1# Owner(s): ["module: unknown"] 2 3from functools import partial 4 5import torch 6from torch.testing._internal.common_device_type import ( 7 instantiate_device_type_tests, 8 OpDTypes, 9 ops, 10) 11from torch.testing._internal.common_methods_invocations import op_db 12from torch.testing._internal.common_utils import ( 13 run_tests, 14 TestCase, 15 TestGradients, 16 unMarkDynamoStrictTest, 17) 18from torch.testing._internal.custom_op_db import custom_op_db 19from torch.testing._internal.hop_db import hop_db 20 21 22# gradcheck requires double precision 23_gradcheck_ops = partial( 24 ops, dtypes=OpDTypes.supported, allowed_dtypes=[torch.double, torch.cdouble] 25) 26 27 28@unMarkDynamoStrictTest 29class TestBwdGradients(TestGradients): 30 # Tests that gradients are computed correctly 31 @_gradcheck_ops(op_db + hop_db + custom_op_db) 32 def test_fn_grad(self, device, dtype, op): 33 # This is verified by test_dtypes in test_ops.py 34 if dtype not in op.supported_backward_dtypes(torch.device(device).type): 35 self.skipTest("Skipped! Dtype is not in supported backward dtypes!") 36 else: 37 self._grad_test_helper(device, dtype, op, op.get_op()) 38 39 # Method grad (and gradgrad, see below) tests are disabled since they're 40 # costly and redundant with function grad (and gradgad) tests 41 # @_gradcheck_ops(op_db) 42 # def test_method_grad(self, device, dtype, op): 43 # self._skip_helper(op, device, dtype) 44 # self._grad_test_helper(device, dtype, op, op.get_method()) 45 46 @_gradcheck_ops(op_db + custom_op_db) 47 def test_inplace_grad(self, device, dtype, op): 48 self._skip_helper(op, device, dtype) 49 if not op.inplace_variant: 50 self.skipTest("Op has no inplace variant!") 51 52 # Verifies an operation doesn't support inplace autograd if it claims not to 53 if not op.supports_inplace_autograd: 54 inplace = self._get_safe_inplace(op.get_inplace()) 55 for sample in op.sample_inputs(device, dtype, requires_grad=True): 56 if sample.broadcasts_input: 57 continue 58 with self.assertRaises(Exception): 59 result = inplace(sample) 60 result.sum().backward() 61 else: 62 self._grad_test_helper( 63 device, dtype, op, self._get_safe_inplace(op.get_inplace()) 64 ) 65 66 # Test that gradients of gradients are computed correctly 67 @_gradcheck_ops(op_db + hop_db + custom_op_db) 68 def test_fn_gradgrad(self, device, dtype, op): 69 self._skip_helper(op, device, dtype) 70 if not op.supports_gradgrad: 71 self.skipTest( 72 "Op claims it doesn't support gradgrad. This is not verified." 73 ) 74 else: 75 self._check_helper(device, dtype, op, op.get_op(), "bwgrad_bwgrad") 76 77 # Test that gradients of gradients are properly raising 78 @_gradcheck_ops(op_db + custom_op_db) 79 def test_fn_fail_gradgrad(self, device, dtype, op): 80 self._skip_helper(op, device, dtype) 81 if op.supports_gradgrad: 82 self.skipTest("Skipped! Operation does support gradgrad") 83 84 err_msg = r"derivative for .* is not implemented" 85 with self.assertRaisesRegex(RuntimeError, err_msg): 86 self._check_helper(device, dtype, op, op.get_op(), "bwgrad_bwgrad") 87 88 # Method gradgrad (and grad, see above) tests are disabled since they're 89 # costly and redundant with function gradgrad (and grad) tests 90 # @_gradcheck_ops(op_db) 91 # def test_method_gradgrad(self, device, dtype, op): 92 # self._skip_helper(op, device, dtype) 93 # self._gradgrad_test_helper(device, dtype, op, op.get_method()) 94 95 @_gradcheck_ops(op_db) 96 def test_inplace_gradgrad(self, device, dtype, op): 97 self._skip_helper(op, device, dtype) 98 if not op.inplace_variant or not op.supports_inplace_autograd: 99 self.skipTest("Skipped! Operation does not support inplace autograd.") 100 self._check_helper( 101 device, dtype, op, self._get_safe_inplace(op.get_inplace()), "bwgrad_bwgrad" 102 ) 103 104 105instantiate_device_type_tests(TestBwdGradients, globals()) 106 107if __name__ == "__main__": 108 TestCase._default_dtype_check_enabled = True 109 run_tests() 110