[PT FE] Support aten::atan2 for pytorch models

src/frontends/pytorch/src/op/atan2.cpp

@@ -0,0 +1,101 @@
+// Copyright (C) 2018-2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#define _USE_MATH_DEFINES
+
+#include <math.h>
+
+#include <memory>
+
+#include "openvino/core/type/element_type.hpp"
+#include "openvino/frontend/pytorch/node_context.hpp"
+#include "openvino/op/add.hpp"
+#include "openvino/op/atan.hpp"
+#include "openvino/op/constant.hpp"
+#include "openvino/op/convert_like.hpp"
+#include "openvino/op/divide.hpp"
+#include "openvino/op/equal.hpp"
+#include "openvino/op/greater.hpp"
+#include "openvino/op/greater_eq.hpp"
+#include "openvino/op/less.hpp"
+#include "openvino/op/logical_and.hpp"
+#include "openvino/op/logical_or.hpp"
+#include "openvino/op/multiply.hpp"
+#include "openvino/op/select.hpp"
+#include "openvino/op/subtract.hpp"
+#include "utils.hpp"
+
+namespace ov {
+namespace frontend {
+namespace pytorch {
+namespace op {
+
+using namespace ov::op;
+
+OutputVector translate_atan2(const NodeContext& context) {
+    // atan2(input, other, *) → Tensor
+    num_inputs_check(context, 2, 2);
+    Output<Node> lhs;
+    Output<Node> rhs;
+
+    std::tie(lhs, rhs) = get_inputs_with_promoted_types(context, 0, 1);
+
+    auto div = context.mark_node(std::make_shared<v1::Divide>(lhs, rhs));
+
+    auto atan = context.mark_node(std::make_shared<v0::Atan>(div));
+
+    // create some constants to adjust result according to quadrant.
+    auto zero = context.mark_node(v0::Constant::create(ov::element::i32, Shape{}, {0}));
+    auto pi = context.mark_node(v0::Constant::create(ov::element::f64, Shape{}, {M_PI}));
+    auto neg_pi = context.mark_node(v0::Constant::create(ov::element::f64, Shape{}, {-M_PI}));
+    auto half_pi = context.mark_node(v0::Constant::create(ov::element::f64, Shape{}, {M_PI_2}));
+    auto neg_half_pi = context.mark_node(v0::Constant::create(ov::element::f64, Shape{}, {-M_PI_2}));
+
+    zero = context.mark_node(std::make_shared<v1::ConvertLike>(zero, rhs));
+    pi = context.mark_node(std::make_shared<v1::ConvertLike>(pi, rhs));
+    neg_pi = context.mark_node(std::make_shared<v1::ConvertLike>(neg_pi, rhs));
+    half_pi = context.mark_node(std::make_shared<v1::ConvertLike>(half_pi, rhs));
+    neg_half_pi = context.mark_node(std::make_shared<v1::ConvertLike>(neg_half_pi, rhs));
+
+    //  x > 0, no adjustment needed
+    auto x_greater_than_zero = context.mark_node(std::make_shared<v1::Greater>(rhs, zero));
+
+    // x < 0 and y >= 0, need to plus pi
+    auto y_greater_equal_zero = context.mark_node(std::make_shared<v1::GreaterEqual>(lhs, zero));
+    auto x_less_than_zero = context.mark_node(std::make_shared<v1::Less>(rhs, zero));
+    auto add_pi_condition = context.mark_node(std::make_shared<v1::LogicalAnd>(x_less_than_zero, y_greater_equal_zero));
+
+    // x < 0 and y < 0, need to minus pi
+    auto y_less_than_zero = std::make_shared<v1::Less>(lhs, zero);
+    auto subtract_pi_condition =
+        context.mark_node(std::make_shared<v1::LogicalAnd>(x_less_than_zero, y_less_than_zero));
+
+    // x = 0 and y > 0, pi/2
+    auto x_equal_zero = std::make_shared<v1::Equal>(rhs, zero);
+    auto y_greater_than_zero = std::make_shared<v1::Greater>(lhs, zero);
+    auto half_pi_condition = context.mark_node(std::make_shared<v1::LogicalAnd>(x_equal_zero, y_greater_than_zero));
+
+    // x = 0 and y < 0, -pi/2
+    auto neg_half_pi_condition = context.mark_node(std::make_shared<v1::LogicalAnd>(x_equal_zero, y_less_than_zero));
+
+    auto special_case_condition =
+        context.mark_node(std::make_shared<v1::LogicalOr>(half_pi_condition, neg_half_pi_condition));
+
+    // do adjustment
+    auto atan_plus_pi = context.mark_node(std::make_shared<v1::Add>(atan, pi));
+    auto atan_minus_pi = context.mark_node(std::make_shared<v1::Subtract>(atan, pi));
+
+    // select result
+    auto ajusted_case = context.mark_node(std::make_shared<v1::Select>(add_pi_condition, atan_plus_pi, atan_minus_pi));
+    auto special_case = context.mark_node(std::make_shared<v1::Select>(half_pi_condition, half_pi, neg_half_pi));
+    auto adjusted_atan = context.mark_node(std::make_shared<v1::Select>(x_greater_than_zero, atan, ajusted_case));
+    auto result = context.mark_node(std::make_shared<v1::Select>(special_case_condition, special_case, adjusted_atan));
+
+    return {result};
+}
+
+}  // namespace op
+}  // namespace pytorch
+}  // namespace frontend
+}  // namespace ov

