End of training

ConSec.py

@@ -128,7 +128,7 @@ class ConSecTagger:
                  labeller=json_labeller):
         self.model = model
         self.tokenizer = tokenizer
-        special_tokens = self.tokenizer.get_added_tokens()
+        special_tokens = self.tokenizer.get_added_vocab()
         self.start_token = special_tokens["[START]"]
         self.gloss_token = special_tokens["[GLOSS]"]
         self.sequencer = sequencer
@@ -148,27 +148,38 @@ class ConSecTagger:
             with self.model.device:
                 tokens = self.tokenizer(text,"[GLOSS] ".join(glosses),
                                         return_tensors="pt")
-                length = tokens.input_ids.shape[0]
+                length = tokens.input_ids.shape[1]
                 positions = torch.arange(length)
-                place = (tokens.input_ids==self.start_token).nonzero(as_tuple=True)[0]
-                wordpos = tokens.token_to_word(place+1)
-                gloss_positions = list((tokens.input_ids==self.gloss_token).nonzero(as_tuple=True))
+                place = (tokens.input_ids==self.start_token).nonzero(as_tuple=True)[1].item()
+                wordpos = tokens.token_to_word(place)
+                gloss_positions = [index.item() 
+                                   for index in (tokens.input_ids==self.gloss_token).nonzero(as_tuple=True)[1]]
                 gloss_positions.append(length)
                 n_candidates = len(candidates)
                 for (i,position) in enumerate(gloss_positions[:-1]):
                     if i<n_candidates:
-                        end = place + gloss_positions[i+1]-position
-                        positions [position:gloss_positions[i+1]] = torch.arange(place,end)
+                        end = (place + gloss_positions[i+1]-position)
+                        positions[position:gloss_positions[i+1]] = torch.arange(place,end)
                     else:
                         known = already_tagged[i-n_candidates]
                         start = tokens.word_to_tokens(known["place"]).start
-                        end = start + gloss_positions[i+1] - position
+                        end = (start + gloss_positions[i+1] - position)
                         positions[position:gloss_positions[i+1]] = torch.arange(start,end)
-                prediction = self.model(token_ids=tokens.token_ids,
+                prediction = self.model(input_ids=tokens.input_ids,
                                     attention_mask=tokens.attention_mask,
                                     token_type_ids=tokens.token_type_ids,
-                                    position_ids=positions)
-                label = candidates[prediction.logits.argmax()]
+                                    position_ids=positions.reshape((1,length)))
+                try:
+                    label = candidates[prediction.logits.argmax()]
+                except IndexError:
+                    print(text)
+                    print(gloss_positions)
+                    print([positions[pos].item() for pos in gloss_positions[:-1]])
+                    print(already_tagged)
+                    print(candidates)
+                    print(prediction.logits)
+                    print(prediction.logits.argmax())
+                    raise
                 already_tagged.append({"label":label,
                                        "place":wordpos,
                                        "index":index})

README.md

@@ -22,11 +22,11 @@ should probably proofread and complete it, then remove this comment. -->
 
 This model is a fine-tuned version of [microsoft/deberta-v3-base](https://huggingface.co/microsoft/deberta-v3-base) on the None dataset.
 It achieves the following results on the evaluation set:
-- Loss: 3.9109
-- Precision: 0.5353
-- Recall: 0.5517
-- F1: 0.5434
-- Matthews: 0.5509
+- Loss: 1.5775
+- Precision: 0.4804
+- Recall: 0.4917
+- F1: 0.4860
+- Matthews: 0.4909
 
 ## Model description
 
@@ -58,12 +58,12 @@ The following hyperparameters were used during training:
 
 | Training Loss | Epoch | Step   | Validation Loss | Precision | Recall | F1     | Matthews |
 |:-------------:|:-----:|:------:|:---------------:|:---------:|:------:|:------:|:--------:|
-| No log        | 0     | 0      | 365.7523        | 0.4494    | 0.3098 | 0.3668 | 0.3092   |
-| 4.1345        | 1.0   | 56179  | 3.8725          | 0.5383    | 0.5535 | 0.5458 | 0.5527   |
-| 4.1174        | 2.0   | 112358 | 3.9544          | 0.5360    | 0.5517 | 0.5437 | 0.5509   |
-| 3.9667        | 3.0   | 168537 | 3.9244          | 0.5348    | 0.5517 | 0.5431 | 0.5509   |
-| 4.2556        | 4.0   | 224716 | 3.9431          | 0.5348    | 0.5508 | 0.5427 | 0.5501   |
-| 4.2113        | 5.0   | 280895 | 3.9109          | 0.5353    | 0.5517 | 0.5434 | 0.5509   |
+| No log        | 0     | 0      | 344.1697        | 0.4603    | 0.3243 | 0.3805 | 0.3236   |
+| 6.7210        | 1.0   | 56179  | 1.5766          | 0.4804    | 0.4917 | 0.4860 | 0.4909   |
+| 5.7990        | 2.0   | 112358 | 1.5649          | 0.4859    | 0.4943 | 0.4900 | 0.4935   |
+| 6.3812        | 3.0   | 168537 | 1.5669          | 0.4804    | 0.4926 | 0.4864 | 0.4918   |
+| 5.8106        | 4.0   | 224716 | 1.5847          | 0.4834    | 0.4921 | 0.4877 | 0.4913   |
+| 6.0390        | 5.0   | 280895 | 1.5775          | 0.4804    | 0.4917 | 0.4860 | 0.4909   |
 
 
 ### Framework versions

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:51944abe32d6d50692f53f7ff5146b58b23aa7e5a015d85ea6b5c3cc14a5fb54
+oid sha256:cc8e2038fc80452fdf46c78fad96e984da050257f1fed05392bbbb4511d9f4a1
 size 367690992