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
25 import parallelize as par
27 logger = logging.getLogger(__file__)
29 parser = config.add_commandline_args(
30 f"ML Model Trainer ({__file__})",
31 "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(
135 modelid_to_params[model.get_id()] = str(params)
139 best_test_score = None
140 best_training_score = None
142 for model in smart_future.wait_any(models):
143 params = modelid_to_params[model.get_id()]
144 if isinstance(model, smart_future.SmartFuture):
145 model = model._resolve()
146 if model is not None:
147 training_score, test_score = self.evaluate_model(
154 score = (training_score + test_score * 20) / 21
155 if not self.spec.quiet:
157 f"{bold()}{params}{reset()}: "
158 f"Training set score={training_score:.2f}%, "
159 f"test set score={test_score:.2f}%",
162 if best_score is None or score > best_score:
164 best_test_score = test_score
165 best_training_score = training_score
168 if not self.spec.quiet:
170 f"New best score {best_score:.2f}% with params {params}"
173 if not self.spec.quiet:
174 msg = f"Done training; best test set score was: {best_test_score:.1f}%"
177 scaler_filename, model_filename, model_info_filename = (
178 self.maybe_persist_scaler_and_model(
188 model_filename = model_filename,
189 model_info_filename = model_info_filename,
190 scaler_filename = scaler_filename,
191 training_score = best_training_score,
192 test_score = best_test_score,
195 @par.parallelize(method=par.Method.THREAD)
196 def read_files_from_list(
200 ) -> Tuple[List, List]:
207 for filename in files:
209 with open(filename, "r") as f:
210 lines = f.readlines()
212 # This example's features
216 # We expect lines in features files to be of the form:
221 (key, value) = line.split(self.spec.key_value_delimiter)
223 logger.debug(f"WARNING: bad line in file {filename} '{line}', skipped")
227 value = value.strip()
228 if (self.spec.features_to_skip is not None
229 and key in self.spec.features_to_skip):
230 logger.debug(f"Skipping feature {key}")
233 value = self.normalize_feature(value)
235 if key == self.spec.label:
241 # Make sure we saw a label and the requisite number of features.
242 if len(x) == self.spec.feature_count and wrote_label:
244 self.file_done_count += 1
249 if self.spec.delete_bad_inputs:
250 msg = f"WARNING: {filename}: missing features or label; expected {self.spec.feature_count} but saw {len(x)}. DELETING."
255 msg = f"WARNING: {filename}: missing features or label; expected {self.spec.feature_count} but saw {len(x)}. Skipping."
260 def make_progress_graph(self) -> None:
261 if not self.spec.quiet:
262 from text_utils import progress_graph
264 self.file_done_count,
265 self.total_file_count
269 def read_input_files(self):
280 all_files = glob.glob(self.spec.file_glob)
281 self.total_file_count = len(all_files)
282 for n, files in enumerate(list_utils.shard(all_files, 500)):
283 file_list = list(files)
284 results.append(self.read_files_from_list(file_list, n))
286 for result in smart_future.wait_any(results, callback=self.make_progress_graph):
287 result = result._resolve()
292 if not self.spec.quiet:
293 print(" " * 80 + "\n")
296 def normalize_feature(self, value: str) -> Any:
297 if value in ("False", "None"):
299 elif value == "True":
301 elif isinstance(value, str) and "." in value:
302 ret = round(float(value) * 100.0)
307 def test_train_split(self, X, y) -> List:
308 logger.debug("Performing test/train split")
309 return train_test_split(
312 random_state=random.randrange(0, 1000),
317 X_test: np.ndarray) -> Tuple[Any, np.ndarray, np.ndarray]:
318 logger.debug("Scaling data")
319 scaler = MinMaxScaler()
321 return (scaler, scaler.transform(X_train), scaler.transform(X_test))
323 # Note: children should implement. Consider using @parallelize.
325 def train_model(self,
327 X_train_scaled: np.ndarray,
328 y_train: np.ndarray) -> Any:
334 X_train_scaled: np.ndarray,
336 X_test_scaled: np.ndarray,
337 y_test: np.ndarray) -> Tuple[np.float64, np.float64]:
338 logger.debug("Evaluating the model")
339 training_score = model.score(X_train_scaled, y_train) * 100.0
340 test_score = model.score(X_test_scaled, y_test) * 100.0
342 f"Model evaluation results: test_score={test_score:.5f}, "
343 f"train_score={training_score:.5f}"
345 return (training_score, test_score)
347 def maybe_persist_scaler_and_model(
349 training_score: np.float64,
350 test_score: np.float64,
354 model: Any) -> Tuple[Optional[str], Optional[str], Optional[str]]:
355 if not self.spec.dry_run:
356 import datetime_utils
360 now: datetime.datetime = datetime_utils.now_pacific()
361 info = f"""Timestamp: {datetime_utils.datetime_to_string(now)}
362 Model params: {params}
363 Training examples: {num_examples}
364 Training set score: {training_score:.2f}%
365 Testing set score: {test_score:.2f}%"""
368 (self.spec.persist_percentage_threshold is not None and
369 test_score > self.spec.persist_percentage_threshold)
372 and input_utils.yn_response("Write the model? [y,n]: ") == "y")
374 scaler_filename = f"{self.spec.basename}_scaler.sav"
375 with open(scaler_filename, "wb") as f:
376 pickle.dump(scaler, f)
377 msg = f"Wrote {scaler_filename}"
380 model_filename = f"{self.spec.basename}_model.sav"
381 with open(model_filename, "wb") as f:
382 pickle.dump(model, f)
383 msg = f"Wrote {model_filename}"
386 model_info_filename = f"{self.spec.basename}_model_info.txt"
387 with open(model_info_filename, "w") as f:
389 msg = f"Wrote {model_info_filename}:"
392 print(string_utils.indent(info, 2))
394 return (scaler_filename, model_filename, model_info_filename)
395 return (None, None, None)