src/frontends/pytorch/src/op_table.cpp

@@ -42,6 +42,7 @@ OP_CONVERTER(translate_argmax);
 OP_CONVERTER(translate_argmin);
 OP_CONVERTER(translate_as_strided);
 OP_CONVERTER(translate_as_tensor);
+OP_CONVERTER(translate_atan2);
 OP_CONVERTER(translate_avg_poolnd);
 OP_CONVERTER(translate_bool);
 OP_CONVERTER(translate_batch_norm);
@@ -380,6 +381,7 @@ const std::unordered_map<std::string, CreatorFunction> get_supported_ops_ts() {
         {"aten::atanh",
          op::optional_out<op::translate_1to1_match_1_inputs_with_fp32_type_alignment<opset10::Atanh>, 1>},
         {"aten::atanh_", op::inplace_op<op::translate_1to1_match_1_inputs<opset10::Atanh>>},
+        {"aten::atan2", op::translate_atan2},
         {"aten::avg_pool1d", op::quantizable_op<op::translate_avg_poolnd>},
         {"aten::avg_pool2d", op::quantizable_op<op::translate_avg_poolnd>},
         {"aten::avg_pool3d", op::quantizable_op<op::translate_avg_poolnd>},

tests/layer_tests/pytorch_tests/test_atan2.py

@@ -0,0 +1,80 @@
+# Copyright (C) 2018-2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import numpy as np
+import pytest
+import torch
+
+from pytorch_layer_test_class import PytorchLayerTest
+
+@pytest.mark.parametrize("input_shape_rhs", [
+    [2, 5, 3, 4],
+    [1, 5, 3, 4],
+    [1]
+])
+class TestAtan2(PytorchLayerTest):
+
+    def _prepare_input(self):
+        return (np.random.randn(2, 5, 3, 4).astype(np.float32), self.input_rhs)
+
+    def create_model(self):
+
+        class aten_atan2(torch.nn.Module):
+            def __init__(self):
+                super(aten_atan2, self).__init__()
+
+            def forward(self, lhs, rhs):
+                return torch.arctan2(lhs, rhs)
+
+        ref_net = None
+
+        return aten_atan2(), ref_net, "aten::atan2"
+    @pytest.mark.nightly
+    @pytest.mark.precommit
+    @pytest.mark.precommit_torch_export
+    @pytest.mark.precommit_fx_backend
+    def test_atan2(self, ie_device, precision, ir_version, input_shape_rhs):
+        self.input_rhs = np.random.randn(*input_shape_rhs).astype(np.float32)
+        self._test(*self.create_model(), ie_device, precision, ir_version, use_convert_model=True)
+
+class TestAtan2Types(PytorchLayerTest):
+
+    def _prepare_input(self):
+        return (torch.randn(self.lhs_shape).to(self.lhs_type).numpy(),
+                torch.randn(self.rhs_shape).to(self.rhs_type).numpy())
+
+    def create_model(self, lhs_type, rhs_type):
+
+        class aten_atan2(torch.nn.Module):
+            def __init__(self, lhs_type, rhs_type):
+                super(aten_atan2, self).__init__()
+                self.lhs_type = lhs_type
+                self.rhs_type = rhs_type
+
+            def forward(self, lhs, rhs):
+                return torch.arctan2(lhs.to(self.lhs_type), rhs.to(self.rhs_type))
+
+        ref_net = None
+
+        return aten_atan2(lhs_type, rhs_type), ref_net, "aten::atan2"
+
+    @pytest.mark.parametrize(("lhs_type", "rhs_type"),
+                             [[torch.int, torch.float32],
+                              [torch.int, torch.float64],
+                              [torch.float32, torch.float64],
+                              [torch.int64, torch.float32]
+                              ])
+    @pytest.mark.parametrize(("lhs_shape", "rhs_shape"), [([2, 3], [2, 3]),
+                                                          ([2, 3], [1, 3]),
+                                                          ([3, 2, 3], [2, 3]),
+                                                          ])    
+    @pytest.mark.nightly
+    @pytest.mark.precommit
+    @pytest.mark.precommit_torch_export
+    def test_atan2_types(self, ie_device, precision, ir_version, lhs_type, lhs_shape, rhs_type, rhs_shape):
+        self.lhs_type = lhs_type
+        self.lhs_shape = lhs_shape
+        self.rhs_type = rhs_type
+        self.rhs_shape = rhs_shape
+        self._test(*self.create_model(lhs_type, rhs_type),
+                   ie_device, precision, ir_version, freeze_model=False, trace_model=True)