seed + new tokens for poleval2018

fastai_contrib/utils.py

@@ -79,7 +79,7 @@ def get_sentencepiece(cache_dir:PathOrStr, load_text, pre_rules: ListRules=None,
     pre_rules = pre_rules if pre_rules is not None else defaults.text_pre_rules
     post_rules = post_rules if post_rules is not None else defaults.text_post_rules
 
-    special_cases = defaults.text_spec_tok
+    special_cases = defaults.text_spec_tok + ['xxlink', 'xxuser', 'xxnumber', 'xxemoji', 'yyemoji']
     if not os.path.isfile(cache_dir / 'spm.model') or not os.path.isfile(cache_dir / f'itos.pkl'):
         # load the text from the train tokens file
         text = load_text()

ulmfit/pretrain_lm.py

@@ -159,7 +159,13 @@ class LMHyperParams:
         json_save(self.model_dir/'info.json', vals)
         print("Saving info", self.model_dir / 'info.json')
 
-    def train_lm(self, num_epochs=20, data_lm=None, bs=70, true_wd=False, drop_mult=0.0, lr=5e-3, label_smoothing_eps=0.0):
+    def train_lm(self, num_epochs=20, data_lm=None, bs=70, true_wd=False, drop_mult=0.0, lr=5e-3, label_smoothing_eps=0.0, seed=None):
+        if seed is not None:
+            print(f"Setting seed to {seed}")
+            torch.manual_seed(seed)
+            torch.backends.cudnn.deterministic = True
+            torch.backends.cudnn.benchmark = False
+            np.random.seed(seed)
         self.model_dir.mkdir(exist_ok=True, parents=True)
         data_lm = self.load_wiki_data(bs=bs) if data_lm is None else data_lm
         learn = self.create_lm_learner(data_lm, drop_mult=drop_mult, label_smoothing_eps=label_smoothing_eps)