invoke saturate when Concat/ResizeBilinear/Pooling does forward_cpu

Author: snoopyisadog

include/caffe/layers/concat_layer.hpp

@@ -84,7 +84,7 @@ class ConcatLayer : public Layer<Dtype> {
   vector<int> input_zero_point_; //CUSTOMIZATION
   double output_scale_;  //CUSTOMIZATION
   int output_zero_point_; //CUSTOMIZATION
-  //Dtype saturate_; //CUSTOMIZATION
+  Dtype saturate_; //CUSTOMIZATION
 };
 
 }  // namespace caffe

include/caffe/layers/resize_bilinear_layer.hpp

@@ -55,6 +55,7 @@ class ResizeBilinearLayer : public Layer<Dtype> {
   bool pytorch_half_pixel_;
   double output_scale_; //CUSTOMIZATION
   int output_zero_point_; //CUSTOMIZATION
+  Dtype saturate_; //CUSTOMIZATION
 
   // Compute the interpolation indices only once.
   struct CachedInterpolation {

src/caffe/layers/concat_layer.cpp

@@ -28,6 +28,7 @@ void ConcatLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
   }
   output_scale_ = concat_param.output_scale();
   output_zero_point_ = concat_param.output_zero_point();
+  saturate_ = concat_param.saturate();
 }
 
 template <typename Dtype>
@@ -97,8 +98,12 @@ void ConcatLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
     }
     offset_concat_axis += bottom_concat_axis;
   }
-  if (is_quant) // CUSTOMIZATION
-    caffe_cpu_quantize<Dtype>(top[0]->count(), top[0]->mutable_cpu_data(), output_scale_, output_zero_point_);
+  if (is_quant){ // CUSTOMIZATION
+    const int count_t = top[0]->count();
+    top_data = top[0]->mutable_cpu_data();
+    caffe_cpu_quantize(count_t, top_data, output_scale_, output_zero_point_);
+    caffe_cpu_saturate(count_t, top_data, saturate_); // if None nothing happens
+  }
 
   if (is_quant) {
     for (int i = 0; i < bottom.size(); ++i) {

src/caffe/layers/pooling_layer.cpp

@@ -498,6 +498,7 @@ void PoolingLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   default:
     LOG(FATAL) << "Unknown pooling method.";
   }
+  caffe_cpu_saturate(top[0]->count(), top[0]->mutable_cpu_data(), saturate_); // if None nothing happens
 }
 
 template <typename Dtype>

src/caffe/layers/resize_bilinear_layer.cpp

@@ -26,6 +26,7 @@ void ResizeBilinearLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
         "If pytorch_half_pixel_ is True, half_pixel_centers_ must be False.";
   output_scale_ = this->layer_param_.resize_bilinear_param().output_scale(); //CUSTOMIZATION
   output_zero_point_ = this->layer_param_.resize_bilinear_param().output_zero_point(); //CUSTOMIZATION
+  saturate_ = this->layer_param_.resize_bilinear_param().saturate(); //CUSTOMIZATION
 }
 
 template <typename Dtype>
@@ -327,6 +328,7 @@ void ResizeBilinearLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
       }
     }
   }
+  caffe_cpu_saturate(top[0]->count(), top[0]->mutable_cpu_data(), saturate_); // if None nothing happens
 }
 
 INSTANTIATE_CLASS(ResizeBilinearLayer);

src/caffe/proto/caffe.proto

@@ -1348,10 +1348,26 @@ message ConcatParameter {
   // DEPRECATED: alias for "axis" -- does not support negative indexing.
   optional uint32 concat_dim = 1 [default = 1];
 
+  //<--CUSTOMIZATION
   repeated double input_scale = 3; // CUSTOMIZATION, blob-wise for quantization
   optional double output_scale = 4 [default = 1]; //CUSTOMIZATION
   repeated int32 input_zero_point = 5; //CUSTOMIZATION, blob-wise for quantization
   optional int32 output_zero_point = 6 [default = 0]; //CUSTOMIZATION
+  enum SaturateMethod {
+    None = 0;
+    Signed = 1;
+    Unsigned = 2;
+    Signed_8bit = 3;
+    Unsigned_8bit = 4;
+  }
+  optional SaturateMethod saturate = 7 [default = None]; //control the output in certain range
+  enum QuantizeMethod {
+    tflite = 0;
+    ONNX = 1;
+    Caffe2 = 2;
+  }
+  optional QuantizeMethod quantize_method = 8 [default = tflite];
+  //CUSTOMIZATION-->
 }
 
 message ContrastiveLossParameter {
@@ -3286,10 +3302,26 @@ message ResizeBilinearParameter {
   optional bool pytorch_half_pixel = 8 [default = false];
   optional int32 output_depth = 9;
   optional float scale_depth = 10 [default = 1];
+  //<--CUSTOMIZATION
   optional double input_scale = 20 [default = 1]; //CUSTOMIZATION
   optional double output_scale = 21 [default = 1]; //CUSTOMIZATION
   optional int32 input_zero_point = 22 [default = 0]; //CUSTOMIZATION
   optional int32 output_zero_point = 23 [default = 0]; //CUSTOMIZATION
+  enum SaturateMethod {
+    None = 0;
+    Signed = 1;
+    Unsigned = 2;
+    Signed_8bit = 3;
+    Unsigned_8bit = 4;
+  }
+  optional SaturateMethod saturate = 24 [default = None]; //control the output in certain range
+  enum QuantizeMethod {
+    tflite = 0;
+    ONNX = 1;
+    Caffe2 = 2;
+  }
+  optional QuantizeMethod quantize_method = 25 [default = tflite];
+  //CUSTOMIZATION-->
 }
 
 message ReduceSumParameter {