|
| 1 | +import hashlib |
| 2 | +from pathlib import Path |
| 3 | +from typing import Dict, Union |
| 4 | + |
| 5 | +import arviz as az |
| 6 | +import numpy as np |
| 7 | +import pandas as pd |
| 8 | +import pymc as pm |
| 9 | + |
| 10 | + |
| 11 | +class ModelBuilder(pm.Model): |
| 12 | + """ |
| 13 | + ModelBuilder can be used to provide an easy-to-use API (similar to scikit-learn) for models |
| 14 | + and help with deployment. |
| 15 | +
|
| 16 | + Extends the pymc.Model class. |
| 17 | + """ |
| 18 | + |
| 19 | + _model_type = "BaseClass" |
| 20 | + version = "None" |
| 21 | + |
| 22 | + def __init__( |
| 23 | + self, |
| 24 | + model_config: Dict, |
| 25 | + sampler_config: Dict, |
| 26 | + data: Dict[str, Union[np.ndarray, pd.DataFrame, pd.Series]] = None, |
| 27 | + ): |
| 28 | + """ |
| 29 | + Initializes model configuration and sampler configuration for the model |
| 30 | +
|
| 31 | + Parameters |
| 32 | + ---------- |
| 33 | + model_config : Dictionary |
| 34 | + dictionary of parameters that initialise model configuration. Generated by the user defined create_sample_input method. |
| 35 | + sampler_config : Dictionary |
| 36 | + dictionary of parameters that initialise sampler configuration. Generated by the user defined create_sample_input method. |
| 37 | + data : Dictionary |
| 38 | + It is the data we need to train the model on. |
| 39 | + Examples |
| 40 | + -------- |
| 41 | + >>> class LinearModel(ModelBuilder): |
| 42 | + >>> ... |
| 43 | + >>> model = LinearModel(model_config, sampler_config) |
| 44 | + """ |
| 45 | + |
| 46 | + super().__init__() |
| 47 | + self.model_config = model_config # parameters for priors etc. |
| 48 | + self.sample_config = sampler_config # parameters for sampling |
| 49 | + self.idata = None # inference data object |
| 50 | + self.data = data |
| 51 | + self.build() |
| 52 | + |
| 53 | + def build(self): |
| 54 | + """ |
| 55 | + Builds the defined model. |
| 56 | + """ |
| 57 | + |
| 58 | + with self: |
| 59 | + self.build_model(self.model_config, self.data) |
| 60 | + |
| 61 | + def _data_setter( |
| 62 | + self, data: Dict[str, Union[np.ndarray, pd.DataFrame, pd.Series]], x_only: bool = True |
| 63 | + ): |
| 64 | + """ |
| 65 | + Sets new data in the model. |
| 66 | +
|
| 67 | + Parameters |
| 68 | + ---------- |
| 69 | + data : Dictionary of string and either of numpy array, pandas dataframe or pandas Series |
| 70 | + It is the data we need to set as idata for the model |
| 71 | + x_only : bool |
| 72 | + if data only contains values of x and y is not present in the data |
| 73 | +
|
| 74 | + Examples |
| 75 | + -------- |
| 76 | + >>> def _data_setter(self, data : pd.DataFrame): |
| 77 | + >>> with self.model: |
| 78 | + >>> pm.set_data({'x': data['input'].values}) |
| 79 | + >>> try: # if y values in new data |
| 80 | + >>> pm.set_data({'y_data': data['output'].values}) |
| 81 | + >>> except: # dummies otherwise |
| 82 | + >>> pm.set_data({'y_data': np.zeros(len(data))}) |
| 83 | + """ |
| 84 | + |
| 85 | + raise NotImplementedError |
| 86 | + |
| 87 | + @classmethod |
| 88 | + def create_sample_input(cls): |
| 89 | + """ |
| 90 | + Needs to be implemented by the user in the inherited class. |
| 91 | + Returns examples for data, model_config, sampler_config. |
| 92 | + This is useful for understanding the required |
| 93 | + data structures for the user model. |
| 94 | +
|
| 95 | + Examples |
| 96 | + -------- |
| 97 | + >>> @classmethod |
| 98 | + >>> def create_sample_input(cls): |
| 99 | + >>> x = np.linspace(start=1, stop=50, num=100) |
| 100 | + >>> y = 5 * x + 3 + np.random.normal(0, 1, len(x)) * np.random.rand(100)*10 + np.random.rand(100)*6.4 |
| 101 | + >>> data = pd.DataFrame({'input': x, 'output': y}) |
| 102 | +
|
| 103 | + >>> model_config = { |
| 104 | + >>> 'a_loc': 7, |
| 105 | + >>> 'a_scale': 3, |
