Luo Tao
8 years ago
commit
70cadf9350
@ -0,0 +1,56 @@
|
||||
digraph G {
|
||||
|
||||
rnn [label="1-th level RNN" shape=box]
|
||||
|
||||
subgraph cluster0 {
|
||||
label = "time step 0"
|
||||
|
||||
sent0 [label="sentence"]
|
||||
sent1 [label="sentence"]
|
||||
|
||||
rnn1 [label="2-th level RNN" shape=box]
|
||||
|
||||
sent0 -> rnn1
|
||||
sent1 -> rnn1
|
||||
}
|
||||
|
||||
subgraph cluster1 {
|
||||
label = "time step 1"
|
||||
|
||||
sent2 [label="sentence"]
|
||||
sent3 [label="sentence"]
|
||||
|
||||
rnn2 [label="2-th level RNN" shape=box]
|
||||
|
||||
sent2 -> rnn2
|
||||
sent3 -> rnn2
|
||||
}
|
||||
|
||||
subgraph cluster2 {
|
||||
label = "time step 2"
|
||||
|
||||
sent4 [label="sentence"]
|
||||
sent5 [label="sentence"]
|
||||
|
||||
rnn3 [label="2-th level RNN" shape=box]
|
||||
|
||||
sent4 -> rnn3
|
||||
sent5 -> rnn3
|
||||
}
|
||||
|
||||
|
||||
para0 [label="paragraph info 0"]
|
||||
para1 [label="paragraph info 1"]
|
||||
para2 [label="paragraph info 2"]
|
||||
|
||||
rnn1 -> para0
|
||||
rnn2 -> para1
|
||||
rnn3 -> para2
|
||||
|
||||
para0 -> rnn
|
||||
para1 -> rnn
|
||||
para2 -> rnn
|
||||
|
||||
chapter [label="chapter info"]
|
||||
rnn -> chapter
|
||||
}
|
||||
|
After Width: | Height: | Size: 51 KiB |
@ -0,0 +1,87 @@
|
||||
digraph G {
|
||||
label = "simple RNN implementation"
|
||||
|
||||
ranksep=2;
|
||||
|
||||
//graph [nodesep=1, ranksep=1];
|
||||
|
||||
node[nodesep=1]
|
||||
|
||||
subgraph cluster0 {
|
||||
label = "global scope"
|
||||
rankdir = TB
|
||||
W
|
||||
boot_memory
|
||||
input
|
||||
output
|
||||
}
|
||||
|
||||
subgraph cluster1 {
|
||||
label = "step-scope 0"
|
||||
rankdir = TB
|
||||
memory0[label="memory"]
|
||||
prememory0[label="pre-memory"]
|
||||
step_input0[label="step input"]
|
||||
step_output0[label="step output"]
|
||||
}
|
||||
|
||||
subgraph cluster2 {
|
||||
label = "step-scope 1"
|
||||
rankdir = TB
|
||||
memory1[label="memory"]
|
||||
prememory1[label="pre-memory"]
|
||||
step_input1[label="step input"]
|
||||
step_output1[label="step output"]
|
||||
}
|
||||
|
||||
subgraph cluster3 {
|
||||
label = "step-scope 2"
|
||||
rankdir = TB
|
||||
memory2[label="memory"]
|
||||
prememory2[label="pre-memory"]
|
||||
step_input2[label="step input"]
|
||||
step_output2[label="step output"]
|
||||
}
|
||||
|
||||
stepnet [shape=box]
|
||||
stepnet0 [shape=box, style=dashed]
|
||||
stepnet1 [shape=box, style=dashed]
|
||||
stepnet2 [shape=box, style=dashed]
|
||||
|
||||
|
||||
edge[color=blue]
|
||||
boot_memory -> prememory0 [label="init" color="blue"]
|
||||
memory0 -> prememory1 [label="copy/reference" color="blue"]
|
||||
memory1 -> prememory2 [label="copy/reference" color="blue"]
|
||||
|
||||
edge[color=black]
|
||||
W -> stepnet0[constraint=false, style=dashed]
|
||||
W -> stepnet1[constraint=false, style=dashed]
|
||||
W -> stepnet2[constraint=false, style=dashed]
|
||||
|
||||
memory0 -> stepnet0[style=dashed]
|
||||
