Upload Bilma

config.json

@@ -1,6 +1,6 @@
 {
   "architectures": [
-    "TFAutoModelForMaskedLM"
+    "lma"
   ],
   "auto_map": {
     "AutoConfig": "configuration_bilma.BilmaConfig",

modeling_bilma.py

@@ -9,7 +9,7 @@ from typing import Dict
 import re
 import unicodedata
 
-from .configuration_bilma import BilmaConfig
+from configuration_bilma import BilmaConfig
 
 # copied from preprocessing.py
 BLANK = ' '
@@ -32,9 +32,10 @@ SYMBOLS = set(";:,.@\\-\"/" + SYMBOLS_)
 
 class Bilma(TFPreTrainedModel):
     config_class = BilmaConfig
-    main_input_name = "input_ids"
+    main_input_name = "capt_input"
 
     def __init__(self, config):
+        self.seq_max_length = config.seq_max_length
         super().__init__(config)
         #if config.weights == "spanish":
         #    my_resources = importlib_resources.files("hf_bilma") 
@@ -48,15 +49,33 @@ class Bilma(TFPreTrainedModel):
                            ff_dim=config.embedding_dim,
                            vocab_size=config.vocab_size,
                            rate=config.drop_rate)
-        self.call(np.zeros((1, config.seq_max_length)))
-
+        
+        #self.call(np.zeros((1, config.seq_max_length)))
+    
     @property
     def dummy_inputs(self) -> Dict[str, tf.Tensor]:
-        dummies = {"capt_input":self.model.inputs[0]}
+    
+        dummies = {}
+        for key, spec in self.input_signature.items():
+            dummy_shape = [dim if dim is not None else 2 for dim in spec.shape]
+            if spec.shape[0] is None:
+                dummy_shape[0] = 1
+            dummies[key] = tf.ones(shape=dummy_shape, dtype=spec.dtype)
+        
+        
         return dummies
     
-    def call(self, tensor):
-        return self.model(tensor)
+    @property
+    def input_signature(self) -> Dict[str, tf.TensorSpec]:
+        sig = {}
+        sig["capt_input"] = tf.TensorSpec([None, self.seq_max_length], tf.int32, name="capt_input")
+        return sig
+    
+    
+    def call(self, capt_input):        
+        #if isinstance(tensor, dict) and len(tensor) == 0:
+        #    return self.model(self.dummy_inputs)
+        return self.model(capt_input)
     
 
 # 

tf_model.h5

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:42a62ff751c90b4e9cdf2f33f2744732dcb79a061db2dda2a344477dec95a6ab
+oid sha256:f932984cd1b53af396b362f3b882736143583d47f4c86f356e7ae359b6bcba7c
 size 156561684