[wasm] Fix bug in CropAndResize, and add AutoML models to benchmark

tfjs-backend-wasm/src/cc/kernels/CropAndResize.cc

@@ -34,7 +34,7 @@ namespace {
 void interpolate_nearest(float* out_buf_ptr, const float* images_buf,
                          std::vector<int> images_strides, int crop_width,
                          int image_width, int image_width_m1, int num_channels,
-                         float extrapolation_value, int box_ind, float y_ind,
+                         float extrapolation_value, int box_offset, float y_ind,
                          float width_scale, float x1, float x2) {
   for (int x = 0; x < crop_width; ++x) {
     const float x_ind = (crop_width > 1) ? x1 * image_width_m1 + x * width_scale
@@ -52,7 +52,7 @@ void interpolate_nearest(float* out_buf_ptr, const float* images_buf,
     float closest_y = round(y_ind);
     for (int c = 0; c < num_channels; ++c) {
       const int in_ind = c + closest_x * images_strides[2] +
-                         closest_y * images_strides[1] + box_ind;
+                         closest_y * images_strides[1] + box_offset;
       *out_buf_ptr = images_buf[in_ind];
       out_buf_ptr++;
     }
@@ -123,7 +123,7 @@ void CropAndResize(int images_id, int boxes_id, int box_ind_id, int num_boxes,
       continue;
     }
 
-    const int box_ind = *box_ind_buf * images_strides[0];
+    const int box_offset = *box_ind_buf * images_strides[0];
 
     const float height_scale =
         (crop_height > 1) ? (y2 - y1) * image_height_m1 / (crop_height - 1) : 0;
@@ -175,22 +175,22 @@ void CropAndResize(int images_id, int boxes_id, int box_ind_id, int num_boxes,
 
       if (should_memcpy) {
         int y_ind_int = y_ind;
-        images_buf += (y_ind_int * images_strides[1] + box_ind);
-
-        memcpy(out_buf_ptr, images_buf, sizeof(float) * crop_width);
+        int offset = box_offset + y_ind_int * images_strides[1];
+        memcpy(out_buf_ptr, images_buf + offset,
+               sizeof(float) * crop_width * num_channels);
         continue;
       }
 
       if (method == InterpolationMethod::BILINEAR) {
         tfjs::wasm::interpolate_bilinear(
             out_buf_ptr, images_buf, images_strides, crop_width, image_width,
             image_width_m1, image_height_m1, num_channels, should_extrapolate,
-            extrapolation_value, box_ind, y_ind, width_scale, x1, x2);
+            extrapolation_value, box_offset, y_ind, width_scale, x1, x2);
 
       } else {
         interpolate_nearest(out_buf_ptr, images_buf, images_strides, crop_width,
                             image_width, image_width_m1, num_channels,
-                            extrapolation_value, box_ind, y_ind, width_scale,
+                            extrapolation_value, box_offset, y_ind, width_scale,
                             x1, x2);
       }
     }

tfjs-core/benchmarks/index.html

@@ -88,9 +88,11 @@ <h2>TensorFlow.js Model Benchmark</h2>
   <script src="https://unpkg.com/@tensorflow/tfjs-layers@latest/dist/tf-layers.js"></script>
   <script src="https://unpkg.com/@tensorflow/tfjs-converter@latest/dist/tf-converter.js"></script>
   <script src="https://unpkg.com/@tensorflow/tfjs-backend-wasm@latest/dist/tf-backend-wasm.js"></script>
+  <script src="https://unpkg.com/@tensorflow/tfjs-automl@latest/dist/tf-automl.js"></script>
   <script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/universal-sentence-encoder"></script>
   <script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/posenet@2"></script>
   <script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/body-pix@2"></script>
+
   <script src="./modelConfig.js"></script>
   <script src="./util.js"></script>
   <script>
@@ -214,7 +216,7 @@ <h2>TensorFlow.js Model Benchmark</h2>
         const start = performance.now();
         let res = predict(model);
         if (res instanceof Promise) {
-          await res;
+          res = await res;
         }
 
