Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into clip_op

update-doc-pybind
wanghaoshuang 8 years ago
commit 320df7ade4

@ -1,4 +1,4 @@
# Design Doc: Computations as Graphs
# Design Doc: Computations as a Graph
A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before.
@ -8,6 +8,8 @@ This document explains that the construction of a graph as three steps:
- construct the backward part
- construct the optimization part
## The Construction of a Graph
Let us take the problem of image classification as a simple example. The application program that trains the model looks like:
```python
@ -25,7 +27,9 @@ The first four lines of above program build the forward part of the graph.
![](images/graph_construction_example_forward_only.png)
In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x. `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b.
In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x. `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b, and the initialization operators.
Initialization operators are kind of "run-once" operators -- the `Run` method increments a class data member counter so to run at most once. By doing so, a parameter wouldn't be initialized repeatedly, say, in every minibatch.
In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`. These protobuf messages are saved in a `BlockDesc` protobuf message.
@ -49,3 +53,18 @@ According to the chain rule of gradient computation, `ConstructBackwardGraph` wo
For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient. Here results in the complete graph:
![](images/graph_construction_example_all.png)
## Block and Graph
The word block and graph are interchangable in the desgin of PaddlePaddle. A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions. A graph of operators and variables is a representation of the block.
A Block keeps operators in an array `BlockDesc::ops`
```protobuf
message BlockDesc {
repeated OpDesc ops = 1;
repeated VarDesc vars = 2;
}
```
in the order that there appear in user programs, like the Python program at the beginning of this article. We can imagine that in `ops`, we have some forward operators, followed by some gradient operators, and then some optimization operators.

@ -2,6 +2,8 @@ digraph ImageClassificationGraph {
///////// The forward part /////////
FeedX [label="Feed", color=blue, shape=box];
FeedY [label="Feed", color=blue, shape=box];
InitW [label="Init", color=blue, shape=diamond];
Initb [label="Init", color=blue, shape=diamond];
FC [label="FC", color=blue, shape=box];
MSE [label="MSE", color=blue, shape=box];
@ -14,6 +16,8 @@ digraph ImageClassificationGraph {
FeedX -> x -> FC -> y -> MSE -> cost [color=blue];
FeedY -> l [color=blue];
InitW -> W [color=blue];
Initb -> b [color=blue];
W -> FC [color=blue];
b -> FC [color=blue];
l -> MSE [color=blue];

Binary file not shown.

Before

Width:  |  Height:  |  Size: 54 KiB

After

Width:  |  Height:  |  Size: 58 KiB

Binary file not shown.

Before

Width:  |  Height:  |  Size: 46 KiB

After

Width:  |  Height:  |  Size: 50 KiB

Binary file not shown.

Before

Width:  |  Height:  |  Size: 28 KiB

After

Width:  |  Height:  |  Size: 32 KiB

File diff suppressed because it is too large Load Diff

@ -21,16 +21,16 @@ namespace framework {
/// @cond HIDDEN
template <int i>
Dim<i> make_dim(const int* d) {
Dim<i> make_dim(const int64_t* d) {
return Dim<i>(*d, make_dim<i - 1>(d + 1));
}
template <>
Dim<1> make_dim<1>(const int* d) {
Dim<1> make_dim<1>(const int64_t* d) {
return Dim<1>(*d);
}
void make_ddim(DDim& ddim, const int* dims, int n) {
void make_ddim(DDim& ddim, const int64_t* dims, int n) {
switch (n) {
case 1:
ddim = make_dim<1>(dims);
