3 from __future__ import annotations
5 from abc import ABC, abstractmethod
13 from types import SimpleNamespace
14 from typing import Any, List, NamedTuple, Optional, Set, Tuple
18 from sklearn.model_selection import train_test_split # type:ignore
19 from sklearn.preprocessing import MinMaxScaler # type: ignore
21 from ansi import bold, reset
24 from decorator_utils import timed
26 import parallelize as par
28 logger = logging.getLogger(__file__)
30 parser = config.add_commandline_args(
31 f"ML Model Trainer ({__file__})", "Arguments related to training an ML model"
36 help="Don't prompt the user for anything.",
39 "--ml_trainer_delete",
41 help="Delete invalid/incomplete features files in addition to warning.",
43 group = parser.add_mutually_exclusive_group()
45 "--ml_trainer_dry_run",
47 help="Do not write a new model, just report efficacy.",
50 "--ml_trainer_persist_threshold",
51 type=argparse_utils.valid_percentage,
53 help="Persist the model if the test set score is >= this threshold.",
57 class InputSpec(SimpleNamespace):
60 features_to_skip: Set[str]
61 key_value_delimiter: str
62 training_parameters: List
65 dry_run: Optional[bool]
67 persist_percentage_threshold: Optional[float]
68 delete_bad_inputs: Optional[bool]
71 def populate_from_config() -> InputSpec:
73 dry_run=config.config["ml_trainer_dry_run"],
74 quiet=config.config["ml_trainer_quiet"],
75 persist_percentage_threshold=config.config["ml_trainer_persist_threshold"],
76 delete_bad_inputs=config.config["ml_trainer_delete"],
80 class OutputSpec(NamedTuple):
81 model_filename: Optional[str]
82 model_info_filename: Optional[str]
83 scaler_filename: Optional[str]
88 class TrainingBlueprint(ABC):
92 self.X_test_scaled = None
93 self.X_train_scaled = None
94 self.file_done_count = 0
95 self.total_file_count = 0
98 def train(self, spec: InputSpec) -> OutputSpec:
104 X_, y_ = self.read_input_files()
105 num_examples = len(y_)
107 # Every example's features
110 # Every example's label
113 print("Doing random test/train split...")
114 X_train, X_test, self.y_train, self.y_test = self.test_train_split(
119 print("Scaling training data...")
120 scaler, self.X_train_scaled, self.X_test_scaled = self.scale_data(
125 print("Training model(s)...")
127 modelid_to_params = {}
128 for params in self.spec.training_parameters:
129 model = self.train_model(params, self.X_train_scaled, self.y_train)
131 modelid_to_params[model.get_id()] = str(params)
135 best_test_score = None
136 best_training_score = None
138 for model in smart_future.wait_any(models):
139 params = modelid_to_params[model.get_id()]
140 if isinstance(model, smart_future.SmartFuture):
141 model = model._resolve()
142 if model is not None:
143 training_score, test_score = self.evaluate_model(
150 score = (training_score + test_score * 20) / 21
151 if not self.spec.quiet:
153 f"{bold()}{params}{reset()}: "
154 f"Training set score={training_score:.2f}%, "
155 f"test set score={test_score:.2f}%",
158 if best_score is None or score > best_score:
160 best_test_score = test_score
161 best_training_score = training_score
164 if not self.spec.quiet:
165 print(f"New best score {best_score:.2f}% with params {params}")
167 if not self.spec.quiet:
168 executors.DefaultExecutors().shutdown()
169 msg = f"Done training; best test set score was: {best_test_score:.1f}%"
177 ) = self.maybe_persist_scaler_and_model(
186 model_filename=model_filename,
187 model_info_filename=model_info_filename,
188 scaler_filename=scaler_filename,
189 training_score=best_training_score,
190 test_score=best_test_score,
193 @par.parallelize(method=par.Method.THREAD)
194 def read_files_from_list(self, files: List[str], n: int) -> Tuple[List, List]:
201 for filename in files:
203 with open(filename, "r") as f:
204 lines = f.readlines()
206 # This example's features
210 # We expect lines in features files to be of the form:
215 (key, value) = line.split(self.spec.key_value_delimiter)
218 f"WARNING: bad line in file {filename} '{line}', skipped"
223 value = value.strip()
225 self.spec.features_to_skip is not None
226 and key in self.spec.features_to_skip
228 logger.debug(f"Skipping feature {key}")
231 value = self.normalize_feature(value)
233 if key == self.spec.label:
239 # Make sure we saw a label and the requisite number of features.
