Upsampling2D test case (fastmachinelearning#520)

* using keras upsampling2d method

* Update nnet_image_stream.h

* reindent and remove commends

* add UpSampling2D to the supported layers

* typo

* testing upsampling2d iostream and io parallel

* Delete test_upsampling2d.py

* pytest for upsampling2d iostream and io parallel

both are passed

* Delete test_upsampling2d.py

hls4ml/templates/vivado/nnet_utils/nnet_image_stream.h

@@ -1,43 +1,64 @@
-#ifndef NNET_IMAGE_STREAM_H_
-#define NNET_IMAGE_STREAM_H_
-
-#include "nnet_common.h"
-#include "hls_stream.h"
-
-namespace nnet {
-
-template<class data_T, typename CONFIG_T>
-void resize_nearest(
-    hls::stream<data_T> &image,
-    hls::stream<data_T> &resized
-) {
-    assert(CONFIG_T::new_height % CONFIG_T::height == 0);
-    assert(CONFIG_T::new_width % CONFIG_T::width == 0);
-    constexpr unsigned ratio_height = CONFIG_T::new_height / CONFIG_T::height;
-    constexpr unsigned ratio_width = CONFIG_T::new_width / CONFIG_T::width;
-    constexpr unsigned ii = ratio_height * ratio_width;
-
-    ResizeImage: for (unsigned i = 0; i < CONFIG_T::height * CONFIG_T::width; i++) {
-        #pragma HLS PIPELINE II=ii
-
-        data_T  in_data = image.read();
-
-        ResizeNew: for (unsigned j = 0; j < ratio_height * ratio_width; j++) {
-            #pragma HLS UNROLL
-
-            data_T out_data;
-            #pragma HLS DATA_PACK variable=out_data
-
-            ResizeChan: for (unsigned k = 0; k < CONFIG_T::n_chan; k++) {
-                #pragma HLS UNROLL
-                out_data[k] = in_data[k];
-            }
-
-            resized.write(out_data);
-        }
-    }
-}
-
-}
-
-#endif
+#ifndef NNET_IMAGE_STREAM_H_
+#define NNET_IMAGE_STREAM_H_
+
+#include "nnet_common.h"
+#include "hls_stream.h"
+
+namespace nnet {
+
+template<class data_T, typename CONFIG_T>
+void resize_nearest(
+    hls::stream<data_T> &image,
+    hls::stream<data_T> &resized
+) {
+	assert(CONFIG_T::new_height % CONFIG_T::height == 0);
+	assert(CONFIG_T::new_width % CONFIG_T::width == 0);
+	constexpr unsigned ratio_height = CONFIG_T::new_height / CONFIG_T::height;
+	constexpr unsigned ratio_width = CONFIG_T::new_width / CONFIG_T::width;
+
+
+	ImageHeight: for (unsigned h = 0; h < CONFIG_T::height; h++) {
+		#pragma HLS PIPELINE
+
+		data_T data_in_row[CONFIG_T::width];
+
+		ImageWidth: for (unsigned i = 0; i < CONFIG_T::width; i++) {
+			#pragma HLS UNROLL
+
+			data_T  in_data = image.read();
+
+			ImageChan: for (unsigned j = 0; j < CONFIG_T::n_chan; j++) {
+				#pragma HLS UNROLL
+
+				data_in_row[i][j] = in_data[j];
+			}
+		}
+
+		ResizeHeight: for (unsigned i = 0; i <ratio_height; i++) {
+			#pragma HLS UNROLL
+
+			ImageWidth2: for (unsigned l = 0; l < CONFIG_T::width; l++) {
+				#pragma HLS UNROLL
+
+				ResizeWidth: for (unsigned j = 0; j < ratio_width; j++) {
+					#pragma HLS UNROLL
+
+					data_T out_data;
+					#pragma HLS DATA_PACK variable=out_data
+
+					ResizeChan: for (unsigned k = 0; k < CONFIG_T::n_chan; k++) {
+						#pragma HLS UNROLL
+
+						out_data[k] = data_in_row[l][k];
+					}
+
+					resized.write(out_data);   
+				}
+			}
+		}
+	}
+}
+
+}
+
+#endif

hls4ml/utils/config.py

@@ -100,7 +100,7 @@ def config_from_keras_model(model, granularity='model', default_precision='ap_fi
         model_arch = json.loads(model.to_json())
 
     #Define supported layers
-    core_layers = ['InputLayer', 'Dropout', 'Flatten', 'Reshape', 'Permute']
+    core_layers = ['InputLayer', 'Dropout', 'Flatten', 'Reshape', 'Permute', 'UpSampling2D']
     dense_layers = ['Dense', 'BinaryDense', 'TernaryDense']
     conv_layers = ['Conv1D', 'Conv2D', 'BinaryConv2D']
     pooling_layers = ['MaxPooling1D', 'MaxPooling2D', 'GlobalMaxPooling1D', 'GlobalMaxPooling2D', 'AveragePooling1D', 'AveragePooling2D', 'GlobalAveragePooling1D', 'GlobalAveragePooling2D']