From: 박천교/On-Device Lab(SR)/Engineer/삼성전자 <ch.bahk@samsung.com>
Date: Thu, 17 Oct 2019 06:35:25 +0000 (+0900)
Subject: [exo] Convert canonical TransposedConv2D (#8267)
X-Git-Tag: submit/tizen/20191205.083104~727
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=e0e593e14e2cd061a032a25e9481cacee4ee41e3;p=platform%2Fcore%2Fml%2Fnnfw.git

[exo] Convert canonical TransposedConv2D (#8267)

This commit introduces convert stage for canonical TransposedConv2D.
This conversion would apply to TransposedConv2D to make TFL-dialect
TFLTransposeConv with additional auxiliary nodes.

Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
---

diff --git a/compiler/exo/src/Conversion/TransposedConv2DConverter.cpp b/compiler/exo/src/Conversion/TransposedConv2DConverter.cpp
new file mode 100644
index 0000000..c03b64f
--- /dev/null
+++ b/compiler/exo/src/Conversion/TransposedConv2DConverter.cpp
@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. 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 "TransposedConv2DConverter.h"
+
+#include "Dialect/IR/TFLNodes.h"
+
+#include "GraphBlock.h"
+
+#include <loco.h>
+#include <loco/Service/ShapeInference.h>
+
+namespace exo
+{
+
+bool TransposedConv2DConverter::convert(loco::TransposedConv2D *origin)
+{
+  // Shape is required to set origin->inputSizes()
+  if (not loco::shape_known(origin))
+    return false;
+
+  if ((origin->ifm() == nullptr) or (origin->ker() == nullptr))
+    return false;
+
+  auto *graph = origin->graph();
+
+  auto tfl_tr_conv = graph->nodes()->create<locoex::TFLTransposeConv>();
+  {
+    tfl_tr_conv->stride()->w(origin->stride()->horizontal());
+    tfl_tr_conv->stride()->h(origin->stride()->vertical());
+
+    auto pad = origin->pad();
+    if (pad->bottom() == 0 && pad->top() == 0 && pad->left() == 0 && pad->right() == 0)
+      tfl_tr_conv->padding(locoex::Padding::VALID);
+    else
+      // TODO This is necessary, but not sufficient condition. More rigorous check required
+      tfl_tr_conv->padding(locoex::Padding::SAME);
+  }
+
+  // let's create a new graph connection with tfl_tr_conv
+  {
+    // Make inputSizes from shape of origin
+    auto input_sizes_const = graph->nodes()->create<locoex::TFLConst>();
+    auto origin_shape = loco::shape_get(origin).as<loco::FeatureShape>();
+
+    const loco::DataType S32 = loco::DataType::S32;
+
+    input_sizes_const->dtype(S32);
+    input_sizes_const->rank(1);
+    input_sizes_const->dim(0) = 4;
+    input_sizes_const->size<S32>(4);
+    // Note that NHWC is layout for inputSizes determined by tflite format
+    input_sizes_const->at<S32>(0) = origin_shape.count().value();  // N
+    input_sizes_const->at<S32>(1) = origin_shape.height().value(); // H
+    input_sizes_const->at<S32>(2) = origin_shape.width().value();  // W
+    input_sizes_const->at<S32>(3) = origin_shape.depth().value();  // C
+
+    tfl_tr_conv->inputSizes(input_sizes_const);
+
+    // filter
+    auto filter_dec = make_filter_decode<FilterLayout::OHWI>(origin->ker());
+    tfl_tr_conv->filter(filter_dec);
+
+    // outBackprop
+    auto feature_dec = make_feature_decode<FeatureLayout::NHWC>(origin->ifm());
+    tfl_tr_conv->outBackprop(feature_dec);
+
+    // output
+    auto feature_enc = make_feature_encode<FeatureLayout::NHWC>(tfl_tr_conv);
+
+    // replace canonical node
+    loco::replace(origin).with(feature_enc);
+    origin->ifm(nullptr);
+  }
+
+  return true;
+}
+
+} // namespace exo
diff --git a/compiler/exo/src/Conversion/TransposedConv2DConverter.h b/compiler/exo/src/Conversion/TransposedConv2DConverter.h
new file mode 100644
index 0000000..f51e0a5
--- /dev/null
+++ b/compiler/exo/src/Conversion/TransposedConv2DConverter.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. 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.
+ */
+
+#ifndef __CONVERSION_TRANSPOSEDCONV2D_CONVERTER__
+#define __CONVERSION_TRANSPOSEDCONV2D_CONVERTER__
+
+#include "CanonicalNodeConverter.h"
+
+#include <loco.h>
+
+namespace exo
+{
+
+/**
+ * @brief  Convert loco::TransposedConv2D to locoex::TFLTransposeConv and auxiliary
+ *
+ *
+ * <BEFORE>
+ *
+ * IFM ------- TransposedConv2D --- OFM
+ * (Feature)  / (Feature)
+ *           /
+ * KER ------
+ * (Filter)
+ *
+ *
+ * <AFTER>
+ *
+ * out_backprop :   IFM ------- FeatureDecode --- TFLTransposeConv --- FeatureEncode --- OFM
+ *                  [Feature]   [Tensor]         / / [Tensor]            [Feature]
+ *                                              / /
+ *        filter:   KER ------- FilterDecode --- /
+ *                  [Filter]    [Tensor]        /
+ *                                             /
+ *  input_sizes :   TFLConst (new) ------------
+ *                  [Tensor]
+ */
+class TransposedConv2DConverter : public CanonicalNodeConverter<loco::TransposedConv2D>
+{
+public:
+  const char *name(void) const final { return "exo::TransposedConv2DConverter"; }
+
+public:
+  bool convert(loco::TransposedConv2D *origin) final;
+};
+
+} // namespace exo
+
+#endif // __CONVERSION_TRANSPOSEDCONV2D_CONVERTER__