提交 aa1e92ec 编写于 作者: H hedaoyuan

Merge branch 'develop' of https://github.com/baidu/Paddle into inference

......@@ -49,11 +49,12 @@ if(NOT WITH_GOLANG)
endif(NOT WITH_GOLANG)
if(NOT WITH_GPU)
add_definitions(-DPADDLE_ONLY_CPU)
add_definitions(-DHPPL_STUB_FUNC)
list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)
else()
add_definitions(-DPADDLE_WITH_CUDA)
FIND_PACKAGE(CUDA REQUIRED)
if(${CUDA_VERSION_MAJOR} VERSION_LESS 7)
......
......@@ -21,7 +21,7 @@ Model Config API
trainer_config_helpers/optimizers.rst
trainer_config_helpers/data_sources.rst
trainer_config_helpers/layers.rst
trainer_config_helpers/activations.rst
trainer_config_helpers/activations.rst
trainer_config_helpers/poolings.rst
trainer_config_helpers/networks.rst
trainer_config_helpers/evaluators.rst
......
......@@ -345,6 +345,11 @@ clip
.. autoclass:: paddle.v2.layer.clip
:noindex:
resize
------
.. autoclass:: paddle.v2.layer.resize
:noindex:
slope_intercept
---------------
.. autoclass:: paddle.v2.layer.slope_intercept
......
......@@ -55,17 +55,23 @@ Let us consolidate the discussion by presenting some examples.
The following C++ programs shows how blocks are used with the `if-else` structure:
```c++
namespace pd = paddle;
int x = 10;
int y = 20;
int out;
int y = 1;
int z = 10;
bool cond = false;
int o1, o2;
if (cond) {
int z = x + y;
out = softmax(z);
o1 = z;
o2 = pd::layer::softmax(z);
} else {
int z = fc(x);
out = z;
int d = pd::layer::fc(z);
o1 = d;
o2 = d+1;
}
```
An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows:
......@@ -73,57 +79,55 @@ An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator
```python
import paddle as pd
x = var(10)
y = var(20)
cond = var(false)
ie = pd.create_ifelseop(inputs=[x], output_num=1)
x = minibatch([10, 20, 30]) # shape=[None, 1]
y = var(1) # shape=[1], value=1
z = minibatch([10, 20, 30]) # shape=[None, 1]
cond = larger_than(x, 15) # [false, true, true]
ie = pd.ifelse()
with ie.true_block():
x = ie.inputs(true, 0)
z = operator.add(x, y)
ie.set_output(true, 0, operator.softmax(z))
d = pd.layer.add_scalar(x, y)
ie.output(d, pd.layer.softmax(d))
with ie.false_block():
x = ie.inputs(false, 0)
z = layer.fc(x)
ie.set_output(true, 0, operator.softmax(z))
out = b(cond)
d = pd.layer.fc(z)
ie.output(d, d+1)
o1, o2 = ie(cond)
```
In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`.
In both examples, the left branch computes `x+y` and `softmax(x+y)`, the right branch computes `x+1` and `fc(x)`.
A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances. The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values.
### Blocks with `for` and `RNNOp`
The following RNN model from the [RNN design doc](./rnn.md)
```python
x = sequence([10, 20, 30])
m = var(0)
W = tensor()
U = tensor()
rnn = create_rnn(inputs=[input])
with rnn.stepnet() as net:
x = net.set_inputs(0)
h = net.add_memory(init=m)
fc_out = pd.matmul(W, x)
hidden_out = pd.matmul(U, h.pre(n=1))
sum = pd.add_two(fc_out, hidden_out)
act = pd.sigmoid(sum)
h.update(act) # update memory with act
net.set_outputs(0, act, hidden_out) # two outputs
x = sequence([10, 20, 30]) # shape=[None, 1]
m = var(0) # shape=[1]
W = var(0.314, param=true) # shape=[1]
U = var(0.375, param=true) # shape=[1]
rnn = pd.rnn()
with rnn.step():
h = rnn.memory(init = m)
hh = rnn.previous_memory(h)
a = layer.fc(W, x)
b = layer.fc(U, hh)
s = pd.add(a, b)
act = pd.sigmoid(s)
rnn.update_memory(h, act)
rnn.output(a, b)
o1, o2 = rnn()
print o1, o2
```
has its equivalent C++ program as follows
```c++
int* x = {10, 20, 30};
int m = 0;
int W = some_value();
int U = some_other_value();
int* m = {0};
int* W = {0.314};
int* U = {0.375};
int mem[sizeof(x) / sizeof(x[0]) + 1];
int o1[sizeof(x) / sizeof(x[0]) + 1];
......@@ -131,20 +135,16 @@ int o2[sizeof(x) / sizeof(x[0]) + 1];
for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) {
int x = x[i-1];
if (i == 1) mem[0] = m;
int fc_out = W * x;
int hidden_out = Y * mem[i-1];
int sum = fc_out + hidden_out;
int a = W * x;
int b = Y * mem[i-1];
int s = fc_out + hidden_out;
int act = sigmoid(sum);
mem[i] = act;
o1[i] = act;
o2[i] = hidden_out;
}
print_array(o1);
print_array(o2);
```
## Compilation and Execution
Like TensorFlow programs, a PaddlePaddle program is written in Python. The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference.
......@@ -210,11 +210,11 @@ a = pd.Varaible(shape=[20, 20])
b = pd.fc(a, params=["fc.w", "fc.b"])
rnn = pd.create_rnn()
with rnn.stepnet() as net:
x = net.set_inputs(a)
with rnn.stepnet()
x = a.as_step_input()
# reuse fc's parameter
fc_without_b = pd.get_variable("fc.w")
net.set_outputs(fc_without_b)
rnn.output(fc_without_b)
out = rnn()
```
......
IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
# The `IfElse` Operator
```python
import paddle as pd
PaddlePaddle's `IfElse` operator differs from TensorFlow's:
x = var()
y = var()
cond = var()
default_value = var()
b = pd.create_ifelseop(inputs=[x], output_num=1)
with b.true_block():
x = b.inputs(0)
z = operator.add(x, y)
b.set_output(0, operator.softmax(z))
with b.false_block():
x = b.inputs(0)
z = layer.fc(x)
b.set_output(0, operator.softmax(z))
out = b(cond)
```
- the TensorFlow version takes a scalar boolean value as the condition so that the whole mini-batch goes to either the true or the false branch, whereas
- the PaddlePaddle version takes a vector of boolean value as the condition, and instances corresponding to true values go to the true branch, those corresponding to false values go to the false branch.
## Example
The following PaddlePaddle program shows the usage of the IfElse operator:
If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
```python
import paddle as pd
x = var()
y = var()
cond = var()
default_value = var()
b = pd.create_ifelseop(inputs=[x], output_num=1, default_value)
with b.true_block():
x = b.inputs(0)
z = operator.add(x, y)
b.set_output(0, operator.softmax(z))
x = minibatch([10, 20, 30]) # shape=[None, 1]
y = var(1) # shape=[1], value=1
z = minibatch([10, 20, 30]) # shape=[None, 1]
cond = larger_than(x, 15) # [false, true, true]
ie = pd.ifelse()
with ie.true_block():
d = pd.layer.add(x, y)
ie.output(d, pd.layer.softmax(d))
with ie.false_block():
d = pd.layer.fc(z)
ie.output(d, d+1)
o1, o2 = ie(cond)
```
out = b(cond)
A challenge to implement the `IfElse` operator is to infer those variables to be split, or, say, to identify the variable of the mini-batch or those derived from the mini-batch.
An equivalent C++ program is as follows:
```c++
namespace pd = paddle;
int x = 10;
int y = 1;
int z = 10;
bool cond = false;
int o1, o2;
if (cond) {
int d = x + y;
o1 = z;
o2 = pd::layer::softmax(z);
} else {
int d = pd::layer::fc(z);
o1 = d;
o2 = d+1;
}
```
where default_value is a list of vars for `cond` == False.
# Design Doc: ProgramDesc
# Design Doc: PaddlePaddle Programs
The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program.
## Compile and Execution
A PaddlePaddle program consists of two parts -- the first generates a `ProgramDesc` protobuf message that describes the program, and the second runs this message using a C++ class `Executor`.
As described in [graph.md](./graph.md), the first five lines of the following PaddlePaddle program
A simple example PaddlePaddle program can be found in [graph.md](./graph.md):
```python
x = layer.data("images")
......@@ -13,36 +15,112 @@ optimize(cost)
train(cost, reader=mnist.train())
```
generates, or compiles, a PaddelPaddle program, which is represented by the following protobuf message:
The first five lines of the following PaddlePaddle program generates, or, compiles, the `ProgramDesc` message. The last line runs it.
```protobuf
message ProgramDesc {
repeated BlockDesc blocks = 1;
## Programs and Blocks
The basic structure of a PaddlePaddle program is some nested blocks, as a C++ or Java program.
- program: some nested blocks
- [block](./block.md):
- some local variable definitions, and
- a sequence of operators
The concept of block comes from usual programs. For example, the following C++ program has three blocks:
```c++
int main() { // block 0
int i = 0;
if (i < 10) { // block 1
for (int j = 0; j < 10; j++) { // block 2
}
}
return 0;
}
```
The following PaddlePaddle program has three blocks:
```python
import paddle as pd // block 0
x = minibatch([10, 20, 30]) # shape=[None, 1]
y = var(1) # shape=[1], value=1
z = minibatch([10, 20, 30]) # shape=[None, 1]
cond = larger_than(x, 15) # [false, true, true]
ie = pd.ifelse()
with ie.true_block(): // block 1
d = pd.layer.add_scalar(x, y)
ie.output(d, pd.layer.softmax(d))
with ie.false_block(): // block 2
d = pd.layer.fc(z)
ie.output(d, d+1)
o1, o2 = ie(cond)
```
## `BlockDesc` and `ProgramDesc`
All protobuf messages are defined in `framework.proto`.
`BlockDesc` is straight-forward -- it includes local variable definitions, `vars`, and a sequence of operators, `ops`.
```protobuf
message BlockDesc {
required int32 parent = 1;
repeated VarDesc vars = 2;
repeated OpDesc ops = 3;
}
```
The parent ID indicates the parent block so that operators in a block can refer to variables defined locally and also those defined in their ancestor blocks.
All hierarchical blocks in a program are flattened and stored in an array. The block ID is the index of the block in this array.
```protobuf
message ProgramDesc {
repeated BlockDesc blocks = 1;
}
```
### Global Block
The global block is the first one in the above array.
## Operators that Use Blocks
In the above example, the operator `IfElseOp` has two blocks -- the true branch and the false branch.
The definition of `OpDesc` shows that an operator could have some attributes:
```protobuf
message OpDesc {
AttrDesc attrs = 1;
...
}
```
and an attribute could be of type block, which is, in fact, a block ID as described above:
```
message AttrDesc {
required AttrType type = 1;
required string name = 1;
// index into ProgramDesc::blocks when type==BLOCK
optional int32 block = 2;
enum AttrType {
INT = 1,
STRING = 2,
...
BLOCK = ...
}
required AttrType type = 2;
optional int32 block = 10; // when type == BLOCK
...
}
```
When each of the first five lines runs, related Python function, e.g., `layer.fc`, calls C++ InferShape functions. This InferShape function needs to access the properties of VarDesc's accessed by the current OpDesc. These VarDesc's might not be defined in the current block, but in some ancestor blocks. This requires that we can trace the parent of a block.
A nested block is often an attribute of an operator, most likely, an IfElseOp or a WhileOp. In above solution, all blocks are in `ProgramDesc::blocks`, this implicitly assigns a zero-based ID to each block -- the index of the block in `ProgramDesc::blocks`. So that `AttrDesc::block` could be an integer block ID.
## InferShape
With this design, the InferShape function should take the following parameters:
......
# Design Doc: Python API
Due to the refactorization of the PaddlePaddle core, we need Python classes to construct corresponding protobuf messages that describe a DL program.
| Python classes | Protobuf messages |
| --- | --- |
| Program | ProgramDesc |
| Block | BlockDesc |
| Operator | OpDesc |
| Variable | VarDesc |
Please be aware that these Python classes need to maintain some construction-time information, which are not part of the protobuf messages.
## Core Concepts
### Program
A `ProgramDesc` describes a [DL program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md), which is composed of an array of `BlockDesc`s. The `BlockDesc`s in a `ProgramDesc` can have a tree-like hierarchical structure. However, the `ProgramDesc` onlys stores a flattened array of `BlockDesc`s. A `BlockDesc` refers to its parent block by its index in the array. For example, operators in the step block of an RNN operator need to be able to access variables in its ancestor blocks.
Whenever we create a block, we need to set its parent block to the current block, hence the Python class `Program` needs to maintain a data member `current_block`.
```python
class Program(objects):
def __init__(self):
self.proto = core.NewProgram() # a C++ ProgramDesc pointer.
self.blocks = vector<Block>()
self.blocks.append(Block(self, -1)) # the global block
self.current_block = 0 # initialized to the global block
def global_block():
return self.blocks[0]
def current_block():
return self.get_block(self.current_block)
def rollback():
self.current_block = self.current_block().parent_idx
def create_block():
new_block_idx = len(self.block)
self.blocks.append(Block(self, self.current_block))
self.current_block = new_block_idx
return current_block()
```
`Program` is an accessor to the protobuf message `ProgramDesc`, which is created in C++ space, because the InferShape function is in C++, which manipulates `VarDesc` messages, which are in turn members of `BlockDesc`, which is a member of `ProgramDesc`.
`Program` creates the first block as the global block in its constructor. All parameters and their initializer operators are in the global block.
### Block
A [Block](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md) includes
1. a map from variable names to an instance of the Python `Variable` class, and
1. a list of `Operator` instances.
```python
class Block(objects):
def __init__(self, program, parent_idx):
self.proto = core.NewBlock(program.proto)
self.program = program
self.vars = map<string, Variable>()
self.ops = vector<Operator>()
self.parent_idx = parent_idx
def create_var(self, ...):
return Variable(self, ...)
def _create_global_var(self, ...):
program.global_block().create_var(...)
def create_parameter(self, name, ...):
# Parameter is a subclass of variable. See Parameter section for details.
self.vars[name] = Parameter(self._create_global_var(...), ...)
return self.vars[name]
def append_operator(self, ...):
self.ops.append(Operator(self, ...))
def prepend_operator(self, ...): # Parameter's ctor prepands initialize operators.
self.ops.prepend(Operator(self, ...))
```
`create_parameter` is necessary because parameters are global variables, defined in the global block, but can be created in some sub-blocks. For example, an FC layer in the step block of an RNN operator.
`prepend_operator` is necessary because the constructor of `Parameter` needs to create the initialize (or load) operator of the parameter, and would like to put it in the *preamble* of the global block.
### Operator
The `Operator` class fills in the `OpDesc` message and calls the C++ function `InferShape` to infer the output shapes from the input shapes.
```python
class Operator(object):
def __init__(self,
block, # Block
type, # string
inputs, # dict<string, Variable>
outputs,# dict<stirng, Variable>
attrs # dict<string, Any>
):
self.proto = core.NewOpDesc(block.proto, type, inputs, outputs, attrs)
core.infer_shape(self.proto, inputs, outputs)
def type(self):
return self.proto.type()
```
`Operator` creates the `OpDesc` message in C++ space, so that it can call the `InferShape` function, which is in C++.
### Variable
Operators take Variables as its inputs and outputs.
```python
class Variable(object):
def __init__(self,
block=None, # Block
name=None, # string
shape, # tuple
dtype="float32", # string
lod_level=None # int
):
if name is None:
name = unique_name_generator()
self.name = name
self.block = block
self.proto = core.NewVarDesc(block.proto, name, shape, lod_level)
self.writer = None
```
Please be aware of `self.writer`, that tracks operator who creates the variable. It possible that there are more than one operators who write a variable, but in Python space, each write to a variable is represented by a Variable class. This is guaranteed by the fact that **`core.NewVarDesc` must NOT create a new `VarDesc` message if its name already exists in the specified block**.
### Parameter
A parameter is a global variable with an initializer (or load) operator.
```python
class Parameter(Variable):
def __init__(self,
block=None, # Block
name=None, # string
shape, # tuple
dtype="float32", # string
lod_level=None # int
trainable, # bool
initialize_op_attrs,
optimize_op_attrs):
super(Parameter, self).__init__(block, name, shape, dtype, lod_level)
self.trainable = trainable
self.optimize_op_attrs = optimize_op_attrs
block.prepend(Operator(block, # Block
initialize_op_attrs['type'], # string
None, # no inputs
self, # output is the parameter
initialize_op_attrs)
```
When users create a parameter, they can call
```python
program.create_parameter(
...,
init_attr={
type: "uniform_random",
min: -1.0,
max: 1.0,
})
)
```
In above example, `init_attr.type` names an initialize operator. It can also name the load operator
```python
init_attr={
type: "load",
filename: "something.numpy",
}
```
`optimize_op_attrs` is not in the `VarDesc` message, but kept in the Python instance, as it will be used in the Python space when creating the optimize operator's `OpDesc`, and will be in the `OpDesc` message.
## Layer Functions
A layer is a Python function that creates some operators and variables. Layers simplify the work of application programmers.
### Data Layer
```python
def data_layer(name, type, column_name):
block = the_current_program.glolal_block()
var = block.create_global_var(
name=name,
shape=[None] + type.dims(),
dtype=type.dtype)
block.prepend_operator(block,
type="Feed",
inputs = None,
outputs = [var],
{column_name: column_name})
return var
```
The input to the feed operator is a special variable in the global scope, which is the output of [Python readers](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/reader/README.md).
### FC Layer
```python
def fc_layer(input, size, ...):
block = program.current_block()
w = block.create_parameter(...)
b = block.create_parameter(...)
out = block.create_var()
op = block.append_operator("FC", X=input, W=w, b=b, out=out)
out.writer = op
return out
```
# Design Doc: Session
## Abstract
The *session* object encapsulates the environment in which the
computation graph is executed.
We will have the *local* session and *remote* session, they offer the
same [interface](#interface). The local session encapsulates the local
runtime environment and the remote session encapsulates the cluster
runtime environment.
The local runtime environment contains:
1. computation devices (i.e., CPU, GPU) handles, and
1. the [scope](../scope.md) which holds all variables.
The remote runtime environment contains:
1. computation devices (i.e., CPU and GPU on node 0, 1) in a cluster,
and
1. the distributed [scope](../scope.md) in a cluster which holds all
variables.
The user can create a remote session on Paddle Cloud and evaluate the
computation graph with it. In this way, the user can control the
remote computation resource in a cluster from his local computer.
## Background
The current design has an implicit global session in which
`paddle.eval()` is executed. The pain point is:
Since the user is not able to explicitly switch between runtime
environments, the user cannot run a topology in two independent
environments.
For example, in reinforcement learning, the user may want to have a
stale model for inference and a fresh model for training, and only
replace the stale model with the fresh model periodically.
Furthermore, we have no concept that encapsulates a remote environment
that executes a computation graph.
We need the session object to address above issues.
## Session
A session is an object that owns the runtime environment. All
computations are executed through `session.eval()`.
