[Accuracy diff No.40-41] [BIT] fix cosine_similarity accuracy error when x1, x2 need broadcast (#73014)

Cutelemon6 · web-flow · commit 79f3f5fc4aff · 2025-06-06T13:08:15.000+08:00
* broadcast to ensure correctness when x1.shape != x2.shape

* apply broadcast to numpy impl

* endeavor to fix performance issue

* do multiply only when m1, m2 != 1, 1

* fix coverage. CI: notest,test=coverage
diff --git a/python/paddle/nn/functional/common.py b/python/paddle/nn/functional/common.py
@@ -2227,9 +2227,15 @@ def cosine_similarity(
             [ 0.97689527,  0.99996042, -0.55138415])
 
     """
+    bs = paddle.broadcast_shape([x1.shape[axis]], [x2.shape[axis]])
     w12 = sum(paddle.multiply(x1, x2), axis=axis)
     w1 = sum(paddle.multiply(x1, x1), axis=axis)
     w2 = sum(paddle.multiply(x2, x2), axis=axis)
+    m1, m2 = bs[0] / x1.shape[axis], bs[0] / x2.shape[axis]
+    if m1 != 1:
+        w1 = w1 * m1
+    if m2 != 1:
+        w2 = w2 * m2
     n12 = sqrt(clip(w1 * w2, min=eps * eps))
     cos_sim = w12 / n12
     return cos_sim
diff --git a/test/legacy_test/test_cosine_similarity_api.py b/test/legacy_test/test_cosine_similarity_api.py
@@ -36,9 +36,15 @@ def setUp(self):
             self.places.append(paddle.CUDAPlace(0))
 
     def _get_numpy_out(self, x1, x2, axis=1, eps=1e-8):
+        bs = np.broadcast_shapes([x1.shape[axis]], [x2.shape[axis]])
         w12 = np.sum(x1 * x2, axis=axis)
         w1 = np.sum(x1 * x1, axis=axis)
         w2 = np.sum(x2 * x2, axis=axis)
+        m1, m2 = bs[0] / x1.shape[axis], bs[0] / x2.shape[axis]
+        if m1 != 1:
+            w1 = w1 * m1
+        if m2 != 1:
+            w2 = w2 * m2
         n12 = np.sqrt(np.clip(w1 * w2, eps * eps, None))
         cos_sim = w12 / n12
         return cos_sim
@@ -145,6 +151,25 @@ def test_dygraph_4(self):
 
         np.testing.assert_allclose(y.numpy(), np_out, rtol=1e-05)
 
+    def test_dygraph_5(self):
+        paddle.disable_static()
+
+        shape1 = [23, 1, 10]
+        shape2 = [23, 12, 1]
+        axis = 2
+        eps = 1e-6
+        np.random.seed(1)
+        np_x1 = np.random.rand(*shape1).astype(np.float32)
+        np_x2 = np.random.rand(*shape2).astype(np.float32)
+        np_out = self._get_numpy_out(np_x1, np_x2, axis=axis, eps=eps)
+
+        cos_sim_func = nn.CosineSimilarity(axis=axis, eps=eps)
+        tensor_x1 = paddle.to_tensor(np_x1)
+        tensor_x2 = paddle.to_tensor(np_x2)
+        y = cos_sim_func(tensor_x1, tensor_x2)
+
+        np.testing.assert_allclose(y.numpy(), np_out, rtol=1e-05)
+
 
 if __name__ == '__main__':
     unittest.main()
