From 977b42612c5128640e97c950726b7eaa74b1ae9e Mon Sep 17 00:00:00 2001
From: Ekaterina Aidova <ekaterina.aidova@intel.com>
Date: Thu, 4 Jul 2024 11:52:01 +0400
Subject: [PATCH] [PT FE]: update layer norm for support non-static normalized
 shape (#25365)

### Details:
- *aten::layer_norm support extended for support normalized shapes
represented as non constant and multiple dims*
 tested with stable-audio model

### Tickets:
 - TBD
---
 src/frontends/pytorch/src/op/layer_norm.cpp   | 17 +++---
 .../pytorch_tests/test_layer_norm.py          | 53 +++++++++++++++++++
 2 files changed, 64 insertions(+), 6 deletions(-)
 create mode 100644 tests/layer_tests/pytorch_tests/test_layer_norm.py

diff --git a/src/frontends/pytorch/src/op/layer_norm.cpp b/src/frontends/pytorch/src/op/layer_norm.cpp
index 8775d30440ca2f..79464fa2d6d609 100644
--- a/src/frontends/pytorch/src/op/layer_norm.cpp
+++ b/src/frontends/pytorch/src/op/layer_norm.cpp
@@ -7,6 +7,9 @@
 #include "openvino/op/constant.hpp"
 #include "openvino/op/multiply.hpp"
 #include "openvino/op/mvn.hpp"
+#include "openvino/op/range.hpp"
+#include "openvino/op/shape_of.hpp"
+#include "openvino/op/squeeze.hpp"
 #include "openvino/op/util/framework_node.hpp"
 #include "utils.hpp"
 
@@ -20,12 +23,14 @@ using namespace ov::op;
 OutputVector translate_layer_norm(const NodeContext& context) {
     num_inputs_check(context, 5, 6);
     auto eps = context.const_input<float>(4);
-    auto normalized_shape = context.const_input<Shape>(1);
-    PYTORCH_OP_CONVERSION_CHECK(normalized_shape.size() == 1,
-                                "Translation for aten::layer_norm supports only single normalized_shape value, "
-                                "which means normalizing over the last dimension.");
-    // TODO: support any dimension
-    auto axes = context.mark_node(v0::Constant::create(element::i32, Shape{1}, {-1}));
+    auto normalized_shape = context.get_input(1);
+    auto num_axes = context.mark_node(std::make_shared<v3::ShapeOf>(normalized_shape, element::i32));
+    num_axes = context.mark_node(std::make_shared<v0::Squeeze>(num_axes));
+    auto zero = context.mark_node(v0::Constant::create(element::i32, Shape{}, {0}));
+    auto minus_one = context.mark_node(v0::Constant::create(element::i32, Shape{}, {-1}));
+    auto axes_range = context.mark_node(std::make_shared<v4::Range>(num_axes, zero, minus_one, element::i32));
+
+    auto axes = context.mark_node(std::make_shared<v1::Multiply>(axes_range, minus_one));
     auto out_node =
         context.mark_node(std::make_shared<v6::MVN>(context.get_input(0), axes, true, eps, MVNEpsMode::INSIDE_SQRT));
     if (!context.input_is_none(2)) {
diff --git a/tests/layer_tests/pytorch_tests/test_layer_norm.py b/tests/layer_tests/pytorch_tests/test_layer_norm.py
new file mode 100644
index 00000000000000..3bba4a31dab0a4
--- /dev/null
+++ b/tests/layer_tests/pytorch_tests/test_layer_norm.py
@@ -0,0 +1,53 @@
+# Copyright (C) 2018-2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+
+from pytorch_layer_test_class import PytorchLayerTest
+import numpy as np
+
+
+class TestLayerNorm(PytorchLayerTest):
+    def _prepare_input(self):
+        import numpy as np
+        return (np.random.randn(20, 5, 10, 10).astype(np.float32),)
+
+    def create_model(self, normalized_shape, weight, bias, eps):
+        import torch
+        import torch.nn.functional as F
+
+        if weight == "ones":
+            weight = torch.ones(normalized_shape)
+        
+        if weight == "random":
+            weight = torch.randn(normalized_shape)
+        
+        if bias == "zeros":
+            bias = torch.zeros(normalized_shape)
+
+        if bias == "random":
+            bias = torch.randn(normalized_shape)
+
+        class aten_layer_norm(torch.nn.Module):
+            def __init__(self, normalized_shape, weight, bias, eps):
+                super(aten_layer_norm, self).__init__()
+                self.normalized_shape = normalized_shape
+                self.weight = weight
+                self.bias = bias
+                self.eps = eps
+
+            def forward(self, x):
+                return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+
+        ref_net = None
+
+        return aten_layer_norm(normalized_shape, weight, bias, eps), ref_net, "aten::layer_norm"
+
+    @pytest.mark.parametrize("normalized_shape", [[10,], [10, 10], [5, 10, 10]])
+    @pytest.mark.parametrize("weight", [None, "ones", "random"])
+    @pytest.mark.parametrize("bias", [None, "zeros", "random"])
+    @pytest.mark.parametrize("eps", [1e-5, 0.005])
+    @pytest.mark.nightly
+    @pytest.mark.precommit
+    def test_layer_norm(self, normalized_shape, weight, bias, eps, ie_device, precision, ir_version):
+        self._test(*self.create_model(normalized_shape, weight, bias, eps), ie_device, precision, ir_version)
\ No newline at end of file