### Interface
```python
eval(
targets,
feed_dict=None,
)
```
Evaluates the target Operations or Variables in `targets`.
- *targets*: the evaluation targets. Can be a single Operation or
Variable, or a list with the Operations or Variables as
elements. The value returned by `eval()` has the same shape as the
`target` argument.
The PaddlePaddle program is represented by
the [ProgramDesc](../design/program.md), `eval()` will infer the
ProgramDesc from the given targets and run the PaddlePaddle
program. Please
see
[this graph](./distributed_architecture.md#local-training-architecture) for
the detailed illustration for the local session
and
[this graph](./distributed_architecture.md#distributed-training-architecture) for
the detailed illustration for the remote session.
- *feed_dict*: a dictionary that contains the tensors which override
the edges of the computation graph.
feed_dict not only can provide the input data, it can override any
OP's input as well:
```python
a = pd.constant(2.0, name="a")
b = pd.variable(name="b")
c = pd.mul(a,b)
sess.eval(targets=c, feed_dict={"b":3.0}) # returns 6.0
```
```python
close()
```
Closes the session and releases the scope that the session owns.
### Create a Local Session
```python
session(
devices=None
)
```
Creates a new session. One session owns one global scope, so creating
multiple sessions will create different scopes.
- *devices*: a single `string` or a list of `string` of device names,
the corresponding devices will be the computation devices for
`eval()`. If not specified, all available devices (e.g., all GPUs)
will be used. The user doesn't need to specify the CPU device since
it will be always used. Multiple sessions can use the same device.
#### Example
```Python
a = paddle.constant(1.0)
b = paddle.constant(2.0)
c = a + b
sess = paddle.session(devices=["gpu:0", "gpu:1", "fpga:0"])
sess.eval(c)
sess.close()
```
### Create a Remote Session
```python
create_cloud_job(
name,
num_trainer,
mem_per_trainer,
gpu_per_trainer,
cpu_per_trainer,
num_ps,
mem_per_ps,
cpu_per_ps,
)
```
Creates a Paddle Cloud job. Fails if the job name exists.
```python
get_cloud_job(
name
)
```
Gets a Paddle Cloud job.
```python
remote_session(
job
)
```
- *job*: the Paddle Cloud job.
#### Example
```Python
reader = paddle.reader.recordio("/pfs/home/peter/mnist-train-*") # data stored on Paddle Cloud
image = reader.column(0)
label = reader.column(1)
fc1 = paddle.op.fc(image, size=256, act="sigmoid")
fc2 = paddle.op.fc(fc1, size=10, act="softmax")
cost = paddle.op.cross_entropy(fc2, label)
opt = paddle.optimizer.sgd(cost)
job = paddle.create_cloud_job("test", 3, "1G", 1, 1, 2, "1G", 1)
sess = paddle.remote_ession(job)
for i in range(1000):
sess.eval(opt)
sess.close()
```
......@@ -17,7 +17,7 @@ The goals of refactoring include:
1. A graph is composed of *variables* and *operators*.
1. The description of graphs must be capable of being serialized/deserialized, so that
1. The description of graphs must be capable of being serialized/deserialized, so that:
1. It can to be sent to the cloud for distributed execution, and
1. It can be sent to clients for mobile or enterprise deployment.
......@@ -137,19 +137,18 @@ Compile Time -> IR -> Runtime
* `Eigen::Tensor` contains basic math and element-wise functions.
* Note that `Eigen::Tensor` has broadcast implementation.
* Limit the number of `tensor.device(dev) = ` in your code.
* `thrust::tranform` and `std::transform`.
* `thrust` has the same API as C++ standard library. Using `transform`, one can quickly implement customized elementwise kernels.
* `thrust::transform` and `std::transform`.
* `thrust` has the same API as C++ standard library. Using `transform`, one can quickly implement customized element-wise kernels.
* `thrust` also has more complex APIs, like `scan`, `reduce`, `reduce_by_key`.
* Hand-writing `GPUKernel` and `CPU` code
* Do not write in header (`.h`) files. CPU Kernel should be in cpp source (`.cc`) and GPU kernels should be in cuda (`.cu`) files. (GCC cannot compile GPU code.)
---
# Operator Registration
## Why registration is necessary?
## Why is registration necessary?
We need a method to build mappings between Op type names and Op classes.
## How is registration implemented?
Maintaining a map, whose key is the type name and the value is the corresponding Op constructor.
---
......@@ -170,7 +169,7 @@ Maintaining a map, whose key is the type name and the value is the corresponding
# Related Concepts
### Op_Maker
It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
It's constructor takes `proto` and `checker`. They are completed during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
### Register Macros
```cpp
......@@ -200,7 +199,7 @@ Make sure the registration process is executed and linked.
---
# Backward Module (2/2)
### Build Backward Network
- **Input**: graph of forwarding operators
- **Input**: graph of forward operators
- **Output**: graph of backward operators
- **Corner cases in construction**
- Shared Variables => insert an `Add` operator to combine gradients
......@@ -224,7 +223,7 @@ Make sure the registration process is executed and linked.
---
# Block (in design)
## the difference with original RNNOp
## the difference between original RNNOp and Block
- As an operator is more intuitive than `RNNOp`,
- Offers a new interface `Eval(targets)` to deduce the minimal block to `Run`,
- Fits the compile-time/ runtime separation design paradigm.
......
# Design Doc: Gradient Operators Registration
## The Problem Posed
In our current operator registration mechanism, for each operator, the programmer should register a *gradient operator creator* function, which takes a C++ operator instance, and returns the corresponding gradient instance.
However, as we decided to separate the *compilation* and *execution* of DL models, we need to reshape the creator to take a protobuf `OpDesc` message, and returns a corresponding message.
More than that, the new registration mechanism need to support the fact that an operators' gradient computation might be a composition of operators.
## Current Implementation
OpInfos store in a association map which key is the operator type. The `grad_op_type` indicate associated gradient operator type. Operator can create gradient operator by `OpInfo::creator_` of gradient. The pseudo code is
```cpp
struct OpInfo {
std::function<OperatorBase*(...)> creator_;
std::string grad_op_type_;
...
};
map<string, OpInfo> OpInfoMap;
OperatorBase* CreateGradientOperator(const OperatorBase& op) {
return OpInfoMap.at(op.Type()).creator_(...);
}
```
## Proposed Solution
The mapping relationship between an operator and its gradient operators is a function. The interface of that function is:
```cpp
// (OpDesc) --> vector<OpDesc>
std::function<std::vector<OpDescBind>(const OpDescBind&)>;
```
The function takes an `OpDescBind` of the forward operator and returns one or many gradient operator descriptions. `OpDescBind` is a C++ wrapper for protobuf message `OpDesc` to manipulate `OpDesc` fast.
The `GradOpDescMaker` will be registered in `OpInfo`, to replace `grad_op_type_` field. The `OpInfo` should be
```cpp
struct OpInfo {
std::function<std::vector<std::unique_ptr<OpDescBind>>(const OpDescBind&)> grad_op_maker_;
...
};
```
The `grad_op_maker_ ` is `nullptr` if the operator does not have associated gradient operators.
We propose a base class called `GradOpDescMakerBase` to let operator developers generate `Gradient Operators` easily. The public interface of that class is
```cpp
class GradOpDescMakerBase {
public:
GradOpDescMakerBase(const OpDescBind& );
virtual std::vector<std::unique_ptr<OpDescBind>> operator()()const = 0;
};
```
We can convert `GradOpDescMakerBase` to `std::function<std::vector<std::unique_ptr<OpDescBind>>(const OpDescBind&)>` by
```cpp
using GradOpMaker = ...;
std::function<std::vector<OpDescBind>(const OpDescBind&)> func;
func = [] (const OpDescBind& fwd_op) {
GradOpMaker maker(fwd_op);
return maker();
};
```
We can write many helper functions since the `GradOpDescMakerBase` is a class now. The basic helper functions get the variables of `Input`, `Output`, `InputGradient` and `OutputGradient` in the forwarding operator.
We should chagne register macros at the same time. In the current solution, there is no difference between forwarding operators and backward operators. So `REGISTER_OP` just register one operator. If the `REGISTER_OPERATOR ` contains `OpProtoAndCheckerMaker` and `GradOpDescMaker`, we just list them in the same macro. It can be done by a macro contains `__VA_ARGS__`.
The user interface should be
```cpp
vector<OpDesc> MinusOpGradMaker(OpDesc) {...}
REGISTER_OPERATOR(minus, MinusOp, MinusOpProtoAndCheckerMaker, SumOpGradMaker);
// Developers can still manually implement gradient operator.
REGISTER_OPERATOR(minus_grad, MinusGradOp);
```
The interface of current `REGISTER_OP` macro could not be changed. In `REGISTER_OP`, it will invoke `REGISTER_OPERATOR` two times and generate GradOpDescMaker inside.
```cpp
REGISTER_OP(minus, MinusOp, MinusOpProtoAndCheckerMaker, minus_grad, MinusGradOp);
```
# Design for TensorArray
This design doc presents the necessity of a new C++ class `TensorArray`.
In addition to the very simple C++ implementation
```c++
class TensorArray {
public:
explicit TensorArray(const LoDTensor&);
explicit TensorArray(size_t size);
private:
vector<LoDTensor> values_;
};
```
We also need to expose it to PaddlePaddle's Python API,
because users would want to use it with our very flexible operators `WhileLoop`.
An example for a RNN based on dynamic operators is
```python
input = pd.data(...)
num_steps = Var(12)
TensorArray states(size=num_steps)
TensorArray step_inputs(unstack_from=input)
TensorArray step_outputs(size=num_steps)
W = Tensor(...)
U = Tensor(...)
default_state = some_op()
step = Var(1)
wloop = paddle.create_whileloop(loop_vars=[step])
with wloop.frame():
wloop.break_if(pd.equal(step, num_steps)
pre_state = states.read(step-1, default_state)
step_input = step_inputs.read(step)
state = pd.sigmoid(pd.matmul(U, pre_state) + pd.matmul(W, step_input))
states.write(step, state)
step_outputs.write(step, state) # output state
step.update(state+1)
output = step_outputs.stack()
```
## Background
Steps are one of the core concepts of RNN. In each time step of RNN, there should be several input segments, states, and output segments; all these components act like arrays, for example, call `states[step_id]` will get the state in `step_id`th time step.
An RNN can be implemented with the following pseudocode
```c++
Array states;
Array input_segments;
Array output_segments;
Parameter W, U;
step = 1
seq_len = 12
while_loop {
if (step == seq_len) break;
states[step] = sigmoid(W * states[step-1] + U * input_segments[step]);
output_segments[step] = states[step] // take state as output
step++;
}
```
According to the [RNN roadmap](https://github.com/PaddlePaddle/Paddle/issues/4561), there are several different RNNs that PaddlePaddle will eventually support.
Currently, the basic RNN implementation supported by PaddlePaddle is the `recurrent_op` which takes tensors as input and splits them into `input_segments`.
Since a tensor cannot store variable-length sequences directly, PaddlePaddle implements the tensor with level of details (`LoDTensor` for short).
Segmenting the `LoDTensor` is much more complicated than splitting a tensor, that makes it necessary to refactor the `recurrent_op` with `LoDTensor` segmenting support.
As the next step in RNN support, `dynamic_recurrent_op` should be introduced to handle inputs with variable-length sequences.
The implementation is similar to `recurrent_op`.
The key difference is the way **the original input `LoDTensors` and outupts are split to get the `input_segments` and the `output_segments`.**
Though it can't be built over `recurrent_op` or `dynamic_recurrent_op` directly,
the logic behind splitting a tensor or a LoD tensor into `input_segments` remains the same.
## Why `TensorArray`
The logic behind splitting the inputs to segments, states and outputs is similar and can be shared in a seperate module.
The array of `states`, `input_segments` and `output_segments` would be exposed to users when writing a dynamic RNN model similar to the above pseudo codes.
So there should be an array-like container, which can store the segments of a tensor or LoD tensor.
**This container can store an array of tensors and provides several methods to split a tensor or a LoD tensor** .
This is where the notion of `TensorArray` comes from.
## Introduce TensorArray to uniform all the three RNNs
TensorArray as a new concept is borrowed from TensorFlow,
it is meant to be used with dynamic iteration primitives such as `while_loop` and `map_fn`.
This concept can be used to support our new design of dynamic operations, and help to refactor some existing variant-sentence-related layers,
such as `RecurrentGradientMachine`.
such as `recurrent_op`, `RecurrentGradientMachine`.
In [our design for dynamic RNN](https://github.com/PaddlePaddle/Paddle/pull/4401),
`TensorArray` is used to segment inputs and store states in all time steps.
By providing some methods similar to a C++ array,
the definition of some state-based dynamic models such as RNN could be more natural and highly flexible.
## Dynamic-Related Methods
Some basic methods should be proposed as follows:
### stack()
Pack the values in a `TensorArray` into a tensor with rank one higher than each tensor in `values`.
### unstack(axis=0)
Unpacks the given dimension of a rank-`R` tensor into rank-`(R-1)` tensors.
### concat()
Return the values in the `TensorArray` as a concatenated Tensor.
### write(index, value, data_shared=true)
Write value into index of the TensorArray.
### read(index)
Read the value at location `index` in the `TensorArray`.
### size()
Return the number of values.
the definition of some state-based dynamic models such as RNN can be more natural and highly flexible.
## Dynamic-operations on TensorArray
`TensorArray` will be used directly when defining dynamic models, so some operators listed below should be implemented
```python
# several helper operators for TensorArray
def tensor_array_stack(ta, tensor):
'''
get a tensor array `ta`, return a packed `tensor`.
'''
pass
def tensor_array_unstack(tensor, ta):
'''
get a `tensor`, unstack it and get a tensor array `ta`.
'''
pass
def tensor_array_write(ta, index, tensor, data_shared):
'''
get a `tensor` and a scalar tensor `index`, write `tensor` into index-th
value of the tensor array `ta`.
`data_shared` is an attribute that specifies whether to copy or reference the tensors.
'''
pass
def tensor_array_read(ta, index, tensor):
'''
get a tensor array `ta`, a scalar tensor `index`, read the index-th value of
`ta` and return as the `tensor`.
'''
pass
def tensor_array_size(ta, tensor):
'''
get a tensor array `ta`, return the size of `ta` and return as the scalar `tensor`.
'''
pass
```
It is trivial for users to use so many low-level operators, so some helper methods should be proposed in python wrapper to make `TensorArray` easier to use,
for example
```python
class TensorArray:
def __init__(self, name):
self.name = name
self.desc = TensorArrayDesc()
def stack(self, name=None):
'''
Pack the values in a `TensorArray` into a tensor with rank one higher
than each tensor in `values`.
`stack` can be used to split tensor into time steps for RNN or whileloop.
@name: str
the name of the variable to output.
'''
tensor = NewVar(name)
tensor_array_stack(self.name, tensor)
return tensor
def unstack(self, input):
'''
Unpacks the given dimension of a rank-`R` tensor into rank-`(R-1)` tensors.
`unstack` can be used to concatenate all the time steps for RNN or whileloop.
@input: str
the name of input tensor
'''
tensor_array_unstack(tensor, self.name)
def write(self, index, value, data_shared=True):
'''
Write value into index of the TensorArray.
If `data_shared` is set to True, than the index-th value in TensorArray will
be shared with the tensor passed in.
@index: str
name of a scalar tensor
@value: str
name of a tensor
@data_shared: bool
'''
tensor_array_write(self.name, index, value, data_shared)
def read(self, index, output):
'''
Read the value at location `index` in the `TensorArray`.
@index: str
name of a scalar tensor
@output:
name of a output variable
'''
tensor_array_read(self.name, index, output)
def size(self, output):
'''
Return the number of values.
@output: str
name of a scalar tensor
'''
tensor_array_size(self.name, output)
```
## LoDTensor-related Supports
The `RecurrentGradientMachine` in Paddle serves as a flexible RNN layer; it takes variant length sequences as input,
because each step of RNN could only take a tensor-represented batch of data as input,
The `RecurrentGradientMachine` in Paddle serves as a flexible RNN layer; it takes varience-length sequences as input, and output sequences too.
Since each step of RNN can only take a tensor-represented batch of data as input,
some preprocess should be taken on the inputs such as sorting the sentences by their length in descending order and cut each word and pack to new batches.
Such cut-like operations can be embedded into `TensorArray` as general methods called `unpack` and `pack`.
Such cut-like operations can be embedded into `TensorArray` as general methods called `unpack` and `pack`,
these two operations are similar to `stack` and `unstack` except that they operate on variable-length sequences formated as a LoD tensor rather than a tensor.
Some definitions are like
```python
def unpack(level):
'''
Split LodTensor in some `level` and generate batches, if set `sort_by_length`,
will sort by length.
With these two methods, a variant-sentence-RNN can be implemented like
Returns:
- a new `TensorArray`, whose values are LodTensors and represents batches
of data.
- an int32 Tensor, which stores the map from the new batch's indices to
original LoDTensor
'''
pass
def pack(level, indices_map):
'''
Recover the original LoD-arranged LoDTensor with the values in a `TensorArray`
and `level` and `indices_map`.
'''
pass
```
With these two methods, a varience-length sentence supported RNN can be implemented like
```c++
// input is the varient-length data
......@@ -58,16 +269,3 @@ LoDTensor rnn_output = ta.pack(ta, indice_map);
```
the code above shows that by embedding the LoDTensor-related preprocess operations into `TensorArray`,
the implementation of a RNN that supports varient-length sentences is far more concise than `RecurrentGradientMachine` because the latter mixes all the codes together, hard to read and extend.
some details are as follows.
### unpack(level, sort_by_length)
Split LodTensor in some `level` and generate batches, if set `sort_by_length`, will sort by length.
Returns:
- a new `TensorArray`, whose values are LodTensors and represents batches of data.
- an int32 Tensor, which stores the map from the new batch's indices to original LoDTensor
### pack(level, indices_map)
Recover the original LoD-arranged LoDTensor with the values in a `TensorArray` and `level` and `indices_map`.
