Upload model

config.json

@@ -1,10 +1,10 @@
 {
   "_commit_hash": null,
   "architectures": [
-    "LongitudinalPromptVariableCXREncoderDecoderModel"
+    "LongitudinalPromptMultiCXREncoderDecoderModel"
   ],
   "auto_map": {
-    "AutoModel": "modelling_longitudinal.LongitudinalPromptVariableCXREncoderDecoderModel"
+    "AutoModel": "modelling_longitudinal.LongitudinalPromptMultiCXREncoderDecoderModel"
   },
   "decoder": {
     "_name_or_path": "",
@@ -78,7 +78,7 @@
     "top_p": 1.0,
     "torch_dtype": null,
     "torchscript": false,
-    "transformers_version": "4.28.1",
+    "transformers_version": "4.31.0",
     "type_vocab_size": 2,
     "typical_p": 1.0,
     "use_bfloat16": false,
@@ -2243,7 +2243,7 @@
     "top_p": 1.0,
     "torch_dtype": "float32",
     "torchscript": false,
-    "transformers_version": "4.28.1",
+    "transformers_version": "4.31.0",
     "typical_p": 1.0,
     "use_bfloat16": false
   },

generation_config.json

@@ -1,5 +1,5 @@
 {
   "_from_model_config": true,
   "pad_token_id": 0,
-  "transformers_version": "4.28.1"
+  "transformers_version": "4.31.0"
 }

modelling_longitudinal.py

@@ -47,7 +47,7 @@ class CvtProjectionHead(torch.nn.Module):
         return x
 
 
-class VariableCvtWithProjectionHead(transformers.CvtPreTrainedModel):
+class MultiCvtWithProjectionHead(transformers.CvtPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
 
@@ -93,7 +93,7 @@ class VariableCvtWithProjectionHead(transformers.CvtPreTrainedModel):
         )
     
 
-class LongitudinalPromptVariableCXREncoderDecoderModel(VisionEncoderDecoderModel):
+class LongitudinalPromptMultiCXREncoderDecoderModel(VisionEncoderDecoderModel):
 
     config_class = VisionEncoderDecoderConfig
     base_model_prefix = "vision_encoder_decoder"
@@ -127,7 +127,7 @@ class LongitudinalPromptVariableCXREncoderDecoderModel(VisionEncoderDecoderModel
 
         # Encoder:
         if encoder is None:
-            encoder = VariableCvtWithProjectionHead(config=config.encoder)
+            encoder = MultiCvtWithProjectionHead(config=config.encoder)
 
         # Decoder:
         if decoder is None:
@@ -150,7 +150,7 @@ class LongitudinalPromptVariableCXREncoderDecoderModel(VisionEncoderDecoderModel
         self.encoder.config = self.config.encoder
         self.decoder.config = self.config.decoder
 
-        # Load variable checkpoint:
+        # Load multi checkpoint:
         if encoder_decoder_ckpt_name:
             encoder_decoder = AutoModel.from_pretrained(encoder_decoder_ckpt_name, trust_remote_code=True)
             self.load_state_dict(encoder_decoder.state_dict())
@@ -266,6 +266,12 @@ class LongitudinalPromptVariableCXREncoderDecoderModel(VisionEncoderDecoderModel
             https://github.com/huggingface/transformers/blob/main/src/transformers/models/encoder_decoder/modeling_encoder_decoder.py#L660
         """
 
+        # An update to generate() now prepends bos_token_id to each sequence if it does not exist at the start of the input: 
+        #   https://github.com/huggingface/transformers/blob/d533465150532b0c5de167b574e59f64c68b1154/src/transformers/generation/utils.py#L699C13-L699C30
+        # Hence, we remove the prepended bos_token_id from each sequence if it is there:
+        if torch.all(input_ids[:, 0] == 1):
+            input_ids = input_ids[:, 1:]
+
         decoder_inputs = self.decoder.prepare_inputs_for_generation(input_ids, past_key_values=past_key_values)
         decoder_attention_mask = (input_ids != mask_token_id).int()
         decoder_position_ids = torch.nn.functional.relu(

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b04999f2bf3a2d29880d0897eeb3cb38496ecb2d4dc322198bc6b972eca5c85e
-size 450315441
+oid sha256:41b92a8b66e47aca158a4952342fbd63128b5bada04662b807e477c92c61aeb0
+size 450311021