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__})",
32 "Arguments related to training an ML model"
37 help="Don't prompt the user for anything."
40 "--ml_trainer_delete",
42 help="Delete invalid/incomplete features files in addition to warning."
44 group = parser.add_mutually_exclusive_group()
46 "--ml_trainer_dry_run",
48 help="Do not write a new model, just report efficacy.",
51 "--ml_trainer_persist_threshold",
52 type=argparse_utils.valid_percentage,
54 help="Persist the model if the test set score is >= this threshold.",
58 class InputSpec(SimpleNamespace):
61 features_to_skip: Set[str]
62 key_value_delimiter: str
63 training_parameters: List
66 dry_run: Optional[bool]
68 persist_percentage_threshold: Optional[float]
69 delete_bad_inputs: Optional[bool]
72 def populate_from_config() -> InputSpec:
74 dry_run = config.config["ml_trainer_dry_run"],
75 quiet = config.config["ml_trainer_quiet"],
76 persist_percentage_threshold = config.config["ml_trainer_persist_threshold"],
77 delete_bad_inputs = config.config["ml_trainer_delete"],
81 class OutputSpec(NamedTuple):
82 model_filename: Optional[str]
83 model_info_filename: Optional[str]
84 scaler_filename: Optional[str]
89 class TrainingBlueprint(ABC):
93 self.X_test_scaled = None
94 self.X_train_scaled = None
95 self.file_done_count = 0
96 self.total_file_count = 0
99 def train(self, spec: InputSpec) -> OutputSpec:
105 X_, y_ = self.read_input_files()
106 num_examples = len(y_)
108 # Every example's features
111 # Every example's label
114 print("Doing random test/train split...")
115 X_train, X_test, self.y_train, self.y_test = self.test_train_split(
120 print("Scaling training data...")
121 scaler, self.X_train_scaled, self.X_test_scaled = self.scale_data(
126 print("Training model(s)...")
128 modelid_to_params = {}
129 for params in self.spec.training_parameters:
130 model = self.train_model(
136 modelid_to_params[model.get_id()] = str(params)
140 best_test_score = None
141 best_training_score = None
143 for model in smart_future.wait_any(models):
144 params = modelid_to_params[model.get_id()]
145 if isinstance(model, smart_future.SmartFuture):
146 model = model._resolve()
147 if model is not None:
148 training_score, test_score = self.evaluate_model(
155 score = (training_score + test_score * 20) / 21
156 if not self.spec.quiet:
158 f"{bold()}{params}{reset()}: "
159 f"Training set score={training_score:.2f}%, "
160 f"test set score={test_score:.2f}%",
163 if best_score is None or score > best_score:
165 best_test_score = test_score
166 best_training_score = training_score
169 if not self.spec.quiet:
171 f"New best score {best_score:.2f}% with params {params}"
174 if not self.spec.quiet:
175 executors.DefaultExecutors().shutdown()
176 msg = f"Done training; best test set score was: {best_test_score:.1f}%"
180 scaler_filename, model_filename, model_info_filename = (
181 self.maybe_persist_scaler_and_model(
191 model_filename = model_filename,
192 model_info_filename = model_info_filename,
193 scaler_filename = scaler_filename,
194 training_score = best_training_score,
195 test_score = best_test_score,
198 @par.parallelize(method=par.Method.THREAD)
199 def read_files_from_list(
203 ) -> Tuple[List, List]:
210 for filename in files:
212 with open(filename, "r") as f:
213 lines = f.readlines()
215 # This example's features
219 # We expect lines in features files to be of the form:
224 (key, value) = line.split(self.spec.key_value_delimiter)
226 logger.debug(f"WARNING: bad line in file {filename} '{line}', skipped")
230 value = value.strip()
231 if (self.spec.features_to_skip is not None
232 and key in self.spec.features_to_skip):
233 logger.debug(f"Skipping feature {key}")
236 value = self.normalize_feature(value)
238 if key == self.spec.label:
244 # Make sure we saw a label and the requisite number of features.