......@@ -206,7 +206,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
- `REGISTER_OP` : 注册`ops::MulOp`类,类型名为`mul`,该类的`ProtoMaker`为`ops::MulOpMaker`,注册`ops::MulOpGrad`,类型名为`mul_grad`。
- `REGISTER_OP_WITHOUT_GRADIENT` : 用于注册没有反向的Op。
- `REGISTER_OP_CPU_KERNEL` :注册`ops::MulKernel`类,并特化模板参数为`paddle::platform::CPUPlace`和`float`类型,同理,注册`ops::MulKernel`类。
- `REGISTER_OP_CPU_KERNEL` :注册`ops::MulKernel`类,并特化模板参数为`paddle::platform::CPUPlace`和`float`类型,同理,注册`ops::MulGradKernel`类。
-`.cu`文件中注册GPU Kernel。
......@@ -285,41 +285,27 @@ class TestMulGradOp(GradientChecker):
'Y': np.random.random((84, 100)).astype("float32")
}
def test_cpu_gpu_compare(self):
self.compare_grad(self.op, self.inputs)
def test_normal(self):
def test_check_grad_normal(self):
# mul op will enlarge the relative error
self.check_grad(
self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
def test_ignore_x(self):
def test_check_grad_ingore_x(self):
self.check_grad(
self.op,
self.inputs, ["Y"],
"Out",
max_relative_error=0.5,
no_grad_set={"X"})
['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
def test_ignore_y(self):
def test_check_grad_ingore_y(self):
self.check_grad(
self.op,
self.inputs, ["X"],
"Out",
max_relative_error=0.5,
no_grad_set={"Y"})
['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
```
下面解释代码中一些关键的地方:
- 调用`create_op("mul")`创建反向Op对应的前向Op。
- 调用`compare_grad`函数对比CPU、GPU计算结果。
- `test_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
- 第一个参数`self.op` : 前向Op。
- 第二个参数`self.inputs` : 输入词典,词典的Key和`ProtoMaker`定义保持一致。
- 第三个参数`["X", "Y"]` : 指定对输入变量`X``Y`做梯度检测。
- 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
- `test_ignore_x``test_ignore_y`分支用来测试只需要计算一个输入梯度的情况。
- `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
- 第一个参数`["X", "Y"]` : 指定对输入变量`X``Y`做梯度检测。
- 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
- 第三个参数`max_relative_error`:指定检测梯度时能容忍的最大错误值。
- `test_check_grad_ingore_x``test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。
### 编译和执行单元测试
......
......@@ -205,7 +205,7 @@ The definition of its corresponding backward operator, if applicable, is similar
- `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`.
- `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient.
- `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulKernel`.
- `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`.
- Registering GPU Kernel in `.cu` files
......@@ -293,41 +293,27 @@ class TestMulGradOp(GradientChecker):
'Y': np.random.random((84, 100)).astype("float32")
}
def test_cpu_gpu_compare(self):
self.compare_grad(self.op, self.inputs)
def test_normal(self):
def test_check_grad_normal(self):
# mul op will enlarge the relative error
self.check_grad(
self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
def test_ignore_x(self):
def test_check_grad_ingore_x(self):
self.check_grad(
self.op,
self.inputs, ["Y"],
"Out",
max_relative_error=0.5,
no_grad_set={"X"})
['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
def test_ignore_y(self):
def test_check_grad_ingore_y(self):
self.check_grad(
self.op,
self.inputs, ["X"],
"Out",
max_relative_error=0.5,
no_grad_set={"Y"})
['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
```
Some key points in the code above include:
- `create_op("mul")` creates the backward operator's corresponding forward operator.
- `compare_grad` compares results between utilizing the CPU and the GPU.
- `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods.
- The first variable `self.op` denotes the forward operator.
- The second variable `self.inputs` denotes the input dictionary, which has its key value identical to its `ProtoMaker` definitions.
- The third variable `["X", "Y"]` appoints `X` and `Y` to be scale tested.
- The fourth variable `"Out"` points to the network's final output target `Out`.
- `test_ignore_x` and `test_ignore_y`branches test the cases where there is only one scaling input.
- The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested.
- The second variable `"Out"` points to the network's final output target `Out`.
- The third variable `max_relative_error` points to the maximum relative tolerance error during scaling tests.
- `test_check_grad_ingore_x` and `test_check_grad_ingore_y`branches test the cases where there is only one scaling input.
### Compiling and Running
......
......@@ -47,7 +47,7 @@ bool isUsingGpu() { return FLAGS_use_gpu; }
void setUseGpu(bool useGpu) { FLAGS_use_gpu = useGpu; }
bool isGpuVersion() {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
return false;
#else
return true;
......
......@@ -46,7 +46,7 @@ paddle_error paddle_matrix_set_row(paddle_matrix mat,
if (rowID >= ptr->mat->getHeight()) return kPD_OUT_OF_RANGE;
paddle::real* buf = ptr->mat->getRowBuf(rowID);
size_t width = ptr->mat->getWidth();
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
hl_memcpy(buf, rowArray, sizeof(paddle::real) * width);
#else
std::copy(rowArray, rowArray + width, buf);
......
......@@ -22,14 +22,12 @@ cc_library(attribute SRCS attribute.cc DEPS framework_proto)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute)
cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope)
cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto proto_desc)
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope proto_desc)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator proto_desc)
cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder op_proto_maker op_info)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator)
cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
py_proto_compile(framework_py_proto SRCS framework.proto)
# Generate an empty __init__.py to make framework_py_proto as a valid python module.
......@@ -43,3 +41,6 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
cc_library(backward SRCS backward.cc DEPS net_op)
cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
cc_library(tensor_array SRCS tensor_array.cc DEPS lod_tensor)
cc_test(tensor_array_test SRCS tensor_array_test.cc DEPS tensor_array place)
......@@ -21,20 +21,12 @@ limitations under the License. */
#include <vector>
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/type_defs.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/variant.h"
namespace paddle {
namespace framework {
// The order should be as same as framework.proto
typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
std::vector<float>, std::vector<std::string>, bool,
std::vector<bool>, BlockDesc*>
Attribute;
typedef std::unordered_map<std::string, Attribute> AttributeMap;
ProgramDesc& GetProgramDesc();
template <typename T>
......
......@@ -13,10 +13,13 @@
limitations under the License. */
#include "paddle/framework/backward.h"
#include "paddle/operators/net_op.h"
#include <deque>
#include <list>
#include <memory>
#include "paddle/framework/block_desc.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
......@@ -24,6 +27,35 @@
namespace paddle {
namespace framework {
static inline std::unique_ptr<OperatorBase> CreateGradOp(
const OperatorBase& op) {
OpDescBind op_desc;
op_desc.SetInputMap(op.Inputs());
op_desc.SetOutputMap(op.Outputs());
op_desc.SetType(op.Type());
op_desc.SetAttrMap(op.Attrs());
auto& info = OpInfoMap::Instance().Get(op.Type());
auto grad_descs = info.GradOpMaker()(op_desc);
std::vector<std::unique_ptr<OperatorBase>> grad_ops;
grad_ops.reserve(grad_descs.size());
std::transform(grad_descs.begin(), grad_descs.end(),
std::back_inserter(grad_ops),
[](const std::unique_ptr<OpDescBind>& grad_desc) {
return OpRegistry::CreateOp(*grad_desc);
});
PADDLE_ENFORCE(!grad_ops.empty());
if (grad_ops.size() == 1) {
return std::move(grad_ops[0]);
} else {
auto net_op = new operators::NetOp();
for (auto& grad_op : grad_ops) {
net_op->AppendOp(std::move(grad_op));
}
net_op->CompleteAddOp();
return std::unique_ptr<OperatorBase>(net_op);
}
}
template <typename Map, typename T>
static void ForEachVarName(const Map& names, T callback) {
for (auto& name : names) {
......@@ -141,9 +173,26 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
net->ops_[op_offset]->Rename(name, dup_outputs.back());
}
// collect all the offset to append `add` op for each alias
insert_position.push_back(
{dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}},
{{"Out", {name}}}, {})});
//
// one variable is shared between multiple operators.
// insert add operator one by one, then add it to output
for (size_t output_idx = 0; output_idx < dup_outputs.size() - 1;
++output_idx) {
auto insert_add_x = dup_outputs[output_idx];
auto insert_add_y = dup_outputs[output_idx + 1];
auto insert_add_out = name + "@SHARED@" + std::to_string(output_idx);
// first add op inserted
if (output_idx == dup_outputs.size() - 2) {
insert_add_out = name;
}
if (output_idx != 0) {
insert_add_y = name + "@SHARED@" + std::to_string(output_idx - 1);
}
insert_position.push_back(
{dup_op.back(),
OpRegistry::CreateOp("sum", {{"X", {insert_add_x, insert_add_y}}},
{{"Out", {insert_add_out}}}, {})});
}
}
// make sure the inserted `add` ops follow the BFS order.
......@@ -154,7 +203,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
net->InsertOp(pos.first + 1, std::move(pos.second));
}
} else {
std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
std::unique_ptr<OperatorBase> grad_op(CreateGradOp(forwardOp));
ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
const std::string& grad_input) {
......@@ -182,7 +231,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
// process recurrent gradient op as a special operator.
if (forwardOp.Type() == "recurrent") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
// or
// this will result in infinite loop.
const auto& rnnop =
*static_cast<const operators::RecurrentOp*>(&forwardOp);
......@@ -222,5 +272,145 @@ std::unique_ptr<OperatorBase> Backward(
return BackwardRecursive(forwardOp, no_grad_names, uid);
}
// ==================================== //
static bool AllGradInSet(const std::vector<std::string>& names,
const std::unordered_set<std::string>& set) {
for (const std::string& name : names) {
if (!set.count(GradVarName(name))) {
return false;
}
}
return true;
}
std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
const std::unique_ptr<OpDescBind>& op_desc,
std::unordered_set<std::string>& no_grad_vars) {
std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
// All input gradients of forwarding operator do not need to calculat.
const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
if (AllGradInSet(inputs, no_grad_vars)) {
return grad_op_descs; // empty vector
}
// All output gradients of forwarding operator do not need to calculate.
const std::vector<std::string>& outputs = op_desc->OutputArgumentNames();
if (AllGradInSet(outputs, no_grad_vars)) {
for (const std::string& name : inputs) {
no_grad_vars.insert(GradVarName(name));
}
return grad_op_descs; // empty vector
}
grad_op_descs = OpRegistry::CreateGradOpDescs(*op_desc);
std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
for (auto& desc : grad_op_descs) {
for (const std::string& in_name : desc->InputArgumentNames()) {
if (no_grad_vars.count(in_name)) {
std::string prefix = in_name.substr(
0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
std::string new_name = prefix + kZeroVarSuffix;
desc->Rename(in_name, new_name);
std::unique_ptr<OpDescBind> fill_zeros_op(new OpDescBind(
"fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {}));
pending_fill_zeros_ops.push_back(std::move(fill_zeros_op));
}
}
for (const std::string& out_name : desc->OutputArgumentNames()) {
if (no_grad_vars.count(out_name)) {
desc->Rename(out_name, kEmptyVarName);
}
}
}
for (auto& p : pending_fill_zeros_ops) {
grad_op_descs.insert(grad_op_descs.begin(), std::move(p));
}
return grad_op_descs;
}
std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
ProgramDescBind& program_desc, int block_idx,
std::unordered_set<std::string>& no_grad_vars) {
BlockDescBind* cur_block = program_desc.Block(block_idx);
std::deque<std::unique_ptr<OpDescBind>>& op_descs = cur_block->ops_;
std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
size_t grad_desc_idx = 0;
std::vector<std::unique_ptr<OpDescBind>> backward_descs;
for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
std::vector<std::unique_ptr<OpDescBind>> op_grads =
MakeOpGrad(*it, no_grad_vars);
if ((*it)->Type() == "recurrent") {
PADDLE_ENFORCE_EQ(
op_grads.size(), size_t(1),
"rnn_op's gradient process should contain only one op.");
int step_block_idx = (*it)->GetBlockAttr("stop_block");
auto backward_block_op_descs =
MakeBlockBackward(program_desc, step_block_idx, no_grad_vars);
BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block);
for (auto& ptr : backward_block_op_descs) {
backward_block->ops_.push_back(std::move(ptr));
}
op_grads[0]->SetBlockAttr("step_block", *backward_block);
}
for (const auto& desc : op_grads) {
for (const std::string& out_name : desc->OutputArgumentNames()) {
dup_out_ops[out_name].emplace_back(grad_desc_idx);
}
++grad_desc_idx;
}
std::transform(
op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
[](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
}
// Check whether some variables are written more than once
std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
for (const auto& dup : dup_out_ops) {
const std::string& out_name = dup.first;
const std::vector<size_t> dup_op = dup.second;
if (out_name != kEmptyVarName && dup_op.size() > 1) {
std::vector<std::string> sum_op_inputs;
for (size_t i = 0; i < dup_op.size(); ++i) {
std::string new_name = out_name + "@RENAME@" + std::to_string(i);
backward_descs[dup_op[i]]->Rename(out_name, new_name);
sum_op_inputs.emplace_back(new_name);
}
std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
"sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
}
}
pending_sum_ops.sort(
[](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
return a.first > b.first;
});
for (auto& p : pending_sum_ops) {
backward_descs.insert(backward_descs.begin() + p.first + 1,
std::move(p.second));
}
return backward_descs;
}
void AppendBackward(ProgramDescBind& program_desc,
const std::unordered_set<std::string>& no_grad_vars) {
std::unordered_set<std::string> no_grad_var_names;
no_grad_var_names.reserve(no_grad_vars.size() + 1);
no_grad_var_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
for (auto& name : no_grad_vars) {
no_grad_var_names.insert(GradVarName(name));
}
const int root_block_idx = 0;
auto backward_op_descs =
MakeBlockBackward(program_desc, root_block_idx, no_grad_var_names);
auto& forw_op_descs = program_desc.Block(root_block_idx)->ops_;
for (auto& ptr : backward_op_descs) {
forw_op_descs.push_back(std::move(ptr));
}
}
} // namespace framework
} // namespace paddle
......@@ -13,8 +13,11 @@
limitations under the License. */
#pragma once
#include <unordered_set>
#include "operator.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/program_desc.h"
namespace paddle {
namespace framework {
......@@ -23,5 +26,9 @@ namespace framework {
extern std::unique_ptr<OperatorBase> Backward(
const OperatorBase& forwardOp,
const std::unordered_set<std::string>& no_grad_vars);
void AppendBackward(ProgramDescBind& program_desc,
const std::unordered_set<std::string>& no_grad_vars);
} // namespace framework
} // namespace paddle
......@@ -15,30 +15,42 @@
#include "paddle/framework/backward.h"
#include <gtest/gtest.h>
#include "paddle/framework/block_desc.h"
#include "paddle/framework/op_desc.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
namespace paddle {
namespace framework {
using OperatorBase = framework::OperatorBase;
using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
using OpProto = framework::OpProto;
using OpAttrChecker = framework::OpAttrChecker;
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "Input X of Add").NotInGradient();
AddInput("b", "Bias of Add").NotInGradient();
AddOutput("Out", "Out of Add").NotInGradient();
AddInput("X", "Input X of Add");
AddInput("b", "Bias of Add");
AddOutput("Out", "Out of Add");
AddComment("Add Op");
}
};
class RowWiseAddGradMaker : public SingleGradOpDescMaker {
public:
using SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<OpDescBind> Apply() const override {
auto grad_op = new OpDescBind();
grad_op->SetInput(GradVarName("Out"), OutputGrad("Out"));
grad_op->SetOutput(GradVarName("X"), InputGrad("X"));
grad_op->SetOutput(GradVarName("b"), InputGrad("b"));
grad_op->SetType("rowwise_add_grad");
return std::unique_ptr<OpDescBind>(grad_op);
}
};
class MulOpMaker : public OpProtoAndCheckerMaker {
public:
MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
......@@ -133,42 +145,46 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
}
};
class AddOpMaker : public OpProtoAndCheckerMaker {
class SumOpMaker : public framework::OpProtoAndCheckerMaker {
public:
AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "x").AsDuplicable();
AddOutput("Out", "out");
AddInput("X", "the input tensors of sum operator.").AsDuplicable();
AddOutput("Out", "the output tensor of sum operator.");
AddComment("");
}
};
class MultInOutOpMaker : public OpProtoAndCheckerMaker {
public:
MultInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "x");
AddInput("H", "h");
AddOutput("Y", "y");
AddOutput("Z", "z");
AddComment("");
}
};
} // namespace framework
} // namespace paddle
namespace f = paddle::framework;
namespace ops = paddle::operators;
using EnforceNotMet = paddle::platform::EnforceNotMet;
REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
f::NOP);
REGISTER_OPERATOR(rowwise_add, f::NOP, f::RowWiseAddOpMaker,
f::RowWiseAddGradMaker);
REGISTER_OPERATOR(rowwise_add_grad, f::NOP);
REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
f::NOP);
TEST(Backward, simple_op_grad) {
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
ASSERT_NE(fwd, nullptr);
auto gop = f::OpRegistry::CreateGradOp(*fwd);
ASSERT_EQ(1UL, gop->Inputs().size());
ASSERT_EQ("rowwise_add_grad", gop->Type());
ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
}
REGISTER_OP(mult_in_out, f::NOP, f::MultInOutOpMaker, mult_in_out_grad, f::NOP);
TEST(Backward, simple_op_not_need_grad) {
auto fwd = f::OpRegistry::CreateOp(
......@@ -283,18 +299,7 @@ TEST(Backward, net_shared_weight) {
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
ASSERT_EQ(3UL, bwd_net->ops_.size());
ASSERT_EQ("add", bwd_net->ops_[2]->Type());
}
TEST(Backward, op_register_grad_not_for_network) {
auto fwd =
f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
{{"mul_result", {"mul_out"}},
{"add_result", {"add_out"}},
{"Out", {"out1"}}},
{{"temporary_index", std::vector<int>{0, 1}}});
ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet);
ASSERT_EQ("sum", bwd_net->ops_[2]->Type());
}
TEST(Backward, op_all_input_are_not_need) {
......@@ -399,3 +404,293 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
}
// =================================== //
f::ProgramDesc *GetNewProgramDesc() {
auto *program_desc = new f::ProgramDesc();
auto *root_block = program_desc->add_blocks();
root_block->set_idx(0);
root_block->set_parent_idx(-1);
return program_desc;
}
TEST(Backward, simple_single_op) {
f::ProgramDesc *program_desc = GetNewProgramDesc();