@ -67,13 +67,13 @@ void make_ddim(DDim& ddim, const int* dims, int n) {
/// @endcond
DDim make_ddim(std::initializer_list<int> dims) {
DDim make_ddim(std::initializer_list<int64_t> dims) {
DDim result(make_dim(0));
make_ddim(result, dims.begin(), dims.size());
return result;
}
DDim make_ddim(const std::vector<int>& dims) {
DDim make_ddim(const std::vector<int64_t>& dims) {
DDim result(make_dim(0));
make_ddim(result, &dims[0], dims.size());
return result;
@ -81,12 +81,12 @@ DDim make_ddim(const std::vector<int>& dims) {
/// @cond HIDDEN
// XXX For some reason, putting this in an anonymous namespace causes errors
class DynamicMutableIndexer : public boost::static_visitor<int&> {
class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {
public:
explicit DynamicMutableIndexer(int idx) : idx_(idx) {}
template <int D>
int& operator()(Dim<D>& dim) const {
int64_t& operator()(Dim<D>& dim) const {
return dim[idx_];
}
@ -94,12 +94,12 @@ class DynamicMutableIndexer : public boost::static_visitor<int&> {
int idx_;
};
class DynamicConstIndexer : public boost::static_visitor<int> {
class DynamicConstIndexer : public boost::static_visitor<int64_t> {
public:
explicit DynamicConstIndexer(int idx) : idx_(idx) {}
template <int D>
int operator()(const Dim<D>& dim) const {
int64_t operator()(const Dim<D>& dim) const {
return dim[idx_];
}
@ -109,22 +109,22 @@ class DynamicConstIndexer : public boost::static_visitor<int> {
/// @endcond
int& DDim::operator[](int idx) {
int64_t& DDim::operator[](int idx) {
return boost::apply_visitor(DynamicMutableIndexer(idx), var);
}
int DDim::operator[](int idx) const {
int64_t DDim::operator[](int idx) const {
return boost::apply_visitor(DynamicConstIndexer(idx), var);
}
ssize_t DDim::size() const { return arity(*this); }
int64_t DDim::size() const { return arity(*this); }
bool DDim::operator==(DDim d) const {
if (var.which() != d.getVar().which()) {
return false;
} else {
std::vector<int> v1 = vectorize(*this);
std::vector<int> v2 = vectorize(d);
std::vector<int64_t> v1 = vectorize(*this);
std::vector<int64_t> v2 = vectorize(d);
for (unsigned int i = 0; i < v1.size(); i++) {
if (v1[i] != v2[i]) {
@ -139,10 +139,10 @@ bool DDim::operator==(DDim d) const {
bool DDim::operator!=(DDim d) const { return !(*this == d); }
DDim DDim::operator+(DDim d) const {
std::vector<int> v1 = vectorize(*this);
std::vector<int> v2 = vectorize(d);
std::vector<int64_t> v1 = vectorize(*this);
std::vector<int64_t> v2 = vectorize(d);
std::vector<int> v3;
std::vector<int64_t> v3;
assert(v1.size() == v2.size());
@ -154,10 +154,10 @@ DDim DDim::operator+(DDim d) const {
}
DDim DDim::operator*(DDim d) const {
std::vector<int> v1 = vectorize(*this);
std::vector<int> v2 = vectorize(d);
std::vector<int64_t> v1 = vectorize(*this);
std::vector<int64_t> v2 = vectorize(d);
std::vector<int> v3;
std::vector<int64_t> v3;
assert(v1.size() == v2.size());
@ -168,15 +168,15 @@ DDim DDim::operator*(DDim d) const {
return make_ddim(v3);
}
int get(const DDim& ddim, int idx) { return ddim[idx]; }
int64_t get(const DDim& ddim, int idx) { return ddim[idx]; }
void set(DDim& ddim, int idx, int value) { ddim[idx] = value; }
/// @cond HIDDEN
struct VectorizeVisitor : public boost::static_visitor<> {
std::vector<int>& vector;
std::vector<int64_t>& vector;
explicit VectorizeVisitor(std::vector<int>& v) : vector(v) {}
explicit VectorizeVisitor(std::vector<int64_t>& v) : vector(v) {}
template <typename T>
void operator()(const T& t) {
@ -188,31 +188,31 @@ struct VectorizeVisitor : public boost::static_visitor<> {
};
/// @endcond
std::vector<int> vectorize(const DDim& ddim) {
std::vector<int> result;
std::vector<int64_t> vectorize(const DDim& ddim) {