245 if len(x) == self.spec.feature_count and wrote_label:
247 self.file_done_count += 1
252 if self.spec.delete_bad_inputs:
253 msg = f"WARNING: {filename}: missing features or label; expected {self.spec.feature_count} but saw {len(x)}. DELETING."
258 msg = f"WARNING: {filename}: missing features or label; expected {self.spec.feature_count} but saw {len(x)}. Skipping."
263 def make_progress_graph(self) -> None:
264 if not self.spec.quiet:
265 from text_utils import progress_graph
267 self.file_done_count,
268 self.total_file_count
272 def read_input_files(self):
283 all_files = glob.glob(self.spec.file_glob)
284 self.total_file_count = len(all_files)
285 for n, files in enumerate(list_utils.shard(all_files, 500)):
286 file_list = list(files)
287 results.append(self.read_files_from_list(file_list, n))
289 for result in smart_future.wait_any(results, callback=self.make_progress_graph):
290 result = result._resolve()
295 if not self.spec.quiet:
296 print(" " * 80 + "\n")
299 def normalize_feature(self, value: str) -> Any:
300 if value in ("False", "None"):
302 elif value == "True":
304 elif isinstance(value, str) and "." in value:
305 ret = round(float(value) * 100.0)
310 def test_train_split(self, X, y) -> List:
311 logger.debug("Performing test/train split")
312 return train_test_split(
315 random_state=random.randrange(0, 1000),
320 X_test: np.ndarray) -> Tuple[Any, np.ndarray, np.ndarray]:
321 logger.debug("Scaling data")
322 scaler = MinMaxScaler()
324 return (scaler, scaler.transform(X_train), scaler.transform(X_test))
326 # Note: children should implement. Consider using @parallelize.
328 def train_model(self,
330 X_train_scaled: np.ndarray,
331 y_train: np.ndarray) -> Any:
337 X_train_scaled: np.ndarray,
339 X_test_scaled: np.ndarray,
340 y_test: np.ndarray) -> Tuple[np.float64, np.float64]:
341 logger.debug("Evaluating the model")
342 training_score = model.score(X_train_scaled, y_train) * 100.0
343 test_score = model.score(X_test_scaled, y_test) * 100.0
345 f"Model evaluation results: test_score={test_score:.5f}, "
346 f"train_score={training_score:.5f}"
348 return (training_score, test_score)
350 def maybe_persist_scaler_and_model(
352 training_score: np.float64,
353 test_score: np.float64,
357 model: Any) -> Tuple[Optional[str], Optional[str], Optional[str]]:
358 if not self.spec.dry_run:
359 import datetime_utils
363 now: datetime.datetime = datetime_utils.now_pacific()
364 info = f"""Timestamp: {datetime_utils.datetime_to_string(now)}
365 Model params: {params}
366 Training examples: {num_examples}
367 Training set score: {training_score:.2f}%
368 Testing set score: {test_score:.2f}%"""
371 (self.spec.persist_percentage_threshold is not None and
372 test_score > self.spec.persist_percentage_threshold)
375 and input_utils.yn_response("Write the model? [y,n]: ") == "y")
377 scaler_filename = f"{self.spec.basename}_scaler.sav"
378 with open(scaler_filename, "wb") as f:
379 pickle.dump(scaler, f)
380 msg = f"Wrote {scaler_filename}"
383 model_filename = f"{self.spec.basename}_model.sav"
384 with open(model_filename, "wb") as f:
385 pickle.dump(model, f)
386 msg = f"Wrote {model_filename}"
389 model_info_filename = f"{self.spec.basename}_model_info.txt"
390 with open(model_info_filename, "w") as f:
392 msg = f"Wrote {model_info_filename}:"
395 print(string_utils.indent(info, 2))
397 return (scaler_filename, model_filename, model_info_filename)
398 return (None, None, None)