Merge pull request #143 from Rjasuja/float16_resize_1_3

Add Float16 support for Resize Bilinear and Resize Nearest Neighbor op

Author: anikulk

ngraph_creator/operations/src/ResizeBilinear.cpp

@@ -12,18 +12,6 @@ ResizeBilinear::ResizeBilinear(int operationIndex) : OperationsBase(operationInd
 }
 
 bool ResizeBilinear::validate() {
-    // TODO Add FLOAT16 check when VPUX plugin is supported
-    if (!checkOutputOperandType(0, (int32_t)OperandType::TENSOR_FLOAT32) &&
-        !checkOutputOperandType(0, (int32_t)OperandType::TENSOR_QUANT8_ASYMM)) {
-        ALOGE("%s check for output types failed", __func__);
-        return false;
-    }
-
-    if (!checkInputOperandType(0, (int32_t)OperandType::TENSOR_FLOAT32) &&
-        !checkInputOperandType(0, (int32_t)OperandType::TENSOR_QUANT8_ASYMM)) {
-        return false;
-    }
-
     const auto& inputDimensionsSize = getInputOperandDimensions(0).size();
     if (inputDimensionsSize != 4) {
         ALOGE("%s Invalid dimensions size for input(%lu)", __func__, inputDimensionsSize);
@@ -74,8 +62,7 @@ std::shared_ptr<ngraph::Node> ResizeBilinear::createNode() {
 
     if (!useNchw) inputNode = transpose(NHWC_NCHW, inputNode);
     // FLOAT16 type check added for future when VPUX plugin support is added
-    if (checkInputOperandType(1, (int32_t)OperandType::FLOAT32) ||
-        checkInputOperandType(1, (int32_t)OperandType::FLOAT16)) {
+    if (checkInputOperandType(1, (int32_t)OperandType::FLOAT32)) {
         // In tensorflow lite, resizing by size is supported. Scaling factors are
         // calculated based on output shape.
         attrs.shape_calculation_mode = ngraph::op::v4::Interpolate::ShapeCalcMode::sizes;
@@ -87,6 +74,14 @@ std::shared_ptr<ngraph::Node> ResizeBilinear::createNode() {
         // integer
         width_scale = (float)out_width / (float)input_width;
         height_scale = (float)out_height / (float)input_height;
+    } else if (checkInputOperandType(1, (int32_t)OperandType::FLOAT16)) {
+        attrs.shape_calculation_mode = ngraph::op::v4::Interpolate::ShapeCalcMode::sizes;
+        width_scale = sModelInfo->ParseOperationInput<_Float16>(mNnapiOperationIndex, 1);
+        height_scale = sModelInfo->ParseOperationInput<_Float16>(mNnapiOperationIndex, 2);
+        out_width = (int)(input_width * width_scale);
+        out_height = (int)(input_height * height_scale);
+        width_scale = (float)out_width / (float)input_width;
+        height_scale = (float)out_height / (float)input_height;
     } else if (checkInputOperandType(1, (int32_t)OperandType::INT32)) {
         attrs.shape_calculation_mode = ngraph::op::v4::Interpolate::ShapeCalcMode::sizes;
         out_width = sModelInfo->ParseOperationInput<int>(mNnapiOperationIndex, 1);

ngraph_creator/operations/src/ResizeNearestNeighbor.cpp

@@ -12,18 +12,6 @@ ResizeNearestNeighbor::ResizeNearestNeighbor(int operationIndex) : OperationsBas
 }
 
 bool ResizeNearestNeighbor::validate() {
-    // TODO Add FLOAT16 check when VPUX plugin is supported
-    if (!checkOutputOperandType(0, (int32_t)OperandType::TENSOR_FLOAT32) &&
-        !checkOutputOperandType(0, (int32_t)OperandType::TENSOR_QUANT8_ASYMM)) {
-        ALOGE("%s check for output types failed", __func__);
-        return false;
-    }
-
-    if (!checkInputOperandType(0, (int32_t)OperandType::TENSOR_FLOAT32) &&
-        !checkInputOperandType(0, (int32_t)OperandType::TENSOR_QUANT8_ASYMM)) {
-        return false;
-    }
-
     const auto& inputDimensionsSize = getInputOperandDimensions(0).size();
     if (inputDimensionsSize != 4) {
         ALOGE("%s Invalid dimensions size for input(%lu)", __func__, inputDimensionsSize);
@@ -73,45 +61,46 @@ std::shared_ptr<ngraph::Node> ResizeNearestNeighbor::createNode() {
     }
 
