Skip to content

Commit 5d48e9c

Browse files
xuezhongxuezhong
authored
Merge pull request #16884 from liuwei1031/release/1.4
test=release/1.4 cherry-pick (#16845)
2 parents a5ef6bf + 055a5de commit 5d48e9c

File tree

4 files changed

+81
-24
lines changed

4 files changed

+81
-24
lines changed

‎paddle/fluid/operators/lstmp_op.cc

Lines changed: 43 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
1313
limitations under the License. */
1414

1515
#include "paddle/fluid/operators/lstmp_op.h"
16+
#include <memory>
1617
#include <string>
1718

1819
namespace paddle {
@@ -45,6 +46,7 @@ class LSTMPOp : public framework::OperatorWithKernel {
4546
"Output(BatchHidden) of LSTMP operator should not be null.");
4647

4748
auto in_dims = ctx->GetInputDim("Input");
49+
4850
PADDLE_ENFORCE_EQ(in_dims.size(), 2,
4951
"Input(X)'s rank of LSTMP operator must be 2.");
5052

@@ -269,13 +271,47 @@ Users can choose to use fully-connected operator before LSTMP operator.
269271
}
270272
};
271273

274+
class LSTMPGradMaker : public framework::SingleGradOpDescMaker {
275+
public:
276+
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
277+
278+
protected:
279+
std::unique_ptr<framework::OpDesc> Apply() const override {
280+
auto* grad_op = new framework::OpDesc();
281+
grad_op->SetType("lstmp_grad");
282+
grad_op->SetInput("Weight", Input("Weight"));
283+
grad_op->SetInput("ProjWeight", Input("ProjWeight"));
284+
grad_op->SetInput("Bias", Input("Bias"));
285+
286+
grad_op->SetInput("Projection", Output("Projection"));
287+
grad_op->SetInput("Cell", Output("Cell"));
288+
grad_op->SetInput("BatchGate", Output("BatchGate"));
289+
grad_op->SetInput("BatchCellPreAct", Output("BatchCellPreAct"));
290+
grad_op->SetInput("BatchHidden", Output("BatchHidden"));
291+
grad_op->SetInput("H0", Input("H0"));
292+
grad_op->SetInput("C0", Input("C0"));
293+
294+
grad_op->SetInput(framework::GradVarName("Projection"),
295+
OutputGrad("Projection"));
296+
297+
grad_op->SetOutput(framework::GradVarName("Input"), InputGrad("Input"));
298+
grad_op->SetOutput(framework::GradVarName("Weight"), InputGrad("Weight"));
299+
grad_op->SetOutput(framework::GradVarName("ProjWeight"),
300+
InputGrad("ProjWeight"));
301+
grad_op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias"));
302+
grad_op->SetOutput(framework::GradVarName("H0"), InputGrad("H0"));
303+
grad_op->SetOutput(framework::GradVarName("C0"), InputGrad("C0"));
304+
305+
grad_op->SetAttrMap(Attrs());
306+
return std::unique_ptr<framework::OpDesc>(grad_op);
307+
}
308+
};
309+
272310
class LSTMPGradOp : public framework::OperatorWithKernel {
273311
public:
274312
using framework::OperatorWithKernel::OperatorWithKernel;
275313

276314
void InferShape(framework::InferShapeContext* ctx) const override {
277-
PADDLE_ENFORCE(ctx->HasInput("Input"),
278-
"Input(Input) of LSTMP operator should not be null.");
279315
PADDLE_ENFORCE(ctx->HasInput("Projection"),
280316
"Input(Projection) of LSTMP operator should not be null.");
281317
PADDLE_ENFORCE(ctx->HasInput("Cell"),
@@ -298,7 +334,8 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
298334
ctx->SetOutputDim(g_name, ctx->GetInputDim(name));
299335
};
300336

301-
SetOutGradDim("Input");
337+
ctx->SetOutputDim(framework::GradVarName("Input"),
338+
ctx->GetInputDim("BatchGate"));
302339
SetOutGradDim("Weight");
303340
SetOutGradDim("ProjWeight");
304341
SetOutGradDim("Bias");
@@ -310,16 +347,16 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
310347
framework::OpKernelType GetExpectedKernelType(
311348
const framework::ExecutionContext& ctx) const override {
312349
return framework::OpKernelType(
313-
ctx.Input<framework::LoDTensor>("Input")->type(), ctx.device_context());
350+
ctx.Input<framework::LoDTensor>("BatchGate")->type(),
351+
ctx.device_context());
314352
}
315353
};
316354