prememory0 -> stepnet0 -> step_output0[style=dashed]
|
||||
|
||||
memory1 -> stepnet1[style=dashed]
|
||||
prememory1 -> stepnet1 -> step_output1[style=dashed]
|
||||
|
||||
memory2 -> stepnet2[style=dashed]
|
||||
prememory2 -> stepnet2 -> step_output2[style=dashed]
|
||||
|
||||
input -> step_input0
|
||||
input -> step_input1
|
||||
input -> step_input2
|
||||
|
||||
step_input0 -> stepnet0 [style=dashed]
|
||||
step_input1 -> stepnet1[style=dashed]
|
||||
step_input2 -> stepnet2[style=dashed]
|
||||
|
||||
step_output0 -> output
|
||||
step_output1 -> output
|
||||
step_output2 -> output
|
||||
|
||||
stepnet0 -> stepnet[style=dashed]
|
||||
stepnet1 -> stepnet[style=dashed]
|
||||
stepnet2 -> stepnet[style=dashed]
|
||||
|
||||
}
|
||||
|
After Width: | Height: | Size: 43 KiB |
|
After Width: | Height: | Size: 181 KiB |
@ -0,0 +1,75 @@
|
||||
digraph G {
|
||||
chapter [label="chapter"]
|
||||
|
||||
subgraph cluster0 {
|
||||
label = "paragraph 0"
|
||||
|
||||
top_rnn0[label="top rnn step 0" shape=box]
|
||||
|
||||
p0 [label="paragraph 0"]
|
||||
p1 [label="paragraph 1"]
|
||||
}
|
||||
|
||||
subgraph cluster1{
|
||||
label = "paragraph 1"
|
||||
|
||||
top_rnn1[label="top rnn step 1" shape=box]
|
||||
|
||||
p2 [label="paragraph 0"]
|
||||
p3 [label="paragraph 1"]
|
||||
}
|
||||
|
||||
subgraph cluster_p0 {
|
||||
label = "sentence 0"
|
||||
|
||||
low_rnn0 [label="low rnn step 0" shape=box]
|
||||
s00 [label="sentence 0"]
|
||||
s01 [label="sentence 1"]
|
||||
|
||||
low_rnn0 -> s00
|
||||
low_rnn0 -> s01
|
||||
}
|
||||
|
||||
subgraph cluster_p1 {
|
||||
label = "sentence 1"
|
||||
low_rnn1 [label="low rnn step 1" shape=box]
|
||||
s10 [label="sentence 0"]
|
||||
s11 [label="sentence 1"]
|
||||
low_rnn1 -> s10
|
||||
low_rnn1 -> s11
|
||||
}
|
||||
|
||||
subgraph cluster_p2 {
|
||||
label = "sentence 1"
|
||||
low_rnn2 [label="low rnn step 0" shape=box]
|
||||
s20 [label="sentence 0"]
|
||||
s21 [label="sentence 1"]
|
||||
low_rnn2 -> s20
|
||||
low_rnn2 -> s21
|
||||
}
|
||||
|
||||
subgraph cluster_p3 {
|
||||
label = "sentence 1"
|
||||
low_rnn3 [label="low rnn step 1" shape=box]
|
||||
s30 [label="sentence 0"]
|
||||
s31 [label="sentence 1"]
|
||||
low_rnn3 -> s30
|
||||
low_rnn3 -> s31
|
||||
}
|
||||
|
||||
|
||||
chapter -> top_rnn0
|
||||
chapter -> top_rnn1
|
||||
|
||||
top_rnn0 -> p0
|
||||
top_rnn0 -> p1
|
||||
top_rnn1 -> p2
|
||||
top_rnn1 -> p3
|
||||
|
||||
|
||||
p0 -> low_rnn0
|
||||
p1 -> low_rnn1
|
||||
p2 -> low_rnn2
|
||||
p3 -> low_rnn3
|
||||
|
||||
}
|
||||
|
After Width: | Height: | Size: 67 KiB |
@ -0,0 +1,153 @@
|
||||
# RNNOp design
|
||||
|
||||
This document is about an RNN operator which requires that instances in a mini-batch have the same length. We will have a more flexible RNN operator.
|
||||
|
||||
## RNN Algorithm Implementation
|
||||
|
||||
<p aligh="center">
|
||||
<img src="./images/rnn.jpg"/>
|
||||
</p>
|
||||
|
||||
The above diagram shows an RNN unrolled into a full network.