std::vector<int64_t> result;
VectorizeVisitor visitor(result);
boost::apply_visitor(visitor, ddim);
return result;
}
struct ProductVisitor : public boost::static_visitor<ssize_t> {
struct ProductVisitor : public boost::static_visitor<int64_t> {
template <int D>
ssize_t operator()(const Dim<D>& dim) {
int64_t operator()(const Dim<D>& dim) {
return product(dim);
}
};
ssize_t product(const DDim& ddim) {
int64_t product(const DDim& ddim) {
ProductVisitor visitor;
return boost::apply_visitor(visitor, ddim);
}
struct SliceVectorizeVisitor : public boost::static_visitor<> {
std::vector<int>& vector;
std::vector<int64_t>& vector;
int begin;
int end;
SliceVectorizeVisitor(std::vector<int>& v, int b, int e)
SliceVectorizeVisitor(std::vector<int64_t>& v, int b, int e)
: vector(v), begin(b), end(e) {
PADDLE_ENFORCE(begin < end,
"Begin index must be less than end index in ddim slice.");
@ -240,7 +240,7 @@ struct SliceVectorizeVisitor : public boost::static_visitor<> {
};
DDim slice_ddim(const DDim& dim, int begin, int end) {
std::vector<int> vec;
std::vector<int64_t> vec;
vec.reserve(end - begin);
SliceVectorizeVisitor visitor(vec, begin, end);
boost::apply_visitor(visitor, dim);
@ -280,7 +280,7 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
return os;
}
DDim::DDim(std::initializer_list<int> init_list) {
DDim::DDim(std::initializer_list<int64_t> init_list) {
*this = make_ddim(init_list);
}
} // namespace framework

@ -40,7 +40,7 @@ struct DDim {
template <int D>
explicit DDim(const Dim<D>& in) : var(in) {}
/*implicit*/ DDim(std::initializer_list<int> init_list);
/*implicit*/ DDim(std::initializer_list<int64_t> init_list);
template <int D>
DDim& operator=(const Dim<D>& in) {
@ -48,8 +48,8 @@ struct DDim {
return *this;
}
int& operator[](int idx);
int operator[](int idx) const;
int64_t& operator[](int idx);
int64_t operator[](int idx) const;
template <typename Visitor>
typename Visitor::result_type apply_visitor(Visitor& visitor) {
@ -71,15 +71,15 @@ struct DDim {
DDim operator*(DDim d) const;
ssize_t size() const;
int64_t size() const;
};
/**
* \brief Make a DDim from std::vector<int>
* \brief Make a DDim from std::vector<int64_t>
*
* \param dims An vector of ints. Must be sized between [1, 9]
*/
DDim make_ddim(const std::vector<int>& dims);
DDim make_ddim(const std::vector<int64_t>& dims);
/**
* \brief Make a DDim from an initializer list
@ -87,14 +87,14 @@ DDim make_ddim(const std::vector<int>& dims);
* \param dims An initializer list of ints. Must be sized between [1, 9]
*
*/
DDim make_ddim(std::initializer_list<int> dims);
DDim make_ddim(std::initializer_list<int64_t> dims);
int get(const DDim& dim, int idx);
int64_t get(const DDim& dim, int idx);
void set(DDim& dim, int idx, int val);
std::vector<int> vectorize(const DDim& ddim);
std::vector<int64_t> vectorize(const DDim& ddim);
ssize_t product(const DDim& ddim);
int64_t product(const DDim& ddim);
/**
* \brief Slice a ddim

@ -12,7 +12,7 @@ TEST(DDim, Equality) {
EXPECT_EQ(ddim[2], 5);
// construct a DDim from a vector
std::vector<int> vec({9, 1, 5});
std::vector<int64_t> vec({9, 1, 5});
paddle::framework::DDim vddim = paddle::framework::make_ddim(vec);
EXPECT_EQ(ddim[0], 9);
EXPECT_EQ(ddim[1], 1);
@ -25,7 +25,7 @@ TEST(DDim, Equality) {
EXPECT_EQ(paddle::framework::get(ddim, 0), 6);
// vectorize a DDim
std::vector<int> res_vec = paddle::framework::vectorize(vddim);
std::vector<int64_t> res_vec = paddle::framework::vectorize(vddim);
EXPECT_EQ(res_vec[0], 9);
EXPECT_EQ(res_vec[1], 1);
EXPECT_EQ(res_vec[2], 5);

@ -17,13 +17,13 @@ struct Dim {
static constexpr int dimensions = i;
template <typename... Args>
HOSTDEVICE Dim(int _head, Args... _tail) : head(_head), tail(_tail...) {
HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) {
static_assert(sizeof...(_tail) == i - 1,
"Dim initialized with the wrong number of parameters");
}
HOSTDEVICE
Dim(int _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
Dim(int64_t _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
HOSTDEVICE
Dim() : head(0), tail() {}
@ -31,12 +31,12 @@ struct Dim {
/** Construct a Dim from a linear index and size. Uses Fortran order
* indexing. */
HOSTDEVICE
Dim(int idx, const Dim<i>& size)
Dim(int64_t idx, const Dim<i>& size)
: head(idx % size.head), tail(idx / size.head, size.tail) {}
/** Construct a Dim with each dimension set to the given index */
HOSTDEVICE
Dim(int idx) : head(idx), tail(idx) {}
Dim(int64_t idx) : head(idx), tail(idx) {}
HOSTDEVICE
bool operator==(const Dim<i>& o) const {
@ -47,13 +47,13 @@ struct Dim {
bool operator!=(const Dim<i>& o) const { return !(*this == o); }
HOSTDEVICE
int& operator[](int idx);
int64_t& operator[](int idx);
HOSTDEVICE
int operator[](int idx) const;
int64_t operator[](int idx) const;
HOST std::string to_string() const;
int head;
int64_t head;
Dim<i - 1> tail;
};
@ -63,7 +63,7 @@ struct Dim<1> {
static constexpr int dimensions = 1;
HOSTDEVICE
Dim(int _head) : head(_head) {}
Dim(int64_t _head) : head(_head) {}
HOSTDEVICE
Dim() : head(0) {}
@ -86,11 +86,11 @@ struct Dim<1> {
bool operator!=(const Dim<1>& o) const { return !(*this == o); }
HOSTDEVICE
int& operator[](int idx);
int64_t& operator[](int idx);
HOSTDEVICE
int operator[](int idx) const;
int64_t operator[](int idx) const;
int head;
int64_t head;
};
namespace {
@ -100,12 +100,12 @@ template <int i>
struct DimGetter {
// Return a copy if Dim is const
template <typename D>
HOSTDEVICE static int impl(const D& d) {
HOSTDEVICE static int64_t impl(const D& d) {
return DimGetter<i - 1>::impl(d.tail);
}
// Return a reference if Dim is mutable
template <typename D>
HOSTDEVICE static int& impl(D& d) {
HOSTDEVICE static int64_t& impl(D& d) {
return DimGetter<i - 1>::impl(d.tail);
}
};
@ -115,18 +115,18 @@ template <>
struct DimGetter<0> {
// Return a copy if Dim is const
template <typename D>
HOSTDEVICE static int impl(const D& d) {
HOSTDEVICE static int64_t impl(const D& d) {
return d.head;
}
// Return a reference if Dim is mutable
template <typename D>
HOSTDEVICE static int& impl(D& d) {
HOSTDEVICE static int64_t& impl(D& d) {
return d.head;
}
};
template <int D>
HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
HOSTDEVICE int64_t& indexer(Dim<D>& dim, int idx) {
#ifndef __CUDA_ARCH__
if (idx < 0) {
throw std::invalid_argument("Tried to access a negative dimension");
@ -141,7 +141,7 @@ HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
}
template <>
HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
HOSTDEVICE int64_t& indexer<1>(Dim<1>& dim, int idx) {
#ifndef __CUDA_ARCH__
if (idx != 0) {
throw std::invalid_argument("Invalid index");
@ -153,7 +153,7 @@ HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
}
template <int D>
HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
HOSTDEVICE int64_t indexer(const Dim<D>& dim, int idx) {
#ifndef __CUDA_ARCH__
if (idx < 0) {
throw std::invalid_argument("Tried to access a negative dimension");
@ -168,7 +168,7 @@ HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
}
template <>
HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
HOSTDEVICE int64_t indexer<1>(const Dim<1>& dim, int idx) {
#ifndef __CUDA_ARCH__
if (idx != 0) {
throw std::invalid_argument("Invalid index");
@ -182,73 +182,76 @@ HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
} // namespace
// Static access to constant Dim
template <int i, int l>
