diff --git a/paddle/fluid/operators/lod_reset_op.cc b/paddle/fluid/operators/lod_reset_op.cc index 6a66297cb843ead1a507a6867c1c562224861cbf..6599e183efcfb617d13c145c6eceb5cdb552adf9 100644 --- a/paddle/fluid/operators/lod_reset_op.cc +++ b/paddle/fluid/operators/lod_reset_op.cc @@ -22,17 +22,16 @@ class LoDResetOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - // input check PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of LoDResetOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of LoDResetOp should not be null."); - // If target LoD is not set form Input(), then it must be set from Attr(). - if (!ctx->HasInput("TargetLoD")) { + + if (!ctx->HasInput("Y")) { auto level0 = ctx->Attrs().Get>("target_lod"); - PADDLE_ENFORCE(level0.size() > 1, - "Target LoD is not found, should be set to be a valid one " - "through Input() or Attr()."); + PADDLE_ENFORCE_GT(level0.size(), 1, + "If Input(Y) is not provided, the target lod should be " + "specified by attribute `target_lod`."); } ctx->SetOutputDim("Out", ctx->GetInputDim("X")); } @@ -50,36 +49,42 @@ class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker { public: LoDResetOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "(LoDTensor) The input tensor of lod_reset operator."); - AddInput("TargetLoD", - "(Tensor, optional) The target level 0 LoD from Input().") + AddInput("X", + "(Tensor, LoDTensor) Input variable of LoDResetOp which " + "could be a Tensor or LoDTensor, where the data of output " + "variable inherits from."); + AddInput("Y", + "(Tensor, LoDTensor, optional) If provided, lod of Input(Y) would " + "be considered as the target lod first, otherwise data of " + "Input(Y) would be considered as the target lod.") .AsDispensable(); - AddOutput("Out", "(LoDTensor) The output tensor of lod_reset operator."); + AddOutput("Out", + "(LoDTensor) Output variable of LoDResetOp which should be a " + "LoDTensor."); AddAttr>("target_lod", "The target level 0 LoD from Attr().") .SetDefault(std::vector{}); AddComment(R"DOC(LoDReset operator -Reset LoD of Input(X) into a new one specified by Input(TargetLoD) or -Attr(target_lod), or set LoD for Input(X) if it doesn't have one. -Currently the lod_reset operator only supports the reset of level 0 LoD. -At least one of Input(TargetLoD) and Attr(target_lod) must be set, -and if both of them are set, Input(TargetLoD) will be chosen as the -target LoD. +Set LoD of `X` to a new one specified by `Y` or attribute `target_lod`. When `Y` +provided, `Y.lod` would be considered as target LoD first, otherwise `Y.data` +would be considered as target LoD. If `Y` is not provided, target LoD should be +specified by attribute `target_lod`. If target LoD is specified by `Y.data` or +`target_lod`, only one level LoD is supported. An example: -Given a float LoDTensor X with shape (6, 1), its transpose form represents - - [1.0, 2.0, 3.0, 4.0, 5.0, 6.0], -with LoD = [[0, 2, 5, 6]] and the three (transposed) sequences look like +Given a 1-level LoDTensor input(X) + X.lod = [[ 0, 2, 5 6 ]] + X.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + X.dims = [6, 1] - [1.0, 2.0], [3.0, 4.0, 5.0], [6.0]. +target_lod: [0, 4, 6] -If target LoD = [0, 4, 6], the lod_reset operator will reset the LoD and -the sequences that the LoDTensor Output(Out) contains becomes: - - [1.0, 2.0, 3.0, 4.0], [5.0, 6.0]. +then we get an 1-level LoDTensor + Out.lod = [[ 0, 4, 6 ]] + Out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + Out.dims = [6, 1] )DOC"); } @@ -90,10 +95,16 @@ class LoDResetGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of LoDResetGradOp should not be null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), - "Input(Out@GRAD) shouldn't be null."); - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + "Input(Out@Grad) of LoDResetGradOp should not be null."); + + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ x_grad_name); + } } protected: @@ -111,9 +122,13 @@ class LoDResetGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(lod_reset, ops::LoDResetOp, ops::LoDResetOpMaker, lod_reset_grad, ops::LoDResetGradOp); -REGISTER_OP_CPU_KERNEL(lod_reset, - ops::LoDResetKernel, - ops::LoDResetKernel); +REGISTER_OP_CPU_KERNEL( + lod_reset, ops::LoDResetKernel, + ops::LoDResetKernel, + ops::LoDResetKernel, + ops::LoDResetKernel); REGISTER_OP_CPU_KERNEL( lod_reset_grad, ops::LoDResetGradKernel, - ops::LoDResetGradKernel); + ops::LoDResetGradKernel, + ops::LoDResetGradKernel, + ops::LoDResetGradKernel); diff --git a/paddle/fluid/operators/lod_reset_op.cu b/paddle/fluid/operators/lod_reset_op.cu index b0e87a851a77a1cc98d419a63d4d9e5e1b9dd163..888d4c12eb4e3f4fd94d8dd4178c59acd0abb23b 100644 --- a/paddle/fluid/operators/lod_reset_op.cu +++ b/paddle/fluid/operators/lod_reset_op.cu @@ -18,8 +18,12 @@ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( lod_reset, ops::LoDResetKernel, - ops::LoDResetKernel); + ops::LoDResetKernel, + ops::LoDResetKernel, + ops::LoDResetKernel); REGISTER_OP_CUDA_KERNEL( lod_reset_grad, ops::LoDResetGradKernel, - ops::LoDResetGradKernel); + ops::LoDResetGradKernel, + ops::LoDResetGradKernel, + ops::LoDResetGradKernel); diff --git a/paddle/fluid/operators/lod_reset_op.h b/paddle/fluid/operators/lod_reset_op.h index 8186d4f8262101edc723af390eee1aec4fa6f3a5..99f01c2a255ade81421c2bba95ff3d38ced6f87c 100644 --- a/paddle/fluid/operators/lod_reset_op.h +++ b/paddle/fluid/operators/lod_reset_op.h @@ -26,35 +26,46 @@ class LoDResetKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const { auto* out = ctx.Output("Out"); auto* in = ctx.Input("X"); - auto* lod_t = ctx.Input("TargetLoD"); + auto* lod_t = ctx.Input("Y"); + + out->ShareDataWith(*in); std::vector level0; if (lod_t) { - auto* lod = lod_t->data(); - if (platform::is_gpu_place(ctx.GetPlace())) { - framework::Tensor lod_cpu; - framework::TensorCopy(*lod_t, platform::CPUPlace(), - ctx.device_context(), &lod_cpu); - lod = lod_cpu.data(); + if (lod_t->lod().size() > 0) { + auto y_lod = lod_t->lod(); + auto last_level = y_lod[y_lod.size() - 1]; + PADDLE_ENFORCE_EQ(last_level.back(), in->dims()[0], + "Last value of `Y`'s last level LoD should be equal " + "to the first dimension of `X`"); + out->set_lod(y_lod); + return; // early return, since lod already set + } else { + auto* lod = lod_t->data(); + if (platform::is_gpu_place(ctx.GetPlace())) { + framework::Tensor lod_cpu; + framework::TensorCopy(*lod_t, platform::CPUPlace(), + ctx.device_context(), &lod_cpu); + lod = lod_cpu.data(); + } + level0 = std::vector(lod, lod + lod_t->numel()); } - level0 = std::vector(lod, lod + lod_t->numel()); } else { level0 = ctx.Attr>("target_lod"); } - PADDLE_ENFORCE(level0.size() > 1UL, - "The size of target LoD should be greater than 1."); - PADDLE_ENFORCE(level0[0] == 0, - "Target LoD should be a vector starting from 0."); - PADDLE_ENFORCE(level0.back() == in->dims()[0], - "Target LoD should be a vector end with the " - "first dimension of Input(X)."); + PADDLE_ENFORCE_GT(level0.size(), 1UL, + "Size of target LoD should be greater than 1."); + PADDLE_ENFORCE_EQ(level0[0], 0, + "Target LoD should be a vector starting from 0."); + PADDLE_ENFORCE_EQ(level0.back(), in->dims()[0], + "Target LoD should be a vector end with the " + "first dimension of Input(X)."); for (size_t i = 0; i < level0.size() - 1; ++i) { PADDLE_ENFORCE(level0[i + 1] > level0[i], "Target LoD should be an ascending vector."); } - out->ShareDataWith(*in); // cast level0 to size_t std::vector ulevel0(level0.size(), 0); std::transform(level0.begin(), level0.end(), ulevel0.begin(), diff --git a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py index 3bf8230f8748dd87ec3c85b0cbd78df2e695a96b..6b6d4c824aeae319dacf224408ce96a0d9c5bb35 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py +++ b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py @@ -42,7 +42,7 @@ class TestLodResetOpByInput(OpTest): target_lod_0 = [0, 4, 7, 10] self.inputs = { 'X': (x, lod), - 'TargetLoD': np.array([target_lod_0]).astype('int32') + 'Y': np.array([target_lod_0]).astype('int32') } self.outputs = {'Out': (x, [target_lod_0])} @@ -50,7 +50,7 @@ class TestLodResetOpByInput(OpTest): self.check_output() def test_check_grad(self): - self.check_grad(["X"], "Out", no_grad_set=set("TargetLoD")) + self.check_grad(["X"], "Out", no_grad_set=set("Y")) class TestLodResetOpBoth(OpTest): @@ -62,7 +62,7 @@ class TestLodResetOpBoth(OpTest): target_lod_0_in = [0, 4, 7, 10] self.inputs = { 'X': (x, lod), - 'TargetLoD': np.array(target_lod_0_in).astype('int32') + 'Y': np.array(target_lod_0_in).astype('int32') } self.attrs = {'target_lod': target_lod_0_attr} self.outputs = {'Out': (x, [target_lod_0_in])} @@ -71,7 +71,24 @@ class TestLodResetOpBoth(OpTest): self.check_output() def test_check_grad(self): - self.check_grad(["X"], "Out", no_grad_set=set("TargetLoD")) + self.check_grad(["X"], "Out", no_grad_set=set("Y")) + + +class TestLodResetOpYIsLoDTensor(OpTest): + def setUp(self): + self.op_type = "lod_reset" + x = np.random.random((10, 20)).astype("float32") + lod = [[0, 3, 5, 10]] + y = np.random.random((10, 10)).astype("float32") + target_lod_0 = [[0, 4, 7, 10]] + self.inputs = {'X': (x, lod), 'Y': (y, target_lod_0)} + self.outputs = {'Out': (x, target_lod_0)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out", no_grad_set=set("Y")) if __name__ == '__main__':