240 if len(x) == self.spec.feature_count and wrote_label:
242 self.file_done_count += 1
247 if self.spec.delete_bad_inputs:
248 msg = f"WARNING: {filename}: missing features or label; expected {self.spec.feature_count} but saw {len(x)}. DELETING."
253 msg = f"WARNING: {filename}: missing features or label; expected {self.spec.feature_count} but saw {len(x)}. Skipping."
258 def make_progress_graph(self) -> None:
259 if not self.spec.quiet:
260 from text_utils import progress_graph
262 progress_graph(self.file_done_count, self.total_file_count)
265 def read_input_files(self):
276 all_files = glob.glob(self.spec.file_glob)
277 self.total_file_count = len(all_files)
278 for n, files in enumerate(list_utils.shard(all_files, 500)):
279 file_list = list(files)
280 results.append(self.read_files_from_list(file_list, n))
282 for result in smart_future.wait_any(results, callback=self.make_progress_graph):
283 result = result._resolve()
288 if not self.spec.quiet:
289 print(" " * 80 + "\n")
292 def normalize_feature(self, value: str) -> Any:
293 if value in ("False", "None"):
295 elif value == "True":
297 elif isinstance(value, str) and "." in value:
298 ret = round(float(value) * 100.0)
303 def test_train_split(self, X, y) -> List:
304 logger.debug("Performing test/train split")
305 return train_test_split(
308 random_state=random.randrange(0, 1000),
312 self, X_train: np.ndarray, X_test: np.ndarray
313 ) -> Tuple[Any, np.ndarray, np.ndarray]:
314 logger.debug("Scaling data")
315 scaler = MinMaxScaler()
317 return (scaler, scaler.transform(X_train), scaler.transform(X_test))
319 # Note: children should implement. Consider using @parallelize.
322 self, parameters, X_train_scaled: np.ndarray, y_train: np.ndarray
329 X_train_scaled: np.ndarray,
331 X_test_scaled: np.ndarray,
333 ) -> Tuple[np.float64, np.float64]:
334 logger.debug("Evaluating the model")
335 training_score = model.score(X_train_scaled, y_train) * 100.0
336 test_score = model.score(X_test_scaled, y_test) * 100.0
338 f"Model evaluation results: test_score={test_score:.5f}, "
339 f"train_score={training_score:.5f}"
341 return (training_score, test_score)
343 def maybe_persist_scaler_and_model(
345 training_score: np.float64,
346 test_score: np.float64,
351 ) -> Tuple[Optional[str], Optional[str], Optional[str]]:
352 if not self.spec.dry_run:
353 import datetime_utils
357 now: datetime.datetime = datetime_utils.now_pacific()
358 info = f"""Timestamp: {datetime_utils.datetime_to_string(now)}
359 Model params: {params}
360 Training examples: {num_examples}
361 Training set score: {training_score:.2f}%
362 Testing set score: {test_score:.2f}%"""
365 self.spec.persist_percentage_threshold is not None
366 and test_score > self.spec.persist_percentage_threshold
369 and input_utils.yn_response("Write the model? [y,n]: ") == "y"
371 scaler_filename = f"{self.spec.basename}_scaler.sav"
372 with open(scaler_filename, "wb") as f:
373 pickle.dump(scaler, f)
374 msg = f"Wrote {scaler_filename}"
377 model_filename = f"{self.spec.basename}_model.sav"
378 with open(model_filename, "wb") as f:
379 pickle.dump(model, f)
380 msg = f"Wrote {model_filename}"
383 model_info_filename = f"{self.spec.basename}_model_info.txt"
384 with open(model_info_filename, "w") as f:
386 msg = f"Wrote {model_info_filename}:"
389 print(string_utils.indent(info, 2))
391 return (scaler_filename, model_filename, model_info_filename)
392 return (None, None, None)