|
| 1 | +# Design Doc: Python API |
| 2 | + |
| 3 | +Due to the refactorization of the PaddlePaddle core, we need Python classes to construct corresponding protobuf messages that describe a DL program. |
| 4 | + |
| 5 | +| Python classes | Protobuf messages | |
| 6 | +| --- | --- | |
| 7 | +| Program | ProgramDesc | |
| 8 | +| Block | BlockDesc | |
| 9 | +| Operator | OpDesc | |
| 10 | +| Variable | VarDesc | |
| 11 | + |
| 12 | +Please be aware that these Python classes need to maintain some construction-time information, which are not part of the protobuf messages. |
| 13 | + |
| 14 | +## Core Concepts |
| 15 | + |
| 16 | +### Program |
| 17 | + |
| 18 | +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. A `BlockDesc` refers to its parent block by its index in the array. For example, operators in the step block of an RNN operator needs to be able to access variables in its ancessor blocks. |
| 19 | + |
| 20 | +Whenever we create a block, we need set its parent block to the current block, so the Python class `Program` needs to maintain a data member `current_block`. |
| 21 | + |
| 22 | +```python |
| 23 | +class Program(objects): |
| 24 | + def __init__(self): |
| 25 | + self.proto = core.NewProgram() # a C++ ProgramDesc pointer. |
| 26 | + self.blocks = vector<Block>() |
| 27 | + self.blocks.append(Block(self, -1)) # the global block |
| 28 | + self.current_block = 0 # initialized to the global block |
| 29 | + |
| 30 | + def global_block(): |
| 31 | + return self.blocks[0] |
| 32 | + |
| 33 | + def current_block(): |
| 34 | + return self.get_block(self.current_block) |
| 35 | + |
| 36 | + def rollback(): |
| 37 | + self.current_block = self.current_block().parent_idx |
| 38 | + |
| 39 | + def create_block(): |
| 40 | + new_block_idx = len(self.block) |
| 41 | + self.blocks.append(Block(self, self.current_block)) |
| 42 | + self.current_block = new_block_idx |
| 43 | + return current_block() |
| 44 | +``` |
| 45 | + |
| 46 | +`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`. |
| 47 | + |
| 48 | +`Program` creates the first block as the global block in its constructor. All parameters and their initializer operators are in the global block. |
| 49 | + |
| 50 | +### Block |
| 51 | + |
| 52 | +A [Block](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md) includes |
| 53 | + |
| 54 | +1. a map from variable names to an instance of the Python `Variable` class, and |
| 55 | +1. a list of `Operator` instances. |
| 56 | + |
| 57 | +```python |
| 58 | +class Block(objects): |
| 59 | + def __init__(self, program, parent_idx): |
| 60 | + self.proto = core.NewBlock(program.proto) |
| 61 | + self.program = program |
| 62 | + self.vars = map<string, Variable>() |
| 63 | + self.ops = vector<Operator>() |
| 64 | + self.parent_idx = parent_idx |
| 65 | + |
| 66 | + def create_var(self, ...): |
| 67 | + return Variable(self, ...) |
| 68 | + |
| 69 | + def _create_global_var(self, ...): |
| 70 | + program.global_block().create_var(...) |
| 71 | + |
| 72 | + def create_parameter(self, name, ...): |
| 73 | + # Parameter is a subclass of variable. See Parameter section for details. |
| 74 | + self.vars[name] = Parameter(self._create_global_var(...), ...) |
| 75 | + return self.vars[name] |
| 76 | + |
| 77 | + def append_operator(self, ...): |
| 78 | + self.ops.append(Operator(self, ...)) |
| 79 | + |
| 80 | + def prepend_operator(self, ...): # Parameter's ctor prepands initialize operators. |
| 81 | + self.ops.prepend(Operator(self, ...)) |
| 82 | +``` |
| 83 | + |
| 84 | +`create_parameter` is necessary because parameters are global variables, those defined in the global block, but can be created in some sub-blocks, e.g., an FC layer in the step block of an RNN operator. |
| 85 | + |
| 86 | +`prepand_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. |
| 87 | + |
| 88 | +### Operator |
| 89 | + |
| 90 | +The `Operator` class fills in the `OpDesc` message and calls the C++ function `InferShape` to infer output shape from input shape. |
| 91 | + |
| 92 | +```python |
| 93 | +class Operator(object): |
| 94 | + def __init__(self, |