|
||||
|
||||
There are several important concepts:
|
||||
|
||||
- *step-net*: the sub-graph to run at each step,
|
||||
- *memory*, $h_t$, the state of the current step,
|
||||
- *ex-memory*, $h_{t-1}$, the state of the previous step,
|
||||
- *initial memory value*, the ex-memory of the first step.
|
||||
|
||||
### Step-scope
|
||||
|
||||
There could be local variables defined in step-nets. PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step.
|
||||
|
||||
<p aligh="center">
|
||||
<img src="./images/rnn.png"/><br/>
|
||||
Figure 2 the RNN's data flow
|
||||
</p>
|
||||
|
||||
Please be aware that all steps run the same step-net. Each step
|
||||
|
||||
1. creates the step-scope,
|
||||
2. realizes local variables, including step-outputs, in the step-scope, and
|
||||
3. runs the step-net, which could use these variables.
|
||||
|
||||
The RNN operator will compose its output from step outputs in step scopes.
|
||||
|
||||
### Memory and Ex-memory
|
||||
|
||||
Let's give more details about memory and ex-memory via a simply example:
|
||||
|
||||
$$
|
||||
h_t = U h_{t-1} + W x_t
|
||||
$$,
|
||||
|
||||
where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively.
|
||||
|
||||
In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step,
|
||||
or copy the value of the previous memory value to the current ex-memory variable.
|
||||
|
||||
### Usage in Python
|
||||
|
||||
For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
|
||||
|
||||
We can define an RNN's step-net using Block:
|
||||
|
||||
```python
|
||||
import paddle as pd
|
||||
|
||||
X = some_op() # x is some operator's output, and is a LoDTensor
|
||||
a = some_op()
|
||||
|
||||
# declare parameters
|
||||
W = pd.Variable(shape=[20, 30])
|
||||
U = pd.Variable(shape=[20, 30])
|
||||
|
||||
rnn = pd.create_rnn_op(output_num=1)
|
||||
with rnn.stepnet():
|
||||
x = rnn.add_input(X)
|
||||
# declare a memory (rnn's step)
|
||||
h = rnn.add_memory(init=a)
|
||||
# h.pre_state() means previous memory of rnn
|
||||
new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state()))
|
||||
# update current memory
|
||||
h.update(new_state)
|
||||
# indicate that h variables in all step scopes should be merged
|
||||
rnn.add_outputs(h)
|
||||
|
||||
out = rnn()
|
||||
```
|
||||
|
||||
Python API functions in above example:
|
||||
|
||||
- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs.
|
||||
- `rnn.add_memory` creates a variable used as the memory.
|
||||
- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output.
|
||||
|
||||
### Nested RNN and LoDTensor
|
||||
|
||||
An RNN whose step-net includes other RNN operators is known as an *nested RNN*.
|
||||
|
||||
For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences.
|
||||
|
||||
The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text.
|
||||
|
||||
<p aligh="center">
|
||||
<img src="./images/2_level_rnn.png"/>
|
||||
</p>
|
||||
|
||||
```python
|
||||
import paddle as pd
|
||||
|
||||
W = pd.Variable(shape=[20, 30])
|
||||
U = pd.Variable(shape=[20, 30])
|
||||
|
||||
W0 = pd.Variable(shape=[20, 30])
|
||||
U0 = pd.Variable(shape=[20, 30])
|
||||
|
||||
# a is output of some op
|
||||
a = some_op()
|
||||
|
||||
# chapter_data is a set of 128-dim word vectors
|
||||
# the first level of LoD is sentence
|
||||
# the second level of LoD is chapter
|
||||
chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2)
|
||||
|
||||
def lower_level_rnn(paragraph):
|
||||
'''
|
||||
x: the input
|
||||
'''
|
||||
rnn = pd.create_rnn_op(output_num=1)
|
||||
with rnn.stepnet():
|
||||
sentence = rnn.add_input(paragraph, level=0)
|
||||
h = rnn.add_memory(shape=[20, 30])
|
||||
h.update(
|
||||
pd.matmul(W, sentence) + pd.matmul(U, h.pre_state()))
|
||||
# get the last state as sentence's info
|
||||
rnn.add_outputs(h)
|
||||
return rnn
|
||||
|
||||
top_level_rnn = pd.create_rnn_op(output_num=1)
|
||||
with top_level_rnn.stepnet():
|
||||
paragraph_data = rnn.add_input(chapter_data, level=1)
|
||||
low_rnn = lower_level_rnn(paragraph_data)
|
||||
paragraph_out = low_rnn()
|
||||
|
||||
h = rnn.add_memory(init=a)
|
||||
h.update(
|
||||
pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state()))
|
||||
top_level_rnn.add_outputs(h)
|
||||
|
||||
# just output the last step
|
||||
chapter_out = top_level_rnn(output_all_steps=False)
|
||||
```
|
||||
|
||||
in above example, the construction of the `top_level_rnn` calls `lower_level_rnn`. The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
|
||||
|
||||
By default, the `RNNOp` will concatenate the outputs from all the time steps,
|
||||
if the `output_all_steps` set to False, it will only output the final time step.
