GH-212: Add file handler to logger in training.

flairNLP · Nov 28, 2018 · ec6b77c · ec6b77c
1 parent cf17507
commit ec6b77c
Show file tree

Hide file tree

Showing 3 changed files with 20 additions and 12 deletions.
diff --git a/flair/__init__.py b/flair/__init__.py
@@ -5,7 +5,6 @@
 
 import sys
 import logging
-import warnings
 
 logger = logging.getLogger(__name__)
 

diff --git a/flair/trainers/trainer.py b/flair/trainers/trainer.py
@@ -120,7 +120,13 @@ def train(self,
               **kwargs
               ) -> dict:
 
-        log_line()
+        fh = logging.FileHandler(base_path / 'training.log')
+        fh.setLevel(logging.INFO)
+        formatter = logging.Formatter('%(asctime)-15s %(message)s')
+        fh.setFormatter(formatter)
+        log.addHandler(fh)
+
+        log_line(log)
         log.info(f'Evaluation method: {evaluation_metric.name}')
 
         # cast string to Path
@@ -170,7 +176,7 @@ def train(self,
             previous_learning_rate = learning_rate
 
             for epoch in range(0 + self.epoch, max_epochs + self.epoch):
-                log_line()
+                log_line(log)
 
                 # bad_epochs = scheduler.num_bad_epochs
                 bad_epochs = 0
@@ -187,9 +193,9 @@ def train(self,
 
                 # stop training if learning rate becomes too small
                 if learning_rate < 0.0001:
-                    log_line()
+                    log_line(log)
                     log.info('learning rate too small - quitting training!')
-                    log_line()
+                    log_line(log)
                     break
 
                 if not test_mode:
@@ -227,7 +233,7 @@ def train(self,
 
                 self.model.eval()
 
-                log_line()
+                log_line(log)
                 log.info(f'EPOCH {epoch + 1}: lr {learning_rate:.4f} - bad epochs {bad_epochs}')
 
                 dev_metric = None
@@ -288,7 +294,7 @@ def train(self,
                 self.model.save(base_path / 'final-model.pt')
 
         except KeyboardInterrupt:
-            log_line()
+            log_line(log)
             log.info('Exiting from training early.')
             if not param_selection_mode:
                 log.info('Saving model ...')
@@ -307,7 +313,7 @@ def final_test(self,
                    evaluation_metric: EvaluationMetric,
                    mini_batch_size: int):
 
-        log_line()
+        log_line(log)
         log.info('Testing using best model ...')
 
         self.model.eval()
@@ -329,12 +335,12 @@ def final_test(self,
                      f'{test_metric.precision(class_name):.4f} - recall: {test_metric.recall(class_name):.4f} - '
                      f'accuracy: {test_metric.accuracy(class_name):.4f} - f1-score: '
                      f'{test_metric.f_score(class_name):.4f}')
-        log_line()
+        log_line(log)
 
         # if we are training over multiple datasets, do evaluation for each
         if type(self.corpus) is MultiCorpus:
             for subcorpus in self.corpus.corpora:
-                log_line()
+                log_line(log)
                 self._calculate_evaluation_results_for(subcorpus.name, subcorpus.test, evaluation_metric,
                                                        embeddings_in_memory, mini_batch_size, base_path / 'test.tsv')
 

diff --git a/flair/training_utils.py b/flair/training_utils.py
@@ -232,5 +232,8 @@ def convert_labels_to_one_hot(label_list: List[List[str]], label_dict: Dictionar
     return [[1 if l in labels else 0 for l in label_dict.get_items()] for labels in label_list]
 
 
-def log_line():
-    log.info('-' * 100)
+def log_line(logger = None):
+    if logger is not None:
+        logger.info('-' * 100)
+    else:
+        log.info('-' * 100)