_OUT = simnn(NET,IM_TRAIN_V);VALID_OUT = simnn(NET,IM_VALID_V);PERF = [];ERR  = [];PERF = [PERF; 0, (1-(sum(abs(TRAIN_OUT-IM_TRAIN_R)./1.8)./(size(TRAIN_OUT,2)))), ...                 (1-(sum(abs(VALID_OUT-IM_VALID_R)./1.8)./(size(VALID_OUT,2))))];ERR  = [ERR;  0, ((sum(abs(TRAIN_OUT-IM_TRAIN_R))./(size(TRAIN_OUT,2))).^2), ...                 ((sum(abs(VALID_OUT-IM_VALID_R))./(size(VALID_OUT,2))).^2)];% Train for specified number of iterationsfor i=1:iter,  fprintf(1,'Starting iteration %d\n', i);  drawnow;  [NET, TR] = train(NET, IM_TRAIN_V, IM_TRAIN_R);  TRAIN_OUT = simnn(NET,IM_TRAIN_V);  VALID_OUT = simnn(NET,IM_VALID_V);  %[TRAIN_OUT; IM_TRAIN_R]  %[VALID_OUT; IM_VALID_R]  PERF = [PERF; i, (1-(sum(abs(TRAIN_OUT-IM_TRAIN_R)./1.8)./(size(TRAIN_OUT,2)))), ...                   (1-(sum(abs(VALID_OUT-IM_VALID_R)./1.8)./(size(VALID_OUT,2))))];  ERR  = [ERR;  i, ((sum(abs(TRAIN_OUT-IM_TRAIN_R))./(size(TRAIN_OUT,2))).^2), ...                   ((sum(abs(VALID_OUT-IM_VALID_R))./(size(VALID_OUT,2))).^2)];end