|
||||
|
||||
|
||||
<p align="center">
|
||||
<img src="images/rnn_2level_data.png"/>
|
||||
</p>
|
||||
@ -0,0 +1,66 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/operators/accuracy_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
class AccuracyOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(const framework::InferShapeContext &ctx) const override {
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Inference"),
|
||||
"Input of Inference must be initialized.");
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
|
||||
"Input of Inference must be initialized.");
|
||||
auto *inference = ctx.Input<framework::Tensor>("Inference");
|
||||
auto *label = ctx.Input<framework::Tensor>("Label");
|
||||
|
||||
PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector");
|
||||
PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
|
||||
"inference size must be the same as label size");
|
||||
|
||||
ctx.Output<Tensor>("Accuracy")->Resize({1});
|
||||
}
|
||||
};
|
||||
|
||||
class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
AccuracyOpMaker(framework::OpProto *proto,
|
||||
framework::OpAttrChecker *op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
// TODO(typhoonzero): support both inference value and indices.
|
||||
AddInput("Inference", "topk(indices) the network output");
|
||||
AddInput("Label", "Label of the training data");
|
||||
// TODO(typhoonzero): AddInput("Weight", ...
|
||||
AddOutput("Accuracy", "The accuracy of current batch");
|
||||
|
||||
AddComment(
|
||||
R"DOC(Accuracy. It will print accuracy rate for classification.
|
||||
The accuracy is:
|
||||
.. math::
|
||||
accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
|
||||
REGISTER_OP_CPU_KERNEL(accuracy,
|
||||
ops::AccuracyKernel<paddle::platform::CPUPlace, float>);
|
||||
@ -0,0 +1,69 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/operators/accuracy_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
__global__ void AccuracySingleKernel(const int N, const int D, const int top_k,
|
||||
const int* Xdata, const int* labelData,
|
||||
float* accuracy) {
|
||||
int correct = 0;
|
||||
for (int row = 0; row < N; row++) {
|
||||
const int label = labelData[row];
|
||||
for (int col = 0; col < D; col++) {
|
||||
const int pred = Xdata[row * D + col];
|
||||
if (pred == label) {
|
||||
++correct;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
*accuracy = static_cast<float>(correct) / static_cast<float>(N);
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
class AccuracyOpCUDAKernel : public framework::OpKernel {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
|
||||
"It must use GPUPlace.");
|
||||
auto* inference = ctx.Input<Tensor>("Inference");
|
||||
auto* label = ctx.Input<Tensor>("Label");
|
||||
auto* accuracy = ctx.Output<Tensor>("Accuracy");
|
||||
// FIXME(typhoonzero): only support indices currently
|
||||
// if add support for output values, how to detect the data type?
