AnnManager example

Example for ANN (matlab and python), lsq, and ga

source_ann/ann_example/get_ann_data.m → source_ann/ann_example/ann_data/get_ann_data.m

@@ -2,13 +2,13 @@
 
 n_sol = 10000;
 
-inp.x_1 = 6.0+3.0.*rand(1, n_sol);
+inp.x_1 = 7.0+3.0.*rand(1, n_sol);
 inp.x_2 = 1.0+5.0.*rand(1, n_sol);
 
 out_ref.y_1 = inp.x_1+inp.x_2+0.1.*rand(1, n_sol);
 out_ref.y_2 = inp.x_1-inp.x_2+0.1.*rand(1, n_sol);
 
-out_nrm.y_1 = 11.0.*ones(1, n_sol);
-out_nrm.y_2 = 4.0.*ones(1, n_sol);
+out_nrm.y_1 = 12.0.*ones(1, n_sol);
+out_nrm.y_2 = 5.0.*ones(1, n_sol);
 
 end

source_ann/ann_example/ann_data/get_ann_manager.m

@@ -0,0 +1,52 @@
+function get_ann_manager(ann_type)
+
+% name
+fprintf('################## master_train : %s\n', ann_type)
+
+% data
+[ann_input, tag_train] = get_ann_param(ann_type);
+[n_sol, inp, out_ref, out_nrm] = get_ann_data();
+
+% test class
+fprintf('constructor\n')
+obj = AnnManager(ann_input);
+
+fprintf('train\n')
+obj.train(tag_train, n_sol, inp, out_ref, out_nrm);
+
+fprintf('get_fom\n')
+fom = obj.get_fom();
+assert(isstruct(fom), 'invalid fom')
+
+fprintf('disp\n')
+obj.disp();
+
+fprintf('dump\n')
+[ann_input, ann_data] = obj.dump();
+
+fprintf('delete\n')
+obj.delete();
+
+fprintf('predict\n')
+predict(ann_input, ann_data, n_sol, inp, out_nrm)
+
+fprintf('################## master_train : %s\n', ann_type)
+
+end
+
+function predict(ann_input, ann_data, n_sol, inp, out_nrm)
+
+obj = AnnManager(ann_input);
+obj.load(ann_data);
+
+[is_valid_tmp, out_nrm_tmp] = obj.predict_nrm(n_sol, inp, out_nrm);
+assert(islogical(is_valid_tmp), 'invalid fom')
+assert(isstruct(out_nrm_tmp), 'invalid fom')
+
+[is_valid_tmp, out_nrm_tmp] = obj.predict_ann(n_sol, inp, out_nrm);
+assert(islogical(is_valid_tmp), 'invalid fom')
+assert(isstruct(out_nrm_tmp), 'invalid fom')
+
+obj.delete();
+
+end

source_ann/ann_example/ann_data/get_ann_param.m

@@ -0,0 +1,132 @@
+function [ann_input, tag_train] = get_ann_param(ann_type)
+
+% var_inp
+var_inp = {};
+var_inp{end+1} = struct('name', 'x_1', 'var_trf', 'lin', 'var_norm', 'min_max', 'min', 0.99.*7.0, 'max', 1.01.*10.0);
+var_inp{end+1} = struct('name', 'x_2', 'var_trf', 'lin', 'var_norm', 'min_max', 'min', 0.99.*1.0, 'max', 1.01.*6.0);
+
+% var_out
+var_out = {};
+var_out{end+1} = struct('name', 'y_1', 'var_trf', 'lin', 'var_norm', 'min_max', 'use_nrm', true, 'var_err', 'rel');
+var_out{end+1} = struct('name', 'y_2', 'var_trf', 'lin', 'var_norm', 'min_max', 'use_nrm', true, 'var_err', 'rel');
+
+% split_train_test
+split_train_test.ratio_train = 0.5;
+split_train_test.n_train_min = 5;
+split_train_test.n_test_min = 5;
+split_train_test.type = 'no_overlap';
+
+% split the variable
+split_var = false;
+
+% ann_info
+switch ann_type
+    case 'matlab_ann'
+        ann_info.type = ann_type;
+        ann_info.fct_model = @fct_model;
+        ann_info.fct_train = @fct_train;
+    case 'python_ann'
+        ann_info.type = ann_type;
+        ann_info.hostname = 'localhost';
+        ann_info.port = 10000;
+        ann_info.timeout = 240;
+    case 'matlab_lsq'
+        ann_info.type = ann_type;
+        ann_info.options = struct(...
+            'Display', 'off',...
+            'FunctionTolerance', 1e-6,...
+            'StepTolerance', 1e-6,...
+            'MaxIterations', 1e3,...
+            'MaxFunctionEvaluations', 10e3);
+        ann_info.x_value = struct(...
+            'x0', [0.0 0.0 0.0 0.0 0.0 0.0],...
+            'ub', [+20.0 +20.0 +20.0 +20.0 +20.0 +20.0],...
+            'lb', [-20.0 -20.0 -20.0 -20.0 -20.0 -20.0]);
+        ann_info.fct_fit = @fct_fit;
+        ann_info.fct_err = @fct_err_vec;
+    case 'matlab_ga'
+        ann_info.type = ann_type;
+        ann_info.options = struct(...
+            'Display', 'off',...
+            'TolFun', 1e-6,...
+            'ConstraintTolerance', 1e-3,...
+            'Generations', 40,...
+            'PopulationSize', 1000);
+        ann_info.x_value = struct(...
+            'n', 6,...
+            'ub', [+20.0 +20.0 +20.0 +20.0 +20.0 +20.0],...
+            'lb', [-20.0 -20.0 -20.0 -20.0 -20.0 -20.0]);
+        ann_info.fct_fit = @fct_fit;
+        ann_info.fct_err = @fct_err_sum;
+    otherwise
+        error('invalid data')
+end
+
+% assign
+ann_input.var_inp = var_inp;
+ann_input.var_out = var_out;
+ann_input.split_train_test = split_train_test;
+ann_input.split_var = split_var;
+ann_input.ann_info = ann_info;
+
+% tag_train
+tag_train = 'none';
+
+end
+
+function model = fct_model(tag_train, n_sol, n_inp, n_out)
+
+assert(ischar(tag_train), 'invalid output')
+assert(isfinite(n_sol), 'invalid input')
+assert(isfinite(n_inp), 'invalid input')
+assert(isfinite(n_out), 'invalid output')
+
+model = fitnet(4);
+model.trainParam.min_grad = 1e-8;
+model.trainParam.epochs = 300;
+model.trainParam.max_fail = 25;
+model.divideParam.trainRatio = 0.8;
+model.divideParam.valRatio = 0.2;
+model.divideParam.testRatio = 0.0;
+
+end
+
+function [model, history] = fct_train(tag_train, model, inp, out)
+
+assert(ischar(tag_train), 'invalid output')
+[model, history] = train(model, inp, out);
+
+end
+
+function out_mat_fit = fct_fit(tag_train, x, inp_mat)
+
+assert(ischar(tag_train), 'invalid output');
+
+x_1 = inp_mat(1, :);
+x_2 = inp_mat(2, :);
+
+y_1 = x(1)+x(2).*x_1+x(3).*x_2;
+y_2 = x(4)+x(5).*x_1+x(6).*x_2;
+
+out_mat_fit = [y_1 ; y_2];
+
+end
+
+function err_vec = fct_err_vec(tag_train, x, inp_mat, out_mat_ref)
+
+assert(ischar(tag_train), 'invalid output')
+
+out_mat_fit = fct_fit(tag_train, x, inp_mat);
+err_vec = out_mat_ref-out_mat_fit;
+err_vec = err_vec(:);
+
+end
+
+function err = fct_err_sum(tag_train, x, inp, out)
+
+assert(ischar(tag_train), 'invalid output')
+
+err_vec = fct_err_vec(tag_train, x, inp, out);
+err = sum(err_vec.^2);
+
+end

