Upload EncT5ForSequenceClassification

README.md

@@ -1,14 +1,13 @@
 ---
-language: 
+language:
 - en
 - fr
 - ro
 - de
+license: apache-2.0
+library_name: transformers
 datasets:
 - c4
-library_name: transformers
-
-license: apache-2.0
 ---
 
 # Model Card for EncT5

config.json

@@ -59,6 +59,7 @@
       "prefix": "translate English to Romanian: "
     }
   },
+  "tie_word_embeddings": false,
   "torch_dtype": "float32",
   "transformers_version": "4.37.1",
   "use_cache": true,

configuration_enct5.py

@@ -131,3 +131,6 @@ class EncT5Config(PretrainedConfig):
             is_encoder_decoder=is_encoder_decoder,
             **kwargs,
         )
+
+        # Override the default behavior to tie word embeddings.
+        self.tie_word_embeddings = False

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:882a6b27f24f427c3deb66b85eef1d2d2c04ef441030948945333e02ad31c334
-size 476297904
+oid sha256:1e9cc0194fa5bfc256b2e2d47affe664f166cdaf29430947220e1606223691cc
+size 476301088

modeling_enct5.py

@@ -93,7 +93,6 @@ class EncT5ForSequenceClassification(EncT5PreTrainedModel):
 
         # Initiate decoder embedding from scratch and define the corresponding latent vector vocabulary size.
         self.decoder_embeddings = nn.Embedding(config.decoder_vocab_size, config.d_model)
-        self.transformer.get_decoder().set_input_embeddings(self.decoder_embeddings)
 
         # Initiate decoder projection head from scratch.
         if config.problem_type == "multi_label_classification":
@@ -115,6 +114,7 @@ class EncT5ForSequenceClassification(EncT5PreTrainedModel):
         Prepares the model for fine-tuning by re-initializing the necessary weights for fine-tuning. This step should be
         performed after loading the pre-trained T5 model but before fine-tuning.
         """
+        self.transformer.get_decoder().set_input_embeddings(self.decoder_embeddings)
         self.transformer.get_decoder().apply(self._init_weights)
         self._init_weights(self.classification_head)