f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
f::BlockDescBind *block = program.Block(0);
f::OpDescBind *op = block->AppendOp();
op->SetType("rowwise_add");
op->SetInput("X", {"x"});
op->SetInput("b", {"b"});
op->SetOutput("Out", {"out"});
AppendBackward(program, {});
ASSERT_EQ(block->AllOps().size(), 2UL);
f::OpDescBind *grad_op = block->AllOps()[1];
EXPECT_EQ(grad_op->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op->InputNames().size(), 1UL);
ASSERT_EQ(grad_op->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b")}));
}
TEST(Backward, simple_mult_op) {
f::ProgramDesc *program_desc = GetNewProgramDesc();
f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
f::BlockDescBind *block = program.Block(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
op1->SetInput("b", {"b1"});
op1->SetOutput("Out", {"out1"});
f::OpDescBind *op2 = block->AppendOp();
op2->SetType("mul");
op2->SetInput("X", {"out1"});
op2->SetInput("Y", {"y2"});
op2->SetOutput("Out", {"out2"});
f::OpDescBind *op3 = block->AppendOp();
op3->SetType("rowwise_add");
op3->SetInput("X", {"out2"});
op3->SetInput("b", {"b3"});
op3->SetOutput("Out", {"out3"});
AppendBackward(program, {});
ASSERT_EQ(block->AllOps().size(), 6UL);
f::OpDescBind *grad_op1 = block->AllOps()[5];
EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b1")}));
f::OpDescBind *grad_op2 = block->AllOps()[4];
EXPECT_EQ(grad_op2->Type(), "mul_grad");
ASSERT_EQ(grad_op2->InputNames().size(), 4UL);
ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op2->Input("X"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op2->Input("Y"), std::vector<std::string>({"y2"}));
EXPECT_EQ(grad_op2->Input("Out"), std::vector<std::string>({"out2"}));
EXPECT_EQ(grad_op2->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out2")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y2")}));
f::OpDescBind *grad_op3 = block->AllOps()[3];
EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out3")}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out2")}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b3")}));
}
TEST(Backward, intermedia_var_no_grad) {
f::ProgramDesc *program_desc = GetNewProgramDesc();
f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
f::BlockDescBind *block = program.Block(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
op1->SetInput("b", {"b1"});
op1->SetOutput("Out", {"out1"});
f::OpDescBind *op2 = block->AppendOp();
op2->SetType("mul");
op2->SetInput("X", {"x2"});
op2->SetInput("Y", {"y2"});
op2->SetOutput("Out", {"out2"});
f::OpDescBind *op3 = block->AppendOp();
op3->SetType("rowwise_add");
op3->SetInput("X", {"out2"});
op3->SetInput("b", {"b3"});
op3->SetOutput("Out", {"out3"});
f::OpDescBind *op4 = block->AppendOp();
op4->SetType("mul");
op4->SetInput("X", {"out1"});
op4->SetInput("Y", {"out3"});
op4->SetOutput("Out", {"out4"});
AppendBackward(program, {"out3"});
ASSERT_EQ(block->AllOps().size(), 6UL);
f::OpDescBind *grad_op1 = block->AllOps()[5];
EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b1")}));
f::OpDescBind *grad_op4 = block->AllOps()[4];
EXPECT_EQ(grad_op4->Type(), "mul_grad");
ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op4->Input("X"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op4->Input("Y"), std::vector<std::string>({"out3"}));
EXPECT_EQ(grad_op4->Input("Out"), std::vector<std::string>({"out4"}));
EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out4")}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
std::vector<std::string>({f::kEmptyVarName}));
}
TEST(Backward, var_no_grad) {
f::ProgramDesc *program_desc = GetNewProgramDesc();
f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
f::BlockDescBind *block = program.Block(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("mult_in_out");
op1->SetInput("X", {"x1"});
op1->SetInput("H", {"h1"});
op1->SetOutput("Y", {"y1"});
op1->SetOutput("Z", {"z1"});
f::OpDescBind *op2 = block->AppendOp();
op2->SetType("mult_in_out");
op2->SetInput("X", {"y1"});
op2->SetInput("H", {"z1"});
op2->SetOutput("Y", {"y2"});
op2->SetOutput("Z", {"z2"});
AppendBackward(program, {"z1"});
ASSERT_EQ(block->AllOps().size(), 5UL);
f::OpDescBind *grad_op2 = block->AllOps()[2];
ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad");
ASSERT_EQ(grad_op2->InputNames().size(), 6UL);
ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op2->Input("X"), std::vector<std::string>({"y1"}));
EXPECT_EQ(grad_op2->Input("H"), std::vector<std::string>({"z1"}));
EXPECT_EQ(grad_op2->Input("Y"), std::vector<std::string>({"y2"}));
EXPECT_EQ(grad_op2->Input("Z"), std::vector<std::string>({"z2"}));
EXPECT_EQ(grad_op2->Input(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y2")}));
EXPECT_EQ(grad_op2->Input(f::GradVarName("Z")),
std::vector<std::string>({f::GradVarName("z2")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("y1")}));
EXPECT_EQ(grad_op2->Output(f::GradVarName("H")),
std::vector<std::string>({f::kEmptyVarName}));
f::OpDescBind *fill_zero_op = block->AllOps()[3];
ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like");
ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL);
ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL);
EXPECT_EQ(fill_zero_op->Input("X"), std::vector<std::string>({"z1"}));
EXPECT_EQ(fill_zero_op->Output("Y"),
std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
f::OpDescBind *grad_op1 = block->AllOps()[4];
ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 6UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input("X"), std::vector<std::string>({"x1"}));
EXPECT_EQ(grad_op1->Input("H"), std::vector<std::string>({"h1"}));
EXPECT_EQ(grad_op1->Input("Y"), std::vector<std::string>({"y1"}));
EXPECT_EQ(grad_op1->Input("Z"), std::vector<std::string>({"z1"}));
EXPECT_EQ(grad_op1->Input(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y1")}));
EXPECT_EQ(grad_op1->Input(f::GradVarName("Z")),
std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("H")),
std::vector<std::string>({f::GradVarName("h1")}));
}
TEST(Backward, shared_var) {
f::ProgramDesc *program_desc = GetNewProgramDesc();
f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
f::BlockDescBind *block = program.Block(0);
f::OpDescBind *op1 = block->AppendOp();
op1->SetType("rowwise_add");
op1->SetInput("X", {"x1"});
op1->SetInput("b", {"b1"});
op1->SetOutput("Out", {"out1"});
f::OpDescBind *op2 = block->AppendOp();
op2->SetType("mul");
op2->SetInput("X", {"out1"});
op2->SetInput("Y", {"y2"});
op2->SetOutput("Out", {"out2"});
f::OpDescBind *op3 = block->AppendOp();
op3->SetType("rowwise_add");
op3->SetInput("X", {"out1"});
op3->SetInput("b", {"b3"});
op3->SetOutput("Out", {"out3"});
AppendBackward(program, {});
ASSERT_EQ(block->AllOps().size(), 7UL);
f::OpDescBind *grad_op3 = block->AllOps()[3];
ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op3->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out3")}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1") + "@RENAME@0"}));
EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b3")}));
f::OpDescBind *grad_op4 = block->AllOps()[4];
ASSERT_EQ(grad_op4->Type(), "mul_grad");
ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op4->Input("X"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op4->Input("Y"), std::vector<std::string>({"y2"}));
EXPECT_EQ(grad_op4->Input("Out"), std::vector<std::string>({"out2"}));
EXPECT_EQ(grad_op4->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out2")}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("out1") + "@RENAME@1"}));
EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
std::vector<std::string>({f::GradVarName("y2")}));
f::OpDescBind *sum_op = block->AllOps()[5];
ASSERT_EQ(sum_op->Type(), "sum");
ASSERT_EQ(sum_op->InputNames().size(), 1UL);
ASSERT_EQ(sum_op->OutputNames().size(), 1UL);
EXPECT_EQ(sum_op->Input("X"),
std::vector<std::string>({f::GradVarName("out1") + "@RENAME@0",
f::GradVarName("out1") + "@RENAME@1"}));
EXPECT_EQ(sum_op->Output("Out"),
std::vector<std::string>({f::GradVarName("out1")}));
f::OpDescBind *grad_op1 = block->AllOps()[6];
ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad");
ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
EXPECT_EQ(grad_op1->Input(f::GradVarName("Out")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("X")),
std::vector<std::string>({f::GradVarName("x1")}));
EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
std::vector<std::string>({f::GradVarName("b1")}));
}
\ No newline at end of file
......@@ -34,6 +34,10 @@ VarDescBind *BlockDescBind::Var(const std::string &name) const {
return it->second.get();
}
bool BlockDescBind::HasVar(const std::string &name) const {
return vars_.find(name) != vars_.end();
}
std::vector<VarDescBind *> BlockDescBind::AllVars() const {
std::vector<VarDescBind *> res;
for (const auto &p : vars_) {
......
......@@ -19,6 +19,7 @@ limitations under the License. */
#include <vector>
#include "paddle/framework/op_desc.h"
#include "paddle/framework/var_desc.h"
#include "paddle/platform/macros.h"
namespace paddle {
namespace framework {
......@@ -31,12 +32,17 @@ class ProgramDescBind;
class BlockDescBind {
public:
friend std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
ProgramDescBind &program_desc, int block_idx,
std::unordered_set<std::string> &no_grad_vars);
friend void AppendBackward(
ProgramDescBind &program_desc,
const std::unordered_set<std::string> &no_grad_vars);
BlockDescBind(ProgramDescBind *prog, BlockDesc *desc)
: prog_(prog), desc_(desc), need_update_(false) {}
BlockDescBind(const BlockDescBind &o) = delete;
BlockDescBind &operator=(const BlockDescBind &o) = delete;
int32_t ID() const { return desc_->idx(); }
int32_t Parent() const { return desc_->parent_idx(); }
......@@ -45,6 +51,8 @@ class BlockDescBind {
VarDescBind *Var(const std::string &name_bytes) const;
bool HasVar(const std::string &var_name) const;
std::vector<VarDescBind *> AllVars() const;
BlockDescBind *ParentBlock() const;
......@@ -66,6 +74,8 @@ class BlockDescBind {
std::deque<std::unique_ptr<OpDescBind>> ops_;
std::unordered_map<std::string, std::unique_ptr<VarDescBind>> vars_;
DISABLE_COPY_AND_ASSIGN(BlockDescBind);
};
} // namespace framework
} // namespace paddle
/* 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 "paddle/framework/grad_op_desc_maker.h"
#include "paddle/framework/op_info.h"
#include "paddle/framework/op_proto_maker.h"
#include "paddle/framework/operator.h"
namespace paddle {
namespace framework {
namespace details {
enum OpInfoFillType {
kOperator = 0,
kOpProtoAndCheckerMaker = 1,
kGradOpDescMaker = 2
};
template <typename T>
struct OpInfoFillTypeID {
static constexpr OpInfoFillType ID() {
return std::is_base_of<OperatorBase, T>::value
? kOperator
: (std::is_base_of<OpProtoAndCheckerMaker, T>::value
? kOpProtoAndCheckerMaker
: (std::is_base_of<GradOpDescMakerBase, T>::value
? kGradOpDescMaker
: static_cast<OpInfoFillType>(-1)));
}
};
template <typename T, OpInfoFillType = OpInfoFillTypeID<T>::ID()>
struct OpInfoFiller;
template <size_t I, bool at_end, typename... ARGS>
class OperatorRegistrarRecursive;
template <size_t I, typename... ARGS>
class OperatorRegistrarRecursive<I, false, ARGS...> {
public:
using T = typename std::tuple_element<I, std::tuple<ARGS...>>::type;
OperatorRegistrarRecursive(const char* op_type, OpInfo* info) {
OpInfoFiller<T> fill;
fill(op_type, info);
constexpr auto size = sizeof...(ARGS);
OperatorRegistrarRecursive<I + 1, I + 1 == size, ARGS...> reg(op_type,
info);
(void)(reg);
}
};
template <size_t I, typename... ARGS>
class OperatorRegistrarRecursive<I, true, ARGS...> {
public:
OperatorRegistrarRecursive(const char* op_type, OpInfo* info) {}
};
template <typename T>
struct OpInfoFiller<T, kOperator> {
void operator()(const char* op_type, OpInfo* info) const {
info->creator_ = [](const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs,
const AttributeMap& attrs) {
return new T(type, inputs, outputs, attrs);
};
}
};
template <typename T>
struct OpInfoFiller<T, kOpProtoAndCheckerMaker> {
void operator()(const char* op_type, OpInfo* info) const {
info->proto_ = new OpProto;
info->checker_ = new OpAttrChecker();
auto maker = T(info->proto_, info->checker_);
maker.Validate();
info->proto_->set_type(op_type);
PADDLE_ENFORCE(
info->proto_->IsInitialized(),
"Fail to initialize %s's OpProto, because %s is not initialized",
op_type, info->proto_->InitializationErrorString());
}
};
template <typename T>
struct OpInfoFiller<T, kGradOpDescMaker> {
void operator()(const char* op_type, OpInfo* info) const {
info->grad_op_maker_ = [](const OpDescBind& fwd_op) {
T maker(fwd_op);
return maker();
};
}
};
} // namespace details
} // namespace framework
} // namespace paddle
......@@ -66,7 +66,6 @@ message OpProto {
optional bool duplicable = 3 [ default = false ];
optional bool intermediate = 4 [ default = false ];
optional bool not_in_gradient = 5 [ default = false ];
}
// AttrProto describes the C++ type Attribute.
......@@ -106,6 +105,7 @@ message LoDTensorDesc {
message VarDesc {
required string name = 1;
optional LoDTensorDesc lod_tensor = 2;
optional bool persistable = 3 [ default = false ];
}
message BlockDesc {
......@@ -115,4 +115,7 @@ message BlockDesc {
repeated OpDesc ops = 4;
}
// Please refer to
// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md
// for more details.
message ProgramDesc { repeated BlockDesc blocks = 1; }
/* 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,
WITHOpArgType::OUT WARRANTIES OR CONDITIONS OF ANY KOpArgType::IND, either
express or implied. See the License for the specific language governing
permissions and limitations under the License. */
#include "paddle/framework/grad_op_builder.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace framework {
enum class OpArgType { IN, OUT };
static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
bool is_grad, VariableNameMap* vars) {
const auto& src_inout =
src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
auto& dst_inout = *vars;
auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
const auto& src_arg_list =
src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
for (const auto& arg : src_arg_list) {
if (arg.not_in_gradient() && !is_grad) continue;
const std::string src_name = arg.name();
std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
dst_inout[dst_name].reserve(src_inout.at(src_name).size());
for (auto& var_name : src_inout.at(src_name)) {
std::string s = is_grad ? GradVarName(var_name) : var_name;
dst_inout[dst_name].emplace_back(s);
}
}
}
OperatorBase* BuildGradOp(const OperatorBase* op) {
auto& info = OpInfoMap::Instance().Get(op->Type());
PADDLE_ENFORCE(info.HasGradientOp());
VariableNameMap inputs;
VariableNameMap outputs;
TransOpArg(op, OpArgType::IN, false, &inputs); // I
TransOpArg(op, OpArgType::OUT, false, &inputs); // O
TransOpArg(op, OpArgType::OUT, true, &inputs); // OG
TransOpArg(op, OpArgType::IN, true, &outputs); // IG
auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_);
return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs());
}
static void TransOpDescArg(const OpDescBind* src_op, const OpArgType& src_type,
bool is_grad, OpDescBind* dst_op,
const OpArgType& dst_type) {
PADDLE_ENFORCE(dst_op != nullptr,
"Protobuf desc of gradient op must be initialized first.");
const auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
const auto& src_arg_list =
src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
for (const auto& arg : src_arg_list) {
if (arg.not_in_gradient() && !is_grad) continue;
const std::string src_name = arg.name();
std::vector<std::string> vars = src_type == OpArgType::IN
? src_op->Input(src_name)
: src_op->Output(src_name);
if (is_grad) {
for (std::string& var : vars) {
var = GradVarName(var);
}
}
std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
dst_type == OpArgType::IN ? dst_op->SetInput(dst_name, vars)
: dst_op->SetOutput(dst_name, vars);
}
}
void CompleteGradOpDesc(const OpDescBind* forw_op, OpDescBind* grad_op) {
auto& info = OpInfoMap::Instance().Get(forw_op->Type());
PADDLE_ENFORCE(info.HasGradientOp());
grad_op->SetType(info.grad_op_type_);
TransOpDescArg(forw_op, OpArgType::IN, false, grad_op, OpArgType::IN);
TransOpDescArg(forw_op, OpArgType::OUT, false, grad_op, OpArgType::IN);
TransOpDescArg(forw_op, OpArgType::OUT, true, grad_op, OpArgType::IN);
TransOpDescArg(forw_op, OpArgType::IN, true, grad_op, OpArgType::OUT);
grad_op->SetAttrMap(forw_op->GetAttrMap());
}
} // namespace framework
} // namespace paddle
#include "paddle/framework/grad_op_builder.h"
#include <gtest/gtest.h>
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
USE_OP(add);
namespace paddle {
namespace framework {
class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
public:
MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("In1", "a single input");
AddInput("In2_mult", "a multiple input").AsDuplicable();
AddInput("In3", "another single input");
AddOutput("Out1", "a single output");
AddOutput("Out2_mult", "a multiple output").AsDuplicable();
AddComment("test op with multiple inputs and outputs");
}
};
class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
public:
IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("In1", "a single input");
AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient();
AddInput("In3_mult", "another multiple input").AsDuplicable();
AddOutput("Out1_mult", "a multiple output").AsDuplicable();
AddOutput("Out2", "a single output").NotInGradient();
AddComment("op with inputs and outputs ignored in gradient calculating");
}
};
} // namespace framework
} // namespace paddle
namespace f = paddle::framework;
TEST(GradOpBuilder, AddTwo) {
std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
"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);
EXPECT_EQ(grad_add_op->Outputs().size(), 2UL);
EXPECT_EQ(grad_add_op->Input("X"), "x");
EXPECT_EQ(grad_add_op->Input("Y"), "y");
EXPECT_EQ(grad_add_op->Input("Out"), "out");
EXPECT_EQ(grad_add_op->Input(f::GradVarName("Out")), f::GradVarName("out"));
EXPECT_EQ(grad_add_op->Output(f::GradVarName("X")), f::GradVarName("x"));
EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
}
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
TEST(GradOpBuilder, MutiInOut) {
std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
"mult_io", {{"In1", {"in1"}},
{"In2_mult", {"in2_1", "in2_2", "in2_3"}},
{"In3", {"in3"}}},
{{"Out1", {"out1"}}, {"Out2_mult", {"out2_1", "out2_2"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_test_op =
f::OpRegistry::CreateGradOp(*test_op);
ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
EXPECT_EQ(grad_test_op->Input("In3"), "in3");
EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
std::vector<std::string>({"out2_1", "out2_2"}));
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")),
f::GradVarName("out1"));
EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")),
std::vector<std::string>(
{f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>({f::GradVarName("in2_1"),
f::GradVarName("in2_2"),
f::GradVarName("in2_3")}));
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3"));