317355
} // namespace operators
318356
} // namespace paddle
319357

320358
namespace ops = paddle::operators;
321-
REGISTER_OPERATOR(lstmp, ops::LSTMPOp, ops::LSTMPOpMaker,
322-
paddle::framework::DefaultGradOpDescMaker<true>);
359+
REGISTER_OPERATOR(lstmp, ops::LSTMPOp, ops::LSTMPOpMaker, ops::LSTMPGradMaker);
323360
REGISTER_OPERATOR(lstmp_grad, ops::LSTMPGradOp);
324361
REGISTER_OP_CPU_KERNEL(
325362
lstmp, ops::LSTMPKernel<paddle::platform::CPUDeviceContext, float>,

‎paddle/fluid/operators/lstmp_op.h

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -267,7 +267,6 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
267267
}
268268

269269
void Compute(const framework::ExecutionContext& ctx) const override {
270-
auto* input = ctx.Input<LoDTensor>("Input");
271270
auto* weight = ctx.Input<Tensor>("Weight");
272271
auto* proj_weight = ctx.Input<Tensor>("ProjWeight");
273272
auto* bias = ctx.Input<Tensor>("Bias");
@@ -323,7 +322,8 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
323322
ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
324323
}
325324

326-
auto in_dims = input->dims();
325+
// batch_gate dims equal to input dims
326+
auto in_dims = batch_gate->dims();
327327
auto out_dims = cell_out->dims();
328328
framework::DDim proj_dims({in_dims[0], proj_weight->dims()[1]});
329329
int frame_size = static_cast<int>(in_dims[1] / 4);

‎paddle/fluid/operators/sample_logits_op.cc

Lines changed: 31 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
1313
limitations under the License. */
1414

1515
#include "paddle/fluid/operators/sample_logits_op.h"
16+
#include <memory>
1617
#include "paddle/fluid/operators/math/sample_prob.h"
1718

1819
namespace paddle {
@@ -60,6 +61,10 @@ class SampleLogitsOpMaker : public framework::OpProtoAndCheckerMaker {
6061
"(Tensor, default: Tensor<float>), A 2-D tensor with shape [N, NT + S]."
6162
"The probabilites of sampled positive and negtive labels.")
6263
.AsIntermediate();
64+
AddOutput("LogitsDim", "Store dim information of Logits for gradient op")
65+
.AsIntermediate();
66+
AddOutput("LabelsDim", "Store dim information of Logits for gradient op")
67+
.AsIntermediate();
6368
AddOutput("SampledLogits",
6469
"(Tensor, default: Tensor<float>), A 2-D tensor with shape"
6570
"[N, NT + S]. The outputs value of sampled logits, which will be"
@@ -121,6 +126,10 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
121126
"Output(SampledLogits) should be not null.");
122127
PADDLE_ENFORCE(ctx->HasOutput("SampledLabels"),
123128
"Output(SampledLabels) should be not null.");
129+
PADDLE_ENFORCE(ctx->HasOutput("LogitsDim"),
130+
"Output(LogitsDim) should be not null.");
131+
PADDLE_ENFORCE(ctx->HasOutput("LabelsDim"),
132+
"Output(LabelsDim) should be not null.");
124133

125134
auto logits_dims = ctx->GetInputDim("Logits");
126135
auto labels_dims = ctx->GetInputDim("Labels");
@@ -137,6 +146,15 @@ class SampleLogitsOp : public framework::OperatorWithKernel {
137146
ctx->SetOutputDim("Probabilities", {logits_dims[0], num_sampled_classes});
138147
ctx->SetOutputDim("SampledLogits", {logits_dims[0], num_sampled_classes});
139148
ctx->SetOutputDim("SampledLabels", {logits_dims[0], labels_dims[1]});
149+
150+
// append 0 to shape variable to avoid optimized by memory optimize pass
151+
auto logits_dim_vec = framework::vectorize(logits_dims);
152+
logits_dim_vec.push_back(0);
153+
ctx->SetOutputDim("LogitsDim", framework::make_ddim(logits_dim_vec));
154+
155+
auto labels_dim_vec = framework::vectorize(labels_dims);
156+
labels_dim_vec.push_back(0);
157+
ctx->SetOutputDim("LabelsDim", framework::make_ddim(labels_dim_vec));
140158
}
141159