HOSTDEVICE int get(const Dim<l>& d) {
HOSTDEVICE int64_t get(const Dim<l>& d) {
return DimGetter<i>::impl(d);
}
// Static access to mutable Dim
template <int i, int l>
HOSTDEVICE int& get(Dim<l>& d) {
HOSTDEVICE int64_t& get(Dim<l>& d) {
return DimGetter<i>::impl(d);
}
// Dynamic access to constant Dim
template <int l>
HOSTDEVICE int Dim<l>::operator[](int i) const {
HOSTDEVICE int64_t Dim<l>::operator[](int i) const {
return indexer(*this, i);
}
// Dynamic access to mutable Dim
template <int l>
HOSTDEVICE int& Dim<l>::operator[](int i) {
HOSTDEVICE int64_t& Dim<l>::operator[](int i) {
return indexer(*this, i);
}
// Dynamic access to constant Dim
inline HOSTDEVICE int Dim<1>::operator[](int i) const {
inline HOSTDEVICE int64_t Dim<1>::operator[](int i) const {
return indexer(*this, i);
}
// Dynamic access to mutable Dim
inline HOSTDEVICE int& Dim<1>::operator[](int i) { return indexer(*this, i); }
inline HOSTDEVICE int64_t& Dim<1>::operator[](int i) {
return indexer(*this, i);
}
// Dynamic access to constant Dim
// without std::enable_if will try to instantiate this on get<0>(d)
template <int l>
HOSTDEVICE typename std::enable_if<(l > 0), int>::type get(const Dim<l>& d,
int i) {
HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim<l>& d,
int i) {
return d[i];
}
// Dynamic access to mutable Dim
template <int l>
HOSTDEVICE typename std::enable_if<(l > 0), int&>::type get(Dim<l>& d, int i) {
HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim<l>& d,
int i) {
return d[i];
}
// Dot product of two dims
template <int i>
HOSTDEVICE int linearize(const Dim<i>& a, const Dim<i>& b) {
HOSTDEVICE int64_t linearize(const Dim<i>& a, const Dim<i>& b) {
return a.head * b.head + linearize(a.tail, b.tail);
}
// Base case dot product of two Dims
// Notice it is inline because it is no longer a template
template <>
HOSTDEVICE inline int linearize(const Dim<1>& a, const Dim<1>& b) {
HOSTDEVICE inline int64_t linearize(const Dim<1>& a, const Dim<1>& b) {
return a.head * b.head;
}
// Product of a Dim
template <int i>
HOSTDEVICE int product(const Dim<i>& a, int prod = 1) {
HOSTDEVICE int64_t product(const Dim<i>& a, int prod = 1) {
return prod * a.head * product(a.tail);
}
// Base case product of a Dim
// Notice it is inline because it is no longer a template
template <>
HOSTDEVICE inline int product(const Dim<1>& a, int prod) {
HOSTDEVICE inline int64_t product(const Dim<1>& a, int prod) {
return prod * a.head;
}

@ -8,7 +8,7 @@ __global__ void test(paddle::framework::Dim<2>* o) {
o[0] = paddle::framework::make_dim(5, 6);
}
__global__ void dyn_idx_gpu(int* o) {
__global__ void dyn_idx_gpu(int64_t* o) {
auto d = paddle::framework::make_dim(5, 6);
o[0] = d[1];
}
@ -47,9 +47,9 @@ TEST(Dim, Equality) {
EXPECT_EQ(b[1], 11);
// dynamic access on GPU
thrust::device_vector<int> r(1);
thrust::device_vector<int64_t> r(1);
dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data()));
int res = r[0];
int64_t res = r[0];
EXPECT_EQ(res, 6);
// ex_prefix_mul

@ -28,7 +28,7 @@ struct EigenDim {
static Type From(const DDim& dims) {
PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)");
Type ret;
for (int d = 0; d < arity(dims); d++) {
for (int64_t d = 0; d < arity(dims); d++) {
ret[d] = dims[d];
}
return ret;

@ -87,3 +87,24 @@ message OpProto {
repeated Attr attrs = 4;
required string comment = 5;
}
enum DataType {
BOOL = 0;
INT16 = 1;
INT32 = 2;
INT64 = 3;
FP16 = 4;
FP32 = 5;
FP64 = 6;
}
message LoDTensorDesc {
required DataType data_type = 1;
repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
optional int32 lod_level = 3 [ default = 0 ];
}
message VarDesc {
required string name = 1;
optional LoDTensorDesc lod_tensor = 2;
}

@ -3,7 +3,7 @@
