From c3328288627713a1f4dc75c27e328ac6c7bb6fe4 Mon Sep 17 00:00:00 2001 From: wuyefeilin <30919197+wuyefeilin@users.noreply.github.com> Date: Thu, 22 Apr 2021 19:45:07 +0800 Subject: [PATCH] support int32 and int64 kernel for clip operator (#32373) support int32 and int64 kernel for clip operator --- paddle/fluid/operators/clip_op.cc | 8 ++++-- paddle/fluid/operators/clip_op.cu | 8 ++++-- .../fluid/tests/unittests/test_clip_op.py | 28 +++++++++++++++++-- python/paddle/tensor/math.py | 26 +++++++++++------ 4 files changed, 55 insertions(+), 15 deletions(-) diff --git a/paddle/fluid/operators/clip_op.cc b/paddle/fluid/operators/clip_op.cc index eb27df8a36..7176a0466b 100644 --- a/paddle/fluid/operators/clip_op.cc +++ b/paddle/fluid/operators/clip_op.cc @@ -145,10 +145,14 @@ REGISTER_OPERATOR(clip_grad, ops::ClipOpGrad, ops::ClipGradInplaceInferer, ops::ClipDoubleGradOpMaker); REGISTER_OP_CPU_KERNEL( clip, ops::ClipKernel, - ops::ClipKernel); + ops::ClipKernel, + ops::ClipKernel, + ops::ClipKernel); REGISTER_OP_CPU_KERNEL( clip_grad, ops::ClipGradKernel, - ops::ClipGradKernel); + ops::ClipGradKernel, + ops::ClipGradKernel, + ops::ClipGradKernel); REGISTER_OP_VERSION(clip) .AddCheckpoint( diff --git a/paddle/fluid/operators/clip_op.cu b/paddle/fluid/operators/clip_op.cu index d31b81c13c..fd61e4ea61 100644 --- a/paddle/fluid/operators/clip_op.cu +++ b/paddle/fluid/operators/clip_op.cu @@ -17,8 +17,12 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( clip, ops::ClipKernel, - ops::ClipKernel); + ops::ClipKernel, + ops::ClipKernel, + ops::ClipKernel); REGISTER_OP_CUDA_KERNEL( clip_grad, ops::ClipGradKernel, - ops::ClipGradKernel); + ops::ClipGradKernel, + ops::ClipGradKernel, + ops::ClipGradKernel); diff --git a/python/paddle/fluid/tests/unittests/test_clip_op.py b/python/paddle/fluid/tests/unittests/test_clip_op.py index 2946798a82..b05100fc7b 100644 --- a/python/paddle/fluid/tests/unittests/test_clip_op.py +++ b/python/paddle/fluid/tests/unittests/test_clip_op.py @@ -50,10 +50,14 @@ class TestClipOp(OpTest): self.outputs = {'Out': np.clip(self.inputs['X'], min_v, max_v)} def test_check_output(self): + paddle.enable_static() self.check_output() + paddle.disable_static() def test_check_grad_normal(self): + paddle.enable_static() self.check_grad(['X'], 'Out') + paddle.disable_static() def initTestCase(self): self.shape = (4, 10, 10) @@ -102,6 +106,7 @@ class TestCase5(TestClipOp): class TestClipOpError(unittest.TestCase): def test_errors(self): + paddle.enable_static() with program_guard(Program(), Program()): input_data = np.random.random((2, 4)).astype("float32") @@ -115,6 +120,7 @@ class TestClipOpError(unittest.TestCase): fluid.layers.clip(x=x2, min=-1.0, max=1.0) self.assertRaises(TypeError, test_dtype) + paddle.disable_static() class TestClipAPI(unittest.TestCase): @@ -140,7 +146,10 @@ class TestClipAPI(unittest.TestCase): out_8 = paddle.clip(images) out_9 = paddle.clip(paddle.cast(images, 'float64'), min=0.2, max=0.9) - res1, res2, res3, res4, res5, res6, res7, res8, res9 = exe.run( + out_10 = paddle.clip(paddle.cast(images * 10, 'int32'), min=2, max=8) + out_11 = paddle.clip(paddle.cast(images * 10, 'int64'), min=2, max=8) + + res1, res2, res3, res4, res5, res6, res7, res8, res9, res10, res11 = exe.run( fluid.default_main_program(), feed={ "image": data, @@ -148,7 +157,8 @@ class TestClipAPI(unittest.TestCase): "max": np.array([0.8]).astype('float32') }, fetch_list=[ - out_1, out_2, out_3, out_4, out_5, out_6, out_7, out_8, out_9 + out_1, out_2, out_3, out_4, out_5, out_6, out_7, out_8, out_9, + out_10, out_11 ]) self.assertTrue(np.allclose(res1, data.clip(0.2, 0.8))) @@ -161,8 +171,14 @@ class