Upload Bilma

config.json

@@ -4,13 +4,14 @@
   ],
   "auto_map": {
     "AutoConfig": "configuration_bilma.BilmaConfig",
-    "TFAutoModelForMaskedLM": "modeling_bilma.Bilma"
+    "TFAutoModel": "modeling_bilma.Bilma"
   },
   "drop_rate": 0.1,
   "embedding_dim": 512,
-  "model_type": "bilma",
+  "model_type": "bert",
   "num_attention_heads": 4,
   "num_encoders": 2,
+  "seq_max_length": 280,
   "transformers_version": "4.30.2",
   "vocab_size": 28949,
   "weights": "spanish"

configuration_bilma.py

@@ -1,14 +1,14 @@
 from transformers import PretrainedConfig
 
 class BilmaConfig(PretrainedConfig):
-    model_type = "bilma"
+    model_type = "bert"
 
     def __init__(
         self,
         weights="spanish",
         num_attention_heads: int = 4,
         num_encoders: int = 2,
-        max_length: int = 280,
+        seq_max_length: int = 280,
         embedding_dim: int = 512,
         vocab_size: int = 28949,
         drop_rate: float = 0.1,        
@@ -20,7 +20,7 @@ class BilmaConfig(PretrainedConfig):
             self.weights = weights
             self.num_attention_heads = 4
             self.num_encoders = 2
-            self.max_length = 280
+            self.seq_max_length = 280
             self.embedding_dim = 512
             self.vocab_size = 28949
             self.drop_rate = 0.1
@@ -30,7 +30,7 @@ class BilmaConfig(PretrainedConfig):
         self.weights = weights
         self.num_attention_heads = num_attention_heads
         self.num_encoders = num_encoders
-        self.max_length = max_length
+        self.seq_max_length = seq_max_length
         self.embedding_dim = embedding_dim
         self.vocab_size = vocab_size
         self.drop_rate = drop_rate        

modeling_bilma.py

@@ -4,6 +4,8 @@ from tensorflow.keras.layers import Layer, Dense, concatenate, Input, add, Dropo
 import tensorflow as tf
 import numpy as np
 
+from typing import Dict
+
 import re
 import unicodedata
 
@@ -41,13 +43,18 @@ class Bilma(TFPreTrainedModel):
         #else:
         self.model = bilma(num_enc=config.num_encoders,
                            embed_dim=config.embedding_dim, 
-                           max_length=config.max_length,
+                           max_length=config.seq_max_length,
                            num_heads=config.num_attention_heads,
                            ff_dim=config.embedding_dim,
                            vocab_size=config.vocab_size,
                            rate=config.drop_rate)
-        
+        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]}
+        return dummies
+    
     def call(self, tensor):
         return self.model(tensor)
     
@@ -459,7 +466,7 @@ def accuracy_function(ignore_id=0):
 
 def bilma(num_enc=6, embed_dim=300, max_length=50, num_heads=6, ff_dim=512, vocab_size=9739, rate=0.1):
     capt_inputs_ids = Input(shape=(max_length, ), name='capt_input')
-    capt_embedding = Embedding(vocab_size, embed_dim, mask_zero=False, name="embedding")
+    capt_embedding = Embedding(vocab_size, embed_dim, mask_zero=False)
     capt_inputs = capt_embedding(capt_inputs_ids)
     
     enc = Encoder(num_enc, embed_dim, max_length, num_heads, ff_dim, rate=rate)

tf_model.h5

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cd8a5be22346774126afb0615bc5c9250d345ca5a451e9de30f847eb19b3536b
+oid sha256:b7113392a0c53e8294014d1c64e57f7a25553b037fb880370f9f68c76261d7dc
 size 156561684