|
||||
const int* inference_data = inference->data<int>();
|
||||
const int* label_data = label->data<int>();
|
||||
float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
|
||||
|
||||
size_t num_samples = inference->dims()[0];
|
||||
size_t infer_width = inference->dims()[1];
|
||||
cudaMemset((void**)&accuracy_data, 0, sizeof(float));
|
||||
|
||||
if (num_samples == 0) {
|
||||
return;
|
||||
}
|
||||
|
||||
AccuracySingleKernel<<<1, 1>>>(num_samples, infer_width, 1, inference_data,
|
||||
label_data, accuracy_data);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
REGISTER_OP_GPU_KERNEL(accuracy,
|
||||
paddle::operators::AccuracyOpCUDAKernel<float>);
|
||||
@ -0,0 +1,77 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#pragma once
|
||||
#include <algorithm>
|
||||
#include "paddle/framework/eigen.h"
|
||||
#include "paddle/framework/op_registry.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
|
||||
template <typename T, int MajorType = Eigen::RowMajor,
|
||||
typename IndexType = Eigen::DenseIndex>
|
||||
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
|
||||
|
||||
template <typename T, int MajorType = Eigen::RowMajor,
|
||||
typename IndexType = Eigen::DenseIndex>
|
||||
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
|
||||
|
||||
template <typename T, int MajorType = Eigen::RowMajor,
|
||||
typename IndexType = Eigen::DenseIndex>
|
||||
using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
|
||||
|
||||
template <typename Place, typename T>
|
||||
class AccuracyKernel : public framework::OpKernel {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
auto* inference = ctx.Input<Tensor>("Inference");
|
||||
auto* label = ctx.Input<Tensor>("Label");
|
||||
auto* accuracy = ctx.Output<Tensor>("Accuracy");
|
||||
|
||||
float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
|
||||
|
||||
const T* inference_data = inference->data<T>();
|
||||
const T* label_data = label->data<T>();
|
||||
|
||||
size_t num_samples = inference->dims()[0];
|
||||
size_t class_dim = inference->dims()[1];
|
||||
*accuracy_data = 0.0f;
|
||||
|
||||
if (num_samples == 0) {
|
||||
return;
|
||||
}
|
||||
|
||||
int num_correct = 0;
|
||||
// assume inference is already the topk of the output
|
||||
for (size_t i = 0; i < num_samples; ++i) {
|
||||
PADDLE_ENFORCE_GE(label_data[i], 0, "label must >= 0");
|
||||
for (size_t j = 0; j < class_dim; ++j) {
|
||||
if (inference_data[i * class_dim + j] == label_data[i]) {
|
||||
++num_correct;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// FIXME(typhoonzero): we don't accumulate the accuracy for now.
|
||||
*accuracy_data =
|
||||
static_cast<float>(num_correct) / static_cast<float>(num_samples);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,109 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/operators/elementwise_mul_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
|
||||
class ElementWiseMulOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(const framework::InferShapeContext &ctx) const override {
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
|
||||
auto x_dim = ctx.Input<Tensor>("X")->dims();
|
||||
auto y_dim = ctx.Input<Tensor>("Y")->dims();
|
||||
PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
|
||||
"Rank of first input must >= rank of second input.")
|
||||
ctx.Output<Tensor>("Out")->Resize(x_dim);
|
||||
}
|
||||
};
|
||||
|
||||
class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
ElementWiseMulOpMaker(framework::OpProto *proto,
|
||||
framework::OpAttrChecker *op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
AddInput("X", "The first input of elementwise mul op");
|
||||
AddInput("Y", "The second input of elementwise mul op");
|
||||
AddAttr<int>("axis",
|
||||
R"DOC(
|
||||
When shape(Y) does not equal shape(X),Y will be broadcasted
|
||||
to match the shape of X and axis should be dimension index Y in X
|
||||
)DOC")
|
||||
.SetDefault(-1)
|
||||
.EqualGreaterThan(-1);
|
||||
|
||||
AddOutput("Out", "The output of elementwise mul op");
|
||||
AddComment(R"DOC(
|
||||
Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
|
||||
1. The shape of Y should be same with X or
|
||||
2. Y's shape is a subset of X.
|
||||
Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
|
||||
example:
|
||||
shape(X) = (2, 3, 4, 5), shape(Y) = (,)
|
||||
shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
|
||||
shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
|
||||
shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
|
||||
shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
class ElementWiseMulOpGrad : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(const framework::InferShapeContext &ctx) const override {
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
|
||||
"Input(Out@GRAD) should not be null");
|
||||
|
||||
auto x_dims = ctx.Input<Tensor>("X")->dims();
|
||||
auto y_dims = ctx.Input<Tensor>("Y")->dims();
|
||||
auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
|
||||
auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
|
||||
auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
|
||||
|
||||
PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
|
||||
"Rank of first input must >= rank of second input.")