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
USE_OP(add_two);
USE_OP(add);
namespace paddle {
namespace framework {
@ -41,7 +41,7 @@ namespace f = paddle::framework;
TEST(GradOpBuilder, AddTwo) {
std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
"add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
"add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_add_op =
f::OpRegistry::CreateGradOp(*add_op);
EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);

@ -94,7 +94,7 @@ Let's go on slicing this slice. Its <1,1>-slice is
|||
```
### The General Slicing Algorithm
### The Slicing Algorithm
The algorithm, with over-simplified data structure, is defined as
@ -106,17 +106,41 @@ struct LoDTensor {
float* tensor_;
};
LoDTensor Slice(const LoDTensor& lodt, int level, int sequence) {
LoDTensor Slice(const LoDTensor& lodt, int level, int sequence);
```
Let us revisit the example above
}
```
3
3 1 2
3 2 4 1 2 3
||| || |||| | || |||
```
### Slicing the Top Level
Suppose that we want to retrieve the <1,2>-slice
Please be aware that an RNN operator only slices the top level of a LoD Tensor to get the step inputs.
```
2
2 3
|| |||
```
```c++
LoDTensor Slice(const LoDTensor& lodt, int sequence) {
we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10.
To avoid the traversal of the LoD tree at slcing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into
```
0
0 9 10
0 3 5 9 10 12
||| || |||| | || |||
```
We don't really need the 0 on top, so the LoD Tensor could be
}
```
0 9 10
0 3 5 9 10 12
||| || |||| | || |||
```

@ -80,7 +80,7 @@ TEST(OpRegistry, CreateOp) {
paddle::framework::Scope scope;
paddle::platform::CPUDeviceContext dev_ctx;
op->Run(scope, dev_ctx);
float scale_get = op->GetAttr<float>("scale");
float scale_get = op->Attr<float>("scale");
ASSERT_EQ(scale_get, scale);
}
@ -121,7 +121,7 @@ TEST(OpRegistry, DefaultValue) {
paddle::framework::Scope scope;
paddle::platform::CPUDeviceContext dev_ctx;
op->Run(scope, dev_ctx);
ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
ASSERT_EQ(op->Attr<float>("scale"), 1.0);
}
TEST(OpRegistry, CustomChecker) {
@ -172,6 +172,6 @@ TEST(OpRegistry, CustomChecker) {
paddle::platform::CPUDeviceContext dev_ctx;
paddle::framework::Scope scope;
op->Run(scope, dev_ctx);
int test_attr = op->GetAttr<int>("test_attr");
int test_attr = op->Attr<int>("test_attr");
ASSERT_EQ(test_attr, 4);
}

@ -69,7 +69,7 @@ class OperatorBase {
virtual ~OperatorBase() {}
template <typename T>
inline const T& GetAttr(const std::string& name) const {
inline const T& Attr(const std::string& name) const {
PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
name);
return boost::get<T>(attrs_.at(name));
@ -238,8 +238,8 @@ class InferShapeContext {
const Scope& scope() const { return scope_; }
template <typename T>
inline const T& GetAttr(const std::string& name) const {
return op_.GetAttr<T>(name);
inline const T& Attr(const std::string& name) const {
return op_.Attr<T>(name);
}
size_t InputSize(const std::string& name) const {

@ -58,7 +58,7 @@ inline T* Tensor::mutable_data(platform::Place place) {
"Tensor's numel must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
/* some versions of boost::variant don't have operator!= */
size_t size = product(dims_) * sizeof(T);
int64_t size = product(dims_) * sizeof(T);
if (holder_ == nullptr || !(holder_->place() == place) ||
holder_->size() < size + offset_) {
if (platform::is_cpu_place(place)) {
@ -131,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
PADDLE_ENFORCE_LT(begin_idx, end_idx,
"Begin index must be less than end index.");
PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
int base = product(dims_) / dims_[0];
