[PT FE]: support aten::take_along_dim (openvinotoolkit#21625)

src/frontends/pytorch/src/op/take_along_dim.cpp

@@ -0,0 +1,56 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "openvino/frontend/pytorch/node_context.hpp"
+#include "openvino/op/broadcast.hpp"
+#include "openvino/op/constant.hpp"
+#include "openvino/op/convert.hpp"
+#include "openvino/op/gather_elements.hpp"
+#include "openvino/op/reshape.hpp"
+#include "openvino/op/scatter_update.hpp"
+#include "openvino/op/shape_of.hpp"
+#include "utils.hpp"
+
+namespace ov {
+namespace frontend {
+namespace pytorch {
+namespace op {
+
+OutputVector translate_take_along_dim(const NodeContext& context) {
+    // aten::take_along_dim(Tensor self, Tensor indices, int? dim=None) -> Tensor
+    // aten::take_along_dim.out(Tensor self, Tensor indices, int? dim=None, *, Tensor(a!) out) -> Tensor(a!)
+    num_inputs_check(context, 3, 4);
+    auto x = context.get_input(0);
+    auto index = context.get_input(1);
+    index = context.mark_node(std::make_shared<ov::op::v0::Convert>(index, element::i32));
+    int64_t axis = 0;
+
+    if (context.input_is_none(2)) {
+        // if dimension is not provided, flattenize input first
+        auto minus_1 = context.mark_node(ov::op::v0::Constant::create(element::i32, Shape{1}, {-1}));
+        x = context.mark_node(std::make_shared<ov::op::v1::Reshape>(x, minus_1, false));
+    } else {
+        axis = context.const_input<int64_t>(2);
+        // OpenVINO GatherElements requires to have equal dims between index and input except specified axis
+        // while PyTorch requires to have them broadcastable
+        auto axis_node = context.mark_node(ov::op::v0::Constant::create(element::i32, Shape{1}, {axis}));
+        auto const_1 = context.mark_node(ov::op::v0::Constant::create(element::i32, Shape{1}, {1}));
+        auto const_0 = context.mark_node(ov::op::v0::Constant::create(element::i32, Shape{1}, {0}));
+        auto x_shape = context.mark_node(std::make_shared<ov::op::v3::ShapeOf>(x, element::i32));
+        auto broadcast_shape =
+            context.mark_node(std::make_shared<ov::op::v3::ScatterUpdate>(x_shape, axis_node, const_1, const_0));
+        index = context.mark_node(
+            std::make_shared<ov::op::v3::Broadcast>(index, broadcast_shape, ov::op::BroadcastType::BIDIRECTIONAL));
+    }
+    auto gather_elements = context.mark_node(std::make_shared<ov::op::v6::GatherElements>(x, index, axis));
+    if (!context.input_is_none(3)) {
+        context.mutate_input(3, gather_elements);
+    }
+    return {gather_elements};
+};
+
+}  // namespace op
+}  // namespace pytorch
+}  // namespace frontend
+}  // namespace ov

src/frontends/pytorch/src/op_table.cpp

@@ -197,6 +197,7 @@ OP_CONVERTER(translate_sub);
 OP_CONVERTER(translate_sub_);
 OP_CONVERTER(translate_sum);
 OP_CONVERTER(translate_t);
+OP_CONVERTER(translate_take_along_dim);
 OP_CONVERTER(translate_to);
 OP_CONVERTER(translate_topk);
 OP_CONVERTER(translate_transpose);
@@ -536,6 +537,7 @@ const std::map<std::string, CreatorFunction> get_supported_ops_ts() {
         {"aten::swapaxes", op::quantizable_op<op::translate_transpose>},
         {"aten::t", op::translate_t},
         {"aten::t_", op::inplace_op<op::translate_t>},
+        {"aten::take_along_dim", op::translate_take_along_dim},
         {"aten::tan", op::translate_1to1_match_1_inputs_with_fp32_type_alignment<opset10::Tan>},
         {"aten::tan_", op::inplace_op<op::translate_1to1_match_1_inputs<opset10::Tan>>},
         {"aten::tanh", op::translate_1to1_match_1_inputs_with_fp32_type_alignment<opset10::Tanh>},

tests/layer_tests/pytorch_tests/test_take_along_dim.py

@@ -0,0 +1,56 @@
+import pytest
+
+from pytorch_layer_test_class import PytorchLayerTest
+
+class TestTakeAlongDim(PytorchLayerTest):
+    def _prepare_input(self, m, n, max_val, out=False, flattenize=False):
+        import numpy as np
+        index = np.random.randint(0, max_val, (m, n) if not flattenize else (m*n, ))
+        inp = np.random.randn(m, n).astype(np.float32)
+        if out:
+            axis = int(max_val == n)
+            if flattenize:
+                out =  np.zeros_like(np.take(inp, index))
+            else:
+                out = np.zeros_like(np.take(inp, index, axis))
+            return (inp, index, out)
+        return (inp, index)
+
+    def create_model(self, axis, out):
+        import torch
+
+        class aten_take_along_dim(torch.nn.Module):
+            def __init__(self, axis, out=False):
+                super(aten_take_along_dim, self).__init__()
+                self.axis = axis
+                if self.axis is None:
+                    self.forward = self.forward_no_dim
+                if out:
+                    self.forward = self.forward_out if self.axis is not None else self.forward_no_dim_out
+
+            def forward(self, x, index):
+                return torch.take_along_dim(x, index, dim=self.axis)
+
+            def forward_out(self, x, index, out):
+                return torch.take_along_dim(x, index, dim=self.axis, out=out), out
+
+            def forward_no_dim(self, x, index):
+                return torch.take_along_dim(x, index)
+
+            def forward_no_dim_out(self, x, index, out):
+                return torch.take_along_dim(x, index, out=out)
+
+        ref_net = None
+
+        return aten_take_along_dim(axis, out), ref_net, "aten::take_along_dim"
+
+    @pytest.mark.nightly
+    @pytest.mark.precommit
+    @pytest.mark.parametrize("m", [2, 10, 100])
+    @pytest.mark.parametrize("n", [2, 10, 100])
+    @pytest.mark.parametrize("axis", [0, 1, None])
+    @pytest.mark.parametrize("out", [True, False])
+    def test_gather(self, m, n, axis, out, ie_device, precision, ir_version):
+        self._test(*self.create_model(axis, out), ie_device, precision, ir_version, kwargs_to_prepare_input={
+            "m": m, "n": n, "max_val": m if axis == 0 else n, "out": out, "flattenize": axis is None
+        })