| 95 | + block, # Block |
| 96 | + type, # string |
| 97 | + inputs, # dict<string, Variable> |
| 98 | + outputs,# dict<stirng, Variable> |
| 99 | + attrs # dict<string, Any> |
| 100 | + ): |
| 101 | + self.proto = core.NewOpDesc(block.proto, type, inputs, outputs, attrs) |
| 102 | + core.infer_shape(self.proto, inputs, outputs) |
| 103 | + |
| 104 | + def type(self): |
| 105 | + return self.proto.type() |
| 106 | +``` |
| 107 | + |
| 108 | +`Operator` creates the `OpDesc` message in C++ space, so could it call the `InferShape` function, which is in C++. |
| 109 | + |
| 110 | +### Variable |
| 111 | + |
| 112 | +Operators take Variables as its inputs and outputs. |
| 113 | + |
| 114 | +```python |
| 115 | +class Variable(object): |
| 116 | + def __init__(self, |
| 117 | + block=None, # Block |
| 118 | + name=None, # string |
| 119 | + shape, # tuple |
| 120 | + dtype="float32", # string |
| 121 | + lod_level=None # int |
| 122 | + ): |
| 123 | + if name is None: |
| 124 | + name = unique_name_generator() |
| 125 | + self.name = name |
| 126 | + self.block = block |
| 127 | + self.proto = core.NewVarDesc(block.proto, name, shape, lod_level) |
| 128 | + self.writer = None |
| 129 | +``` |
| 130 | + |
| 131 | +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 writes 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**. |
| 132 | + |
| 133 | +### Parameter |
| 134 | + |
| 135 | +A parameter is a global variable with an initializer (or load) operator. |
| 136 | + |
| 137 | +```python |
| 138 | +class Parameter(Variable): |
| 139 | + def __init__(self, |
| 140 | + block=None, # Block |
| 141 | + name=None, # string |
| 142 | + shape, # tuple |
| 143 | + dtype="float32", # string |
| 144 | + lod_level=None # int |
| 145 | + trainable, # bool |
| 146 | + initialize_op_attrs, |
| 147 | + optimize_op_attrs): |
| 148 | + super(Parameter, self).__init__(block, name, shape, dtype, lod_level) |
| 149 | + self.trainable = trainable |
| 150 | + self.optimize_op_attrs = optimize_op_attrs |
| 151 | + block.prepend(Operator(block, # Block |
| 152 | + initialize_op_attrs['type'], # string |
| 153 | + None, # no inputs |
| 154 | + self, # output is the parameter |
| 155 | + initialize_op_attrs) |
| 156 | +``` |
| 157 | + |
| 158 | +When users create a parameter, s/he can call |
| 159 | + |
| 160 | +```python |
| 161 | +program.create_parameter( |
| 162 | + ..., |
| 163 | + init_attr={ |
| 164 | + type: "uniform_random", |
| 165 | + min: -1.0, |
| 166 | + max: 1.0, |
| 167 | + }) |
| 168 | +) |
| 169 | +``` |
| 170 | + |
| 171 | +In above example, `init_attr.type` names an initialize operator. It can also name the load operator |
| 172 | + |
| 173 | +```python |
| 174 | +init_attr={ |
| 175 | + type: "load", |
| 176 | + filename: "something.numpy", |
| 177 | +} |
| 178 | +``` |
| 179 | + |
| 180 | +`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. |
| 181 | + |
| 182 | +## Layer Functions |
| 183 | + |
| 184 | +A layer is a Python function that creates some operators and variables. Layers simplify the work of application programmers. |
| 185 | + |
| 186 | +### Data Layer |
| 187 | + |
| 188 | +```python |
| 189 | +def data_layer(name, type, column_name): |
| 190 | + block = the_current_program.glolal_block() |
| 191 | + var = block.create_global_var( |
| 192 | + name=name, |
| 193 | + shape=[None] + type.dims(), |
| 194 | + dtype=type.dtype) |
| 195 | + block.prepend_operator(block, |
| 196 | + type="Feed", |
| 197 | + inputs = None, |
| 198 | + outputs = [var], |
| 199 | + {column_name: column_name}) |
| 200 | + return var |
| 201 | +``` |
| 202 | + |
| 203 | +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). |
| 204 | + |
| 205 | +### FC Layer |
| 206 | + |
| 207 | +```python |
| 208 | +def fc_layer(input, size, ...): |
| 209 | + block = program.current_block() |
| 210 | + w = block.create_parameter(...) |
| 211 | + b = block.create_parameter(...) |
| 212 | + out = block.create_var() |
| 213 | + op = block.append_operator("FC", X=input, W=w, b=b, out=out) |
| 214 | + out.writer = op |
| 215 | + return out |
| 216 | +``` |
0 commit comments