size_t base = product(dims_) / dims_[0];
Tensor dst;
dst.holder_ = holder_;
DDim dst_dims = dims_;

@ -83,8 +83,8 @@ void Conv3DLayer::forward(PassType passType) {
int outWidth = getSize();
resetOutput(batchSize, outWidth);
REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
for (size_t i = 0; i != inputLayers_.size(); ++i) {
REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
const MatrixPtr &inMat = getInputValue(i);
const MatrixPtr &outMat = getOutputValue();
int M = M_[i];
@ -120,7 +120,6 @@ void Conv3DLayer::forward(PassType passType) {
}
}
if (nullptr != this->biasParameter_) {
REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
this->addBias();
}
forwardActivation();
@ -134,15 +133,14 @@ void Conv3DLayer::backward(const UpdateCallback &callback) {
biases_->getParameterPtr()->incUpdate(callback);
}
REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
for (size_t i = 0; i != inputLayers_.size(); ++i) {
REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
if (weights_[i]->getWGrad()) {
bpropWeights(i);
}
if (getInputGrad(i)) {
bpropData(i);
}
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
weights_[i]->getParameterPtr()->incUpdate(callback);
}
}

@ -84,8 +84,8 @@ void DeConv3DLayer::forward(PassType passType) {
resetOutput(batchSize, outWidth);
const MatrixPtr outMat = getOutputValue();
REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
for (size_t i = 0; i != inputLayers_.size(); ++i) {
REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
const MatrixPtr &inMat = getInputValue(i);
int M = M_[i];
int N = N_[i];
@ -120,7 +120,6 @@ void DeConv3DLayer::forward(PassType passType) {
}
}
if (nullptr != this->biasParameter_) {
REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
this->addBias();
}
forwardActivation();
@ -133,12 +132,12 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
bpropBiases();
biases_->getParameterPtr()->incUpdate(callback);
}
REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
for (size_t i = 0; i < inputLayers_.size(); ++i) {
if (weights_[i]->getWGrad() || this->needGradient_) {
int M = M_[i];
int N = N_[i];
int K = K_[i];
REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
const MatrixPtr &inMat = getInputValue(i);
for (int n = 0; n < batchSize; ++n) {
@ -182,7 +181,6 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
}
}
}
REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
weights_[i]->getParameterPtr()->incUpdate(callback);
}
}

@ -14,27 +14,31 @@ function(op_library TARGET)
cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
"${multiValueArgs}" ${ARGN})
foreach(src ${op_library_SRCS})
if (${src} MATCHES ".*\\.cu$")
list(APPEND cu_srcs ${src})
elseif(${src} MATCHES ".*\\.cc$")
list(APPEND cc_srcs ${src})
else()
message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
list(LENGTH op_library_SRCS op_library_SRCS_len)
if (${op_library_SRCS_len} EQUAL 0)
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
list(APPEND cc_srcs ${TARGET}.cc)
endif()
endforeach()
if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
list(APPEND cu_srcs ${TARGET}.cu)
endif()
else()
foreach(src ${op_library_SRCS})
if (${src} MATCHES ".*\\.cu$")
list(APPEND cu_srcs ${src})
elseif(${src} MATCHES ".*\\.cc$")
list(APPEND cc_srcs ${src})
else()
message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
endif()
endforeach()
endif()
list(LENGTH cc_srcs cc_srcs_len)
if (${cc_srcs_len} EQUAL 0)
message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
endif()
list(LENGTH cu_srcs cu_srcs_len)
list(LENGTH op_library_DEPS dep_len)
if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
message(WARNING "The op library ${TARGET} not support GPU!")
