Fix unit tests to avoid check_grad checking failures (#21554)

* fix python API tests that do not need to inherit OpTest, test=develop

* fix fp16 cases that will only be enabled in GPU mode, test=develop

* remove TestSoftmaxFP16Op from test cases of softmax_mkldnn_op, test=develop

* fix tests so that the cases are only created in GPU mode, test=develop

python/paddle/fluid/tests/unittests/mkldnn/test_softmax_mkldnn_op.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 from paddle.fluid.tests.unittests.op_test import OpTest
 import paddle.fluid.core as core
-from paddle.fluid.tests.unittests.test_softmax_op import *
+from paddle.fluid.tests.unittests.test_softmax_op import TestSoftmaxOp, TestSoftmaxOp2, TestSoftmaxOp3, TestSoftmaxOp4, TestSoftmaxOp5, TestSoftmaxOp6
 from mkldnn_op_test import check_if_mkldnn_primitives_exist_in_bwd
 
 

python/paddle/fluid/tests/unittests/test_assign_op.py

@@ -37,7 +37,7 @@ class TestAssignOp(op_test.OpTest):
         self.check_grad(['X'], 'Out')
 
 
-class TestAssignOpError(op_test.OpTest):
+class TestAssignOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
             # The type of input must be Variable or numpy.ndarray.

python/paddle/fluid/tests/unittests/test_coalesce_tensor_op.py

@@ -22,6 +22,8 @@ from paddle.fluid import core
 alignment = 256
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestAllocContinuousSpace(OpTest):
     def setUp(self):
         self.op_type = "coalesce_tensor"
@@ -78,13 +80,12 @@ class TestAllocContinuousSpace(OpTest):
         return outputs, coalesce_tensor_var
 
     def test_check_output(self):
-        if core.is_compiled_with_cuda():
         self.check_output_with_place(
-                place=core.CUDAPlace(0),
-                no_check_set=["FusedOutput"],
-                atol=1e-5)
+            place=core.CUDAPlace(0), no_check_set=["FusedOutput"], atol=1e-5)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestAllocContinuousSpace2(TestAllocContinuousSpace):
     def init_attr(self):
         return {
@@ -95,11 +96,8 @@ class TestAllocContinuousSpace2(TestAllocContinuousSpace):
         }
 
     def test_check_output(self):
-        if core.is_compiled_with_cuda():
         self.check_output_with_place(
-                place=core.CUDAPlace(0),
-                no_check_set=["FusedOutput"],
-                atol=1e-5)
+            place=core.CUDAPlace(0), no_check_set=["FusedOutput"], atol=1e-5)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/unittests/test_concat_op.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 from op_test import OpTest
 import paddle.fluid as fluid
-from paddle.fluid import compiler, Program, program_guard
+from paddle.fluid import compiler, Program, program_guard, core
 
 
 class TestConcatOp(OpTest):
@@ -134,6 +134,8 @@ create_test_AxisTensor(TestConcatOp5)
 
 
 def create_test_fp16(parent):
+    @unittest.skipIf(not core.is_compiled_with_cuda(),
+                     "core is not compiled with CUDA")
     class TestConcatFp16(parent):
         def get_dtype(self):
             return np.float16

python/paddle/fluid/tests/unittests/test_conv2d_op.py

@@ -697,7 +697,7 @@ class TestCUDNNExhaustiveSearch(TestConv2dOp):
         self.exhaustive_search = True
 
 
-class TestConv2dOpError(OpTest):
+class TestConv2dOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
 

python/paddle/fluid/tests/unittests/test_conv3d_op.py

@@ -210,6 +210,8 @@ def create_test_channel_last_class(parent):
 
 
 def create_test_cudnn_channel_last_class(parent):
+    @unittest.skipIf(not core.is_compiled_with_cuda(),
+                     "core is not compiled with CUDA")
     class TestCudnnChannelLastCase(parent):
         def init_kernel_type(self):
             self.use_cudnn = True
@@ -403,11 +405,15 @@ class TestWithDilation(TestConv3dOp):
 #---------------- Conv3dCUDNN ----------------
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCUDNN(TestConv3dOp):
     def init_kernel_type(self):
         self.use_cudnn = True
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16CUDNN(TestConv3dOp):
     def init_kernel_type(self):
         self.use_cudnn = True
@@ -420,11 +426,15 @@ class TestFP16CUDNN(TestConv3dOp):
                 self.check_output_with_place(place, atol=2e-2)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestWithGroup1CUDNN(TestWithGroup1):
     def init_kernel_type(self):
         self.use_cudnn = True
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16WithGroup1CUDNN(TestWithGroup1):
     def init_kernel_type(self):
         self.use_cudnn = True
@@ -437,11 +447,15 @@ class TestFP16WithGroup1CUDNN(TestWithGroup1):
                 self.check_output_with_place(place, atol=2e-2)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestWithGroup2CUDNN(TestWithGroup2):
     def init_kernel_type(self):
         self.use_cudnn = True
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16WithGroup2CUDNN(TestWithGroup2):
     def init_kernel_type(self):
         self.use_cudnn = True
@@ -454,11 +468,15 @@ class TestFP16WithGroup2CUDNN(TestWithGroup2):
                 self.check_output_with_place(place, atol=2e-2)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestWith1x1CUDNN(TestWith1x1):
     def init_kernel_type(self):
         self.use_cudnn = True
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16With1x1CUDNN(TestWith1x1):
     def init_kernel_type(self):
         self.use_cudnn = True
@@ -471,11 +489,15 @@ class TestFP16With1x1CUDNN(TestWith1x1):
                 self.check_output_with_place(place, atol=2e-2)
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestWithInput1x1Filter1x1CUDNN(TestWithInput1x1Filter1x1):
     def init_kernel_type(self):
         self.use_cudnn = True
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16WithInput1x1Filter1x1CUDNN(TestWithInput1x1Filter1x1):
     def init_kernel_type(self):
         self.use_cudnn = True

python/paddle/fluid/tests/unittests/test_cross_entropy_op.py

@@ -359,7 +359,7 @@ create_test_class(TestCrossEntropyOp7RemoveLastDim,
                   "TestCrossEntropyF16Op7RemoveLastDim")
 
 
-class TestCrossEntropyOpError(OpTest):
+class TestCrossEntropyOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
 

python/paddle/fluid/tests/unittests/test_data_norm_op.py

@@ -278,7 +278,7 @@ class TestDataNormOpWithSlotDim(OpTest):
         self.check_grad(['X'], 'Y', no_grad_set=set([]), check_dygraph=False)
 
 
-class TestDataNormOpWithSyncStats(OpTest):
+class TestDataNormOpWithSyncStats(unittest.TestCase):
     """
     test class for data norm op
     test forward and backward

python/paddle/fluid/tests/unittests/test_elementwise_add_op.py

@@ -89,6 +89,8 @@ class TestElementwiseAddOp(OpTest):
         pass
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16ElementwiseAddOp(TestElementwiseAddOp):
     def init_dtype(self):
         self.dtype = np.float16

python/paddle/fluid/tests/unittests/test_elementwise_div_op.py

@@ -15,6 +15,7 @@
 from __future__ import print_function
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 from op_test import OpTest
 
 
@@ -204,6 +205,8 @@ class TestElementwiseDivOp_INT(OpTest):
         pass
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestElementwiseDivOpFp16(ElementwiseDivOp):
     def init_dtype(self):
         self.dtype = np.float16

python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py

@@ -177,6 +177,8 @@ class TestElementwiseMulOp_broadcast_5(ElementwiseMulOp):
         self.init_kernel_type()
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestElementwiseMulOpFp16(ElementwiseMulOp):
     def init_dtype(self):
         self.dtype = np.float16

python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py

@@ -90,7 +90,7 @@ class TestFillConstantBatchSizeLikeWithLoDTensor(OpTest):
 
 
 # Test python API
-class TestFillConstantBatchSizeLikeAPI(OpTest):
+class TestFillConstantBatchSizeLikeAPI(unittest.TestCase):
     def test_api(self):
         like = fluid.layers.fill_constant(
             shape=[1, 200], value=10, dtype='int64')

python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py

@@ -325,6 +325,25 @@ for mode in {0, 1}:
             'functor_list': ["scale", "elementwise_add"],
             'save_intermediate_out': save_intermediate_out,
         })
+        create_test_class('add_scale' + suffix, add_scale_func, {
+            'scale': scale,
+            'functor_list': ["elementwise_add", "scale"],
+            'save_intermediate_out': save_intermediate_out,
+        })
+        create_test_class('add_relu' + suffix, add_relu_func, {
+            'functor_list': ["elementwise_add", "relu"],
+            'save_intermediate_out': save_intermediate_out,
+        })
+        create_test_class('relu_add' + suffix, relu_add_func, {
+            'functor_list': ["relu", "elementwise_add"],
+            'save_intermediate_out': save_intermediate_out,
+        })
+        create_test_class('mul_scale' + suffix, mul_scale_func, {
+            'scale': scale,
+            'functor_list': ["elementwise_mul", "scale"],
+            'save_intermediate_out': save_intermediate_out,
+        })
+        if core.is_compiled_with_cuda():
             create_test_class(
                 'scale_add_fp16' + suffix,
                 scale_add_func, {
@@ -334,11 +353,6 @@ for mode in {0, 1}:
                 },
                 dtype=np.float16,
                 grad_chek=False)
-        create_test_class('add_scale' + suffix, add_scale_func, {
-            'scale': scale,
-            'functor_list': ["elementwise_add", "scale"],
-            'save_intermediate_out': save_intermediate_out,
-        })
             create_test_class(
                 'add_scale_fp16' + suffix,
                 add_scale_func, {
@@ -348,14 +362,7 @@ for mode in {0, 1}:
                 },
                 dtype=np.float16,
                 grad_chek=False)
-        create_test_class('add_relu' + suffix, add_relu_func, {
-            'functor_list': ["elementwise_add", "relu"],
-            'save_intermediate_out': save_intermediate_out,
-        })
-        create_test_class('relu_add' + suffix, relu_add_func, {
-            'functor_list': ["relu", "elementwise_add"],
-            'save_intermediate_out': save_intermediate_out,
-        })
+
             create_test_class(
                 'add_relu_fp16' + suffix,
                 add_relu_func, {
@@ -372,13 +379,8 @@ for mode in {0, 1}:
                 },
                 dtype=np.float16,
                 grad_chek=False)
-        create_test_class('mul_scale' + suffix, mul_scale_func, {
-            'scale': scale,
-            'functor_list': ["elementwise_mul", "scale"],
-            'save_intermediate_out': save_intermediate_out,
-        })
             create_test_class(
-            'mul_scale' + suffix,
+                'mul_scale_fp16' + suffix,
                 mul_scale_func, {
                     'scale': scale,
                     'functor_list': ["elementwise_mul", "scale"],

python/paddle/fluid/tests/unittests/test_fusion_transpose_flatten_concat_op.py

@@ -20,6 +20,8 @@ from op_test import OpTest
 import paddle.fluid.core as core
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFusionTransposeFlattenConcationOp(OpTest):
     def setUp(self):
         self.init_test_case()
@@ -48,11 +50,8 @@ class TestFusionTransposeFlattenConcationOp(OpTest):
         self.outputs = {'Out': out}
 
     def test_check_output(self):
-        if core.is_compiled_with_cuda():
         place = core.CUDAPlace(0)
         self.check_output_with_place(place, 1e-6)
-        else:
-            pass
 
     def init_test_case(self):
         self.shapes = [(3, 4, 17, 17), (3, 8, 7, 7), (3, 12, 5, 5)]
@@ -61,6 +60,8 @@ class TestFusionTransposeFlattenConcationOp(OpTest):
         self.concat_axis = 1
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCase1(TestFusionTransposeFlattenConcationOp):
     def init_test_case(self):
         self.shapes = [(3, 4, 18, 17), (3, 8, 18, 7), (6, 12, 9, 5)]
@@ -69,6 +70,8 @@ class TestCase1(TestFusionTransposeFlattenConcationOp):
         self.concat_axis = 1
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCase2(TestFusionTransposeFlattenConcationOp):
     def init_test_case(self):
         self.shapes = [(3, 8, 20, 17), (3, 8, 19, 17), (3, 8, 40, 17)]
@@ -77,6 +80,8 @@ class TestCase2(TestFusionTransposeFlattenConcationOp):
         self.concat_axis = 0
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCase3(TestFusionTransposeFlattenConcationOp):
     def init_test_case(self):
         self.shapes = [(3, 8, 20, 17), (3, 8, 19, 17), (3, 8, 40, 17)]
@@ -85,6 +90,8 @@ class TestCase3(TestFusionTransposeFlattenConcationOp):
         self.concat_axis = 1
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCase4(TestFusionTransposeFlattenConcationOp):
     def init_test_case(self):
         self.shapes = [(3, 8, 9, 17), (8, 3, 9, 17), (4, 6, 9, 17)]
@@ -93,6 +100,8 @@ class TestCase4(TestFusionTransposeFlattenConcationOp):
         self.concat_axis = 1
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCase5(TestFusionTransposeFlattenConcationOp):
     def init_test_case(self):
         self.shapes = [(3, 8, 9, 17, 2), (3, 8, 2, 17, 9), (3, 17, 9, 8, 2)]

python/paddle/fluid/tests/unittests/test_isfinite_op.py

@@ -14,6 +14,7 @@
 
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 from op_test import OpTest
 
 
@@ -37,6 +38,8 @@ class TestInf(OpTest):
         self.check_output()
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16Inf(TestInf):
     def init_dtype(self):
         self.dtype = np.float16
@@ -62,6 +65,8 @@ class TestNAN(OpTest):
         self.check_output()
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16NAN(TestNAN):
     def init_dtype(self):
         self.dtype = np.float16
@@ -88,6 +93,8 @@ class TestIsfinite(OpTest):
         self.check_output()
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestFP16Isfinite(TestIsfinite):
     def init_dtype(self):
         self.dtype = np.float16

python/paddle/fluid/tests/unittests/test_lstm_cudnn_op.py

@@ -116,6 +116,8 @@ def lstm_naive(
     return output, pre_h, pre_c
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestCUDNNLstmOp(OpTest):
     def setUp(self):
         self.op_type = "cudnn_lstm"
@@ -172,24 +174,18 @@ class TestCUDNNLstmOp(OpTest):
 
     def test_output_with_place(self):
         # depend on the scope structure
-        if self.has_cuda():
         place = core.CUDAPlace(0)
         self.check_output_with_place(place, atol=1e-5, check_dygraph=False)
 
     def test_grad_with_place(self):
         # depend on the scope structure
-        if core.is_compiled_with_cuda():
         place = core.CUDAPlace(0)
         self.check_grad_with_place(
             place,
-                set(['Input', 'W', 'InitH', 'InitC']),
-                ['Out', 'last_h', 'last_c'],
+            set(['Input', 'W', 'InitH', 'InitC']), ['Out', 'last_h', 'last_c'],
             max_relative_error=0.02,
             check_dygraph=False)
 
-    def has_cuda(self):
-        return core.is_compiled_with_cuda()
-
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_matmul_op.py

@@ -114,7 +114,7 @@ class Generator(object):
             ['X'], 'Out', max_relative_error=1e-3, no_grad_set=set('Y'))
 
 
-class TestMatmulOpError(OpTest):
+class TestMatmulOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
             # The inputs type of matmul_op must be Variable.
@@ -173,7 +173,7 @@ def api_test(dim_x, dim_y, trans_x, trans_y):
         dim_x, dim_y, trans_x, trans_y))
     shape_x, shape_y = generate_compatible_shapes(dim_x, dim_y, trans_x,
                                                   trans_y)
-    globals()[test_name] = type(test_name, (OpTest, ), {
+    globals()[test_name] = type(test_name, (unittest.TestCase, ), {
         'shape_X': shape_x,
         'shape_Y': shape_y,
         'transpose_X': trans_x,

python/paddle/fluid/tests/unittests/test_nce.py

@@ -220,7 +220,7 @@ class TestNCECase1SelectedRows(unittest.TestCase):
         self.assertEqual(rets[0], rets[1])
 
 
-class TestNCE_OpError(OpTest):
+class TestNCE_OpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
             input1 = fluid.create_lod_tensor(

python/paddle/fluid/tests/unittests/test_pad_op.py

@@ -17,6 +17,7 @@ from __future__ import print_function
 import unittest
 import numpy as np
 from op_test import OpTest
+import paddle.fluid.core as core
 
 
 class TestPadOp(OpTest):
@@ -75,6 +76,8 @@ class TestCase3(TestPadOp):
 
 
 def create_test_fp16(parent):
+    @unittest.skipIf(not core.is_compiled_with_cuda(),
+                     "core is not compiled with CUDA")
     class TestPadFp16(parent):
         def get_dtype(self):
             return np.float16

python/paddle/fluid/tests/unittests/test_sign_op.py

@@ -36,7 +36,7 @@ class TestSignOp(OpTest):
         self.check_grad(['X'], 'Out')
 
 
-class TestSignOpError(OpTest):
+class TestSignOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
             # The input type of sign_op must be Variable or numpy.ndarray.

python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 
 from op_test import OpTest
 from test_softmax_op import stable_softmax
@@ -106,6 +107,8 @@ class TestSoftmaxWithCrossEntropyOpNoCudnn(TestSoftmaxWithCrossEntropyOp):
         self.dtype = np.float64
 
 
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
 class TestSoftmaxWithCrossEntropyOpFp16(TestSoftmaxWithCrossEntropyOp):
     def initParams(self):
         self.op_type = "softmax_with_cross_entropy"

python/paddle/fluid/tests/unittests/test_split_op.py

@@ -18,7 +18,7 @@ import unittest
 import numpy as np
 from op_test import OpTest
 import paddle.fluid as fluid
-from paddle.fluid import compiler, Program, program_guard
+from paddle.fluid import compiler, Program, program_guard, core
 
 
 class TestSplitOp(OpTest):
@@ -210,6 +210,8 @@ class TestSplitByrefOp(OpTest):
 
 
 def create_test_fp16(parent):
+    @unittest.skipIf(not core.is_compiled_with_cuda(),
+                     "core is not compiled with CUDA")
     class TestSplitFp16(parent):
         def get_dtype(self):
             return np.float16

python/paddle/fluid/tests/unittests/test_transpose_op.py

@@ -80,7 +80,7 @@ class TestCase4(TestTransposeOp):
         self.axis = (4, 2, 3, 1, 0, 5)
 
 
-class TestTransposeOpError(OpTest):
+class TestTransposeOpError(unittest.TestCase):
     def test_errors(self):
         with program_guard(Program(), Program()):
             x = fluid.layers.data(name='x', shape=[10, 5, 3], dtype='float32')