Add test for global pooling

code/nnv/engine/nn/layers/Conv2DLayer.m

@@ -519,19 +519,19 @@ function set_name(obj, name)
             end
 
             % compute output sets
-            if ~isa(input.V, 'gpuArray')
-                c = vl_nnconv(input.V(:,:,:,1), obj.Weights, obj.Bias, 'Stride', obj.Stride, 'Pad', obj.PaddingSize, 'Dilate', obj.DilationFactor);
-                V = vl_nnconv(input.V(:,:,:,2:input.numPred + 1), obj.Weights, [], 'Stride', obj.Stride, 'Pad', obj.PaddingSize, 'Dilate', obj.DilationFactor);   
+            % if ~isa(input.V, 'gpuArray')
+                % c = vl_nnconv(input.V(:,:,:,1), obj.Weights, obj.Bias, 'Stride', obj.Stride, 'Pad', obj.PaddingSize, 'Dilate', obj.DilationFactor);
+                % V = vl_nnconv(input.V(:,:,:,2:input.numPred + 1), obj.Weights, [], 'Stride', obj.Stride, 'Pad', obj.PaddingSize, 'Dilate', obj.DilationFactor);   
                 % c = dlconv(dlarray(input.V(:,:,:,1), "SSC"), obj.Weights, obj.Bias, 'Stride', obj.Stride, 'Padding', obj.PaddingSize, 'DilationFactor', obj.DilationFactor);
                 % V = dlconv(dlarray(input.V(:,:,:,2:input.numPred + 1), "SSCB"), obj.Weights, 0, 'Stride', obj.Stride, 'Padding', obj.PaddingSize, 'DilationFactor', obj.DilationFactor);
                 % c = extractdata(c);
                 % V = extractdata(V);
-            else
-                c = dlconv(dlarray(input.V(:,:,:,1), "SSC"), obj.Weights, obj.Bias, 'Stride', obj.Stride, 'Padding', obj.PaddingSize, 'DilationFactor', obj.DilationFactor);
-                V = dlconv(dlarray(input.V(:,:,:,2:input.numPred + 1), "SSCB"), obj.Weights, 0, 'Stride', obj.Stride, 'Padding', obj.PaddingSize, 'DilationFactor', obj.DilationFactor);
-                c = extractdata(c);
-                V = extractdata(V);
-            end
+            % else
+            c = dlconv(dlarray(input.V(:,:,:,1), "SSC"), obj.Weights, obj.Bias, 'Stride', obj.Stride, 'Padding', obj.PaddingSize, 'DilationFactor', obj.DilationFactor);
+            V = dlconv(dlarray(input.V(:,:,:,2:input.numPred + 1), "SSCB"), obj.Weights, 0, 'Stride', obj.Stride, 'Padding', obj.PaddingSize, 'DilationFactor', obj.DilationFactor);
+            c = extractdata(c);
+            V = extractdata(V);
+            % end
             Y = cat(4, c, V);
             S = ImageStar(Y, input.C, input.d, input.pred_lb, input.pred_ub);
 

code/nnv/engine/nn/layers/GlobalAveragePooling2DLayer.m

@@ -38,8 +38,14 @@
                     numOutputs = 1;
                     inputNames = {'in1'};
                     outputNames = {'out'};
+                case 0
+                    name = 'global_average_pooling_2d';
+                    numInputs = 1;
+                    numOutputs = 1;
+                    inputNames = {'in1'};
+                    outputNames = {'out'};
                 otherwise
-                    error('Invalid number of input arguments, should be 1 or 5');        
+                    error('Invalid number of input arguments, should be 0, 1 or 5');        
             end
 
             if ~ischar(name)

code/nnv/engine/nn/layers/TransposedConv2DLayer.m

@@ -255,13 +255,13 @@
             end
 
             % compute output sets
-            if isa(input.V, 'gpuArray')
-                c = extractdata(dltranspconv(dlarray(input.V(:,:,:,1)), obj.Weights, obj.Bias, "Stride", obj.Stride, "Cropping", obj.CroppingSize,"DataFormat",'SSCU'));
-                V = extractdata(dltranspconv(dlarray(input.V(:,:,:,2:input.numPred + 1)), obj.Weights, 0, "Stride", obj.Stride, "Cropping", obj.CroppingSize,"DataFormat",'SSCU'));
-            else
-                c = vl_nnconvt(input.V(:,:,:,1), obj.Weights, obj.Bias, 'Upsample', obj.Stride, 'Crop', obj.CroppingSize);
-                V = vl_nnconvt(input.V(:,:,:,2:input.numPred + 1), obj.Weights, [], 'Upsample', obj.Stride, 'Crop', obj.CroppingSize);
-            end
+            % if isa(input.V, 'gpuArray')
+            c = extractdata(dltranspconv(dlarray(input.V(:,:,:,1)), obj.Weights, obj.Bias, "Stride", obj.Stride, "Cropping", obj.CroppingSize,"DataFormat",'SSCU'));
+            V = extractdata(dltranspconv(dlarray(input.V(:,:,:,2:input.numPred + 1)), obj.Weights, 0, "Stride", obj.Stride, "Cropping", obj.CroppingSize,"DataFormat",'SSCU'));
+            % else
+                % c = vl_nnconvt(input.V(:,:,:,1), obj.Weights, obj.Bias, 'Upsample', obj.Stride, 'Crop', obj.CroppingSize);
+                % V = vl_nnconvt(input.V(:,:,:,2:input.numPred + 1), obj.Weights, [], 'Upsample', obj.Stride, 'Crop', obj.CroppingSize);
+            % end
             Y = cat(4, c, V);
             S = ImageStar(Y, input.C, input.d, input.pred_lb, input.pred_ub);
 