endif()
if (WITH_GPU)
nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
${op_common_deps})
@ -46,22 +50,22 @@ endfunction()
add_subdirectory(math)
list(REMOVE_ITEM GENERAL_OPS
net_op
minus_op
mul_op
recurrent_op
scale_op)
op_library(net_op SRCS net_op.cc)
op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
set(DEPS_OPS
identity_op
minus_op
mul_op
recurrent_op
scale_op)
op_library(identity_op DEPS scale_op)
op_library(minus_op DEPS scale_op)
op_library(mul_op DEPS math_function)
op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS framework_proto tensor operator net_op)
op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
op_library(scale_op DEPS net_op)
list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
foreach(src ${GENERAL_OPS})
op_library(${src} SRCS ${src}.cc ${src}.cu)
op_library(${src})
endforeach()
set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")

@ -57,7 +57,6 @@ class AddOpGrad : public framework::OperatorWithKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad);
REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad);
REGISTER_OP_CPU_KERNEL(add_two,
ops::AddKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(add, ops::AddKernel<paddle::platform::CPUPlace, float>);

@ -12,10 +12,7 @@
See the License for the specific language governing permissions and
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/framework/op_registry.h"
#include "paddle/operators/add_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(add_two,
ops::AddKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(add, ops::AddKernel<paddle::platform::GPUPlace, float>);

@ -0,0 +1,107 @@
/* 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/cos_sim_op.h"
namespace paddle {
namespace operators {
using framework::Tensor;
class CosSimOp : 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) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
ctx.Input<Tensor>("Y")->dims(),
"Dimensions of Input(X) and Input(Y) must be the same.");
auto dims = ctx.Input<Tensor>("X")->dims();
ctx.Output<Tensor>("Out")->Resize({dims[0], 1});
ctx.Output<Tensor>("XNorm")->Resize({dims[0], 1});
ctx.Output<Tensor>("YNorm")->Resize({dims[0], 1});
}
};
class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
public:
CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of cos_sim op.");
AddInput("Y", "The second input of cos_sim op.");
AddOutput("Out", "The output of cos_sim op.");
AddOutput("XNorm", "Row norm of the first input.").AsIntermediate();
AddOutput("YNorm", "Row norm of the second input.").AsIntermediate();
AddComment(R"DOC(
Cosine Similarity Operator.
The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y))
)DOC");
}
};
class CosSimOpGrad : 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) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"),
"Input(XNorm) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"),
"Input(YNorm) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
"Input(Out@GRAD) must not be null.");
auto x_dims = ctx.Input<Tensor>("X")->dims();
auto y_dims = ctx.Input<Tensor>("Y")->dims();
auto xnorm_dims = ctx.Input<Tensor>("XNorm")->dims();
auto ynorm_dims = ctx.Input<Tensor>("YNorm")->dims();
auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
PADDLE_ENFORCE_EQ(x_dims, y_dims,
"Dimensions of Input(X) and Input(Y) must be the same.");
PADDLE_ENFORCE_EQ(xnorm_dims[0], x_dims[0],
"1st dimension of XNorm must equal that of Input(X).");
PADDLE_ENFORCE_EQ(xnorm_dims[1], 1, "2st dimension of XNorm must be one.");
PADDLE_ENFORCE_EQ(ynorm_dims[0], y_dims[0],
"1st dimension of YNorm must equal that of Input(Y).");
PADDLE_ENFORCE_EQ(ynorm_dims[1], 1, "2st dimension of YNorm must be one.");
PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0],
"1st dimension of Out@GRAD must equal that of Input(X)");
PADDLE_ENFORCE_EQ(out_dims[1], 1, "1st dimension of Out@GRAD must be one.");
auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
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(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad,
ops::CosSimOpGrad);
REGISTER_OP_CPU_KERNEL(cos_sim,
ops::CosSimKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
cos_sim_grad, ops::CosSimGradKernel<paddle::platform::CPUPlace, float>);

@ -13,8 +13,10 @@
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/operators/gather_op.h"
#include "paddle/operators/cos_sim_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(gather,
ops::GatherOpKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(cos_sim,
ops::CosSimKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
cos_sim_grad, ops::CosSimGradKernel<paddle::platform::GPUPlace, float>);

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save