Merge pull request #13 from matlab-deep-learning/bugfix/12

Fixes #12

code/mtcnn/+mtcnn/+util/DagNetworkStrategy.m

@@ -5,10 +5,16 @@
         Pnet
         Rnet
         Onet
+        ExecutionEnvironment
     end
 
     methods
-        function obj = DagNetworkStrategy()
+        function obj = DagNetworkStrategy(useGpu)
+            if useGpu
+                obj.ExecutionEnvironment = "gpu";
+            else
+                obj.ExecutionEnvironment = "cpu";
+            end
         end
 
         function load(obj)
@@ -18,19 +24,24 @@ function load(obj)
             obj.Onet = importdata(fullfile(mtcnnRoot(), "weights", "dagONet.mat"));
         end
 
-        function pnet = getPNet(obj)
-            pnet = obj.Pnet;
+        function [probability, correction] = applyPNet(obj, im)
+            % need to use activations as we don't know what size it will be
+            result = obj.Pnet.activations(im, "concat", ...
+                "ExecutionEnvironment", obj.ExecutionEnvironment);
+
+            probability = result(:,:,1:2,:);
+            correction = result(:,:,3:end,:);
         end
 
         function [probs, correction] = applyRNet(obj, im)
-            output = obj.Rnet.predict(im);
+            output = obj.Rnet.predict(im, "ExecutionEnvironment", obj.ExecutionEnvironment);
 
             probs = output(:,1:2);
             correction = output(:,3:end);
         end
 
         function [probs, correction, landmarks] = applyONet(obj, im)
-            output = obj.Onet.predict(im);
+            output = obj.Onet.predict(im, "ExecutionEnvironment", obj.ExecutionEnvironment);
 
             probs = output(:,1:2);
             correction = output(:,3:6);

code/mtcnn/+mtcnn/+util/DlNetworkStrategy.m

@@ -26,8 +26,13 @@ function load(obj)
             end
         end
 
-        function pnet = getPNet(obj)
-            pnet = obj.PnetWeights;
+        function [probability, correction] = applyPNet(obj, im)
+            im = dlarray(im, "SSCB");
+
+            [probability, correction] = mtcnn.pnet(im, obj.PnetWeights);
+
+            probability = extractdata(gather(probability));
+            correction = extractdata(gather(correction));
         end
 
         function [probs, correction] = applyRNet(obj, im)

code/mtcnn/+mtcnn/Detector.m

@@ -49,7 +49,7 @@
             end
 
             if obj.UseDagNet
-                obj.Networks = mtcnn.util.DagNetworkStrategy();
+                obj.Networks = mtcnn.util.DagNetworkStrategy(obj.UseGPU);
             else
                 obj.Networks = mtcnn.util.DlNetworkStrategy(obj.UseGPU);
             end
@@ -89,7 +89,7 @@
                 [thisBox, thisScore] = ...
                     mtcnn.proposeRegions(im, scale, ...
                                             obj.ConfidenceThresholds(1), ...
-                                            obj.Networks.getPNet());
+                                            obj.Networks);
                 bboxes = cat(1, bboxes, thisBox);
                 scores = cat(1, scores, thisScore);
             end
@@ -183,7 +183,7 @@
                         "Input image is of unsupported type '%s'", class(im));
             end
 
-            if obj.UseGPU()
+            if obj.UseGPU && ~obj.UseDagNet
                 outIm = gpuArray(outIm);
             end
 

code/mtcnn/+mtcnn/proposeRegions.m

@@ -1,4 +1,4 @@
-function [bboxes, scores] = proposeRegions(im, scale, threshold, weightsOrNet)
+function [bboxes, scores] = proposeRegions(im, scale, threshold, networkStrategy)
 % proposeRegions    Generate region proposals at a given scale.
 %
 % Args:
@@ -13,34 +13,15 @@
 
 % Copyright 2019 The MathWorks, Inc.
 
-    useDagNet = isa(weightsOrNet, "DAGNetwork");
-    if isa(im, "gpuArray")
-        imClass = classUnderlying(im);
-    else
-        imClass = class(im);
-    end
-    assert(imClass == "single", "mtcnn:proposeRegions:wrongImageType", ...
-        "Input image should be a single scale -1 to 1");
 
     % Stride of the proposal network
     stride = 2;
     % Field of view of the proposal network in pixels
     pnetSize = 12;
 
     im = imresize(im, 1/scale);
-
-    if useDagNet
-        % need to use activations as we don't know what size it will be
-        result = weightsOrNet.activations(im, "concat");
-        probability = gather(result(:,:,1:2,:));
-        correction = gather(result(:,:,3:end,:));
-    else
-        im = dlarray(im, "SSCB");
-        [probability, correction] = mtcnn.pnet(im, weightsOrNet);
-        probability = extractdata(gather(probability));
-        correction = extractdata(gather(correction));
-    end
-
+
+    [probability, correction] = networkStrategy.applyPNet(im);
 
     faces = probability(:,:,2) > threshold;
     if sum(faces, 'all') == 0

test/+tests/DetectorTest.m

@@ -112,12 +112,14 @@ function testNmsThresholds(test)
         end
 
         %% GPU
-        function testGpuDetect(test)
+        function testGpuDetect(test, imageTypeConversion, useDagNet)
+
             % filter if no GPU present
             test.assumeGreaterThan(gpuDeviceCount, 0);
 
-            detector = mtcnn.Detector("UseGPU", true);
-            [bboxes, scores, landmarks] = detector.detect(test.Image);
+            inputImage = imageTypeConversion(test.Image);
+            detector = mtcnn.Detector("UseGPU", true, "UseDagNet", useDagNet);
+            [bboxes, scores, landmarks] = detector.detect(inputImage);
 
             test.verifyEqual(size(bboxes), [6, 4]);
             test.verifyEqual(size(scores), [6, 1]);

test/+tests/ProposeRegionsTest.m

@@ -8,17 +8,18 @@
     end
 
     properties (TestParameter)
-        getNet = struct("weights", @() load(fullfile(mtcnnRoot, "weights", "pnet.mat")), ...
-            "net", @() importdata(fullfile(mtcnnRoot, "weights", "dagPNet.mat")));
+        getNet = struct("dl", @() mtcnn.util.DlNetworkStrategy(false) , ...
+                        "dag", @() mtcnn.util.DagNetworkStrategy(false));
     end
 
     methods (Test)
         function testOutputs(test, getNet)
             scale = 2;
             conf = 0.5;
-            weights = getNet();
+            strategy = getNet();
+            strategy.load();
 
-            [box, score] = mtcnn.proposeRegions(test.Image, scale, conf, weights);
+            [box, score] = mtcnn.proposeRegions(test.Image, scale, conf, strategy);
 
             test.verifyOutputs(box, score);
         end
@@ -29,9 +30,10 @@ function test1DActivations(test, getNet)
             cropped = imcrop(test.Image, [300, 42, 65, 38]);
             scale = 3;
             conf = 0.5;
-            weights = getNet();
+            strategy = getNet();
+            strategy.load();
 
-            [box, score] = mtcnn.proposeRegions(cropped, scale, conf, weights);
+            [box, score] = mtcnn.proposeRegions(cropped, scale, conf, strategy);
 
             test.verifyOutputs(box, score);
         end