npedrazzini commited on
Commit
a41f6a7
·
verified ·
1 Parent(s): 50b8766

Update continuous_time_embedding.py

Browse files
Files changed (1) hide show
  1. continuous_time_embedding.py +8 -14
continuous_time_embedding.py CHANGED
@@ -88,13 +88,13 @@ class ContinuousTimeBertForMLM(nn.Module):
88
 
89
  def save_pretrained(self, save_dir):
90
  os.makedirs(save_dir, exist_ok=True)
91
- self.model.save_pretrained(os.path.join(save_dir, "full_model"))
92
  torch.save(self.time_embed.state_dict(), os.path.join(save_dir, "time_embed.pt"))
93
  torch.save(self.post_inject_norm.state_dict(), os.path.join(save_dir, "post_inject_norm.pt"))
94
-
95
  @classmethod
96
  def load_pretrained(cls, save_dir, hidden_size=768, **kwargs):
97
- model = AutoModelForMaskedLM.from_pretrained(os.path.join(save_dir, "full_model"))
98
  obj = cls(model, hidden_size, **kwargs)
99
  obj.time_embed.load_state_dict(torch.load(os.path.join(save_dir, "time_embed.pt"), map_location="cpu"))
100
  obj.post_inject_norm.load_state_dict(torch.load(os.path.join(save_dir, "post_inject_norm.pt"), map_location="cpu"))
@@ -107,16 +107,11 @@ def load_continuous_time_model(repo_id_or_path, device=None, **kwargs):
107
  else:
108
  local_dir = snapshot_download(repo_id_or_path)
109
 
110
- tokenizer = None
111
- for candidate in (os.path.join(local_dir, "full_model"), local_dir):
112
- try:
113
- candidate_tokenizer = AutoTokenizer.from_pretrained(candidate)
114
- except Exception:
115
- continue
116
- if len(candidate_tokenizer) < 1000:
117
- continue
118
- tokenizer = candidate_tokenizer
119
- break
120
 
121
  if tokenizer is None:
122
  print(f"[load_continuous_time_model] No valid tokenizer found in {repo_id_or_path}, "
@@ -124,7 +119,6 @@ def load_continuous_time_model(repo_id_or_path, device=None, **kwargs):
124
  tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
125
 
126
  model = ContinuousTimeBertForMLM.load_pretrained(local_dir, **kwargs)
127
-
128
  device = device or ("cuda" if torch.cuda.is_available() else "cpu")
129
  model.to(device).eval()
130
  return tokenizer, model
 
88
 
89
  def save_pretrained(self, save_dir):
90
  os.makedirs(save_dir, exist_ok=True)
91
+ self.model.save_pretrained(save_dir)
92
  torch.save(self.time_embed.state_dict(), os.path.join(save_dir, "time_embed.pt"))
93
  torch.save(self.post_inject_norm.state_dict(), os.path.join(save_dir, "post_inject_norm.pt"))
94
+
95
  @classmethod
96
  def load_pretrained(cls, save_dir, hidden_size=768, **kwargs):
97
+ model = AutoModelForMaskedLM.from_pretrained(save_dir)
98
  obj = cls(model, hidden_size, **kwargs)
99
  obj.time_embed.load_state_dict(torch.load(os.path.join(save_dir, "time_embed.pt"), map_location="cpu"))
100
  obj.post_inject_norm.load_state_dict(torch.load(os.path.join(save_dir, "post_inject_norm.pt"), map_location="cpu"))
 
107
  else:
108
  local_dir = snapshot_download(repo_id_or_path)
109
 
110
+ try:
111
+ candidate_tokenizer = AutoTokenizer.from_pretrained(local_dir)
112
+ tokenizer = candidate_tokenizer if len(candidate_tokenizer) >= 1000 else None
113
+ except Exception:
114
+ tokenizer = None
 
 
 
 
 
115
 
116
  if tokenizer is None:
117
  print(f"[load_continuous_time_model] No valid tokenizer found in {repo_id_or_path}, "
 
119
  tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
120
 
121
  model = ContinuousTimeBertForMLM.load_pretrained(local_dir, **kwargs)
 
122
  device = device or ("cuda" if torch.cuda.is_available() else "cpu")
123
  model.to(device).eval()
124
  return tokenizer, model