}
TEST(GradOpBuilder, IOIgnoredInGradient) {
std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
"io_ignored", {{"In1", {"in1"}},
{"In2_mult", {"in2_1", "in2_2"}},
{"In3_mult", {"in3_1", "in3_2"}}},
{{"Out1_mult", {"out1_1", "out1_2"}}, {"Out2", {"out2"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_test_op =
f::OpRegistry::CreateGradOp(*test_op);
// 'In2' and 'Out2' are ignored in gradient calculating
ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
std::vector<std::string>({"in3_1", "in3_2"}));
EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
std::vector<std::string>({"out1_1", "out1_2"}));
EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
std::vector<std::string>(
{f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
f::GradVarName("out2"));
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>(
{f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")),
std::vector<std::string>(
{f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
}
TEST(GradOpDescBuilder, MutiInOut) {
f::OpDescBind *forw_op = new f::OpDescBind();
forw_op->SetType("mult_io");
forw_op->SetInput("In1", {"in1"});
forw_op->SetInput("In2_mult", {"in2_1", "in2_2", "in2_3"});
forw_op->SetInput("In3", {"in3"});
forw_op->SetOutput("Out1", {"out1"});
forw_op->SetOutput("Out2_mult", {"out2_1", "out2_2"});
f::OpDescBind *grad_op = new f::OpDescBind();
f::CompleteGradOpDesc(forw_op, grad_op);
EXPECT_EQ(grad_op->Type(), "mult_io_grad");
ASSERT_EQ(grad_op->InputNames().size(), 3UL + 2UL + 2UL);
EXPECT_EQ(grad_op->Input("In1"), std::vector<std::string>({"in1"}));
EXPECT_EQ(grad_op->Input("In2_mult"),
std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
EXPECT_EQ(grad_op->Input("In3"), std::vector<std::string>({"in3"}));
EXPECT_EQ(grad_op->Input("Out1"), std::vector<std::string>({"out1"}));
EXPECT_EQ(grad_op->Input("Out2_mult"),
std::vector<std::string>({"out2_1", "out2_2"}));
EXPECT_EQ(grad_op->Input(f::GradVarName("Out1")),
std::vector<std::string>({f::GradVarName("out1")}));
EXPECT_EQ(grad_op->Input(f::GradVarName("Out2_mult")),
std::vector<std::string>(
{f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
ASSERT_EQ(grad_op->OutputNames().size(), 3UL);
EXPECT_EQ(grad_op->Output(f::GradVarName("In1")),
std::vector<std::string>({f::GradVarName("in1")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("In2_mult")),
std::vector<std::string>({f::GradVarName("in2_1"),
f::GradVarName("in2_2"),
f::GradVarName("in2_3")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("In3")),
std::vector<std::string>({f::GradVarName("in3")}));
delete forw_op;
delete grad_op;
}
TEST(GradOpDescBuilder, IOIgnoredInGradient) {
f::OpDescBind *forw_op = new f::OpDescBind();
forw_op->SetType("io_ignored");
forw_op->SetInput("In1", {"in1"});
forw_op->SetInput("In2_mult", {"in2_1", "in2_2"});
forw_op->SetInput("In3_mult", {"in3_1", "in3_2"});
forw_op->SetOutput("Out1_mult", {"out1_1", "out1_2"});
forw_op->SetOutput("Out2", {"out2"});
f::OpDescBind *grad_op = new f::OpDescBind();
f::CompleteGradOpDesc(forw_op, grad_op);
EXPECT_EQ(grad_op->Type(), "io_ignored_grad");
// 'In2' and 'Out2' are ignored in gradient calculating
ASSERT_EQ(grad_op->InputNames().size(), 2UL + 1UL + 2UL);
EXPECT_EQ(grad_op->Input("In1"), std::vector<std::string>({"in1"}));
EXPECT_EQ(grad_op->Input("In3_mult"),
std::vector<std::string>({"in3_1", "in3_2"}));
EXPECT_EQ(grad_op->Input("Out1_mult"),
std::vector<std::string>({"out1_1", "out1_2"}));
EXPECT_EQ(grad_op->Input(f::GradVarName("Out1_mult")),
std::vector<std::string>(
{f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
EXPECT_EQ(grad_op->Input(f::GradVarName("Out2")),
std::vector<std::string>({f::GradVarName("out2")}));
ASSERT_EQ(grad_op->OutputNames().size(), 3UL);
EXPECT_EQ(grad_op->Output(f::GradVarName("In1")),
std::vector<std::string>({f::GradVarName("in1")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("In2_mult")),
std::vector<std::string>(
{f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
EXPECT_EQ(grad_op->Output(f::GradVarName("In3_mult")),
std::vector<std::string>(
{f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
delete forw_op;
delete grad_op;
}
\ No newline at end of file
/* 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 "paddle/framework/op_desc.h"
#include "paddle/framework/operator.h"
namespace paddle {
namespace framework {
class GradOpDescMakerBase {
public:
explicit GradOpDescMakerBase(const OpDescBind& fwd_op) : fwd_op_(fwd_op) {}
virtual ~GradOpDescMakerBase() = default;
virtual std::vector<std::unique_ptr<OpDescBind>> operator()() const = 0;
protected:
static std::vector<std::string> ToGradNames(
const std::vector<std::string>& var_names) {
std::vector<std::string> ret_val;
ret_val.reserve(var_names.size());
std::transform(var_names.begin(), var_names.end(),
std::back_inserter(ret_val), GradVarName);
return ret_val;
}
std::vector<std::string> InputGrad(const std::string& name) const {
return ToGradNames(fwd_op_.Input(name));
}
std::vector<std::string> OutputGrad(const std::string& name) const {
return ToGradNames(fwd_op_.Output(name));
}
std::vector<std::string> InputNames() const {
return this->fwd_op_.InputNames();
}
std::vector<std::string> OutputNames() const {
return this->fwd_op_.OutputNames();
}
std::vector<std::string> Input(const std::string& name) const {
return fwd_op_.Input(name);
}
std::vector<std::string> Output(const std::string& name) const {
return fwd_op_.Output(name);
}
const std::unordered_map<std::string, Attribute>& Attrs() const {
return fwd_op_.GetAttrMap();
}
const Attribute& GetAttr(const std::string& name) const {
auto& map = fwd_op_.GetAttrMap();
auto it = map.find(name);
PADDLE_ENFORCE(it != map.end(), "Cannot find attribute %s", name);
return it->second;
}
std::string ForwardOpType() const { return this->fwd_op_.Type(); }
private:
const OpDescBind& fwd_op_;
};
class SingleGradOpDescMaker : public GradOpDescMakerBase {
public:
using GradOpDescMakerBase::GradOpDescMakerBase;
std::vector<std::unique_ptr<OpDescBind>> operator()() const {
std::vector<std::unique_ptr<OpDescBind>> retv;
retv.emplace_back(this->Apply());
return retv;
}
protected:
virtual std::unique_ptr<OpDescBind> Apply() const = 0;
};
class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
public:
using SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
virtual std::unique_ptr<OpDescBind> Apply() const {
auto* grad = new OpDescBind();
grad->SetType(this->GradOpType());
for (auto& input_param : this->InputNames()) {
grad->SetInput(input_param, this->Input(input_param));
grad->SetOutput(GradVarName(input_param), this->InputGrad(input_param));
}
for (auto& output_param : this->OutputNames()) {
grad->SetInput(output_param, this->Output(output_param));
grad->SetInput(GradVarName(output_param), this->OutputGrad(output_param));
}
grad->SetAttrMap(this->Attrs());
return std::unique_ptr<OpDescBind>(grad);
}
virtual std::string GradOpType() const {
return this->ForwardOpType() + "_grad";
}
};
} // namespace framework
} // namespace paddle
......@@ -15,7 +15,7 @@
#pragma once
#include <memory>
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <thrust/system/cuda/experimental/pinned_allocator.h>
......@@ -29,7 +29,7 @@
namespace paddle {
namespace framework {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
template <typename T>
using Vector = std::vector<T>;
#else
......
# Design Doc: LoD (Level-of-Detail) Tensor
PaddlePaddle's RNN doesn't require that all instances have the same length. To do so, we introduce an extension to Tensor, namely, LoD Tensor.
Like other deep learning systems, PaddlePaddle supports training models from sequence data. Also, like other systems, PaddlePaddle represent a mini-batch of sequences as a Tensor. What is different is that PaddlePaddle doesn't require all sequences in a mini-batch to be of the same length. Thus no need for padding zeros.
## Challenge of Variable-length Inputs
| | TensorFlow | PaddlePaddle |
|-----------------------|------------|--------------|
| RNN | Support | Support |
| recursive RNN | Support | Support |
| padding zeros | Must | No need |
| blob data type | Tensor | LoDTensor |
People usually represent a mini-batch by a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. So a transformation, T, of all images can be a matrix multiplication of the 10xOx32-dimensional tensor T and the 10x32x32 Tensor.
PaddlePaddle achieves this flexibility by passing through a new data type, *LoD Tensor*, which is a Tensor attached with segmentation index known as *LoD*, between operators. The LoD index doesn't only segment a tensor, but also recursively segments sub-sequences. This document presents the design of LoD and LoDTensor.
Another example is that each mini-batch contains 32 sentences, where each word is a D-dimensional one-hot vector. If all sentences have the same length L, we can represent this mini-batch by a 32xLxD tensor. However, in most cases, sentences have variable lengths, and we will need an index data structure to record these variable lengths.
## LoD as a Solution
## The Challenge: Variable-length Sequences
### Mini-Batch of variable-length sentences
Most deep learning systems represent a mini-batch as a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. Another example is that each mini-batch contains N sentences, where each word is a D-dimensional one-hot vector. Suppose that all sentences have the same length L, we can represent this mini-batch by a NxLxD tensor.
Let's imagine a mini-batch of 3 variable lengths sentences, containing 3, 1, and 2 words respectively. We can represent it by a (3+1+2)xD tensor plus some index information:
Both examples show that the elements of sequences are usually of the same size. In the first example, all images are 32x32, and in the second one, all words are D-dimensional vectors. It doesn't make sense to allow variable-sized images, as that would require transformations like convolution to handle variable-sized Tensors.
The real challenge is that in most cases, sentences have variable lengths, and we will need an index data structure to segment the tensor into sequences. Also, sequences might consist of sub-sequences.
## A Solution: The LoD Index
To understand our solution, it is best to look at some examples.
### A Mini-Batch of Sentences
Let's imagine a mini-batch of 3 variable lengths sentences composed of 3, 1, and 2 words, respectively. We can represent the mini-batch by a (3+1+2)xD tensor plus some index information:
```
3
3 1 2
||| | ||
```
Each `|` represents a D-dimensional word vectors. The number 3 on top indicate 3 sentences, and numbers 3, 1, and 2 on the second level represent the number of words in each sentence.
where each `|` represents a D-dimensional word vector. The numbers, 3, 1, and 2, form a 1-level LoD.
### Recursive Sequences
Let check another example of a 2-level LoD Tensor. Consider a mini-batch of three articles with 3, 1, and 2 sentences, and each sentence consists of a variable number of words:
```
3 1 2
3 2 4 1 2 3
||| || |||| | || |||
```
### Mini-Batch of variable-length videos
### A Mini-Batch of Videos
This approach generalizes to the case where elements are not words, but higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. If a mini-batch contains 3 videos of 3, 1, and 2 frames respectively. The underlying tensor is of size (3+1+2)x640x480. The index information illustrates as:
LoD tensors generalize to the case where elements are higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. Here is a mini-batch of 3 videos with 3, 1, and 2 frames, respectively.
```
3
3 1 2
口口口 口 口口
```
where each `口` represents an image.
The underlying tensor is of size (3+1+2)x640x480, and each `口` represents a 640x480 image.
### Mini-Batch of fixed-size images
### A Mini-Batch of Images
Let's get back to a typical example, image classification, where each mini-batch has M fixed-sized images. The LoD Tensor representation is
In traditional cases like a mini-batch with N fixed-sized images, the LoD Tensor representation is as
```
M
1 1 1 1 1
口口口口 ... 口
```
The many 1's on the second level seem duplicated. For this particular case of 2 levels and the second level always have length 1, we can ignore the LoD index.
### Design and summarization
In this case, we don't lose any information by ignoring the many 1's in the index and simply considering this LoD Tensor as a usual Tensor:
In summary, as long as that the essential elements (words or images) have the same size, we can represent mini-batches by a LoD Tensor:
```
口口口口 ... 口
```
- The underlying tensor has size LxD1xD2x..., where D1xD2... is the size of the essential elements, and
- The first dimension size L has an additonal property -- a LoD index as a nested vector:
### Model Parameters
```c++
typedef std::vector<std::<vector>> LoD;
```
A model parameter is just a usual Tensor, which, just like the above example, is a **0-level LoD Tensor**.
- The LoD index is not necessary when there are only two levels and all elements of the second level have length 1.
## Slicing of LoD Tensor
## The LoD Tensor
Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words. This network requires that mini-batches represented by 3 level LoD Tensor, for example,
Let us revisit above example of the 2-level LoD Tensor
```
3
3 1 2
3 2 4 1 2 3
||| || |||| | || |||
```
To allow each level of RNN to handle its input, we define **the slicing of a LoD Tensor is defined as getting the j-th sequence on level i, or the <i,j>-slice**
It is indeed a tree, where leaves are elementary sequences identified by **branches**.
For example, the third sentence in above example is identified by branch <0,2>, where 0 indicates the first article with length 3, and 2 indicates the third sentence in this article with length 4.
### The LoD Index
For example, the <2,1>-slice of above slice is
We can save the LoD index in the above example
```
2
||
3 1 2
3 2 4 1 2 3
```
and the <1,2>-slice of above example is
in a not-full 2D matrix:
```c++
typedef std::vector<std::vector<int> > LoD;
```
2
2 3
|| |||
```
Let's go on slicing this slice. Its <1,1>-slice is
where
- `LoD.size()` is the number of levels, or the maximum length of branches,
- `LoD[i][j]` is the length of the j-th segment at the i-th level.
## The Offset Representation
To quickly access elementary sequences, we adopt an offset representation -- instead of saving the lengths, we save the beginning and ending elements of sequences.
In the above example, we accumulate the length of elementary sequences:
```
1
1
|
3 2 4 1 2 3
```
### The Slicing Algorithm
into offsets
The algorithm, with over-simplified data structure, is defined as
```
0 3 5 9 10 12 15
= = = = = =
3 2+3 4+5 1+9 2+10 3+12
```
```c++
typedef std::vector<std::vector<int>> LoD;
so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5.
struct LoDTensor {
LoD lod_;
float* tensor_;
};
Similarly, the lengths in the top level LoD
LoDTensor Slice(const LoDTensor& lodt, int level, int sequence);
```
3 1 2
```
Let us revisit the example above
are transformed into offsets of elements/words as follows:
```
3
3 1 2
3 2 4 1 2 3
||| || |||| | || |||
0 9 10 15
= = =
3+2+4 1+9 2+3+10
```
Suppose that we want to retrieve the <1,2>-slice
so we can tell that the first article is from word 0 to word 9, and the second article is from word 9 to word 10.
The complete offset representation is as follows:
```
2
2 3
|| |||
0 9 10 15
0 3 5 9 10 12 15
||| || |||| | || |||
```
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.
## Slicing of LoD Tensors
When we use the above 2-level LoD Tensor as the input to a nested-RNN, we need to retrieve certain sequences. Here we define the sequence identified by branch <i,j,...> as the **<i,j,...>-slice**.
To avoid the traversal of the LoD tree at slicing 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
For example, the <2>-slice of above example is
```
0
0 9 10
0 3 5 9 10 12
||| || |||| | || |||
10 15
10 12 15
|| |||
```
We don't really need the 0 on top, so the LoD Tensor could be
and the <2,0>-slice of above slice is
```
0 9 10
0 3 5 9 10 12
||| || |||| | || |||
10 12
||
```
......@@ -18,6 +18,15 @@ limitations under the License. */
namespace paddle {
namespace framework {
OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs,
const AttributeMap &attrs) {
op_desc_.set_type(type);
inputs_ = inputs;
outputs_ = outputs;
attrs_ = attrs;
}
OpDesc *OpDescBind::Proto() {
Sync();
return &op_desc_;
......@@ -31,11 +40,10 @@ const std::vector<std::string> &OpDescBind::Input(
return it->second;
}
std::vector<std::string> OpDescBind::InputNames() const {
std::vector<std::string> OpDescBind::InputArgumentNames() const {
std::vector<std::string> retv;
retv.reserve(this->inputs_.size());
for (auto &ipt : this->inputs_) {
retv.push_back(ipt.first);
retv.insert(retv.end(), ipt.second.begin(), ipt.second.end());
}
return retv;
}
......@@ -54,11 +62,10 @@ const std::vector<std::string> &OpDescBind::Output(
return it->second;
}
std::vector<std::string> OpDescBind::OutputNames() const {
std::vector<std::string> OpDescBind::OutputArgumentNames() const {
std::vector<std::string> retv;
retv.reserve(this->outputs_.size());
for (auto &ipt : this->outputs_) {
retv.push_back(ipt.first);
retv.insert(retv.end(), ipt.second.begin(), ipt.second.end());
}
return retv;
}
......@@ -112,6 +119,42 @@ const std::unordered_map<std::string, Attribute> &OpDescBind::GetAttrMap()
return attrs_;
}
void OpDescBind::Rename(const std::string &old_name,
const std::string &new_name) {
for (auto &input : inputs_) {
std::replace(input.second.begin(), input.second.end(), old_name, new_name);
}
for (auto &output : outputs_) {
std::replace(output.second.begin(), output.second.end(), old_name,
new_name);
}
need_update_ = true;
}
struct SetAttrDescVisitor : public boost::static_visitor<void> {
explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
mutable OpDesc::Attr *attr_;
void operator()(int v) const { attr_->set_i(v); }
void operator()(float v) const { attr_->set_f(v); }
void operator()(const std::string &v) const { attr_->set_s(v); }
void operator()(bool b) const { attr_->set_b(b); }
void operator()(const std::vector<int> &v) const {
VectorToRepeated(v, attr_->mutable_ints());
}
void operator()(const std::vector<float> &v) const {
VectorToRepeated(v, attr_->mutable_floats());
}
void operator()(const std::vector<std::string> &v) const {
VectorToRepeated(v, attr_->mutable_strings());
}
void operator()(const std::vector<bool> &v) const {
VectorToRepeated(v, attr_->mutable_bools());
}
void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->idx()); }
void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
};
void OpDescBind::Sync() {
if (need_update_) {
this->op_desc_.mutable_inputs()->Clear();
......@@ -134,7 +177,8 @@ void OpDescBind::Sync() {
attr_desc->set_name(attr.first);
attr_desc->set_type(
static_cast<framework::AttrType>(attr.second.which() - 1));
boost::apply_visitor(SetAttrDescVisitor(attr_desc), attr.second);
SetAttrDescVisitor visitor(attr_desc);
boost::apply_visitor(visitor, attr.second);
}
need_update_ = false;
......
