| 1 |
# Decision Forest Code Completion Model |
| 2 |
|
| 3 |
## Decision Forest |
| 4 |
A **decision forest** is a collection of many decision trees. A **decision tree** is a full binary tree that provides a quality prediction for an input (code completion item). Internal nodes represent a **binary decision** based on the input data, and leaf nodes represent a prediction. |
| 5 |
|
| 6 |
In order to predict the relevance of a code completion item, we traverse each of the decision trees beginning with their roots until we reach a leaf. |
| 7 |
|
| 8 |
An input (code completion candidate) is characterized as a set of **features**, such as the *type of symbol* or the *number of existing references*. |
| 9 |
|
| 10 |
At every non-leaf node, we evaluate the condition to decide whether to go left or right. The condition compares one *feature** of the input against a constant. The condition can be of two types: |
| 11 |
- **if_greater**: Checks whether a numerical feature is **>=** a **threshold**. |
| 12 |
- **if_member**: Check whether the **enum** feature is contained in the **set** defined in the node. |
| 13 |
|
| 14 |
A leaf node contains the value **score**. |
| 15 |
To compute an overall **quality** score, we traverse each tree in this way and add up the scores. |
| 16 |
|
| 17 |
## Model Input Format |
| 18 |
The input model is represented in json format. |
| 19 |
|
| 20 |
### Features |
| 21 |
The file **features.json** defines the features available to the model. |
| 22 |
It is a json list of features. The features can be of following two kinds. |
| 23 |
|
| 24 |
#### Number |
| 25 |
``` |
| 26 |
{ |
| 27 |
"name": "a_numerical_feature", |
| 28 |
"kind": "NUMBER" |
| 29 |
} |
| 30 |
``` |
| 31 |
#### Enum |
| 32 |
``` |
| 33 |
{ |
| 34 |
"name": "an_enum_feature", |
| 35 |
"kind": "ENUM", |
| 36 |
"enum": "fully::qualified::enum", |
| 37 |
"header": "path/to/HeaderDeclaringEnum.h" |
| 38 |
} |
| 39 |
``` |
| 40 |
The field `enum` specifies the fully qualified name of the enum. |
| 41 |
The maximum cardinality of the enum can be **32**. |
| 42 |
|
| 43 |
The field `header` specifies the header containing the declaration of the enum. |
| 44 |
This header is included by the inference runtime. |
| 45 |
|
| 46 |
|
| 47 |
### Decision Forest |
| 48 |
The file `forest.json` defines the decision forest. It is a json list of **DecisionTree**. |
| 49 |
|
| 50 |
**DecisionTree** is one of **IfGreaterNode**, **IfMemberNode**, **LeafNode**. |
| 51 |
#### IfGreaterNode |
| 52 |
``` |
| 53 |
{ |
| 54 |
"operation": "if_greater", |
| 55 |
"feature": "a_numerical_feature", |
| 56 |
"threshold": A real number, |
| 57 |
"then": {A DecisionTree}, |
| 58 |
"else": {A DecisionTree} |
| 59 |
} |
| 60 |
``` |
| 61 |
#### IfMemberNode |
| 62 |
``` |
| 63 |
{ |
| 64 |
"operation": "if_member", |
| 65 |
"feature": "an_enum_feature", |
| 66 |
"set": ["enum_value1", "enum_value2", ...], |
| 67 |
"then": {A DecisionTree}, |
| 68 |
"else": {A DecisionTree} |
| 69 |
} |
| 70 |
``` |
| 71 |
#### LeafNode |
| 72 |
``` |
| 73 |
{ |
| 74 |
"operation": "boost", |
| 75 |
"score": A real number |
| 76 |
} |
| 77 |
``` |
| 78 |
|
| 79 |
## Code Generator for Inference |
| 80 |
The implementation of inference runtime is split across: |
| 81 |
|
| 82 |
### Code generator |
| 83 |
The code generator `CompletionModelCodegen.py` takes input the `${model}` dir and generates the inference library: |
| 84 |
- `${output_dir}/{filename}.h` |
| 85 |
- `${output_dir}/{filename}.cpp` |
| 86 |
|
| 87 |
Invocation |
| 88 |
``` |
| 89 |
python3 CompletionModelCodegen.py \ |
| 90 |
--model path/to/model/dir \ |
| 91 |
--output_dir path/to/output/dir \ |
| 92 |
--filename OutputFileName \ |
| 93 |
--cpp_class clang::clangd::YourExampleClass |
