【PaddlePaddle Hackathon 4】add paddle flip op

src/frontends/paddle/src/op/flip.cpp

@@ -0,0 +1,17 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "reverse_op.hpp"
+
+namespace ov {
+namespace frontend {
+namespace paddle {
+namespace op {
+NamedOutputs flip(const NodeContext& node) {
+    return reverse_op(node);
+}
+}  // namespace op
+}  // namespace paddle
+}  // namespace frontend
+}  // namespace ov

src/frontends/paddle/src/op/reverse.cpp

@@ -2,32 +2,14 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
-#include "default_opset.hpp"
-#include "openvino/frontend/paddle/node_context.hpp"
-#include "openvino/opsets/opset1.hpp"
+#include "reverse_op.hpp"
 
 namespace ov {
 namespace frontend {
 namespace paddle {
 namespace op {
-
-using namespace default_opset;
-
 NamedOutputs reverse(const NodeContext& node) {
-    auto x = node.get_input("X");
-    // axis is a vector
-    auto axis = node.get_attribute<std::vector<int32_t>>("axis");
-    // try to keep the axis positive since reverse IR doesn't support negative axis.
-    const auto dims = static_cast<int32_t>(x.get_partial_shape().rank().get_length());
-    std::for_each(axis.begin(), axis.end(), [&dims](int32_t& value) {
-        if (value < 0) {
-            value += dims;
-        }
-    });
-
-    auto axis_node = std::make_shared<Constant>(ngraph::element::i32, Shape{axis.size()}, axis);
-    auto reverse_op = std::make_shared<ov::opset1::Reverse>(x, axis_node, ov::opset1::Reverse::Mode::INDEX);
-    return node.default_single_output_mapping({reverse_op}, {"Out"});
+    return reverse_op(node);
 }
 }  // namespace op
 }  // namespace paddle

src/frontends/paddle/src/op/reverse_op.hpp

@@ -0,0 +1,35 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "default_opset.hpp"
+#include "openvino/frontend/paddle/node_context.hpp"
+
+namespace ov {
+namespace frontend {
+namespace paddle {
+namespace op {
+NamedOutputs reverse_op(const NodeContext& node) {
+    const auto data_node = node.get_input("X");
+    const auto axes = node.get_attribute<std::vector<int32_t>>("axis");
+    auto axes_length = axes.size();
+    const auto starts =
+        default_opset::Constant::create(element::i32,
+                                        {axes_length},
+                                        std::vector<int32_t>(axes_length, std::numeric_limits<int32_t>::max()));
+    const auto stops =
+        default_opset::Constant::create(element::i32,
+                                        {axes_length},
+                                        std::vector<int32_t>(axes_length, std::numeric_limits<int32_t>::min()));
+    const auto steps =
+        default_opset::Constant::create(element::i32, {axes_length}, std::vector<int32_t>(axes_length, -1));
+    const auto axes_node = default_opset::Constant::create(element::i32, {axes_length}, axes);
+
+    return node.default_single_output_mapping(
+        {std::make_shared<default_opset::Slice>(data_node, starts, stops, steps, axes_node)},
+        {"Out"});
+}
+}  // namespace op
+}  // namespace paddle
+}  // namespace frontend
+}  // namespace ov

src/frontends/paddle/src/op_table.cpp

@@ -41,6 +41,7 @@ OP_CONVERTER(elementwise_sub);
 OP_CONVERTER(embedding);
 OP_CONVERTER(exp);
 OP_CONVERTER(expand_v2);
+OP_CONVERTER(flip);
 OP_CONVERTER(fill_any_like);
 OP_CONVERTER(fill_constant_batch_size_like);
 OP_CONVERTER(fill_constant);
@@ -158,6 +159,7 @@ std::map<std::string, CreatorFunction> get_supported_ops() {
             {"fill_constant_batch_size_like", op::fill_constant_batch_size_like},
             {"fill_constant", op::fill_constant},
             {"flatten_contiguous_range", op::flatten_contiguous_range},
+            {"flip", op::flip},
             {"floor", op::floor},
             {"gather", op::gather},
             {"gather_nd", op::gather_nd},

src/frontends/paddle/tests/op_fuzzy.cpp

@@ -188,6 +188,13 @@ static const std::vector<std::string> models{
     std::string("expand_v2_tensor_list"),
     std::string("expand_v2_tensor_list2"),
     std::string("exp_test_float32"),
+    std::string("flip_1"),
+    std::string("flip_2"),
+    std::string("flip_3"),
+    std::string("flip_4"),
+    std::string("flip_5"),
+    std::string("flip_dynamic_1"),
+    std::string("flip_dynamic_2"),
     std::string("fill_any_like"),
     std::string("fill_any_like_f16"),
     std::string("fill_any_like_f32"),

src/frontends/paddle/tests/test_models/gen_scripts/generate_flip.py

@@ -0,0 +1,73 @@
+#
+# flip paddle model generator
+#
+import numpy as np
+from save_model import saveModel
+import paddle
+import sys
+
+
+def flip(name: str, x, axis, is_dynamic=False):
+    paddle.enable_static()
+
+    with paddle.static.program_guard(paddle.static.Program(), paddle.static.Program()):
+        if is_dynamic:
+            data = paddle.static.data(name='x', shape=(-1, ) * len(x.shape), dtype=x.dtype)
+        else:
+            data = paddle.static.data(name='x', shape=x.shape, dtype=x.dtype)
+        out = paddle.flip(data, axis=axis)
+
+        cpu = paddle.static.cpu_places(1)
+        exe = paddle.static.Executor(cpu[0])
+        # startup program will call initializer to initialize the parameters.
+        exe.run(paddle.static.default_startup_program())
+
+        outs = exe.run(
+            feed={'x': x},
+            fetch_list=[out])
+
+        saveModel(name, exe, feedkeys=['x'], fetchlist=[out], inputs=[
+                  x], outputs=[outs[0]], target_dir=sys.argv[1])
+
+    return outs[0]
+
+
+def main():
+    data_type = 'int32'
+    axis = [2, 3]
+    x = np.random.randint(0, 5, (2, 3, 4, 5)).astype(data_type)
+    flip("flip_1", x, axis)
+
+    data_type = 'float32'
+    axis = [-1, -3]
+    # axis = [3, 1]
+    x = np.random.randn(3, 2, 1, 5).astype(data_type)
+    flip("flip_2", x, axis)
+
+    data_type = 'float32'
+    axis = [0, 1]
+    x = np.random.randn(1, 1, 1, 1).astype(data_type)
+    flip("flip_3", x, axis)
+
+    data_type = 'int64'
+    axis = 1
+    x = np.random.randint(-1, 3, (5, 3, 1, 1)).astype(data_type)
+    flip("flip_4", x, axis)
+
+    data_type = 'float32'
+    axis = -1
+    x = np.random.randn(1).astype(data_type)
+    flip("flip_5", x, axis)
+
+    data_type = 'int64'
+    axis = 3
+    x = np.random.randint(-5, 5, (1, 1, 4, 1)).astype(data_type)
+    flip("flip_dynamic_1", x, axis, True)
+
+    data_type = 'float32'
+    axis = [-1, -2]
+    x = np.random.randn(1, 4, 1).astype(data_type)
+    flip("flip_dynamic_2", x, axis, True)
+
+if __name__ == "__main__":
+    main()