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <unordered_map>
#include <vector>
#include "paddle/framework/attribute.h"
#include "paddle/framework/type_defs.h"
#include "paddle/framework/var_desc.h"
namespace paddle {
......@@ -26,6 +27,11 @@ class BlockDescBind;
class OpDescBind {
public:
OpDescBind() {}
OpDescBind(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs);
OpDesc *Proto();
std::string Type() const { return op_desc_.type(); }
......@@ -34,14 +40,14 @@ class OpDescBind {
const std::vector<std::string> &Input(const std::string &name) const;
std::vector<std::string> InputNames() const;
std::vector<std::string> InputArgumentNames() const;
void SetInput(const std::string &param_name,
const std::vector<std::string> &args);
const std::vector<std::string> &Output(const std::string &name) const;
std::vector<std::string> OutputNames() const;
std::vector<std::string> OutputArgumentNames() const;
void SetOutput(const std::string &param_name,
const std::vector<std::string> &args);
......@@ -60,49 +66,52 @@ class OpDescBind {
void SetBlockAttr(const std::string &name, BlockDescBind &block);
// Only be used in C++
void SetAttrMap(const std::unordered_map<std::string, Attribute> &attr_map);
Attribute GetAttr(const std::string &name) const;
int GetBlockAttr(const std::string &name) const;
void Rename(const std::string &old_name, const std::string &new_name);
// Only be used in C++
const std::unordered_map<std::string, Attribute> &GetAttrMap() const;
const AttributeMap &GetAttrMap() const;
private:
struct SetAttrDescVisitor : public boost::static_visitor<void> {
explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
mutable OpDesc::Attr *attr_;
void operator()(int v) const { attr_->set_i(v); }
void operator()(float v) const { attr_->set_f(v); }
void operator()(const std::string &v) const { attr_->set_s(v); }
void operator()(bool b) const { attr_->set_b(b); }
void operator()(const std::vector<int> &v) const {
VectorToRepeated(v, attr_->mutable_ints());
}
void operator()(const std::vector<float> &v) const {
VectorToRepeated(v, attr_->mutable_floats());
}
void operator()(const std::vector<std::string> &v) const {
VectorToRepeated(v, attr_->mutable_strings());
}
void operator()(const std::vector<bool> &v) const {
VectorToRepeated(v, attr_->mutable_bools());
}
void operator()(BlockDesc *desc) const {
attr_->set_block_idx(desc->idx());
}
void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
};
// Only be used in C++
void SetAttrMap(const AttributeMap &attr_map);
std::vector<std::string> InputNames() const { return MapKeys(inputs_); }
std::vector<std::string> OutputNames() const { return MapKeys(outputs_); }
void SetInputMap(const VariableNameMap &input) {
this->inputs_ = input;
this->need_update_ = true;
}
void SetOutputMap(const VariableNameMap &output) {
this->outputs_ = output;
this->need_update_ = true;
}
void Sync();
const VariableNameMap &Inputs() const { return inputs_; }
const VariableNameMap &Outputs() const { return outputs_; }
private:
template <typename MapType>
static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {
std::vector<typename MapType::key_type> ret_val;
ret_val.reserve(map.size());
std::transform(
map.begin(), map.end(), std::back_inserter(ret_val),
[](const typename MapType::value_type &pair) { return pair.first; });
return ret_val;
}
OpDesc op_desc_;
std::unordered_map<std::string, std::vector<std::string>> inputs_;
std::unordered_map<std::string, std::vector<std::string>> outputs_;
std::unordered_map<std::string, Attribute> attrs_;
VariableNameMap inputs_;
VariableNameMap outputs_;
AttributeMap attrs_;
// need_update_ indicate there some local changes not be synchronized. If
// local changes should be synchronized, need_update_ should be set to true.
......
......@@ -19,21 +19,18 @@
#include <unordered_map>
#include "paddle/framework/attribute.h"
#include "paddle/framework/op_desc.h"
#include "paddle/framework/type_defs.h"
#include "paddle/platform/macros.h"
namespace paddle {
namespace framework {
class OperatorBase;
using VariableNameMap = std::map<std::string, std::vector<std::string>>;
using OpCreator = std::function<OperatorBase*(
const std::string& /*type*/, const VariableNameMap& /*inputs*/,
const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
struct OpInfo {
OpCreator creator_;
std::string grad_op_type_;
OpProto* proto_;
OpAttrChecker* checker_;
GradOpMakerFN grad_op_maker_;
OpProto* proto_{nullptr};
OpAttrChecker* checker_{nullptr};
bool HasOpProtoAndChecker() const {
return proto_ != nullptr && checker_ != nullptr;
......@@ -46,30 +43,25 @@ struct OpInfo {
return *proto_;
}
const OpAttrChecker& Checker() const {
PADDLE_ENFORCE_NOT_NULL(checker_,
"Operator Checker has not been registered");
return *checker_;
}
const OpCreator& Creator() const {
PADDLE_ENFORCE_NOT_NULL(creator_,
"Operator Creator has not been registered");
return creator_;
}
bool HasGradientOp() const { return !grad_op_type_.empty(); }
const GradOpMakerFN& GradOpMaker() const {
PADDLE_ENFORCE_NOT_NULL(grad_op_maker_,
"Operator GradOpMaker has not been registered.");
return grad_op_maker_;
}
const OpAttrChecker* Checker() const { return checker_; }
};
class OpInfoMap {
public:
static OpInfoMap& Instance();
OpInfoMap(const OpInfoMap& o) = delete;
OpInfoMap(OpInfoMap&& o) = delete;
OpInfoMap& operator=(const OpInfoMap& o) = delete;
OpInfoMap& operator=(OpInfoMap&& o) = delete;
bool Has(const std::string& op_type) const {
return map_.find(op_type) != map_.end();
}
......@@ -105,6 +97,8 @@ class OpInfoMap {
private:
OpInfoMap() = default;
std::unordered_map<std::string, const OpInfo> map_;
DISABLE_COPY_AND_ASSIGN(OpInfoMap);
};
} // namespace framework
......
......@@ -44,11 +44,6 @@ class OpProtoAndCheckerMaker {
var_->set_intermediate(true);
return *this;
}
VariableBuilder& NotInGradient() {
var_->set_not_in_gradient(true);
return *this;
}
};
VariableBuilder AddInput(const std::string& name, const std::string& comment);
......
......@@ -48,4 +48,4 @@ TEST(ProtoMaker, DuplicatedInOut) {
paddle::framework::OpAttrChecker op_checker;
auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
}
\ No newline at end of file
}
......@@ -23,7 +23,9 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, AttributeMap attrs) {
auto& info = OpInfoMap::Instance().Get(type);
info.Checker().Check(attrs);
if (info.Checker() != nullptr) {
info.Checker()->Check(attrs);
}
auto op = info.Creator()(type, inputs, outputs, attrs);
return std::unique_ptr<OperatorBase>(op);
}
......@@ -52,9 +54,15 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
return CreateOp(op_desc.type(), inputs, outputs, attrs);
}
std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
return CreateOp(op_desc.Type(), op_desc.Inputs(), op_desc.Outputs(),
op_desc.GetAttrMap());
}
std::vector<std::unique_ptr<OpDescBind>> OpRegistry::CreateGradOpDescs(
const OpDescBind& op_desc) {
auto& info = OpInfoMap::Instance().Get(op_desc.Type());
return info.grad_op_maker_(op_desc);
}
} // namespace framework
......
......@@ -21,49 +21,54 @@ limitations under the License. */
#include <unordered_map>
#include <unordered_set>
#include "paddle/framework/attribute.h"
#include "paddle/framework/details/op_registry.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/grad_op_builder.h"
#include "paddle/framework/op_info.h"
#include "paddle/framework/op_proto_maker.h"
#include "paddle/framework/grad_op_desc_maker.h"
#include "paddle/framework/op_desc.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/scope.h"
namespace paddle {
namespace framework {
class Registrar {
public:
// In our design, various kinds of classes, e.g., operators and kernels,
// have their corresponding registry and registrar. The action of
// registration is in the constructor of a global registrar variable, which,
// however, are not used in the code that calls package framework, and would
// be removed from the generated binary file by the linker. To avoid such
// removal, we add Touch to all registrar classes and make USE_OP macros to
// call this method. So, as long as the callee code calls USE_OP, the global
// registrar variable won't be removed by the linker.
void Touch() {}
};
template <typename... ARGS>
struct OperatorRegistrar : public Registrar {
explicit OperatorRegistrar(const char* op_type) : op_type(op_type) {
PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
"'%s' is registered more than once.", op_type);
static_assert(sizeof...(ARGS) != 0,
"OperatorRegistrar should be invoked at least by OpClass");
details::OperatorRegistrarRecursive<0, false, ARGS...>(op_type, &info);
OpInfoMap::Instance().Insert(op_type, info);
}
const char* op_type;
OpInfo info;
};
class OpRegistry {
public:
template <typename OpType, typename ProtoMakerType, typename GradOpType>
static void RegisterOp(const std::string& op_type,
const std::string& grad_op_type) {
PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
"'%s' is registered more than once.", op_type);
OpInfo op_info;
op_info.creator_ = [](
const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs) {
return new OpType(type, inputs, outputs, attrs);
};
op_info.grad_op_type_ = grad_op_type;
if (std::type_index(typeid(ProtoMakerType)) !=
std::type_index(typeid(NOPMaker))) {
op_info.proto_ = new OpProto;
op_info.checker_ = new OpAttrChecker;
auto maker = ProtoMakerType(op_info.proto_, op_info.checker_);
maker.Validate();
op_info.proto_->set_type(op_type);
PADDLE_ENFORCE(
op_info.proto_->IsInitialized(),
"Fail to initialize %s's OpProto, because %s is not initialized",
op_type, op_info.proto_->InitializationErrorString());
} else {
op_info.proto_ = nullptr;
op_info.checker_ = nullptr;
}
OpInfoMap::Instance().Insert(op_type, op_info);
OperatorRegistrar<OpType, ProtoMakerType> reg(op_type.c_str());
reg.info.grad_op_type_ = grad_op_type;
// register gradient op
if (!grad_op_type.empty()) {
RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
OperatorRegistrar<GradOpType> grad_reg(grad_op_type.c_str());
}
}
......@@ -74,20 +79,10 @@ class OpRegistry {
static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
};
static std::vector<std::unique_ptr<OpDescBind>> CreateGradOpDescs(
const OpDescBind& op_desc);
class Registrar {
public:
// In our design, various kinds of classes, e.g., operators and kernels,
// have their corresponding registry and registrar. The action of
// registration is in the constructor of a global registrar variable, which,
// however, are not used in the code that calls package framework, and would
// be removed from the generated binary file by the linker. To avoid such
// removal, we add Touch to all registrar classes and make USE_OP macros to
// call this method. So, as long as the callee code calls USE_OP, the global
// registrar variable won't be removed by the linker.
void Touch() {}
static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
};
template <typename OpType, typename ProtoMakerType, typename GradOpType>
......@@ -145,33 +140,41 @@ class OpKernelRegistrar : public Registrar {
__test_global_namespace_##uniq_name##__>::value, \
msg)
#define REGISTER_OPERATOR(op_type, op_class, ...) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op__##op_type, \
"REGISTER_OPERATOR must be called in global namespace"); \
class _OpClass_##op_type##_ : public op_class { \
public: \
DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \
DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \
}; \
static ::paddle::framework::OperatorRegistrar<_OpClass_##op_type##_, \
##__VA_ARGS__> \
__op_registrar_##op_type##__(#op_type); \
int TouchOpRegistrar_##op_type() { \
__op_registrar_##op_type##__.Touch(); \
return 0; \
}
/**
* Macro to register Operator.
*/
#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \
grad_op_class) \
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
class _OpClass_##op_type##_ : public op_class { \
public: \
DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_); \
DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class); \
}; \
class _OpGradClass_##op_type##_ : public grad_op_class { \
public: \
DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_); \
DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class); \
}; \
static ::paddle::framework::OpRegistrar< \
_OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_> \
__op_registrar_##op_type##__(#op_type, #grad_op_type); \
int TouchOpRegistrar_##op_type() { \
__op_registrar_##op_type##__.Touch(); \
return 0; \
}
#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \
grad_op_class) \
REGISTER_OPERATOR(grad_op_type, grad_op_class); \
class _GradOpDescMaker_##grad_op_type##_ \
: public ::paddle::framework::DefaultGradOpDescMaker { \
using ::paddle::framework::DefaultGradOpDescMaker::DefaultGradOpDescMaker; \
\
protected: \
virtual std::string GradOpType() const { return #grad_op_type; } \
}; \
REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \
op_maker_class);
#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP)
REGISTER_OPERATOR(op_type, op_class, op_maker_class)
/**
* Macro to register OperatorKernel.
......@@ -218,7 +221,7 @@ class OpKernelRegistrar : public Registrar {
// TODO(fengjiayi): The following macros
// seems ugly, do we have better method?
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
#define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)
#else
#define USE_OP_KERNEL(op_type) \
......
......@@ -173,3 +173,14 @@ TEST(OpRegistry, CustomChecker) {
int test_attr = op->Attr<int>("test_attr");
ASSERT_EQ(test_attr, 4);
}
class CosineOpComplete : public paddle::framework::CosineOp {
public:
DEFINE_OP_CONSTRUCTOR(CosineOpComplete, paddle::framework::CosineOp);
DEFINE_OP_CLONE_METHOD(CosineOpComplete);
};
TEST(OperatorRegistrar, Test) {
using namespace paddle::framework;
OperatorRegistrar<CosineOpComplete, CosineOpProtoAndCheckerMaker> reg("cos");
}
......@@ -25,7 +25,7 @@ Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
return *device_context_.GetEigenDevice<platform::CPUPlace>();
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
template <>
Eigen::GpuDevice&
ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
......@@ -245,5 +245,12 @@ std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
return res;
}
std::ostream& operator<<(std::ostream& os,
const OperatorWithKernel::OpKernelKey& kernel_key) {
os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_
<< "]";
return os;
}
} // namespace framework
} // namespace paddle
......@@ -22,6 +22,7 @@ limitations under the License. */
#include "op_info.h"
#include "paddle/framework/attribute.h"
#include "paddle/framework/block_desc.h"
#include "paddle/framework/data_type.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/lod_tensor.h"
......@@ -317,46 +318,170 @@ class ExecutionContext : public InferShapeContext {
const platform::DeviceContext& device_context_;
};
class CompileTimeInferShapeContext : public InferShapeContextBase {
public:
CompileTimeInferShapeContext(const OpDescBind& op, const BlockDescBind& block)
: op_(op), block_(block) {}
bool HasInput(const std::string& name) const override {
const std::vector<std::string>& input_names = op_.Input(name);
auto length = input_names.size();
PADDLE_ENFORCE_EQ(length, 1UL,
"Input(%s) should have only one value, "
"but it have %d now",
name, length);
return block_.HasVar(input_names[0]);
}
bool HasOutput(const std::string& name) const override {
const std::vector<std::string>& output_names = op_.Output(name);
auto length = output_names.size();
PADDLE_ENFORCE_EQ(length, 1UL,
"Output(%s) should have only one value, "
"but it have %d now",
name, length);
return block_.HasVar(output_names[0]);
}
bool HasInputs(const std::string& name) const override {
const std::vector<std::string>& input_names = op_.Input(name);
PADDLE_ENFORCE(!input_names.empty(), "Inputs(%s) length is 0", name);
for (auto& input : input_names) {
if (!block_.HasVar(input)) return false;
}
return true;
}
bool HasOutputs(const std::string& name) const override {
const std::vector<std::string>& output_names = op_.Output(name);
PADDLE_ENFORCE(!output_names.empty(), "Inputs(%s) length is 0", name);
for (auto& output : output_names) {
if (!block_.HasVar(output)) return false;
}
return true;
}
DDim GetInputDim(const std::string& name) const override {
std::vector<DDim> ddims = GetInputsDim(name);
auto length = ddims.size();
PADDLE_ENFORCE_EQ(length, 1UL,
"Input(%s) should have 1 value, "
"but it has %d now",
name, length);
return ddims[0];
}
void SetInputDim(const std::string& name, const DDim& dim) override {
SetInputsDim(name, {dim});
}
DDim GetOutputDim(const std::string& name) const override {
std::vector<DDim> ddims = GetOutputsDim(name);
auto length = ddims.size();
PADDLE_ENFORCE_EQ(length, 1UL,
"Output(%s) should have 1 value, "
"but it has %d now",
name, length);
return ddims[0];
}
void SetOutputDim(const std::string& name, const DDim& dim) override {
SetOutputsDim(name, {dim});
}
AttrReader Attrs() const override { return AttrReader(op_.GetAttrMap()); }
const std::vector<std::string>& Inputs(
const std::string& name) const override {
return op_.Input(name);
}
const std::vector<std::string>& Outputs(
const std::string& name) const override {
return op_.Output(name);
}
private:
DDim GetDim(const std::string& name) const override {
return framework::make_ddim(block_.Var(name)->Shape());
}
void SetDim(const std::string& name, const DDim& dim) override {
block_.Var(name)->SetShape(framework::vectorize(dim));
}
const OpDescBind& op_;
const BlockDescBind& block_;
};
class RuntimeInferShapeContext : public InferShapeContextBase {
public:
RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
: op_(op), scope_(scope) {}
bool HasInput(const std::string& name) const {
bool HasInput(const std::string& name) const override {
auto ipt = op_.Input(name);
auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
return var != nullptr;
}
bool HasOutput(const std::string& name) const {
bool HasOutput(const std::string& name) const override {
auto ipt = op_.Output(name);
auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
return var != nullptr;
}
DDim GetInputDim(const std::string& name) const {
bool HasInputs(const std::string& name) const override {
auto inputs = op_.Inputs(name);
if (inputs.empty()) {
return false;
}
for (auto& input : inputs) {
if (scope_.FindVar(input) == nullptr) {
return false;
}
}
return true;
}
bool HasOutputs(const std::string& name) const override {
auto outputs = op_.Outputs(name);
if (outputs.empty()) {
return false;
}
for (auto& output : outputs) {
if (scope_.FindVar(output) == nullptr) {
return false;
}
}
return true;
}
DDim GetInputDim(const std::string& name) const override {
return GetDim(op_.Input(name));
}
void SetInputDim(const std::string& name, const DDim& dim) {
void SetInputDim(const std::string& name, const DDim& dim) override {
SetDim(op_.Input(name), dim);
}
DDim GetOutputDim(const std::string& name) const {
DDim GetOutputDim(const std::string& name) const override {
return GetDim(op_.Output(name));
}
void SetOutputDim(const std::string& name, const DDim& dim) {
void SetOutputDim(const std::string& name, const DDim& dim) override {
SetDim(op_.Output(name), dim);
}
AttrReader Attrs() const { return AttrReader(op_.Attrs()); }
AttrReader Attrs() const override { return AttrReader(op_.Attrs()); }
const std::vector<std::string>& Inputs(const std::string& name) const {
const std::vector<std::string>& Inputs(
const std::string& name) const override {
return op_.Inputs(name);
}
const std::vector<std::string>& Outputs(const std::string& name) const {
const std::vector<std::string>& Outputs(
const std::string& name) const override {
return op_.Outputs(name);
}
......@@ -377,11 +502,11 @@ class RuntimeInferShapeContext : public InferShapeContextBase {
return t;
}
DDim GetDim(const std::string& name) const {
DDim GetDim(const std::string& name) const override {
return GetTensor<false>(name)->dims();
}
void SetDim(const std::string& name, const DDim& dim) {
void SetDim(const std::string& name, const DDim& dim) override {
GetTensor<true>(name)->Resize(dim);
}
......@@ -452,9 +577,25 @@ class OperatorWithKernel : public OperatorBase {
this->InferShape(&infer_shape_ctx);
ExecutionContext ctx(*this, scope, dev_ctx);
auto& opKernel = AllOpKernels().at(type_).at(
OpKernelKey(IndicateDataType(ctx), dev_ctx));
opKernel->Compute(ctx);
// check if op[type] has kernel registered.
auto& all_op_kernels = AllOpKernels();
auto kernels_iter = all_op_kernels.find(type_);
if (kernels_iter == all_op_kernels.end()) {
PADDLE_THROW("op[%s] has no kernel", type_);
}
// check if op[type] have kernel for kernel_key
OpKernelMap& kernels = kernels_iter->second;
auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx);
auto kernel_iter = kernels.find(kernel_key);
if (kernel_iter == kernels.end()) {
PADDLE_THROW("op[%s] has no kernel with kernel_key[%s]", type_,
kernel_key);
}
kernel_iter->second->Compute(ctx);
}
static std::unordered_map<std::string /* op_type */, OpKernelMap>&
......@@ -471,9 +612,9 @@ class OperatorWithKernel : public OperatorBase {
});
}
protected:
virtual void InferShape(InferShapeContextBase* ctx) const = 0;
protected:
// indicate kernel DataType by input data. Defaultly all input data must be
// same.
virtual DataType IndicateDataType(const ExecutionContext& ctx) const {
......@@ -503,5 +644,8 @@ class OperatorWithKernel : public OperatorBase {
}
};
std::ostream& operator<<(std::ostream& os,
const OperatorWithKernel::OpKernelKey& kernel_key);
} // namespace framework
} // namespace paddle
......@@ -16,6 +16,7 @@ limitations under the License. */
#include <vector>
#include "paddle/framework/framework.pb.h"
#include "paddle/platform/macros.h"
namespace paddle {
namespace framework {
......@@ -26,9 +27,6 @@ class ProgramDescBind {
public:
static ProgramDescBind &Instance(ProgramDesc *prog);
ProgramDescBind(const ProgramDescBind &o) = delete;
ProgramDescBind &operator=(const ProgramDescBind &o) = delete;
BlockDescBind *AppendBlock(const BlockDescBind &parent);
BlockDescBind *Block(size_t idx) { return blocks_[idx].get(); }
......@@ -46,6 +44,8 @@ class ProgramDescBind {
ProgramDesc *prog_;
std::vector<std::unique_ptr<BlockDescBind>> blocks_;
DISABLE_COPY_AND_ASSIGN(ProgramDescBind);
};
} // namespace framework
} // namespace paddle
......@@ -19,6 +19,7 @@ limitations under the License. */
#include <unordered_map>
#include "paddle/framework/variable.h"
#include "paddle/platform/macros.h"
namespace paddle {
namespace framework {
......@@ -38,11 +39,6 @@ class Scope {
Scope() {}
~Scope();
// Disable Copy, Assign, Move.