     if (!useNchw) inputNode = transpose(NHWC_NCHW, inputNode);
-    // FLOAT16 type check added for future when VPUX plugin support is added
-    if (checkInputOperandType(1, (int32_t)OperandType::FLOAT32) ||
-        checkInputOperandType(1, (int32_t)OperandType::FLOAT16)) {
+
+    attrs.shape_calculation_mode = ngraph::op::v4::Interpolate::ShapeCalcMode::sizes;
+    // mode is passed as "nearest" for Nearest Neighbor interpolation
+    attrs.mode = ngraph::op::v4::Interpolate::InterpolateMode::nearest;
+    attrs.nearest_mode = ngraph::op::v4::Interpolate::NearestMode::floor;
+
+    if (checkInputOperandType(1, (int32_t)OperandType::FLOAT32)) {
         // In tensorflow lite, resizing by size is supported. Scaling factors are
         // calculated based on output shape.
-        attrs.shape_calculation_mode = ngraph::op::v4::Interpolate::ShapeCalcMode::sizes;
         width_scale = sModelInfo->ParseOperationInput<float>(mNnapiOperationIndex, 1);
         height_scale = sModelInfo->ParseOperationInput<float>(mNnapiOperationIndex, 2);
         out_width = (int)(input_width * width_scale);
         out_height = (int)(input_height * height_scale);
-        // Recalculating scaling factors here because of typecasting output shape to
-        // integer
-        width_scale = (float)out_width / (float)input_width;
-        height_scale = (float)out_height / (float)input_height;
+    } else if (checkInputOperandType(1, (int32_t)OperandType::FLOAT16)) {
+        width_scale = sModelInfo->ParseOperationInput<_Float16>(mNnapiOperationIndex, 1);
+        height_scale = sModelInfo->ParseOperationInput<_Float16>(mNnapiOperationIndex, 2);
+        out_width = (int)(input_width * width_scale);
+        out_height = (int)(input_height * height_scale);
     } else if (checkInputOperandType(1, (int32_t)OperandType::INT32)) {
-        attrs.shape_calculation_mode = ngraph::op::v4::Interpolate::ShapeCalcMode::sizes;
         out_width = sModelInfo->ParseOperationInput<int>(mNnapiOperationIndex, 1);
         out_height = sModelInfo->ParseOperationInput<int>(mNnapiOperationIndex, 2);
-        width_scale = (float)out_width / (float)input_width;
-        height_scale = (float)out_height / (float)input_height;
     }
+    width_scale = (float)out_width / (float)input_width;
+    height_scale = (float)out_height / (float)input_height;
 
     if (align_corners == true) {
         attrs.coordinate_transformation_mode =
             ngraph::op::v4::Interpolate::CoordinateTransformMode::align_corners;
+        attrs.nearest_mode = ngraph::op::v4::Interpolate::NearestMode::round_prefer_ceil;
     } else if (half_pixel == true) {
         attrs.coordinate_transformation_mode =
             ngraph::op::v4::Interpolate::CoordinateTransformMode::half_pixel;
+        attrs.nearest_mode = ngraph::op::v4::Interpolate::NearestMode::round_prefer_ceil;
     } else {
         // If none of the align_corners and half_pixel are true, transformation
         // mode is set to asymmetric
         attrs.coordinate_transformation_mode =
             ngraph::op::v4::Interpolate::CoordinateTransformMode::asymmetric;
     }
 
-    // mode is passed as "nearest" for Nearest Neighbor interpolation
-    attrs.mode = ngraph::op::v4::Interpolate::InterpolateMode::nearest;
-    attrs.nearest_mode = ngraph::op::v4::Interpolate::NearestMode::floor;
-
     std::vector<int32_t> output_shape = {out_height, out_width};
     auto outputShapeNode = createConstNode(ngraph::element::i32, {2}, output_shape);