|
||||
|
||||
if (x_grad) {
|
||||
x_grad->Resize(x_dims);
|
||||
}
|
||||
|
||||
if (y_grad) {
|
||||
y_grad->Resize(y_dims);
|
||||
}
|
||||
}
|
||||
};
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker,
|
||||
elementwise_mul_grad, ops::ElementWiseMulOpGrad);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
elementwise_mul,
|
||||
ops::ElementWiseMulKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
elementwise_mul_grad,
|
||||
ops::ElementWiseMulGradKernel<paddle::platform::CPUPlace, float>);
|
||||
@ -0,0 +1,25 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#define EIGEN_USE_GPU
|
||||
#include "paddle/operators/elementwise_mul_op.h"
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
elementwise_mul,
|
||||
ops::ElementWiseMulKernel<paddle::platform::GPUPlace, float>);
|
||||
REGISTER_OP_GPU_KERNEL(
|
||||
elementwise_mul_grad,
|
||||
ops::ElementWiseMulGradKernel<paddle::platform::GPUPlace, float>);
|
||||
@ -0,0 +1,185 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#pragma once
|
||||
#include <iostream>
|
||||
#include "paddle/framework/eigen.h"
|
||||
#include "paddle/framework/op_registry.h"
|
||||
#include "paddle/operators/math/math_function.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
/*
|
||||
* Out = X ⊙ Y
|
||||
* 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
|
||||
* pre=2, n=3*4, post=5
|
||||
* 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
|
||||
* pre=2*3, n=4*5, post=1
|
||||
*/
|
||||
|
||||
inline void get_mid_dims(const framework::DDim& x_dims,
|
||||
const framework::DDim& y_dims, const int axis,
|
||||
int& pre, int& n, int& post) {
|
||||
pre = 1;
|
||||
n = 1;
|
||||
post = 1;
|
||||
for (int i = 0; i < axis; ++i) {
|
||||
pre *= x_dims[i];
|
||||
}
|
||||
|
||||
for (int i = 0; i < y_dims.size(); ++i) {
|
||||
PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
|
||||
"Broadcast dimension mismatch.");
|
||||
n *= y_dims[i];
|
||||
}
|
||||
|
||||
for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
|
||||
post *= x_dims[i];
|
||||
}
|
||||
}
|
||||
|
||||
template <typename Place, typename T>
|
||||
class ElementWiseMulKernel : public framework::OpKernel {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
using Tensor = framework::Tensor;
|
||||
|
||||
auto* x = ctx.Input<Tensor>("X");
|
||||
auto* y = ctx.Input<Tensor>("Y");
|
||||
auto* z = ctx.Output<Tensor>("Out");
|
||||
z->mutable_data<T>(ctx.GetPlace());
|
||||
|
||||
auto x_e = framework::EigenVector<T>::Flatten(*x);
|
||||
auto y_e = framework::EigenVector<T>::Flatten(*y);
|
||||
auto z_e = framework::EigenVector<T>::Flatten(*z);
|
||||
|
||||
auto x_dims = x->dims();
|
||||
auto y_dims = y->dims();
|
||||
PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
|
||||
"Rank of first input must >= rank of second input.")
|
||||
|
||||
if (x_dims == y_dims || product(y_dims) == 1) {
|
||||
z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_e;
|
||||
return;
|
||||
}
|
||||
|
||||
int axis = ctx.Attr<int>("axis");
|
||||
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
|
||||
PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
|
||||
"Axis should be in range [0, x_dims)");
|
||||
|
||||
int pre, n, post;
|
||||
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
|
||||
if (post == 1) {
|
||||
auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
|
||||
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
|
||||
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
|
||||
z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
|
||||
return;
|
||||
} else {
|
||||
auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
|
||||
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
|
||||
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
|
||||
z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
|
||||
return;
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template <typename Place, typename T>
|
||||
class ElementWiseMulGradKernel : public framework::OpKernel {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
using Tensor = framework::Tensor;
|
||||
|
||||
auto* x = ctx.Input<Tensor>("X");
|
||||
auto* y = ctx.Input<Tensor>("Y");
|
||||
auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
|
||||
|
||||
auto x_e = framework::EigenVector<T>::Flatten(*x);
|
||||
auto y_e = framework::EigenVector<T>::Flatten(*y);
|
||||
auto dout_e = framework::EigenVector<T>::Flatten(*dout);
|
||||
|
||||
auto x_dims = x->dims();
|
||||
auto y_dims = y->dims();
|
||||
|
||||
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
|
||||
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
|
||||
if (dx) {
|
||||
dx->mutable_data<T>(ctx.GetPlace());
|
||||
}
|
||||
if (dy) {
|
||||
dy->mutable_data<T>(ctx.GetPlace());
|
||||
}
|
||||
|
||||
if (x_dims == y_dims || product(y_dims) == 1) {
|
||||
if (dx) {
|
||||
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
|
||||
dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
|
||||
}
|
||||
|
||||
if (dy) {
|
||||
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
|
||||
dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;
|
||||
}
|
||||
return;
|
||||
}
|
||||
|
||||
int axis = ctx.Attr<int>("axis");
|
||||
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
|
||||
|
||||
int pre, n, post;
|
||||
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
|
||||
|
||||
// TODO(gongweibao): wrap reshape to a function.
