diff --git a/paddle/fluid/operators/nll_loss_op.cc b/paddle/fluid/operators/nll_loss_op.cc index a4e1f7b3091a9f692e479300310333bfdd359096..8f14bc10d50942f55e29f196e9ca3f35e8f71d14 100644 --- a/paddle/fluid/operators/nll_loss_op.cc +++ b/paddle/fluid/operators/nll_loss_op.cc @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/infershape_utils.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/phi/infermeta/backward.h" #include "paddle/phi/infermeta/ternary.h" namespace paddle { @@ -94,68 +95,6 @@ class NLLLossGradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { - OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "NLLLoss"); - OP_INOUT_CHECK(ctx->HasInput("Label"), "Input", "Label", "NLLLoss"); - OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input", - framework::GradVarName("Out"), "NLLLoss"); - OP_INOUT_CHECK(ctx->HasOutput(framework::GradVarName("X")), "Output", - framework::GradVarName("X"), "NLLLoss"); - - auto reduction = ctx->Attrs().Get("reduction"); - auto x_dims = ctx->GetInputDim("X"); - auto label_dims = ctx->GetInputDim("Label"); - auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out")); - bool contain_unknown_dim = - phi::contain_unknown_dim(x_dims) || phi::contain_unknown_dim(dout_dims); - bool check = ctx->IsRuntime() || !contain_unknown_dim; - - if (check) { - auto batch_size = x_dims[0]; - if (x_dims.size() == 2) { - PADDLE_ENFORCE_EQ(dout_dims.size(), 1, - platform::errors::InvalidArgument( - "The dimensions of Input(Out@Grad) must be 1")); - if (reduction == "none") { - PADDLE_ENFORCE_EQ( - dout_dims[0], batch_size, - platform::errors::InvalidArgument( - "The unreduced size ofInput(Out@Grad) must be the " - "same as batch_size.")); - } else { - PADDLE_ENFORCE_EQ( - dout_dims[0], 1, - platform::errors::InvalidArgument( - "The reduced size of Input(Out@Grad) must be 1")); - } - } else if (x_dims.size() == 4) { - if (reduction == "none") { - PADDLE_ENFORCE_EQ( - dout_dims.size(), 3, - platform::errors::InvalidArgument( - "The dimensions of Input(Out@Grad) must be 3,But got [%s].", - dout_dims.size())); - PADDLE_ENFORCE_EQ( - dout_dims[0] == label_dims[0] && dout_dims[1] == label_dims[1] && - dout_dims[2] == label_dims[2], - true, platform::errors::InvalidArgument( - "The dimensions of Input(Out@Grad) must be match " - "to Input(Label) dimensions.")); - } else { - PADDLE_ENFORCE_EQ( - dout_dims[0], 1, - platform::errors::InvalidArgument( - "The reduced size of Input(Out@Grad) must be 1")); - } - } - } - - auto x_grad_name = framework::GradVarName("X"); - if (ctx->HasOutput(x_grad_name)) { - ctx->SetOutputDim(x_grad_name, x_dims); - } - } - protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { @@ -192,9 +131,12 @@ class NLLLossGradMaker : public framework::SingleGradOpMaker { DECLARE_INFER_SHAPE_FUNCTOR(nll_loss, NllLossRawInferShapeFunctor, PD_INFER_META(phi::NllLossRawInferMeta)); +DECLARE_INFER_SHAPE_FUNCTOR(nll_loss_grad, NllLossGradInferShapeFunctor, + PD_INFER_META(phi::NllLossGradInferMeta)); namespace ops = paddle::operators; REGISTER_OPERATOR(nll_loss, ops::NLLLossOp, ops::NLLLossOpMaker, ops::NLLLossGradMaker, ops::NLLLossGradMaker, NllLossRawInferShapeFunctor); -REGISTER_OPERATOR(nll_loss_grad, ops::NLLLossGradOp); +REGISTER_OPERATOR(nll_loss_grad, ops::NLLLossGradOp, + NllLossGradInferShapeFunctor); diff --git a/paddle/phi/api/lib/data_transform.cc b/paddle/phi/api/lib/data_transform.cc index 4e6ebe33aec8fb4a218caed6426e9331f15b3b86..c1fc0fd907bba836886e26002f03562ae578ef6b 100644 --- a/paddle/phi/api/lib/data_transform.cc +++ b/paddle/phi/api/lib/data_transform.cc @@ -180,20 +180,23 @@ std::shared_ptr PrepareData( const phi::TensorArgDef& target_args_def, const TransformFlag& transform_flag) { const auto& tensor_in = input.impl(); - phi::DenseTensor& dense_tensor = - *static_cast(tensor_in.get()); - if (!transform_flag.NeedTransform() || !dense_tensor.initialized() || - (!NeedTransformPlace( - dense_tensor.place(), target_args_def.backend, transform_flag) && - !NeedTransformDataType( - dense_tensor.dtype(), target_args_def.dtype, transform_flag) && - !NeedTransformLayout( - dense_tensor.layout(), target_args_def.layout, transform_flag))) { - return std::static_pointer_cast(tensor_in); + if (tensor_in) { + phi::DenseTensor& dense_tensor = + *static_cast(tensor_in.get()); + if (!transform_flag.NeedTransform() || !dense_tensor.initialized() || + (!NeedTransformPlace( + dense_tensor.place(), target_args_def.backend, transform_flag) && + !NeedTransformDataType( + dense_tensor.dtype(), target_args_def.dtype, transform_flag) && + !NeedTransformLayout( + dense_tensor.layout(), target_args_def.layout, transform_flag))) { + return std::static_pointer_cast(tensor_in); + } + phi::DenseTensor out = + TransformData(dense_tensor, target_args_def, transform_flag); + return std::make_shared(std::move(out)); } - phi::DenseTensor out = - TransformData(dense_tensor, target_args_def, transform_flag); - return std::make_shared(std::move(out)); + return nullptr; } std::shared_ptr PrepareData( diff --git a/paddle/phi/infermeta/backward.cc b/paddle/phi/infermeta/backward.cc index 5d9ed8e9e8c879487368e2e4b8464c7a810ea076..e7682d78a14a1ecef09ba97d64125cd268fa86e5 100644 --- a/paddle/phi/infermeta/backward.cc +++ b/paddle/phi/infermeta/backward.cc @@ -180,6 +180,72 @@ void MaxPoolWithIndexGradInferMeta(const MetaTensor& x, dx->share_meta(x); } +void NllLossGradInferMeta(const MetaTensor& x, + const MetaTensor& label, + paddle::optional weight, + const MetaTensor& total_weight, + const MetaTensor& out_grad, + int64_t ignore_index, + const std::string& reduction, + MetaTensor* dx, + MetaConfig config) { + const auto& x_dims = x.dims(); + const auto& label_dims = label.dims(); + const auto& dout_dims = out_grad.dims(); + bool contain_unknown_dim = + phi::contain_unknown_dim(x_dims) || phi::contain_unknown_dim(dout_dims); + bool check = config.is_runtime || !contain_unknown_dim; + + if (check) { + auto batch_size = x_dims[0]; + if (x_dims.size() == 2) { + PADDLE_ENFORCE_EQ(dout_dims.size(), + 1, + phi::errors::InvalidArgument( + "The dimensions of Input(Out@Grad) must be 1")); + if (reduction == "none") { + PADDLE_ENFORCE_EQ( + dout_dims[0], + batch_size, + phi::errors::InvalidArgument( + "The unreduced size ofInput(Out@Grad) must be the " + "same as batch_size.")); + } else { + PADDLE_ENFORCE_EQ(dout_dims[0], + 1, + phi::errors::InvalidArgument( + "The reduced size of Input(Out@Grad) must be 1")); + } + } else if (x_dims.size() == 4) { + if (reduction == "none") { + PADDLE_ENFORCE_EQ( + dout_dims.size(), + 3, + phi::errors::InvalidArgument( + "The dimensions of Input(Out@Grad) must be 3,But got [%s].", + dout_dims.size())); + PADDLE_ENFORCE_EQ(dout_dims[0] == label_dims[0] && + dout_dims[1] == label_dims[1] && + dout_dims[2] == label_dims[2], + true, + phi::errors::InvalidArgument( + "The dimensions of Input(Out@Grad) must be match " + "to Input(Label) dimensions.")); + } else { + PADDLE_ENFORCE_EQ(dout_dims[0], + 1, + phi::errors::InvalidArgument( + "The reduced size of Input(Out@Grad) must be 1")); + } + } + } + + if (dx) { + dx->set_dims(x_dims); + dx->set_dtype(x.dtype()); + } +} + void PoolGradInferMeta(const MetaTensor& x, const MetaTensor& out, const MetaTensor& dout, diff --git a/paddle/phi/infermeta/backward.h b/paddle/phi/infermeta/backward.h index 10b3e7cec7d2e5b22ec789f063b10f752f85e4ec..4cdc048b24964796bfd2ebeace0474e8e10f31b5 100644 --- a/paddle/phi/infermeta/backward.h +++ b/paddle/phi/infermeta/backward.h @@ -104,6 +104,16 @@ void MaxPoolWithIndexGradInferMeta(const MetaTensor& x, bool adaptive, MetaTensor* dx); +void NllLossGradInferMeta(const MetaTensor& input, + const MetaTensor& label, + paddle::optional weight, + const MetaTensor& total_weight, + const MetaTensor& out_grad, + int64_t ignore_index, + const std::string& reduction, + MetaTensor* intput_grad, + MetaConfig config = MetaConfig()); + void PsroiPoolGradInferMeta(const MetaTensor& x, const MetaTensor& rois, paddle::optional rois_num, diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index 8d14516374038a87dcf73544c71abcbcf4298b04..be883d243f79585ce697e01d2170a7bdf804e1d2 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -710,6 +710,8 @@ class OpTest(unittest.TestCase): def prepare_python_api_arguments(api, op_proto_ins, op_proto_attrs, kernel_sig): """ map from `op proto inputs and attrs` to `api input list and api attrs dict` + + NOTE: the op_proto_attrs and op_proto_ins is a default dict. default value is [] """ class Empty: @@ -770,7 +772,9 @@ class OpTest(unittest.TestCase): api_params), "Error happens. contack xiongkun03 to solve." inputs_sig, attrs_sig, outputs_sig = kernel_sig inputs_and_attrs = inputs_sig + attrs_sig - input_arguments = [op_proto_ins[name] for name in inputs_sig] + [ + input_arguments = [ + op_proto_ins.get(name, Empty()) for name in inputs_sig + ] + [ parse_attri_value(name, op_proto_ins, op_proto_attrs) for name in attrs_sig ] @@ -814,16 +818,19 @@ class OpTest(unittest.TestCase): transform inputs by the following rules: 1. [Tensor] -> Tensor 2. [Tensor, Tensor, ...] -> list of Tensors + 3. None -> None + 4. Others: raise Error only support "X" is list of Tensor, currently don't support other structure like dict. """ - for inp in args[:inp_num]: + inp_args = [[inp] if inp is None else inp + for inp in args[:inp_num]] # convert None -> [None] + for inp in inp_args: assert isinstance( inp, list ), "currently only support `X` is [Tensor], don't support other structure." - args = [ - inp[0] if len(inp) == 1 else inp for inp in args[:inp_num] - ] + args[inp_num:] + args = [inp[0] if len(inp) == 1 else inp + for inp in inp_args] + args[inp_num:] return args def _get_kernel_signature(eager_tensor_inputs, eager_tensor_outputs, diff --git a/python/paddle/fluid/tests/unittests/test_nll_loss.py b/python/paddle/fluid/tests/unittests/test_nll_loss.py index 0bc5e1cad9acd09995142f6e32d76103efbeaae7..c53fdffe1cf1b1d8b8022b92fb12ab6ea5902007 100644 --- a/python/paddle/fluid/tests/unittests/test_nll_loss.py +++ b/python/paddle/fluid/tests/unittests/test_nll_loss.py @@ -763,6 +763,8 @@ class TestNLLLossOp1DWithReduce(OpTest): def setUp(self): self.init_test_case() self.op_type = "nll_loss" + self.python_api = paddle.nn.functional.nll_loss + self.python_out_sig = ["Out"] self.with_weight = False self.python_api = paddle.nn.functional.nll_loss self.python_out_sig = ["Out"] @@ -786,19 +788,19 @@ class TestNLLLossOp1DWithReduce(OpTest): self.attrs = {'reduction': 'mean', 'ignore_index': -100} def test_check_output(self): - self.check_output(check_eager=False) + self.check_output(check_eager=True) def