last fixes for architecture

src/faceDetectionNet/index.ts

@@ -67,7 +67,6 @@ export function faceDetectionNet(weights: Float32Array) {
     })
   }
 
-  // TODO debug output
   function forward(input: ImageData|ImageData[]|number[]) {
     return tf.tidy(
       () => forwardTensor(getImgTensor(input))
@@ -100,10 +99,10 @@ export function faceDetectionNet(weights: Float32Array) {
       .map(({ score, idx }) => ({
         score,
         box: {
-          left: Math.max(0, width * boxes.get(idx, 0)),
-          right: Math.min(width, width * boxes.get(idx, 1)),
-          top: Math.max(0, height * boxes.get(idx, 2)),
-          bottom: Math.min(height, height * boxes.get(idx, 3))
+          top: Math.max(0, height * boxes.get(idx, 0)),
+          left: Math.max(0, width * boxes.get(idx, 1)),
+          bottom: Math.min(height, height * boxes.get(idx, 2)),
+          right: Math.min(width, width * boxes.get(idx, 3))
         }
       }))
 

src/faceDetectionNet/mobileNetV1.ts

@@ -21,7 +21,7 @@ function depthwiseConvLayer(
       params.batch_norm_scale,
       params.batch_norm_offset
     )
-    return tf.relu(out)
+    return tf.clipByValue(out, 0, 6)
 
   })
 }
@@ -35,6 +35,7 @@ export function mobileNetV1(x: tf.Tensor4D, params: FaceDetectionNet.MobileNetV1
 
     let conv11 = null
     let out = pointwiseConvLayer(x, params.conv_0_params, [2, 2])
+
     params.conv_pair_params.forEach((param, i) => {
       const layerIdx = i + 1
       const depthwiseConvStrides = getStridesForLayerIdx(layerIdx)

src/faceDetectionNet/outputLayer.ts

@@ -30,17 +30,17 @@ function decodeBoxesLayer(x0: tf.Tensor2D, x1: tf.Tensor2D) {
   const vec = tf.unstack(tf.transpose(x1, [1, 0]))
 
   const div0_out = tf.div(tf.mul(tf.exp(tf.div(vec[2], tf.scalar(5))), sizes[0]), tf.scalar(2))
-  const add0_out = tf.add(tf.mul(tf.exp(tf.div(vec[0], tf.scalar(10))), sizes[0]), centers[0])
+  const add0_out = tf.add(tf.mul(tf.div(vec[0], tf.scalar(10)), sizes[0]), centers[0])
 
   const div1_out = tf.div(tf.mul(tf.exp(tf.div(vec[3], tf.scalar(5))), sizes[1]), tf.scalar(2))
-  const add1_out = tf.add(tf.mul(tf.exp(tf.div(vec[1], tf.scalar(10))), sizes[1]), centers[1])
+  const add1_out = tf.add(tf.mul(tf.div(vec[1], tf.scalar(10)), sizes[1]), centers[1])
 
   return tf.transpose(
     tf.stack([
-      tf.sub(div0_out, add0_out),
-      tf.sub(div1_out, add1_out),
-      tf.add(div0_out, add0_out),
-      tf.add(div1_out, add1_out)
+      tf.sub(add0_out, div0_out),
+      tf.sub(add1_out, div1_out),
+      tf.add(add0_out, div0_out),
+      tf.add(add1_out, div1_out)
     ]),
     [1, 0]
   )

src/faceDetectionNet/pointwiseConvLayer.ts

@@ -11,7 +11,7 @@ export function pointwiseConvLayer(
 
     let out = tf.conv2d(x, params.filters, strides, 'same')
     out = tf.add(out, params.batch_norm_offset)
-    return tf.relu(out)
+    return tf.clipByValue(out, 0, 6)
 
   })
 }
\ No newline at end of file