142160
protected:
@@ -155,28 +173,27 @@ class SampleLogitsOpGrad : public framework::OperatorWithKernel {
155173
using framework::OperatorWithKernel::OperatorWithKernel;
156174

157175
void InferShape(framework::InferShapeContext* ctx) const override {
158-
PADDLE_ENFORCE(ctx->HasInput("Logits"),
159-
"Input(Logits) should not be null.");
160-
PADDLE_ENFORCE(ctx->HasInput("Labels"),
161-
"Input(Labels) should be not null.");
176+
PADDLE_ENFORCE(ctx->HasInput("LogitsDim"),
177+
"Input(LogitsDim) should not be null.");
178+
PADDLE_ENFORCE(ctx->HasInput("LabelsDim"),
179+
"Input(LabelsDim) should be not null.");
162180
PADDLE_ENFORCE(ctx->HasInput("Samples"),
163181
"Input(Samples) should be not null.");
164-
PADDLE_ENFORCE(ctx->HasInput("SampledLogits"),
165-
"Input(SampledLogits) should be not null.");
166182
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("SampledLogits")),
167183
"Input(SampledLogits@Grad) should not be null.");
168184
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")),
169185
"Output(Logits@Grad) should be not null.");
170186

171-
auto logit_dims = ctx->GetInputDim("Logits");
172-
auto label_dims = ctx->GetInputDim("Labels");
173-
PADDLE_ENFORCE_EQ(label_dims.size(), 2UL,
187+
auto logits_dims = ctx->GetInputDim("LogitsDim");
188+
logits_dims = framework::DDim(logits_dims.Get(), logits_dims.size() - 1);
189+
auto labels_dims = ctx->GetInputDim("LabelsDim");
190+
labels_dims = framework::DDim(labels_dims.Get(), labels_dims.size() - 1);
191+
PADDLE_ENFORCE_EQ(labels_dims.size(), 2UL,
174192
"The label should be a 2-D tensor.");
175-
PADDLE_ENFORCE_EQ(logit_dims.size(), 2UL,
193+
PADDLE_ENFORCE_EQ(logits_dims.size(), 2UL,
176194
"The logits should be a 2-D tensor.");
177195

178-
ctx->SetOutputDim(framework::GradVarName("Logits"),
179-
ctx->GetInputDim("Logits"));
196+
ctx->SetOutputDim(framework::GradVarName("Logits"), logits_dims);
180197
}
181198

182199
protected:
@@ -199,10 +216,9 @@ class SampleLogitsGradMaker : public framework::SingleGradOpDescMaker {
199216
std::unique_ptr<framework::OpDesc> Apply() const override {
200217
auto* grad_op = new framework::OpDesc();
201218
grad_op->SetType("sample_logits_grad");
202-
grad_op->SetInput("Logits", Input("Logits"));
203-
grad_op->SetInput("Labels", Input("Labels"));
219+
grad_op->SetInput("LogitsDim", Output("LogitsDim"));
220+
grad_op->SetInput("LabelsDim", Output("LabelsDim"));
204221
grad_op->SetInput("Samples", Output("Samples"));
205-
grad_op->SetInput("SampledLogits", Output("SampledLogits"));
206222
grad_op->SetInput(framework::GradVarName("SampledLogits"),
207223
OutputGrad("SampledLogits"));
208224
grad_op->SetOutput(framework::GradVarName("Logits"), InputGrad("Logits"));

‎python/paddle/fluid/layers/nn.py

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6144,6 +6144,8 @@ def sampled_softmax_with_cross_entropy(logits,
61446144
sampled_label = helper.create_variable_for_type_inference(dtype='int64')
61456145
sampled_softlabel = helper.create_variable_for_type_inference(
61466146
dtype=logits.dtype)
6147+
logits_dim = helper.create_variable_for_type_inference(dtype=logits.dtype)
6148+
labels_dim = helper.create_variable_for_type_inference(dtype=label.type)
61476149

61486150
helper.append_op(
61496151
type='sample_logits',
@@ -6157,7 +6159,9 @@ def sampled_softmax_with_cross_entropy(logits,
61576159
'Samples': samples,
61586160
'Probabilities': probabilities,
61596161
'SampledLabels': sampled_label,
6160-
'SampledLogits': sampled_logits
6162+
'SampledLogits': sampled_logits,
6163+
'LogitsDim': logits_dim,
6164+
'LabelsDim': labels_dim
61616165
},
61626166
attrs={
61636167
'use_customized_samples': use_customized_samples,

0 commit comments

Comments
 (0)