Upload TFBilma

config.json

@@ -1,17 +1,17 @@
 {
   "_name_or_path": "w",
   "architectures": [
-    "lma"
+    "Bilma"
   ],
   "auto_map": {
     "AutoConfig": "configuration_bilma.BilmaConfig",
-    "TFAutoModel": "modeling_bilma.Bilma"
+    "TFAutoModel": "modeling_bilma.TFBilma"
   },
-  "drop_rate": 0.1,
-  "embedding_dim": 512,
-  "model_type": "bilma",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 512,
+  "model_type": "TFBilma",
   "num_attention_heads": 4,
-  "num_encoders": 2,
+  "num_hidden_layers": 2,
   "seq_max_length": 280,
   "transformers_version": "4.30.2",
   "vocab_size": 28949,

configuration_bilma.py

@@ -1,17 +1,17 @@
 from transformers import PretrainedConfig
 
 class BilmaConfig(PretrainedConfig):
-    model_type = "bilma"
+    model_type = "TFBilma"
 
     def __init__(
         self,
         weights="spanish",
         num_attention_heads: int = 4,
-        num_encoders: int = 2,
+        num_hidden_layers: int = 2,
         seq_max_length: int = 280,
-        embedding_dim: int = 512,
+        hidden_size: int = 512,
         vocab_size: int = 28949,
-        drop_rate: float = 0.1,        
+        hidden_dropout_prob: float = 0.1,        
         **kwargs,
     ):
         if weights not in ["spanish", ""]:
@@ -19,19 +19,19 @@ class BilmaConfig(PretrainedConfig):
         if weights == "spanish":
             self.weights = weights
             self.num_attention_heads = 4
-            self.num_encoders = 2
+            self.num_hidden_layers = 2
             self.seq_max_length = 280
-            self.embedding_dim = 512
+            self.hidden_size = 512
             self.vocab_size = 28949
-            self.drop_rate = 0.1
+            self.hidden_dropout_prob = 0.1
             super().__init__(**kwargs)
             return
 
         self.weights = weights
         self.num_attention_heads = num_attention_heads
-        self.num_encoders = num_encoders
+        self.num_hidden_layers = num_hidden_layers
         self.seq_max_length = seq_max_length
-        self.embedding_dim = embedding_dim
+        self.hidden_size = hidden_size
         self.vocab_size = vocab_size
-        self.drop_rate = drop_rate        
+        self.hidden_dropout_prob = hidden_dropout_prob        
         super().__init__(**kwargs)

modeling_bilma.py

@@ -30,10 +30,10 @@ SYMBOLS = set(";:,.@\\-\"/" + SYMBOLS_)
 # ------------------
 
 
-class Bilma(TFPreTrainedModel):
+class TFBilma(TFPreTrainedModel):
     config_class = BilmaConfig
     main_input_name = "capt_input"
-    base_model_prefix = "bilma"
+    #base_model_prefix = "bilma"
 
     def __init__(self, config):
         self.seq_max_length = config.seq_max_length
@@ -43,13 +43,13 @@ class Bilma(TFPreTrainedModel):
         #    model_file = str((my_resources / "bilma_dataset_small_epoch-1_part-60.h5").joinpath())
         #    self.model = bm.load(model_file)
         #else:
-        self.model = bilma(num_enc=config.num_encoders,
-                           embed_dim=config.embedding_dim, 
+        self.model = bilma(num_enc=config.num_hidden_layers,
+                           embed_dim=config.hidden_size, 
                            max_length=config.seq_max_length,
                            num_heads=config.num_attention_heads,
-                           ff_dim=config.embedding_dim,
+                           ff_dim=config.hidden_size,
                            vocab_size=config.vocab_size,
-                           rate=config.drop_rate)
+                           rate=config.hidden_dropout_prob)
             
     @property
     def dummy_inputs(self) -> Dict[str, tf.Tensor]:

tf_model.h5

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad2516c4780938848a1837ae0e3ba8c2ca643fbcdee4e7ac1b7b9fec90b529d3
+oid sha256:e5e4fc365e4ff80eba3e9363ebe96b53093ae73fa5a8e48b8e31b43da3a2c6e3
 size 156562964