|
||||
if (post == 1) {
|
||||
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
|
||||
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
|
||||
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
|
||||
if (dx) {
|
||||
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
|
||||
dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
|
||||
}
|
||||
|
||||
if (dy) {
|
||||
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
|
||||
dy_e.device(ctx.GetEigenDevice<Place>()) =
|
||||
(x_e * dout_e)
|
||||
.reshape(Eigen::DSizes<int, 2>(pre, n))
|
||||
.sum(Eigen::array<int, 1>{{0}});
|
||||
}
|
||||
return;
|
||||
} else {
|
||||
auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
|
||||
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
|
||||
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
|
||||
if (dx) {
|
||||
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
|
||||
dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
|
||||
}
|
||||
|
||||
if (dy) {
|
||||
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
|
||||
dy_e.device(ctx.GetEigenDevice<Place>()) =
|
||||
(x_e * dout_e)
|
||||
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
|
||||
.sum(Eigen::array<int, 2>{{0, 2}});
|
||||
}
|
||||
return;
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,112 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/operators/pad_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using framework::Tensor;
|
||||
|
||||
class PadOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(const framework::InferShapeContext &ctx) const override {
|
||||
auto x_dim = ctx.Input<Tensor>("X")->dims();
|
||||
auto paddings = Attr<std::vector<int>>("paddings");
|
||||
PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),
|
||||
"Size of paddings should be equal to 2 * dimension size "
|
||||
"of input tensor.");
|
||||
std::vector<int64_t> out_dims(x_dim.size());
|
||||
for (int i = 0; i < x_dim.size(); ++i) {
|
||||
out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
|
||||
}
|
||||
ctx.Output<Tensor>("Out")->Resize(framework::make_ddim(out_dims));
|
||||
}
|
||||
};
|
||||
|
||||
class PadOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
PadOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
AddInput("X",
|
||||
"The input of pad op. "
|
||||
"The input should be a k-D tensor(k > 0 and k < 7)");
|
||||
AddOutput("Out",
|
||||
"The output of pad op."
|
||||
"A tensor with the same shape as X.")
|
||||
.NotInGradient();
|
||||
AddComment(R"DOC(
|
||||
Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example:
|
||||
|
||||
Given:
|
||||
|
||||
X = [[1, 2],
|
||||
[3, 4]]
|
||||
|
||||
and
|
||||
|
||||
paddings = [0, 1, 1, 2]
|
||||
|
||||
and
|
||||
|
||||
pad_value = 0
|
||||
|
||||
then we get
|
||||
|
||||
Out = [[0, 1, 2, 0, 0]
|
||||
[0, 3, 4, 0, 0]
|
||||
[0, 0, 0, 0, 0]]
|
||||
)DOC");
|
||||
AddAttr<std::vector<int>>(
|
||||
"paddings",
|
||||
"A list<int> to describes padding rules for each dimension."
|
||||
" For 2-D image tensor, paddings=[0, 1, 2, 3] means"
|
||||
" padding 0 row to top, 1 row to bottom, 2 columns to left"
|
||||
" and 3 columns to right.Size of paddings should be equal to"
|
||||
" 2 * dimension size of input tensor.");
|
||||
AddAttr<float>("pad_value",
|
||||
"(float) default to 0; "
|
||||
"The value to fill padded areas.")
|
||||
.SetDefault(0.0f);
|
||||
}
|
||||
};
|
||||
|
||||
class PadOpGrad : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
protected:
|
||||
void InferShape(const framework::InferShapeContext &ctx) const override {
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
|
||||
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
|
||||
"Input(Out@GRAD) should not be null");
|
||||
auto x_dims = ctx.Input<Tensor>("X")->dims();
|
||||
auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
|
||||
if (x_grad != nullptr) {
|
||||
x_grad->Resize(x_dims);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad);
|
||||
REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel<paddle::platform::CPUPlace, float>);
|
||||
REGISTER_OP_CPU_KERNEL(pad_grad,
|
||||
ops::PadGradKernel<paddle::platform::CPUPlace, float>);
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue