Add Image Transformer Library

tensorflow_lite_support/cc/task/processor/BUILD

@@ -40,6 +40,24 @@ cc_library_with_tflite(
     ],
 )
 
+cc_library_with_tflite(
+    name = "image_postprocessor",
+    srcs = ["image_postprocessor.cc"],
+    hdrs = ["image_postprocessor.h"],
+    tflite_deps = [
+        ":processor",
+        "//tensorflow_lite_support/cc/task/vision/utils:image_tensor_specs",
+    ],
+    deps = [
+        "//tensorflow_lite_support/cc/port:status_macros",
+        "//tensorflow_lite_support/cc/port:statusor",
+        "//tensorflow_lite_support/cc/task/vision/core:frame_buffer",
+        "//tensorflow_lite_support/cc/task/vision/utils:frame_buffer_utils",
+        "//tensorflow_lite_support/cc/task/core:task_utils",
+        "@com_google_absl//absl/status",
+    ],
+)
+
 cc_library_with_tflite(
     name = "classification_postprocessor",
     srcs = ["classification_postprocessor.cc"],

tensorflow_lite_support/cc/task/processor/image_postprocessor.cc

@@ -0,0 +1,229 @@
+/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow_lite_support/cc/task/processor/image_postprocessor.h"
+
+namespace tflite {
+namespace task {
+namespace processor {
+
+namespace {
+
+using ::absl::StatusCode;
+using ::tflite::metadata::ModelMetadataExtractor;
+using ::tflite::support::CreateStatusWithPayload;
+using ::tflite::support::StatusOr;
+using ::tflite::support::TfLiteSupportStatus;
+
+StatusOr<absl::optional<vision::NormalizationOptions>>
+GetNormalizationOptionsIfAny(const TensorMetadata& tensor_metadata) {
+  ASSIGN_OR_RETURN(
+      const tflite::ProcessUnit* normalization_process_unit,
+      ModelMetadataExtractor::FindFirstProcessUnit(
+          tensor_metadata, tflite::ProcessUnitOptions_NormalizationOptions));
+  if (normalization_process_unit == nullptr) {
+    return {absl::nullopt};
+  }
+  const tflite::NormalizationOptions* tf_normalization_options =
+      normalization_process_unit->options_as_NormalizationOptions();
+  const auto mean_values = tf_normalization_options->mean();
+  const auto std_values = tf_normalization_options->std();
+  if (mean_values->size() != std_values->size()) {
+    return CreateStatusWithPayload(
+        StatusCode::kInvalidArgument,
+        absl::StrCat("NormalizationOptions: expected mean and std of same "
+                     "dimension, got ",
+                     mean_values->size(), " and ", std_values->size(), "."),
+        TfLiteSupportStatus::kMetadataInvalidProcessUnitsError);
+  }
+  absl::optional<vision::NormalizationOptions> normalization_options;
+  if (mean_values->size() == 1) {
+    normalization_options = vision::NormalizationOptions{
+        .mean_values = {mean_values->Get(0), mean_values->Get(0),
+                        mean_values->Get(0)},
+        .std_values = {std_values->Get(0), std_values->Get(0),
+                       std_values->Get(0)},
+        .num_values = 1};
+  } else if (mean_values->size() == 3) {
+    normalization_options = vision::NormalizationOptions{
+        .mean_values = {mean_values->Get(0), mean_values->Get(1),
+                        mean_values->Get(2)},
+        .std_values = {std_values->Get(0), std_values->Get(1),
+                       std_values->Get(2)},
+        .num_values = 3};
+  } else {
+    return CreateStatusWithPayload(
+        StatusCode::kInvalidArgument,
+        absl::StrCat("NormalizationOptions: only 1 or 3 mean and std "
+                     "values are supported, got ",
+                     mean_values->size(), "."),
+        TfLiteSupportStatus::kMetadataInvalidProcessUnitsError);
+  }
+  return normalization_options;
+}
+}  // namespace
+
+/* static */
+tflite::support::StatusOr<std::unique_ptr<ImagePostprocessor>>
+ImagePostprocessor::Create(core::TfLiteEngine* engine,
+                           const std::initializer_list<int> output_indices,
+                           const std::initializer_list<int> input_indices) { 
+  ASSIGN_OR_RETURN(auto processor, Processor::Create<ImagePostprocessor>(/* num_expected_tensors = */ 1, engine, output_indices, /* requires_metadata = */ false));
+
+  RETURN_IF_ERROR(processor->Init(input_indices));
+  return processor;
+}
+
+absl::Status ImagePostprocessor::Init(const std::vector<int>& input_indices) {
+  if (core::TfLiteEngine::OutputCount(engine_->interpreter()) != 1) {
+    return tflite::support::CreateStatusWithPayload(
+        absl::StatusCode::kInvalidArgument,
+        absl::StrFormat(
+            "Image segmentation models are expected to have only 1 "
+            "output, found %d",
+            core::TfLiteEngine::OutputCount(engine_->interpreter())),
+        tflite::support::TfLiteSupportStatus::kInvalidNumOutputTensorsError);
+  }
+
+  if (GetTensor()->type != kTfLiteUInt8 && GetTensor()->type != kTfLiteFloat32) {
+    return tflite::support::CreateStatusWithPayload(
+        absl::StatusCode::kInvalidArgument,
+        absl::StrFormat("Type mismatch for output tensor %s. Requested one "
+                        "of these types: "
+                        "kTfLiteUint8/kTfLiteFloat32, got %s.",
+                        GetTensor()->name, TfLiteTypeGetName(GetTensor()->type)),
+        tflite::support::TfLiteSupportStatus::kInvalidOutputTensorTypeError);
+  }
+
+  if (GetTensor()->dims->data[0] != 1 || GetTensor()->dims->data[3] != 3) {
+    return CreateStatusWithPayload(
+        absl::StatusCode::kInvalidArgument,
+        absl::StrCat("The input tensor should have dimensions 1 x height x "
+                     "width x 3. Got ",
+                     GetTensor()->dims->data[0], " x ", GetTensor()->dims->data[1],
+                     " x ", GetTensor()->dims->data[2], " x ",
+                     GetTensor()->dims->data[3], "."),
+        tflite::support::TfLiteSupportStatus::
+            kInvalidInputTensorDimensionsError);
+  }
+
+  // Gather metadata
+  auto* output_metadata =
+      engine_->metadata_extractor()->GetOutputTensorMetadata(
+          tensor_indices_.at(0));
+  auto* input_metadata = engine_->metadata_extractor()->GetInputTensorMetadata(
+      input_indices.at(0));
+
+  // Use input metadata for normalization as fallback.
+  auto* processing_metadata =
+      output_metadata != nullptr ? output_metadata : input_metadata;
+
+  absl::optional<vision::NormalizationOptions> normalization_options;
+  ASSIGN_OR_RETURN(normalization_options,
+                   GetNormalizationOptionsIfAny(*processing_metadata));
+
+  if (GetTensor()->type == kTfLiteFloat32) {
+    if (!normalization_options.has_value()) {
+      return CreateStatusWithPayload(
+          absl::StatusCode::kNotFound,
+          "Output tensor has type kTfLiteFloat32: it requires specifying "
+          "NormalizationOptions metadata to preprocess output images.",
+          TfLiteSupportStatus::kMetadataMissingNormalizationOptionsError);
+    } else if (GetTensor()->bytes / sizeof(float) %
+                   normalization_options.value().num_values !=
+               0) {
+      return CreateStatusWithPayload(
+          StatusCode::kInvalidArgument,
+          "The number of elements in the output tensor must be a multiple of "
+          "the number of normalization parameters.",
+          TfLiteSupportStatus::kInvalidArgumentError);
+    }
+
+    options_ = std::make_unique<vision::NormalizationOptions>(
+        normalization_options.value());
+  }
+
+  return absl::OkStatus();
+}
+
+absl::StatusOr<vision::FrameBuffer> ImagePostprocessor::Postprocess() {
+  has_uint8_outputs_ = GetTensor()->type == kTfLiteUInt8;
+  const int kRgbPixelBytes = 3;
+
+  vision::FrameBuffer::Dimension to_buffer_dimension = {
+      GetTensor()->dims->data[2], GetTensor()->dims->data[1]};
+  size_t output_byte_size =
+      GetBufferByteSize(to_buffer_dimension, vision::FrameBuffer::Format::kRGB);
+  std::vector<uint8> postprocessed_data(output_byte_size / sizeof(uint8), 0);
+
+  if (has_uint8_outputs_) {  // No denormalization required.
+    if (GetTensor()->bytes != output_byte_size) {
+      return tflite::support::CreateStatusWithPayload(
+          absl::StatusCode::kInternal,
+          "Size mismatch or unsupported padding bytes between pixel data "
+          "and output tensor.");
+    }
+    const uint8* output_data =
+        core::AssertAndReturnTypedTensor<uint8>(GetTensor()).value();
+    postprocessed_data.insert(postprocessed_data.begin(), &output_data[0],
+                              &output_data[output_byte_size / sizeof(uint8)]);
+  } else {  // Denormalize to [0, 255] range.
+    if (GetTensor()->bytes / sizeof(float) != output_byte_size / sizeof(uint8)) {
+      return tflite::support::CreateStatusWithPayload(
+          absl::StatusCode::kInternal,
+          "Size mismatch or unsupported padding bytes between pixel data "
+          "and output tensor.");
+    }
+
+    uint8* denormalized_output_data = postprocessed_data.data();
+    const float* output_data =
+        core::AssertAndReturnTypedTensor<float>(GetTensor()).value();
+    const auto norm_options = GetNormalizationOptions();
+
+    if (norm_options.num_values == 1) {
+      float mean_value = norm_options.mean_values[0];
+      float std_value = norm_options.std_values[0];
+
+      for (size_t i = 0; i < output_byte_size / sizeof(uint8);
+           ++i, ++denormalized_output_data, ++output_data) {
+        *denormalized_output_data = static_cast<uint8>(std::round(std::min(
+            255.f, std::max(0.f, (*output_data) * std_value + mean_value))));
+      }
+    } else {
+      for (size_t i = 0; i < output_byte_size / sizeof(uint8);
+           ++i, ++denormalized_output_data, ++output_data) {
+        *denormalized_output_data = static_cast<uint8>(std::round(std::min(
+            255.f,
+            std::max(0.f, (*output_data) * norm_options.std_values[i % 3] +
+                              norm_options.mean_values[i % 3]))));
+      }
+    }
+  }
+
+  vision::FrameBuffer::Plane postprocessed_plane = {
+      /*buffer=*/postprocessed_data.data(),
+      /*stride=*/{GetTensor()->dims->data[2] * kRgbPixelBytes, kRgbPixelBytes}};
+  auto postprocessed_frame_buffer =
+      vision::FrameBuffer::Create({postprocessed_plane}, to_buffer_dimension,
+                                  vision::FrameBuffer::Format::kRGB,
+                                  vision::FrameBuffer::Orientation::kTopLeft);
+
+  vision::FrameBuffer postprocessed_result = *postprocessed_frame_buffer.get();
+  return postprocessed_result;
+}
+
+}  // namespace processor
+}  // namespace task
+}  // namespace tflite

tensorflow_lite_support/cc/task/processor/image_postprocessor.h

@@ -0,0 +1,70 @@
+/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either exPostss or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_SUPPORT_CC_TASK_PROCESSOR_IMAGE_POSTPROCESSOR_H_
+#define TENSORFLOW_LITE_SUPPORT_CC_TASK_PROCESSOR_IMAGE_POSTPROCESSOR_H_
+
+#include "tensorflow_lite_support/cc/port/status_macros.h"
+#include "tensorflow_lite_support/cc/task/processor/processor.h"
+#include "tensorflow_lite_support/cc/task/vision/core/frame_buffer.h"
+#include "tensorflow_lite_support/cc/task/vision/utils/frame_buffer_utils.h"
+#include "tensorflow_lite_support/cc/task/core/task_utils.h"
+#include "tensorflow_lite_support/cc/task/vision/utils/image_tensor_specs.h"
+
+namespace tflite {
+namespace task {
+namespace processor {
+
+// Process the associated output image tensor and convert it to a FrameBuffer.
+// Requirement for the output tensor:
+//   (kTfLiteUInt8/kTfLiteFloat32)
+//    - image input of size `[batch x height x width x channels]`.
+//    - batch inference is not supported (`batch` is required to be 1).
+//    - only RGB inputs are supported (`channels` is required to be 3).
+//    - if type is kTfLiteFloat32, NormalizationOptions are required to be
+//      attached to the metadata for output de-normalization. Uses input metadata
+//      as fallback in case output metadata isn't provided.
+class ImagePostprocessor : public Postprocessor {
+ public:
+  static tflite::support::StatusOr<std::unique_ptr<ImagePostprocessor>>
+  Create(core::TfLiteEngine* engine,
+         const std::initializer_list<int> output_indices,
+         const std::initializer_list<int> input_indices);
+
+  // Processes the output tensor to an RGB of FrameBuffer type.
+  // If output tensor is of type kTfLiteFloat32, denormalize it into [0 - 255]
+  // via normalization parameters.
+  absl::StatusOr<vision::FrameBuffer> Postprocess();
+
+ private:
+  using Postprocessor::Postprocessor;
+
+  // Whether the model features quantized inference type (QUANTIZED_UINT8). This
+  // is currently detected by checking if all output tensors data type is uint8.
+  bool has_uint8_outputs_;
+
+  std::unique_ptr<vision::NormalizationOptions> options_;
+
+  absl::Status Init(const std::vector<int>& input_indices);
+
+  const vision::NormalizationOptions& GetNormalizationOptions() {
+    return *options_.get();
+  }
+};
+}  // namespace processor
+}  // namespace task
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_SUPPORT_CC_TASK_PROCESSOR_IMAGE_POSTPROCESSOR_H_

tensorflow_lite_support/cc/task/vision/BUILD

@@ -173,3 +173,31 @@ cc_library_with_tflite(
         "@org_tensorflow//tensorflow/lite/core/api:op_resolver",
     ],
 )
+
+# IMPORTANT: in order to use hardware acceleration delegates, configurable through the
+# `compute_settings` field of the ImageClassifierOptions, you must additionally link to
+# the appropriate delegate plugin target (e.g. `gpu_plugin` for GPU) from:
+# https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/experimental/acceleration/configuration/BUILD
+# To use EDGETPU_CORAL, link to `edgetpu_coral_plugin` from:
+# https://github.com/tensorflow/tflite-support/blob/a58a4f9225c411fa9ba29f821523e6e283988d23/tensorflow_lite_support/acceleration/configuration/BUILD#L11
+cc_library_with_tflite(
+    name = "image_transformer",
+    srcs = ["image_transformer.cc"],
+    hdrs = ["image_transformer.h"],
+    tflite_deps = [
+        "@org_tensorflow//tensorflow/lite/core/shims:builtin_ops",
+        "//tensorflow_lite_support/cc/task/core:task_api_factory",
+        "//tensorflow_lite_support/cc/task/vision/core:base_vision_task_api",
+        "//tensorflow_lite_support/cc/task/processor:image_postprocessor",
+    ],
+    deps = [
+        "//tensorflow_lite_support/cc/port:integral_types",
+        "//tensorflow_lite_support/cc/port:status_macros",
+        "//tensorflow_lite_support/cc/port:statusor",
+        "//tensorflow_lite_support/cc/task/core:external_file_handler",
+        "//tensorflow_lite_support/cc/task/vision/proto:image_transformer_options_proto_inc",
+        "@com_google_absl//absl/strings:str_format",
+        "@flatbuffers",
+        "@org_tensorflow//tensorflow/lite/core/api",
+    ],
+)