         if (res instanceof tf.Tensor) {

tfjs-core/benchmarks/modelConfig.js

@@ -74,7 +74,8 @@ const sentences = [
 const benchmarks = {
   'mobilenet_v2': {
     load: async () => {
-      const url = 'https://storage.googleapis.com/learnjs-data/mobilenet_v2_100_fused/model.json';
+      const url =
+          'https://storage.googleapis.com/learnjs-data/mobilenet_v2_100_fused/model.json';
       return tf.loadGraphModel(url);
     },
     predictFunc: () => {
@@ -97,7 +98,8 @@ const benchmarks = {
   },
   'face_detector': {
     load: async () => {
-      const url = 'https://storage.googleapis.com/learnjs-data/face_detector_front/model.json';
+      const url =
+          'https://storage.googleapis.com/learnjs-data/face_detector_front/model.json';
       return tf.loadGraphModel(url);
     },
     predictFunc: () => {
@@ -107,6 +109,28 @@ const benchmarks = {
       };
     },
   },
+  'AutoML Image': {
+    load: async () => {
+      const url =
+          'https://storage.googleapis.com/tfjs-testing/tfjs-automl/img_classification/model.json';
+      return tf.automl.loadImageClassification(url);
+    },
+    predictFunc: () => {
+      const zeros = tf.zeros([224, 224, 3]);
+      return model => model.classify(zeros);
+    }
+  },
+  'AutoML Object': {
+    load: async () => {
+      const url =
+          'https://storage.googleapis.com/tfjs-testing/tfjs-automl/object_detection/model.json';
+      return tf.automl.loadObjectDetection(url);
+    },
+    predictFunc: () => {
+      const zeros = tf.zeros([224, 224, 3]);
+      return model => model.detect(zeros);
+    }
+  },
   'USE - batchsize 30': {
     load: async () => {
       return use.load();
@@ -160,7 +184,8 @@ const benchmarks = {
   },
 };
 
-const imageBucket = 'https://storage.googleapis.com/tfjs-models/assets/posenet/';
+const imageBucket =
+    'https://storage.googleapis.com/tfjs-models/assets/posenet/';
 async function loadImage(imagePath) {
   const image = new Image();
   const promise = new Promise((resolve, reject) => {

tfjs-core/src/ops/image_ops_test.ts

@@ -229,6 +229,33 @@ describeWithFlags('cropAndResize', ALL_ENVS, () => {
     expect(output.dtype).toBe('float32');
     expectArraysClose(await output.data(), [2.5]);
   });
+
+  it('5x5-bilinear, no change in shape', async () => {
+    const image: tf.Tensor4D = tf.ones([1, 5, 5, 3]);
+    const boxes: tf.Tensor2D = tf.tensor2d([0, 0, 1, 1], [1, 4]);
+    const boxInd: tf.Tensor1D = tf.tensor1d([0], 'int32');
+
+    const output =
+        tf.image.cropAndResize(image, boxes, boxInd, [5, 5], 'bilinear', 0);
+
+    expect(output.shape).toEqual([1, 5, 5, 3]);
+    expect(output.dtype).toBe('float32');
+    expectArraysClose(await output.data(), await image.data());
+  });
+
+  it('5x5-bilinear, just a crop, no resize', async () => {
+    const image: tf.Tensor4D = tf.ones([1, 6, 6, 3]);
+    const boxes: tf.Tensor2D = tf.tensor2d([0.5, 0.5, 1, 1], [1, 4]);
+    const boxInd: tf.Tensor1D = tf.tensor1d([0], 'int32');
+
+    const output =
+        tf.image.cropAndResize(image, boxes, boxInd, [3, 3], 'bilinear', 0);
+
+    expect(output.shape).toEqual([1, 3, 3, 3]);
+    expect(output.dtype).toBe('float32');
+    expectArraysClose(await output.data(), await tf.ones([1, 3, 3, 3]).data());
+  });
+
   it('1x1-nearest', async () => {
     const image: tf.Tensor4D = tf.tensor4d([1, 2, 3, 4], [1, 2, 2, 1]);
     const boxes: tf.Tensor2D = tf.tensor2d([0, 0, 1, 1], [1, 4]);

tfjs-core/src/ops/resize_bilinear_test.ts

@@ -28,6 +28,17 @@ describeWithFlags('resizeBilinear', ALL_ENVS, () => {
         await output.data(), [2, 2, 2, 10 / 3, 10 / 3, 10 / 3, 4, 4, 4]);
   });
 
+  it('5x5-bilinear, no change in shape', async () => {
+    const image: tf.Tensor4D = tf.ones([1, 5, 5, 3]);
+
+    const alignCorners = false;
+    const output = tf.image.resizeBilinear(image, [5, 5], alignCorners);
+
+    expect(output.shape).toEqual([1, 5, 5, 3]);
+    expect(output.dtype).toBe('float32');
+    expectArraysClose(await output.data(), await image.data());
+  });
+
   it('simple alignCorners=true', async () => {
     const input = tf.tensor3d([2, 2, 4, 4], [2, 2, 1]);
     const output = input.resizeBilinear([3, 3], true);