| 94 |
``` |
| 95 |
### Build System |
| 96 |
`CompletionModel.cmake` provides `gen_decision_forest` method . |
| 97 |
Client intending to use the CompletionModel for inference can use this to trigger the code generator and generate the inference library. |
| 98 |
It can then use the generated API by including and depending on this library. |
| 99 |
|
| 100 |
### Generated API for inference |
| 101 |
The code generator defines the Example `class` inside relevant namespaces as specified in option `${cpp_class}`. |
| 102 |
|
| 103 |
Members of this generated class comprises of all the features mentioned in `features.json`. |
| 104 |
Thus this class can represent a code completion candidate that needs to be scored. |
| 105 |
|
| 106 |
The API also provides `float Evaluate(const MyClass&)` which can be used to score the completion candidate. |
| 107 |
|
| 108 |
|
| 109 |
## Example |
| 110 |
### model/features.json |
| 111 |
``` |
| 112 |
[ |
| 113 |
{ |
| 114 |
"name": "ANumber", |
| 115 |
"type": "NUMBER" |
| 116 |
}, |
| 117 |
{ |
| 118 |
"name": "AFloat", |
| 119 |
"type": "NUMBER" |
| 120 |
}, |
| 121 |
{ |
| 122 |
"name": "ACategorical", |
| 123 |
"type": "ENUM", |
| 124 |
"enum": "ns1::ns2::TestEnum", |
| 125 |
"header": "model/CategoricalFeature.h" |
| 126 |
} |
| 127 |
] |
| 128 |
``` |
| 129 |
### model/forest.json |
| 130 |
``` |
| 131 |
[ |
| 132 |
{ |
| 133 |
"operation": "if_greater", |
| 134 |
"feature": "ANumber", |
| 135 |
"threshold": 200.0, |
| 136 |
"then": { |
| 137 |
"operation": "if_greater", |
| 138 |
"feature": "AFloat", |
| 139 |
"threshold": -1, |
| 140 |
"then": { |
| 141 |
"operation": "boost", |
| 142 |
"score": 10.0 |
| 143 |
}, |
| 144 |
"else": { |
| 145 |
"operation": "boost", |
| 146 |
"score": -20.0 |
| 147 |
} |
| 148 |
}, |
| 149 |
"else": { |
| 150 |
"operation": "if_member", |
| 151 |
"feature": "ACategorical", |
| 152 |
"set": [ |
| 153 |
"A", |
| 154 |
"C" |
| 155 |
], |
| 156 |
"then": { |
| 157 |
"operation": "boost", |
| 158 |
"score": 3.0 |
| 159 |
}, |
| 160 |
"else": { |
| 161 |
"operation": "boost", |
| 162 |
"score": -4.0 |
| 163 |
} |
| 164 |
} |
| 165 |
}, |
| 166 |
{ |
| 167 |
"operation": "if_member", |
| 168 |
"feature": "ACategorical", |
| 169 |
"set": [ |
| 170 |
"A", |
| 171 |
"B" |
| 172 |
], |
| 173 |
"then": { |
| 174 |
"operation": "boost", |
| 175 |
"score": 5.0 |
| 176 |
}, |
| 177 |
"else": { |
| 178 |
"operation": "boost", |
| 179 |
"score": -6.0 |
| 180 |
} |
| 181 |
} |
| 182 |
] |
| 183 |
``` |
| 184 |
### DecisionForestRuntime.h |
| 185 |
``` |
| 186 |
... |
| 187 |
namespace ns1 { |
| 188 |
namespace ns2 { |
| 189 |
namespace test { |
| 190 |
class Example { |
| 191 |
public: |
| 192 |
void setANumber(float V) { ... } |
| 193 |
void setAFloat(float V) { ... } |
| 194 |
void setACategorical(unsigned V) { ... } |
| 195 |
|
| 196 |
private: |
| 197 |
... |
| 198 |
}; |
| 199 |
|
| 200 |
float Evaluate(const Example&); |
| 201 |
} // namespace test |
| 202 |
} // namespace ns2 |
| 203 |
} // namespace ns1 |
| 204 |
``` |
| 205 |
|
| 206 |
### CMake Invocation |
| 207 |
Inorder to use the inference runtime, one can use `gen_decision_forest` function |
| 208 |
described in `CompletionModel.cmake` which invokes `CodeCompletionCodegen.py` with the appropriate arguments. |
| 209 |
|
| 210 |
For example, the following invocation reads the model present in `path/to/model` and creates |
| 211 |
`${CMAKE_CURRENT_BINARY_DIR}/myfilename.h` and `${CMAKE_CURRENT_BINARY_DIR}/myfilename.cpp` |
| 212 |
describing a `class` named `MyClass` in namespace `fully::qualified`. |
| 213 |
|
| 214 |
|
| 215 |
|
| 216 |
``` |
| 217 |
gen_decision_forest(path/to/model |
| 218 |
myfilename |
| 219 |
::fully::qualifed::MyClass) |
| 220 |
``` |