code/nnv/tests/nn/layers/GlobalAveragePooling2DLayer/test_GlobalPooling2D.m

@@ -0,0 +1,128 @@
+
+
+%% test 1: constructor
+L1 = GlobalAveragePooling2DLayer();
+
+
+%% test 2: inference
+L1 = GlobalAveragePooling2DLayer();
+x = load('one_image.mat');
+x = x.one_image;
+L1.evaluate(x);
+
+
+%% test 3: equivalence (inference)
+L1 = GlobalAveragePooling2DLayer();
+x = load('one_image.mat');
+x = x.one_image;
+y = L1.evaluate(x);
+
+dlX = dlarray(x, 'SSBC');
+dlY = avgpool(dlX,'global');
+
+assert(all(dlY == y, 'all'));
+
+%% test 4: inference, higher dimension
+
+miniBatchSize = 10;
+inputSize = [5 5];
+numChannels = 3;
+X = rand(inputSize(1),inputSize(2),numChannels,miniBatchSize);
+
+L1 = GlobalAveragePooling2DLayer();
+Y = L1.evaluate(X);
+
+dlX = dlarray(X,'SSCB');
+dlY = avgpool(dlX,'global');
+dlY = extractdata(dlY);
+
+assert(all(dlY == Y, 'all'));
+
+%% test 5: reachability
+
+x = load('one_image.mat');
+X = x.one_image;
+
+lb = X - 0.1;
+ub = X + 0.1;
+IS = ImageStar(lb,ub);
+
+L1 = GlobalAveragePooling2DLayer();
+Y = L1.evaluate(X);
+Yset = L1.reach(IS,'approx-star');
+
+[LB,UB] = Yset.estimateRanges;
+
+assert(all(LB <= Y,'all'))
+assert(all(UB >= Y,'all'))
+
+%% test 6: reach (sound)
+
+N = 100; % random samples
+
+x = load('one_image.mat');
+X = x.one_image;
+
+lb = X - 0.1;
+ub = X + 0.1;
+IS = ImageStar(lb,ub);
+x_samples = IS.sample(N);
+
+L1 = GlobalAveragePooling2DLayer();
+Yset = L1.reach(IS,'approx-star');
+[LB,UB] = Yset.estimateRanges;
+
+for i=1:N
+    xi = x_samples{i};
+    Yi = L1.evaluate(xi);
+    assert(all(LB <= Yi,'all'))
+    assert(all(UB >= Yi,'all'))
+end
+
+
+%% test 7: reachability
+
+miniBatchSize = 1;
+inputSize = [5 5];
+numChannels = 3;
+X = rand(inputSize(1),inputSize(2),numChannels,miniBatchSize);
+
+lb = X - 0.1;
+ub = X + 0.1;
+IS = ImageStar(lb,ub);
+
+L1 = GlobalAveragePooling2DLayer();
+Y = L1.evaluate(X);
+Yset = L1.reach(IS,'approx-star');
+
+[LB,UB] = Yset.estimateRanges;
+
+assert(all(LB <= Y,'all'))
+assert(all(UB >= Y,'all'))
+
+%% test 8: reach (sound)
+
+N = 200; % random samples
+
+miniBatchSize = 1;
+inputSize = [5 5];
+numChannels = 3;
+X = rand(inputSize(1),inputSize(2),numChannels,miniBatchSize);
+
+lb = X - 0.1;
+ub = X + 0.1;
+IS = ImageStar(lb,ub);
+
+x_samples = IS.sample(N);
+
+L1 = GlobalAveragePooling2DLayer();
+Yset = L1.reach(IS,'approx-star');
+[LB,UB] = Yset.estimateRanges;
+
+for i=1:N
+    xi = x_samples{i};
+    Yi = L1.evaluate(xi);
+    assert(all(LB <= Yi,'all'))
+    assert(all(UB >= Yi,'all'))
+end
+