Scope(const Scope& other) = delete;
Scope& operator=(const Scope& other) = delete;
Scope(Scope&& other) = delete;
/// Create a sub-scope. Returns a reference other than a pointer so
/// to prevent from manual deletion.
/// Mark it to const because that new kid scope cannot change parent scope.
......@@ -73,6 +69,8 @@ class Scope {
std::unordered_map<std::string, Variable*> vars_;
mutable std::list<Scope*> kids_;
Scope const* parent_{nullptr};
DISABLE_COPY_AND_ASSIGN(Scope);
};
} // namespace framework
......
......@@ -19,11 +19,18 @@ limitations under the License. */
namespace paddle {
namespace framework {
// TODO(longfei): Once after both CompileTimeInferShapeContext and
// RuntimeInferShapeContext get merged, we can rename InferShapeContextBase into
// InferShapeContext so to replace the current InferShapeContext.
class InferShapeContextBase {
public:
virtual ~InferShapeContextBase() {}
virtual bool HasInput(const std::string &name) const = 0;
virtual bool HasOutput(const std::string &name) const = 0;
virtual bool HasInputs(const std::string &name) const = 0;
virtual bool HasOutputs(const std::string &name) const = 0;
virtual framework::DDim GetInputDim(const std::string &name) const = 0;
std::vector<framework::DDim> GetInputsDim(const std::string &name) const {
const std::vector<std::string> &names = Inputs(name);
......
/* 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/framework/tensor_array.h"
#include <glog/logging.h>
#include <algorithm>
#include <limits>
namespace paddle {
namespace framework {
namespace detail {
/*
* Offer an iterator over the length-sorted lod-tensor's top level. The top
* level of a lod-tensor stores batch-size of sequences, each top-level sequence
* may contains several lower-level sequences, sort top-level lod by the numbers
* of lower-level sequences in descending order, so that during RNN's running,
* the batch-size will keep decreasing, the short sentences will end at the tail
* of each batch.
*
* Let's take a simple lod-tensor for example
*
* |(0) |(1) top-level has two instances
* ||| ||||| lower-level
*
* sort by lower-level's length
*
* |(1) |(0)
* ||||| |||
*
* when RNN runs, it get 5 batches (equals the number of elements the longest
* sequence has)
*
* |||||
* |||
*
* the first three batches has two elements, the last two elements just has 1
* element each.
*/
struct DynamicBatchUnpacker {
using value_type = float;
DynamicBatchUnpacker(const LoDTensor& source, size_t level,
bool descend = true)
: source(&source), level(level) {
BuildLengthSortedMeta(descend);
}
LoDTensor GetBatch(size_t index);
std::vector<DySeqMeta> meta;
LoDTensor const* source;
size_t level;
protected:
void BuildLengthSortedMeta(bool descend);
};
LoDTensor PackDynamicBatch(const std::vector<LoDTensor>& source,
const std::vector<DySeqMeta>& meta, const LoD& lod,
size_t level);
} // namespace detail
const LoDTensor& TensorArray::Read(size_t index) const {
PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index);
if (index >= size()) {
values_.resize(index + 1);
}
return values_[index];
}
void TensorArray::Write(size_t index, const LoDTensor& value) {
PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index);
if (index >= size()) {
values_.resize(index + 1);
}
values_[index].Resize(value.dims());
values_[index].mutable_data<value_type>(platform::CPUPlace());
values_[index].CopyFrom<value_type>(value, platform::CPUPlace());
}
void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
PADDLE_ENFORCE_LE(index, MAX_SIZE, "index[%d] too large", index);
if (index >= size()) {
values_.resize(index + 1);
}
values_[index].ShareDataWith<value_type>(value);
}
LoDTensor TensorArray::Pack(size_t level, const std::vector<DySeqMeta>& meta,
const LoD& lod) const {
return detail::PackDynamicBatch(values_, meta, lod, level);
}
std::vector<DySeqMeta> TensorArray::Unpack(const LoDTensor& source, int level,
bool length_desend) {
detail::DynamicBatchUnpacker unpacker(source, level,
length_desend /*descend*/);
// find max length of all the sequences
size_t max_length = 0;
for (const auto& seq : unpacker.meta) {
max_length = std::max(max_length, seq.end - seq.begin);
}
// write batches to values
for (size_t batch_id = 0; batch_id < max_length; batch_id++) {
Write(batch_id, unpacker.GetBatch(batch_id));
}
return unpacker.meta;
}
LoDTensor TensorArray::Stack() const {
LoDTensor result;
if (size() == 0) return result;
const auto& first_dims = values_.front().dims();
// check all the values have the same shape
// TODO(superjom) check the same dtypes
for (size_t idx = 1; idx < size(); idx++) {
const auto& value_dims = values_[idx].dims();
PADDLE_ENFORCE_EQ(first_dims, value_dims);
}
// copy
auto result_dims = vectorize(first_dims);
result_dims.insert(result_dims.begin(), size());
result.Resize(make_ddim(result_dims));
result.mutable_data<value_type>(platform::CPUPlace());
for (size_t idx = 0; idx < size(); idx++) {
result.Slice<value_type>(idx, idx + 1)
.CopyFrom<value_type>(Read(idx), platform::CPUPlace());
}
return result;
}
void TensorArray::Unstack(const LoDTensor& source) const {
Unstack(source, false /*data_shared*/);
}
void TensorArray::UnstackShared(const LoDTensor& source) const {
Unstack(source, true /*data_shared*/);
}
void TensorArray::Unstack(const LoDTensor& source, bool data_shared) const {
size_t first_dim = source.dims()[0];
DDim value_dims = slice_ddim(source.dims(), 1, source.dims().size());
PADDLE_ENFORCE_GT(first_dim, 0,
"source should have some data to be unstacked");
values_.resize(first_dim);
for (size_t elem = 0; elem < first_dim; elem++) {
// create a new value
auto& value = values_[elem];
if (data_shared) {
// share memory
value.ShareDataWith<value_type>(source.Slice<value_type>(elem, elem + 1));
} else {
// copy
value.Resize(value_dims);
value.CopyFrom<value_type>(source.Slice<value_type>(elem, elem + 1),
platform::CPUPlace());
}
}
}
size_t TensorArray::size() const { return values_.size(); }
namespace detail {
void DynamicBatchUnpacker::BuildLengthSortedMeta(bool descend) {
PADDLE_ENFORCE(meta.empty(), "duplicate build meta");
// collect meta for each sequence in some level
auto lod = SliceLevels(source->lod(), level, level + 1)[0];
for (size_t seq_id = 0; seq_id < lod.size() - 1; seq_id++) {
DySeqMeta seq_meta({lod[seq_id], lod[seq_id + 1], seq_id});
meta.push_back(seq_meta);
}
PADDLE_ENFORCE_GT(meta.size(), 0, "meta is empty");
// sort by length
sort(meta.begin(), meta.end(),
[descend](const DySeqMeta& a, const DySeqMeta& b) {
bool a_ge_b = (a.end - a.begin) > (b.end - b.begin);
return descend ? a_ge_b : !a_ge_b;
});
}
LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) {
PADDLE_ENFORCE(!meta.empty(), "should build meta first");
LoDTensor result;
// collect indice need to copy to the batch
std::vector<size_t> indice;
for (const auto& seq : meta) {
size_t id = seq.begin + index;
if (id >= seq.end) break;
indice.push_back(id);
}
PADDLE_ENFORCE(!indice.empty(), "invalid batch at %d", index);
// copy the indice of records in LoDTensor
auto record_dims = slice_ddim(source->dims(), 1, source->dims().size());
auto record_dims_vec = vectorize(record_dims);
record_dims_vec.insert(record_dims_vec.begin(), indice.size());
result.Resize(make_ddim(record_dims_vec));
result.mutable_data<value_type>(platform::CPUPlace());
for (size_t i = 0; i < indice.size(); i++) {
auto index = indice[i];
auto target = result.Slice<value_type>(i, i + 1);
auto source_ = source->Slice<value_type>(index, index + 1);
target.CopyFrom<value_type>(source_, platform::CPUPlace());
}
return result;
}
// TODO(supejom) to cache lod if reasonable
LoDTensor PackDynamicBatch(const std::vector<LoDTensor>& source,
const std::vector<DySeqMeta>& meta, const LoD& lod,
size_t level) {
PADDLE_ENFORCE(!source.empty());
PADDLE_ENFORCE(!meta.empty());
PADDLE_ENFORCE(!lod.empty());
LoDTensor result;
// init result space
auto record_dims = slice_ddim(source[0].dims(), 1, source[0].dims().size());
auto record_dims_vec = vectorize(record_dims);
auto height = lod[level].back();
record_dims_vec.insert(record_dims_vec.begin(), height);
result.Resize(make_ddim(record_dims_vec));
result.mutable_data<float>(platform::CPUPlace());
for (size_t batch_id = 0; batch_id < source.size(); batch_id++) {
for (size_t seq_id = 0; seq_id < meta.size(); seq_id++) {
const auto& seq_meta = meta[seq_id];
// source is source[batch_id][seq_id]
// target is result[index]
auto index = seq_meta.begin + batch_id;
if (index >= seq_meta.end) break;
auto source_ = source[batch_id].Slice<float>(seq_id, seq_id + 1);
auto target = result.Slice<float>(index, index + 1);
target.CopyFrom<float>(source_, platform::CPUPlace());
}
}
result.set_lod(lod);
return result;
}
} // namespace detail
} // namespace framework
} // namespace paddle
/* 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 <vector>
#include "paddle/framework/lod_tensor.h"
namespace paddle {
namespace framework {
/*
* DyBatchSeqPosition stores indices of the basic element in tensor. It is used
* after lod-tensor's re-assembling, its info can be used to recover the order
* in original lod-tensor.
*/
struct DySeqMeta {
DySeqMeta(size_t begin, size_t end, size_t ori_idx)
: begin(begin), end(end), ori_idx(ori_idx) {}
size_t begin;
size_t end; // not included
size_t ori_idx;
};
/*
* TensorArray is a C-array-like array of tensors, it is meant to be used with
* dynamic iteration primitives such as while_loop. It is used to segment inputs
* and store states in all time steps.
*
* By providing some methods similar to a C++ array, the difinition of some
* state-based dynamic models such as RNN cound be more natural and highly
* flexible.
*/
class TensorArray {
public:
using value_type = float;
// max number of values allowed to store.
const size_t MAX_SIZE{100000};
/*
* Read the value at location `index` in the `TensorArray`.
*/
const LoDTensor &Read(size_t index) const;
/*
* Write value into the index of the TensorArray.
*/
void Write(size_t index, const LoDTensor &value);
/*
* Write value into the index of the TensorArray, with memory shared.
*/
void WriteShared(size_t index, const LoDTensor &value);
/*
* Recover the original LoD-arranged LoDTensor with the `values`, `level` and
* `indice_map`.
*/
LoDTensor Pack(size_t level, const std::vector<DySeqMeta> &meta,
const LoD &lod) const;
/*
* Split LoDTensor in some `level` and write the generated batches to
* `values`, if set `desend`, will sort by length in descending order else in
* ascending order.
*/
std::vector<DySeqMeta> Unpack(const LoDTensor &source, int level,
bool length_desend);
/*
* Pack the values into a tensor with rank one higher than each tensor in
* values.
*/
LoDTensor Stack() const;
/*
* Unpacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors.
*/
void Unstack(const LoDTensor &source) const;
/*
* Unpacks the given division of a rank-`R` tensor into rank-`(R-1)` tensors,
* with memory of tensors shared.
*/
void UnstackShared(const LoDTensor &source) const;
/*
* Return the number of values.
*/
size_t size() const;
protected:
void Unstack(const LoDTensor &source, bool data_shared) const;
private:
mutable std::vector<LoDTensor> values_;
}; // class TensorArray
} // namespace framework
} // namespace paddle
/* 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/framework/tensor_array.h"
#include <gtest/gtest.h>
namespace paddle {
namespace framework {
class TensorArrayTester : public ::testing::Test {
protected:
void SetUp() override {
LoDTensor source;
source.Resize(make_ddim({batch_size, dim}));
int* data = source.mutable_data<int>(platform::CPUPlace());
for (int i = 0; i < 16 * 32; i++) {
data[i] = i;
}
ta.Unstack(source);
}
TensorArray ta;
const int batch_size = 16;
const int dim = 32;
};
TEST_F(TensorArrayTester, Read) {
for (int i = 0; i < batch_size; i++) {
const auto& tensor = ta.Read(i);
ASSERT_EQ(tensor.dims()[0], 1);
ASSERT_EQ(tensor.dims()[1], dim);
}
}
TEST_F(TensorArrayTester, Write) {
LoDTensor source;
source.Resize(make_ddim({1, dim}));
for (int i = 0; i < dim; i++) {
*(source.mutable_data<int>(platform::CPUPlace()) + i) = i;
}
ta.Write(2, source);
const auto& tensor = ta.Read(2);
for (int i = 0; i < dim; i++) {
EXPECT_EQ(*(tensor.data<int>() + i), *(source.data<int>() + i));
}
}
TEST_F(TensorArrayTester, WriteShared) {
LoDTensor source;
source.Resize(make_ddim({1, dim}));
for (int i = 0; i < dim; i++) {
*(source.mutable_data<int>(platform::CPUPlace()) + i) = i;
}
ta.WriteShared(2, source);
const auto& tensor = ta.Read(2);
for (int i = 0; i < dim; i++) {
EXPECT_EQ(*(tensor.data<int>() + i), *(source.data<int>() + i));
}
EXPECT_EQ(source.data<int>(), tensor.data<int>());
}
class TensorArrayPackTester : public ::testing::Test {
protected:
virtual void SetUp() override {
lod.push_back(std::vector<size_t>{0, 2, 9, 13});
source.set_lod(lod);
source.Resize(make_ddim({13, 128}));
source.mutable_data<int>(platform::CPUPlace());
// content of each setence: 0 1 2 3 4
const auto& level = lod.front();
for (size_t i = 0; i < level.size() - 1; i++) {
size_t begin = level[i];
size_t end = level[i + 1];
for (size_t j = begin; j < end; j++) {
auto record = source.Slice<int>(j, j + 1);
for (int dim = 0; dim < 128; dim++) {
record.mutable_data<int>(platform::CPUPlace())[dim] = j - begin;
}
}
}
// unpack
meta = ta.Unpack(source, 0, true);
}
LoD lod;
TensorArray ta;
LoDTensor source;
std::vector<DySeqMeta> meta;
};
TEST_F(TensorArrayPackTester, Unpack) {
ASSERT_EQ(ta.size(), 7UL);
const auto& t0 = ta.Read(0);
const auto& t1 = ta.Read(1);
ASSERT_EQ(t0.data<int>()[0], int(0));
ASSERT_EQ(t1.data<int>()[0], int(1));
}
TEST_F(TensorArrayPackTester, Pack) {
LoDTensor packed = ta.Pack(0, meta, lod);
}
TEST_F(TensorArrayTester, size) {
ASSERT_EQ(ta.size(), static_cast<size_t>(batch_size));
}
} // namespace framework
} // namespace paddle
......@@ -65,7 +65,7 @@ inline T* Tensor::mutable_data(platform::Place place) {
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), size));
} else if (platform::is_gpu_place(place)) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
}
#else
......@@ -103,7 +103,7 @@ inline void Tensor::CopyFrom(const Tensor& src,
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
else if (platform::is_gpu_place(src_place) &&
platform::is_cpu_place(dst_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
......
......@@ -74,7 +74,7 @@ TEST(Tensor, MutableData) {
EXPECT_EQ(p1, p2);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
{
Tensor src_tensor;
float* p1 = nullptr;
......@@ -126,7 +126,7 @@ TEST(Tensor, ShareDataWith) {
ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
{
Tensor src_tensor;
Tensor dst_tensor;
......@@ -163,7 +163,7 @@ TEST(Tensor, Slice) {
EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
{
Tensor src_tensor;
src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
......@@ -218,7 +218,7 @@ TEST(Tensor, CopyFrom) {
EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
}
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
{
Tensor src_tensor;
Tensor gpu_tensor;
......
/* 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 <functional>
#include <map>
#include "paddle/platform/variant.h"
namespace paddle {
namespace framework {
class OperatorBase;
class OpDescBind;
using VariableNameMap = std::map<std::string, std::vector<std::string>>;
// The order should be as same as framework.proto
using Attribute =
boost::variant<boost::blank, int, float, std::string, std::vector<int>,
std::vector<float>, std::vector<std::string>, bool,
std::vector<bool>, BlockDesc*>;
using AttributeMap = std::unordered_map<std::string, Attribute>;
using OpCreator = std::function<OperatorBase*(
const std::string& /*type*/, const VariableNameMap& /*inputs*/,
const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
using GradOpMakerFN =
std::function<std::vector<std::unique_ptr<OpDescBind>>(const OpDescBind&)>;
} // namespace framework
} // namespace paddle
......@@ -194,7 +194,7 @@ public:
REGISTER_TYPED_FUNC(BlockExpand, CPU, BlockExpandForward);
REGISTER_TYPED_FUNC(BlockExpandGrad, CPU, BlockExpandBackward);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(BlockExpand, GPU, BlockExpandForward);
REGISTER_TYPED_FUNC(BlockExpandGrad, GPU, BlockExpandBackward);
#endif
......