TestClipAPI(unittest.TestCase): self.assertTrue(np.allclose(res8, data)) self.assertTrue( np.allclose(res9, data.astype(np.float64).clip(0.2, 0.9))) + self.assertTrue( + np.allclose(res10, (data * 10).astype(np.int32).clip(2, 8))) + self.assertTrue( + np.allclose(res11, (data * 10).astype(np.int64).clip(2, 8))) + paddle.disable_static() def test_clip_dygraph(self): + paddle.disable_static() place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda( ) else fluid.CPUPlace() paddle.disable_static(place) @@ -176,9 +192,16 @@ class TestClipAPI(unittest.TestCase): out_2 = paddle.clip(images, min=0.2, max=0.9) out_3 = paddle.clip(images, min=v_min, max=v_max) + out_4 = paddle.clip(paddle.cast(images * 10, 'int32'), min=2, max=8) + out_5 = paddle.clip(paddle.cast(images * 10, 'int64'), min=2, max=8) + self.assertTrue(np.allclose(out_1.numpy(), data.clip(0.2, 0.8))) self.assertTrue(np.allclose(out_2.numpy(), data.clip(0.2, 0.9))) self.assertTrue(np.allclose(out_3.numpy(), data.clip(0.2, 0.8))) + self.assertTrue( + np.allclose(out_4.numpy(), (data * 10).astype(np.int32).clip(2, 8))) + self.assertTrue( + np.allclose(out_5.numpy(), (data * 10).astype(np.int64).clip(2, 8))) def test_errors(self): paddle.enable_static() @@ -186,6 +209,7 @@ class TestClipAPI(unittest.TestCase): x2 = fluid.data(name='x2', shape=[1], dtype="int8") self.assertRaises(TypeError, paddle.clip, x=x1, min=0.2, max=0.8) self.assertRaises(TypeError, paddle.clip, x=x2, min=0.2, max=0.8) + paddle.disable_static() if __name__ == '__main__': diff --git a/python/paddle/tensor/math.py b/python/paddle/tensor/math.py index 87efa9ac44..215d467828 100755 --- a/python/paddle/tensor/math.py +++ b/python/paddle/tensor/math.py @@ -1475,10 +1475,10 @@ def clip(x, min=None, max=None, name=None): Out = MIN(MAX(x, min), max) Args: - x (Tensor): An N-D Tensor with data type float32 or float64. - min (float32|Tensor): The lower bound with type ``float32`` or a ``Tensor`` + x (Tensor): An N-D Tensor with data type float32, float64, int32 or int64. + min (float|int|Tensor): The lower bound with type ``float`` , ``int`` or a ``Tensor`` with shape [1] and type ``int32``, ``float32``, ``float64``. - max (float32|Tensor): The upper bound with type ``float32`` or a ``Tensor`` + max (float|int|Tensor): The upper bound with type ``float``, ``int`` or a ``Tensor`` with shape [1] and type ``int32``, ``float32``, ``float64``. name (str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please @@ -1503,16 +1503,24 @@ def clip(x, min=None, max=None, name=None): # [[4.5, 6.4] """ - fmin = float(np.finfo(np.float32).min) - fmax = float(np.finfo(np.float32).max) + x_dtype = str(x.dtype) + if x_dtype == 'paddle.int32': + min_ = np.iinfo(np.int32).min + max_ = np.iinfo(np.int32).max - 2**7 + elif x_dtype == 'paddle.int64': + min_ = np.iinfo(np.int64).min + max_ = np.iinfo(np.int64).max - 2**39 + else: + min_ = float(np.finfo(np.float32).min) + max_ = float(np.finfo(np.float32).max) if in_dygraph_mode(): if isinstance(min, Variable): min = min.numpy().item(0) if isinstance(max, Variable): max = max.numpy().item(0) - min = fmin if min is None else min - max = fmax if max is None else max + min = min_ if min is None else min + max = max_ if max is None else max return core.ops.clip(x, "min", min, "max", max) if min is not None: @@ -1526,10 +1534,10 @@ def clip(x, min=None, max=None, name=None): check_dtype(max.dtype, 'max', ['float32', 'float64', 'int32'], 'clip', '(When the type of max in clip is Variable.)') - check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'clip') + check_variable_and_dtype(x, 'x', ['float32', 'float64', 'int32', 'int64'], 'clip') inputs = {'X': x} - attrs = {'min': fmin, 'max': fmax} + attrs = {'min': min_, 'max': max_} if isinstance(min, Variable): min.stop_gradient = True -- GitLab