source_ann/ann_example/run_ann_example.m

@@ -1,61 +1,35 @@
 function run_ann_example()
 
 addpath('../ann_matlab');
+addpath('ann_data');
+close('all')
 
 % master_train
-master_train('matlab_ann')
-
+fprintf('AnnManager Example\n')
+fprintf('    1 - ANN regression with MATLAB Deep Learning\n')
+fprintf('    2 - ANN regression with Python Keras and TensorFlow\n')
+fprintf('    3 - MATLAB regression with nonlinear least-squares\n')
+fprintf('    4 - MATLAB regression with genetic algorithm\n')
+idx = input('Enter your choice >> ');
+
+choice = {'matlab_ann', 'python_ann', 'matlab_lsq', 'matlab_ga'};
+choice = get_choice(choice, idx);
+
+if isempty(choice)
+    fprintf('Invalid input\n')
+else
+    fprintf('\n')
+    get_ann_manager(choice)
 end
 
-function master_train(ann_type)
-
-% name
-fprintf('################## master_train\n')
-
-% data
-[ann_input, tag_train] = get_ann_param(ann_type);
-[n_sol, inp, out_ref, out_nrm] = get_ann_data();
-
-% test class
-fprintf('constructor\n')
-obj = AnnManager(ann_input);
-
-fprintf('train\n')
-obj.train(tag_train, n_sol, inp, out_ref, out_nrm);
-
-fprintf('get_fom\n')
-fom = obj.get_fom();
-assert(isstruct(fom), 'invalid fom')
-
-fprintf('disp\n')
-obj.disp();
-
-fprintf('dump\n')
-[ann_input, ann_data] = obj.dump();
-
-fprintf('delete\n')
-obj.delete();
-
-fprintf('predict\n')
-predict(ann_input, ann_data, n_sol, inp, out_nrm)
-
-fprintf('################## master_train\n')
-
 end
 
-function predict(ann_input, ann_data, n_sol, inp, out_nrm)
+function choice = get_choice(choice, idx)
 
-obj = AnnManager(ann_input);
-obj.load(ann_data);
-
-[is_valid_tmp, out_nrm_tmp] = obj.predict_nrm(n_sol, inp, out_nrm);
-assert(islogical(is_valid_tmp), 'invalid fom')
-assert(isstruct(out_nrm_tmp), 'invalid fom')
-
-[is_valid_tmp, out_nrm_tmp] = obj.predict_ann(n_sol, inp, out_nrm);
-assert(islogical(is_valid_tmp), 'invalid fom')
-assert(isstruct(out_nrm_tmp), 'invalid fom')
-
-obj.delete();
+if isnumeric(idx)&&(length(idx)==1)&&(idx>=1)&&(idx<=length(choice))
+    choice = choice{idx};
+else
+    choice = [];
+end
 
 end

source_ann/ann_example/run_ann_server.py

@@ -17,8 +17,6 @@ def fct_model(tag_train, n_sol, n_inp, n_out):
         keras.layers.Dense(64, activation='relu'),
         keras.layers.Dense(64, activation='relu'),
         keras.layers.Dense(64, activation='relu'),
-        keras.layers.Dense(64, activation='relu'),
-        keras.layers.Dense(64, activation='relu'),
         keras.layers.Dense(activation='linear', units=n_out),
     ])