......@@ -395,7 +395,7 @@ REGISTER_TYPED_FUNC(ContextProjectionForward,
REGISTER_TYPED_FUNC(ContextProjectionBackward,
CPU,
ContextProjectionBackwardFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(ContextProjectionForward,
GPU,
ContextProjectionForwardFunc);
......
......@@ -233,7 +233,7 @@ private:
REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc);
REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc);
REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc);
#endif
......
......@@ -169,7 +169,7 @@ private:
REGISTER_TYPED_FUNC(Crop, CPU, CropFunc);
REGISTER_TYPED_FUNC(CropGrad, CPU, CropGradFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(Crop, GPU, CropFunc);
REGISTER_TYPED_FUNC(CropGrad, GPU, CropGradFunc);
#endif
......
......@@ -336,7 +336,7 @@ private:
REGISTER_TYPED_FUNC(CrossMapNormal, CPU, CrossMapNormalFunc);
REGISTER_TYPED_FUNC(CrossMapNormalGrad, CPU, CrossMapNormalGradFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(CrossMapNormal, GPU, CrossMapNormalFunc);
REGISTER_TYPED_FUNC(CrossMapNormalGrad, GPU, CrossMapNormalGradFunc);
#endif
......
......@@ -292,7 +292,7 @@ REGISTER_TYPED_FUNC(DepthwiseConvGradInput,
REGISTER_TYPED_FUNC(DepthwiseConvGradFilter,
CPU,
DepthwiseConvGradFilterFunction);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(DepthwiseConv, GPU, DepthwiseConvFunction);
REGISTER_TYPED_FUNC(DepthwiseConvGradInput,
GPU,
......
......@@ -17,7 +17,7 @@ limitations under the License. */
namespace paddle {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(DepthwiseConv, Forward) {
DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
"GemmConv-CPU", "DepthwiseConv-GPU", forward);
......
......@@ -340,7 +340,7 @@ public:
REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction);
REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction);
REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction);
REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction);
REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction);
......
......@@ -24,7 +24,7 @@ TEST(GemmConv, NaiveConv) {
"NaiveConv-CPU", "GemmConv-CPU", forward);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(GemmConv, Forward) {
Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
"GemmConv-CPU", "GemmConv-GPU", forward);
......
......@@ -116,7 +116,7 @@ void TestIm2ColFunctor() {
TEST(Im2ColFunctor, CPU) { TestIm2ColFunctor<DEVICE_TYPE_CPU, float>(); }
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(Im2ColFunctor, GPU) { TestIm2ColFunctor<DEVICE_TYPE_GPU, float>(); }
......
......@@ -341,7 +341,7 @@ private:
};
REGISTER_TYPED_FUNC(MulOp, CPU, MulFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(MulOp, GPU, MulFunc);
#endif
} // namespace paddle
......@@ -207,7 +207,7 @@ private:
REGISTER_TYPED_FUNC(Pad, CPU, PadFunc);
REGISTER_TYPED_FUNC(PadGrad, CPU, PadGradFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(Pad, GPU, PadFunc);
REGISTER_TYPED_FUNC(PadGrad, GPU, PadGradFunc);
#endif
......
......@@ -217,7 +217,7 @@ public:
REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc);
REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc);
REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc);
#endif
......
......@@ -132,7 +132,7 @@ public:
REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc);
REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc);
REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc);
#endif
......
......@@ -16,7 +16,7 @@ limitations under the License. */
#include "BatchNormalizationLayer.h"
#include "Layer.h"
#include "paddle/utils/Stat.h"
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include "CudnnBatchNormLayer.h"
#endif
......
......@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/utils/Stat.h"
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include "hl_batch_transpose.h"
#endif
#include "BatchNormalizationLayer.h"
......@@ -90,7 +90,7 @@ void BatchNormalizationLayer::expandMat(const MatrixPtr& in, MatrixPtr& out) {
size_t batchSize = in->getHeight();
CHECK_EQ(out->getHeight(), batchSize * imgPixels_);
if (useGpu_) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
LOG(FATAL) << "paddle is compiled only for cpu";
#else
batchTranspose(
......@@ -127,7 +127,7 @@ void BatchNormalizationLayer::shrinkMat(const MatrixPtr& in, MatrixPtr& out) {
}
CHECK_EQ(in->getHeight(), static_cast<size_t>(batchSize * imgPixels_));
if (useGpu_) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
LOG(FATAL) << "paddle is compiled only for cpu";
#else
batchTranspose(
......
......@@ -15,7 +15,7 @@ limitations under the License. */
#include "PoolLayer.h"
#include "PoolProjectionLayer.h"
#include "paddle/utils/Logging.h"
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include "CudnnPoolLayer.h"
#endif
namespace paddle {
......@@ -53,7 +53,7 @@ Layer* PoolLayer::create(const LayerConfig& config) {
const std::string& pool = config.inputs(0).pool_conf().pool_type();
if (pool == "max-projection" || pool == "avg-projection") {
return new PoolProjectionLayer(config);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
} else if (CudnnPoolLayer::typeCheck(pool)) {
return new CudnnPoolLayer(config);
#endif
......
......@@ -674,7 +674,7 @@ void testLayerGradKernel(TestConfig testConf,
bool useGpu,
bool useWeight,
float epsilon) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) return;
#endif
FLAGS_use_gpu = useGpu;
......
......@@ -118,7 +118,7 @@ TEST(Layer, batchNorm) {
CHECK_EQ(static_cast<int>(convLayer->getOutputValue()->getWidth()), 576);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
void batchNormInference(int n, int c, int h, int w) {
MatrixPtr input = std::make_shared<GpuMatrix>(n, c * h * w);
MatrixPtr cudnnOut = std::make_shared<GpuMatrix>(n, c * h * w);
......
......@@ -116,7 +116,7 @@ MatrixPtr doOneConvTest(size_t imgSize,
}
TEST(Layer, convParaUnified) {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
MatrixPtr input, resultCpu, resultGpu;
/// TEST1 for conv ///
......
......@@ -150,7 +150,7 @@ TEST(Layer, detectionOutputLayerFwd) {
useGpu,
result2);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
// GPU case 1.
useGpu = true;
inputLoc = Matrix::create(1, 16, false, useGpu);
......
......@@ -51,7 +51,7 @@ void testEvaluator(TestConfig testConf,
string testEvaluatorName,
size_t batchSize,
bool useGpu) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) return;
#endif
FLAGS_use_gpu = useGpu;
......
......@@ -96,7 +96,7 @@ TEST(Layer, kmaxSeqScoreLayer) {
Matrix::create(subSeqStartPosition.back(), 1, false, false);
std::vector<bool> mode = {false};
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
mode.push_back(true);
#endif
......
......@@ -12,7 +12,7 @@ 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. */
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include <cudnn.h>
#endif
#include <gtest/gtest.h>
......@@ -257,7 +257,7 @@ void testProjectionConv(size_t groups, bool isDeconv) {
true);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(Projection, conv) {
/// test ConvProjection
testProjectionConv(1, false);
......@@ -421,7 +421,7 @@ TEST(Layer, depthwiseConvLayer) {
// 'depthwise_conv' is a sepecial case of 'exconv' whose
// groups size equals to the input channels size.
testDepthwiseConvLayer("exconv", /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testDepthwiseConvLayer("exconv", /* useGpu= */ true);
#endif
}
......@@ -479,7 +479,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) {
TEST(Layer, convLayer) {
testConvLayer("exconv", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testConvLayer("exconv", /* trans= */ false, /* useGpu= */ true);
testConvLayer("cudnn_conv", /* trans= */ false, /* useGpu= */ true);
#endif
......@@ -524,7 +524,7 @@ TEST(Layer, convTransLayer) {
for (auto useGpu : {false, true}) {
testConvTransLayer("exconvt", /* trans= */ false, /* useGpu= */ useGpu);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testConvTransLayer("cudnn_convt", /* trans= */ false, /* useGpu= */ true);
#endif
}
......@@ -637,7 +637,7 @@ TEST(Layer, SelectiveFullyConnectedLayer) {
/* trans= */ false,
/* useGup= */ false,
false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testLayerGrad(config,
"selective_fc",
100,
......@@ -1209,7 +1209,7 @@ void testPoolLayer(const string& poolType, bool trans, bool useGpu) {
testLayerGrad(config, "pool", 100, trans, useGpu);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
void testPoolLayer2(const string& poolType, bool trans, bool useGpu) {
TestConfig config;
config.inputDefs.push_back({INPUT_DATA, "layer_0", 3200, 0});
......@@ -1235,7 +1235,7 @@ TEST(Layer, PoolLayer) {
testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false);
testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true);
testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true);
testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true);
......@@ -1308,7 +1308,7 @@ void testPool3DLayer(const string& poolType, bool trans, bool useGpu) {
TEST(Layer, Pool3DLayer) {
testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ false);
testPool3DLayer("max", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testPool3DLayer("avg", /* trans= */ false, /* useGpu= */ true);
testPool3DLayer("max", /* trans= */ false, /* useGpu= */ true);
#endif
......@@ -1694,7 +1694,7 @@ void testBatchNormLayer(const string& type, bool trans, bool useGpu) {
TEST(Layer, BatchNormalizationLayer) {
testBatchNormLayer("batch_norm", false, false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testBatchNormLayer("batch_norm", false, true);
if (hl_get_cudnn_lib_version() >= int(4000)) {
testBatchNormLayer("cudnn_batch_norm", false, true);
......@@ -1743,7 +1743,7 @@ void testBatchNorm3DLayer(const string& type, bool trans, bool useGpu) {
TEST(Layer, testBatchNorm3DLayer) {
testBatchNorm3DLayer("batch_norm", false, false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testBatchNorm3DLayer("batch_norm", false, true);
if (hl_get_cudnn_lib_version() >= int(4000)) {
testBatchNorm3DLayer("cudnn_batch_norm", false, true);
......@@ -2261,7 +2261,7 @@ void test3DConvLayer(const string& type, bool trans, bool useGpu) {
TEST(Layer, test3DConvLayer) {
test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
test3DConvLayer("conv3d", /* trans= */ false, /* useGpu= */ true);
#endif
}
......@@ -2338,7 +2338,7 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
TEST(Layer, test3DDeConvLayer) {
test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
test3DDeConvLayer("deconv3d", /* trans= */ false, /* useGpu= */ true);
#endif
}
......
......@@ -215,13 +215,13 @@ struct testActDesc {
static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) {
cfg.biasSize = 0;
cfg.layerConfig.set_type("addto");
size_t layerSize = pm.ih * pm.ih * pm.iw;
size_t layerSize = pm.ic * pm.ih * pm.iw;
cfg.layerConfig.set_size(layerSize);
cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0});
cfg.layerConfig.add_inputs();
}
void testActivation(std::string& actType, const testActDesc& pm) {
void testActivation(std::string actType, const testActDesc& pm) {
// TODO(TJ): remove me when paddle support elu activation
if (actType == "mkldnn_elu") {
return;
......@@ -240,6 +240,7 @@ TEST(MKLDNNActivation, Activations) {
for (auto type : types) {
/* bs, c, h, w*/
testActivation(type, {16, 64, 32, 32});
testActivation(type, {2, 8, 1, 1});
}
}
......
......@@ -243,7 +243,7 @@ TEST(Compare, concat_slice) {
compareNetwork(config_file_a, config_file_b);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(Compare, img_pool) {
std::string config_file_a = "./gserver/tests/img_pool_a.conf";
std::string config_file_b = "./gserver/tests/img_pool_b.conf";
......
......@@ -151,7 +151,7 @@ TEST(Layer, priorBoxLayerFwd) {
useGpu,
result);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
// reset the input parameters
variance[1] = 0.1;
variance[3] = 0.2;
......
......@@ -485,7 +485,7 @@ TEST(ProtoDataProvider, test) {
// Currently in async mode, useGpu is not supported
continue;
}
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) {
continue;
}
......@@ -525,7 +525,7 @@ TEST(ProtoDataProvider, constant_slots) {
for (int numConstantSlots : {1, 2}) {
for (int useGpu : numTwoArray) {
for (int dataCompression : numTwoArray) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) {
continue;
}
......@@ -708,7 +708,7 @@ TEST(ProtoSequenceDataProvider, test) {
// Currently in async mode, useGpu is not supported
continue;
}
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) {
continue;
}
......
......@@ -37,7 +37,7 @@ TEST(PyDataProvider, py_fill_slots) {
config.clear_files();
std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList";
config.set_files(dataFile);
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
bool useGpu = false;
#else
bool useGpu = true;
......@@ -71,7 +71,7 @@ TEST(PyDataProvider, py_fill_nest_slots) {
std::string dataFile = "gserver/tests/pyDataProvider/pyDataProviderList";
config.set_files(dataFile);
EXPECT_EQ(config.IsInitialized(), true);
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
bool useGpu = false;
#else
bool useGpu = true;
......
......@@ -320,7 +320,7 @@ TEST(Layer, SelectiveFcLayer_train_dense_mul) {
"filelist=gserver/tests/SelectiveFcTest/dense_mul_list";
for (auto useGpu : {false, true}) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) {
break;
}
......@@ -387,7 +387,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config,
outMatSelfc->getWidth(),
outMatSelfc->getElementCnt()));
cpuOutMatSelfc->copyFrom(*outMatSelfc, HPPL_STREAM_DEFAULT);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
if (useGpu) {
hl_stream_synchronize(HPPL_STREAM_DEFAULT);
}
......@@ -417,7 +417,7 @@ void testSelectiveFcLayerTrainSparseMul(const LayerConfig& config,
MatrixPtr cpuOutMatFc(
new CpuMatrix(outMatFc->getHeight(), outMatFc->getWidth()));
cpuOutMatFc->copyFrom(*outMatFc, HPPL_STREAM_DEFAULT);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
if (useGpu) {
hl_stream_synchronize(HPPL_STREAM_DEFAULT);
}
......@@ -442,7 +442,7 @@ TEST(Layer, SelectiveFcLayer_train_sparse_mul) {
selLayerConfig.set_size(fcLayerWidth);
testSelectiveFcLayerTrainSparseMul(selLayerConfig, false);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testSelectiveFcLayerTrainSparseMul(selLayerConfig, true);
#endif
}
......
......@@ -194,7 +194,7 @@ TEST(Layer, SeqSliceLayer) {
vector<vector<real>> ends;
std::vector<bool> mode = {false};
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
mode.push_back(true);
#endif
genSeqInfo(seqStartPos, subSeqStartPos);
......
......@@ -199,7 +199,7 @@ TEST(Layer, WarpCTCLayer) {
for (auto batchSize : {1, 10, 32}) {
for (auto normByTimes : {false, true}) {
for (auto useGpu : {false, true}) {
#ifdef PADDLE_ONLY_CPU
#ifndef PADDLE_WITH_CUDA
if (useGpu) continue;
#endif
LOG(INFO) << "layerSize=" << layerSize << " batchSize=" << batchSize
......
......@@ -670,7 +670,7 @@ void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) {
}
void GpuMatrix::selectRows(Matrix& table, IVector& ids) {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
CHECK(dynamic_cast<GpuMatrix*>(&table));
CHECK(table.useGpu());
CHECK(ids.useGpu());
......@@ -694,7 +694,7 @@ void GpuMatrix::selectRows(Matrix& table, IVector& ids) {
}
void GpuMatrix::addToRows(Matrix& table, IVector& ids) {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
CHECK(dynamic_cast<GpuMatrix*>(&table));
CHECK(table.useGpu());
CHECK(ids.useGpu());
......@@ -741,7 +741,7 @@ void GpuMatrix::rowMax(Matrix& max) {
}
void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal";
size_t numSamples = getHeight();
size_t beam = maxVal.getWidth();
......
......@@ -99,7 +99,11 @@ public:
/**
* @brief clear local buffer. It only affect auto-growth buffer.
*/
inline void clear() { rowStore_.clear(); }
inline void clear() {
// swap an empty vector to it to free the memory.
std::vector<real, AlignedAllocator<real, 32>> empty;
rowStore_.swap(empty);
}
/**
* @brief get current number of rows.
......
......@@ -836,7 +836,7 @@ void GpuSparseMatrix::zeroMem() {
}
void GpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal";
size_t numSamples = getHeight();
size_t beam = maxVal.getWidth();
......
......@@ -172,7 +172,7 @@ void GpuVectorT<T>::isEqualTo(const VectorT<T>& b, const T& value) {
template <class T>
void GpuVectorT<T>::selectFrom(const VectorT<T>& src, const VectorT<int>& ids) {
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
hl_vector_select_from<T>(this->getData(),
this->getSize(),
src.getData(),
......@@ -850,7 +850,7 @@ CpuGpuVectorT<T>::CpuGpuVectorT(CpuGpuVectorT<T>& src,
size_t size)
: sync_(nullptr) {
CHECK_LE(offset + size, static_cast<size_t>(src.getSize()));
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
SyncedFlag* flag = src.getSync();
if (*flag == DATA_AT_CPU) {
src.copyToGpu(); // will set synchronous data between CPU and GPU
......@@ -861,7 +861,7 @@ CpuGpuVectorT<T>::CpuGpuVectorT(CpuGpuVectorT<T>& src,
auto cMemHandle = (src.getVector(false))->getMemoryHandle();
cpuVectorT_ = std::make_shared<CpuVectorT<T>>(
size, std::dynamic_pointer_cast<CpuMemoryHandle>(cMemHandle), offset);
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
auto gMemHandle = (src.getVector(true))->getMemoryHandle();
gpuVectorT_ = std::make_shared<GpuVectorT<T>>(
size, std::dynamic_pointer_cast<GpuMemoryHandle>(gMemHandle), offset);
......
......@@ -68,7 +68,7 @@ void testPoolAllocator() {
TEST(Allocator, Pool) {
testPoolAllocator<CpuAllocator>();
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
testPoolAllocator<GpuAllocator>();
#endif
}
......@@ -92,7 +92,7 @@ TEST(MemoryHandle, Cpu) {
EXPECT_EQ(ptr1, ptr2);
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(MemoryHandle, Gpu) {
int numGpu = hl_get_device_count();
......
......@@ -12,7 +12,7 @@ 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. */
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
/**
* This test file use autotest::AutoCompare and cmpWithoutArg to compares the
* implementation of CPU and GPU member function in
......
......@@ -12,7 +12,7 @@ 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. */
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include <gtest/gtest.h>
#include "paddle/math/Vector.h"
......
......@@ -94,7 +94,7 @@ void testWrapper(F&& f) {
}
}
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
TEST(ExecViaCpu, test1) {
testWrapper(f);
testWrapper(&f);
......
......@@ -12,7 +12,7 @@ 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. */
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
#include <gtest/gtest.h>
#include "paddle/math/Matrix.h"
......
......@@ -12,7 +12,7 @@ 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. */
#ifndef PADDLE_ONLY_CPU
#ifdef PADDLE_WITH_CUDA
/**
* This test file use autotest::AutoCompare and cmpWithArg to compares the
* implementation of CPU and GPU member function in Matrix.cpp.
......
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册