test_check_output_with_weight(self): self.with_weight = True - self.check_output() + self.check_output(check_eager=True) def test_check_grad(self): self.with_weight = True place = fluid.CPUPlace() - self.check_grad_with_place(place, ['X'], 'Out', check_eager=False) + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) if fluid.core.is_compiled_with_cuda(): place = fluid.CUDAPlace(0) - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) def init_test_case(self): self.input_shape = [10, 10] @@ -809,6 +811,8 @@ class TestNLLLossOp1DNoReduce(OpTest): def setUp(self): self.init_test_case() self.op_type = "nll_loss" + self.python_api = paddle.nn.functional.nll_loss + self.python_out_sig = ["Out"] self.with_weight = False np.random.seed(200) input_np = np.random.uniform(0.1, 0.8, @@ -831,19 +835,19 @@ class TestNLLLossOp1DNoReduce(OpTest): self.attrs = {'reduction': 'none', 'ignore_index': -100} def test_check_output(self): - self.check_output() + self.check_output(check_eager=True) def test_check_output_with_weight(self): self.with_weight = True - self.check_output() + self.check_output(check_eager=True) def test_check_grad(self): self.with_weight = True place = fluid.CPUPlace() - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) if fluid.core.is_compiled_with_cuda(): place = fluid.CUDAPlace(0) - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) def init_test_case(self): self.input_shape = [10, 10] @@ -854,6 +858,8 @@ class TestNLLLossOp2DWithReduce(OpTest): def setUp(self): self.init_test_case() self.op_type = "nll_loss" + self.python_api = paddle.nn.functional.nll_loss + self.python_out_sig = ["Out"] self.with_weight = False np.random.seed(200) input_np = np.random.uniform(0.1, 0.8, @@ -875,19 +881,19 @@ class TestNLLLossOp2DWithReduce(OpTest): self.attrs = {'reduction': 'mean', 'ignore_index': -100} def test_check_output(self): - self.check_output() + self.check_output(check_eager=True) def test_check_output_with_weight(self): self.with_weight = True - self.check_output() + self.check_output(check_eager=True) def test_check_grad(self): self.with_weight = True place = fluid.CPUPlace() - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) if fluid.core.is_compiled_with_cuda(): place = fluid.CUDAPlace(0) - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) def init_test_case(self): self.input_shape = [2, 3, 5, 5] @@ -898,6 +904,8 @@ class TestNLLLossOp2DNoReduce(OpTest): def setUp(self): self.init_test_case() self.op_type = "nll_loss" + self.python_api = paddle.nn.functional.nll_loss + self.python_out_sig = ["Out"] self.with_weight = False np.random.seed(200) input_np = np.random.uniform(0.1, 0.8, @@ -920,19 +928,19 @@ class TestNLLLossOp2DNoReduce(OpTest): self.attrs = {'reduction': 'none', 'ignore_index': -100} def test_check_output(self): - self.check_output() + self.check_output(check_eager=True) def test_check_output_with_weight(self): self.with_weight = True - self.check_output() + self.check_output(check_eager=True) def test_check_grad(self): self.with_weight = True place = fluid.CPUPlace() - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) if fluid.core.is_compiled_with_cuda(): place = fluid.CUDAPlace(0) - self.check_grad_with_place(place, ['X'], 'Out') + self.check_grad_with_place(place, ['X'], 'Out', check_eager=True) def init_test_case(self): self.input_shape = [5, 3, 5, 5] diff --git a/python/paddle/nn/functional/loss.py b/python/paddle/nn/functional/loss.py index 660e6d35871085e90965534915d6cf19a240c960..ca5629aab6790b54906c3b9835017a4e15ea1a57 100755 --- a/python/paddle/nn/functional/loss.py +++ b/python/paddle/nn/functional/loss.py @@ -784,7 +784,17 @@ def nll_loss(input, input_dims)) n = input_shape[0] c = input_shape[1] - if _non_static_mode(): + if in_dygraph_mode(): + if input_dims != 2 and input_dims != 4: + input, _ = _C_ops.reshape2(input, None, 'shape', [n, c, 1, -1]) + label, _ = _C_ops.reshape2(label, None, 'shape', [n, 1, -1]) + out_shape = [n] + input_shape[2:] + out, total_weight = _C_ops.final_state_nll_loss(input, label, weight, + ignore_index, reduction) + if input_dims != 2 and input_dims != 4 and reduction == 'none': + out, _ = _C_ops.reshape2(out, None, 'shape', out_shape) + return out + if _in_legacy_dygraph(): if input_dims != 2 and input_dims != 4: input, _ = _C_ops.reshape2(input, None, 'shape', [n, c, 1, -1]) label, _ = _C_ops.reshape2(label, None, 'shape', [n, 1, -1]) diff --git a/python/paddle/utils/code_gen/api.yaml b/python/paddle/utils/code_gen/api.yaml index 5bbc64ec44afc00bf67e178c978a34e21ed5a0e5..da79a928dba7ab29266b86930e0b4c441aa5cfc8 100644 --- a/python/paddle/utils/code_gen/api.yaml +++ b/python/paddle/utils/code_gen/api.yaml @@ -806,6 +806,17 @@ func : mv backward : mv_grad +- api : nll_loss + args : (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, str reduction) + output : Tensor(out), Tensor(total_weight) + infer_meta : + func : NllLossRawInferMeta + kernel : + func : nll_loss + data_type : input + optional : weight + backward : nll_loss_grad + - api : not_equal args : (Tensor x, Tensor y, int axis = -1) output : Tensor diff --git a/python/paddle/utils/code_gen/backward.yaml b/python/paddle/utils/code_gen/backward.yaml index aa7fd88285f6f74b18fbd0c3feb8aff1465a83c2..dc7261eef1650c1d7ace40a37630e631f2c3e6fd 100644 --- a/python/paddle/utils/code_gen/backward.yaml +++ b/python/paddle/utils/code_gen/backward.yaml @@ -460,15 +460,14 @@ func : mv_grad - backward_api : nll_loss_grad - forward : nll_loss (Tensor x, Tensor label, Tensor weight, int64_t ignore_index, str reduction) -> Tensor(out), Tensor(total_weight) - args : (Tensor x, Tensor label, Tensor weight, Tensor total_weight, Tensor out_grad, int64_t ignore_index, str reduction) - output : Tensor (x_grad) + forward : nll_loss (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, str reduction) -> Tensor(out), Tensor(total_weight) + args : (Tensor input, Tensor label, Tensor weight, Tensor total_weight, Tensor out_grad, int64_t ignore_index, str reduction) + output : Tensor(input_grad) infer_meta : - func : UnchangedInferMeta - param : [x] + func : NllLossGradInferMeta kernel : func : nll_loss_grad - data_type : out_grad + data_type : input optional : weight - backward_api : psroi_pool_grad diff --git a/python/paddle/utils/code_gen/sparse_bw_api.yaml b/python/paddle/utils/code_gen/sparse_bw_api.yaml index 711b4cedc59a586cdf5bf9e21f90e531cc6cbbc6..1f474d56a9022c9ee63065b115ba23abcf65eb45 100644 --- a/python/paddle/utils/code_gen/sparse_bw_api.yaml +++ b/python/paddle/utils/code_gen/sparse_bw_api.yaml @@ -9,5 +9,5 @@ forward : sparse_relu(Tensor x) -> Tensor(out@SparseCooTensor) args : (Tensor x, Tensor out_grad) output : Tensor(x_grad@SparseCooTensor) - kernel : + kernel : func : sparse_relu_grad diff --git a/tools/infrt/skipped_phi_api.json b/tools/infrt/skipped_phi_api.json index 7e03e01d0fe5b2318cf0f1dbe55446b54e003c66..74650846921b6cc39045aa75c4a4ce1a5dd29d14 100644 --- a/tools/infrt/skipped_phi_api.json +++ b/tools/infrt/skipped_phi_api.json @@ -1,4 +1,4 @@ { -"phi_apis":["conj"], +"phi_apis":["conj", "nll_loss"], "phi_kernels":["equal_all"] }