| 106 | + >>> 'b_loc': 5, |
| 107 | + >>> 'b_scale': 3, |
| 108 | + >>> 'obs_error': 2, |
| 109 | + >>> } |
| 110 | +
|
| 111 | + >>> sampler_config = { |
| 112 | + >>> 'draws': 1_000, |
| 113 | + >>> 'tune': 1_000, |
| 114 | + >>> 'chains': 1, |
| 115 | + >>> 'target_accept': 0.95, |
| 116 | + >>> } |
| 117 | + >>> return data, model_config, sampler_config |
| 118 | + """ |
| 119 | + |
| 120 | + raise NotImplementedError |
| 121 | + |
| 122 | + def save(self, fname): |
| 123 | + """ |
| 124 | + Saves inference data of the model. |
| 125 | +
|
| 126 | + Parameters |
| 127 | + ---------- |
| 128 | + fname : string |
| 129 | + This denotes the name with path from where idata should be saved. |
| 130 | +
|
| 131 | + Examples |
| 132 | + -------- |
| 133 | + >>> class LinearModel(ModelBuilder): |
| 134 | + >>> ... |
| 135 | + >>> data, model_config, sampler_config = LinearModel.create_sample_input() |
| 136 | + >>> model = LinearModel(model_config, sampler_config) |
| 137 | + >>> idata = model.fit(data) |
| 138 | + >>> name = './mymodel.nc' |
| 139 | + >>> model.save(name) |
| 140 | + """ |
| 141 | + |
| 142 | + file = Path(str(fname)) |
| 143 | + self.idata.to_netcdf(file) |
| 144 | + |
| 145 | + @classmethod |
| 146 | + def load(cls, fname): |
| 147 | + """ |
| 148 | + Loads inference data for the model. |
| 149 | +
|
| 150 | + Parameters |
| 151 | + ---------- |
| 152 | + fname : string |
| 153 | + This denotes the name with path from where idata should be loaded from. |
| 154 | +
|
| 155 | + Returns |
| 156 | + ------- |
| 157 | + Returns the inference data that is loaded from local system. |
| 158 | +
|
| 159 | + Raises |
| 160 | + ------ |
| 161 | + ValueError |
| 162 | + If the inference data that is loaded doesn't match with the model. |
| 163 | +
|
| 164 | + Examples |
| 165 | + -------- |
| 166 | + >>> class LinearModel(ModelBuilder): |
| 167 | + >>> ... |
| 168 | + >>> name = './mymodel.nc' |
| 169 | + >>> imported_model = LinearModel.load(name) |
| 170 | + """ |
| 171 | + |
| 172 | + filepath = Path(str(fname)) |
| 173 | + data = az.from_netcdf(filepath) |
| 174 | + idata = data |
| 175 | + # Since there is an issue with attrs getting saved in netcdf format which will be fixed in future the following part of code is commented |
| 176 | + # Link of issue -> https://github.com/arviz-devs/arviz/issues/2109 |
| 177 | + # if model.idata.attrs is not None: |
| 178 | + # if model.idata.attrs['id'] == self.idata.attrs['id']: |
| 179 | + # self = model |
| 180 | + # self.idata = data |
| 181 | + # return self |
| 182 | + # else: |
| 183 | + # raise ValueError( |
| 184 | + # f"The route '{file}' does not contain an inference data of the same model '{self.__name__}'" |
| 185 | + # ) |
| 186 | + return idata |
| 187 | + |
| 188 | + def fit(self, data: Dict[str, Union[np.ndarray, pd.DataFrame, pd.Series]] = None): |
| 189 | + """ |
| 190 | + As the name suggests fit can be used to fit a model using the data that is passed as a parameter. |
| 191 | + Sets attrs to inference data of the model. |
| 192 | +
|
| 193 | + Parameter |
| 194 | + --------- |
| 195 | + data : Dictionary of string and either of numpy array, pandas dataframe or pandas Series |
| 196 | + It is the data we need to train the model on. |
| 197 | +
|
| 198 | + Returns |
| 199 | + ------- |
| 200 | + returns inference data of the fitted model. |
| 201 | +
|
| 202 | + Examples |
| 203 | + -------- |
| 204 | + >>> data, model_config, sampler_config = LinearModel.create_sample_input() |
| 205 | + >>> model = LinearModel(model_config, sampler_config) |
| 206 | + >>> idata = model.fit(data) |
| 207 | + Auto-assigning NUTS sampler... |
| 208 | + Initializing NUTS using jitter+adapt_diag... |
| 209 | + """ |
| 210 | + |
| 211 | + if data is not None: |
| 212 | + self.data = data |
| 213 | + self._data_setter(data) |
| 214 | + |
