Fix resnet50v2_nhwc model (webmachinelearning#162)

* Fix resnet50v2_nhwc model

1. Change 1x1 filter to 2x2 for `maxPool2d` [1].
2. Merge `pad` into `conv2d`.
3. Replace `reduceMean` for axis 1, 2 with `averagePool2d`.

[1]: https://github.com/google/XNNPACK/issues/472s

* Fix lint error

image_classification/resnet50v2_nhwc.js

@@ -83,12 +83,9 @@ export class ResNet50V2Nhwc {
     }
     if (!downsample && shortcut) {
       residual = this.builder_.maxPool2d(
-          input, {windowDimensions: [1, 1], strides, layout, autoPad});
-      const padding = this.builder_.constant(
-          {type: 'int32', dimensions: [4, 2]},
-          new Int32Array([0, 0, 1, 1, 1, 1, 0, 0]));
-      const pad = this.builder_.pad(conv1, padding);
-      conv2 = await this.buildConv_(pad, nameIndices.concat(['2']), {strides});
+          input, {windowDimensions: [2, 2], strides, layout, autoPad});
+      conv2 = await this.buildConv_(
+          conv1, nameIndices.concat(['2']), {strides, padding: [1, 1, 1, 1]});
     } else {
       conv2 = await this.buildConv_(
           conv1, nameIndices.concat(['2']), {autoPad});
@@ -101,14 +98,10 @@ export class ResNet50V2Nhwc {
   async load(contextOptions) {
     this.context_ = await navigator.ml.createContext(contextOptions);
     this.builder_ = new MLGraphBuilder(this.context_);
-    const padding = this.builder_.constant(
-        {type: 'int32', dimensions: [4, 2]},
-        new Int32Array([0, 0, 3, 3, 3, 3, 0, 0]));
-
     const input = this.builder_.input('input',
         {type: 'float32', dimensions: this.inputOptions.inputDimensions});
-    const pad = this.builder_.pad(input, padding);
-    const conv1 = await this.buildConv_(pad, ['', '', '1'], {strides}, false);
+    const conv1 = await this.buildConv_(
+        input, ['', '', '1'], {strides, padding: [3, 3, 3, 3]}, false);
     const pool = this.builder_.maxPool2d(
         conv1, {windowDimensions: [3, 3], strides, layout, autoPad});
     // Block 1
@@ -155,8 +148,7 @@ export class ResNet50V2Nhwc {
 
     const fusedBn =
         await this.buildFusedBatchNorm_(bottleneck13, ['postnorm']);
-    const mean = this.builder_.reduceMean(
-        fusedBn, {keepDimensions: true, axes: [1, 2]});
+    const mean = this.builder_.averagePool2d(fusedBn, {layout});
     const conv2 = await this.buildConv_(
         mean, ['', '', 'logits'], {autoPad}, false);
     const reshape = this.builder_.reshape(conv2, [1, null]);