| 215 | + if self.basic_RVs == []: |
| 216 | + self.build() |
| 217 | + |
| 218 | + with self: |
| 219 | + self.idata = pm.sample(**self.sample_config) |
| 220 | + self.idata.extend(pm.sample_prior_predictive()) |
| 221 | + self.idata.extend(pm.sample_posterior_predictive(self.idata)) |
| 222 | + |
| 223 | + self.idata.attrs["id"] = self.id() |
| 224 | + self.idata.attrs["model_type"] = self._model_type |
| 225 | + self.idata.attrs["version"] = self.version |
| 226 | + self.idata.attrs["sample_conifg"] = self.sample_config |
| 227 | + self.idata.attrs["model_config"] = self.model_config |
| 228 | + return self.idata |
| 229 | + |
| 230 | + def predict( |
| 231 | + self, |
| 232 | + data_prediction: Dict[str, Union[np.ndarray, pd.DataFrame, pd.Series]] = None, |
| 233 | + point_estimate: bool = True, |
| 234 | + ): |
| 235 | + """ |
| 236 | + Uses model to predict on unseen data. |
| 237 | +
|
| 238 | + Parameters |
| 239 | + --------- |
| 240 | + data_prediction : Dictionary of string and either of numpy array, pandas dataframe or pandas Series |
| 241 | + It is the data we need to make prediction on using the model. |
| 242 | + point_estimate : bool |
| 243 | + Adds point like estimate used as mean passed as |
| 244 | +
|
| 245 | + Returns |
| 246 | + ------- |
| 247 | + returns dictionary of sample's posterior predict. |
| 248 | +
|
| 249 | + Examples |
| 250 | + -------- |
| 251 | + >>> data, model_config, sampler_config = LinearModel.create_sample_input() |
| 252 | + >>> model = LinearModel(model_config, sampler_config) |
| 253 | + >>> idata = model.fit(data) |
| 254 | + >>> x_pred = [] |
| 255 | + >>> prediction_data = pd.DataFrame({'input':x_pred}) |
| 256 | + # only point estimate |
| 257 | + >>> pred_mean = model.predict(prediction_data) |
| 258 | + # samples |
| 259 | + >>> pred_samples = model.predict(prediction_data, point_estimate=False) |
| 260 | + """ |
| 261 | + |
| 262 | + if data_prediction is not None: # set new input data |
| 263 | + self._data_setter(data_prediction) |
| 264 | + |
| 265 | + with self.model: # sample with new input data |
| 266 | + post_pred = pm.sample_posterior_predictive(self.idata) |
| 267 | + |
| 268 | + # reshape output |
| 269 | + post_pred = self._extract_samples(post_pred) |
| 270 | + if point_estimate: # average, if point-like estimate desired |
| 271 | + for key in post_pred: |
| 272 | + post_pred[key] = post_pred[key].mean(axis=0) |
| 273 | + |
| 274 | + return post_pred |
| 275 | + |
| 276 | + @staticmethod |
| 277 | + def _extract_samples(post_pred: az.data.inference_data.InferenceData) -> Dict[str, np.array]: |
| 278 | + """ |
| 279 | + This method can be used to extract samples from posterior predict. |
| 280 | +
|
| 281 | + Parameters |
| 282 | + ---------- |
| 283 | + post_pred: arviz InferenceData object |
| 284 | +
|
| 285 | + Returns |
| 286 | + ------- |
| 287 | + Dictionary of numpy arrays from InferenceData object |
| 288 | + """ |
| 289 | + |
| 290 | + post_pred_dict = dict() |
| 291 | + for key in post_pred.posterior_predictive: |
| 292 | + post_pred_dict[key] = post_pred.posterior_predictive[key].to_numpy()[0] |
| 293 | + |
| 294 | + return post_pred_dict |
| 295 | + |
| 296 | + def id(self): |
| 297 | + """ |
| 298 | + It creates a hash value to match the model version using last 16 characters of hash encoding. |
| 299 | +
|
| 300 | + Returns |
| 301 | + ------- |
| 302 | + Returns string of length 16 characters contains unique hash of the model |
| 303 | + """ |
| 304 | + |
| 305 | + hasher = hashlib.sha256() |
| 306 | + hasher.update(str(self.model_config.values()).encode()) |
| 307 | + hasher.update(self.version.encode()) |
| 308 | + hasher.update(self._model_type.encode()) |
| 309 | + hasher.update(str(self.sample_config.values()).encode()) |
| 310 | + return hasher.hexdigest()[:16] |
0 commit comments