mod ds

hft.py

@@ -9,7 +9,7 @@ import pytesseract
 import docx2txt
 import PyPDF2
 import json
-from torch.amp import autocast
+from torch.cuda.amp import autocast
 from collections import defaultdict
 from huggingface_hub import login
 
@@ -21,21 +21,24 @@ os.environ["TOKENIZERS_PARALLELISM"] = "false"
 os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
 
 def free_memory():
-    torch.empty_cache('cuda')
-    torch.ipc_collect('cuda')
+    torch.cuda.empty_cache()
+    torch.cuda.ipc_collect()
 
 class SourceMapper:
     def __init__(self):
-        self.source_to_idx = defaultdict(lambda: len(self.source_to_idx))
-        self.idx_to_source = {}
+        self.source_to_idx = defaultdict(lambda: 0)  # Domyślnie 0 dla nieznanych
+        self.idx_to_source = {0: "Unknown"}
+        self.next_idx = 1  # Indeksy od 1 dla znanych źródeł
         
     def add_source(self, source):
         if source and source not in self.source_to_idx:
-            idx = self.source_to_idx[source]
+            idx = self.next_idx
+            self.source_to_idx[source] = idx
             self.idx_to_source[idx] = source
+            self.next_idx += 1
             
     def get_idx(self, source):
-        return self.source_to_idx[source] if source else -1
+        return self.source_to_idx.get(source, 0)
     
     def get_source(self, idx):
         return self.idx_to_source.get(idx, "Unknown")
@@ -97,7 +100,7 @@ def prepare_dataset(directory, catalog_path, source_mapper):
                 for chunk in chunks:
                     data.append({
                         "text": chunk,
-                        "source_idx": -1
+                        "source_idx": 0
                     })
     return data
 
@@ -126,8 +129,9 @@ class CustomModel(GPTNeoForCausalLM):
         self.source_embedding = nn.Embedding(
             num_embeddings=1000,
             embedding_dim=config.hidden_size,
-            padding_idx=-1
+            padding_idx=0  # Poprawiony padding_idx
         )
+        self.source_proj = nn.Linear(config.hidden_size, config.vocab_size)
         
     def forward(self, input_ids=None, attention_mask=None, labels=None, source_idx=None, **kwargs):
         with autocast():
@@ -138,8 +142,9 @@ class CustomModel(GPTNeoForCausalLM):
                 **kwargs
             )
             if source_idx is not None:
-                source_embeds = self.source_embedding(source_idx).unsqueeze(1)
-                outputs.logits += source_embeds
+                source_embeds = self.source_embedding(source_idx)
+                source_projected = self.source_proj(source_embeds)
+                outputs.logits += source_projected.unsqueeze(1)
         return outputs
 
 source_mapper = SourceMapper()
@@ -198,8 +203,8 @@ def generate_answer(question, model, tokenizer, source_mapper, max_length=200):
     answer = tokenizer.decode(outputs.sequences[0], skip_special_tokens=True)
     
     # Pobierz źródło z ostatniego tokena
-    last_token_id = outputs.sequences[0][-1].item()
-    source_idx = model.source_embedding.weight.shape[0] - 1  # Tymczasowe rozwiązanie
+    last_token_logits = outputs.scores[-1]
+    source_idx = torch.argmax(last_token_logits, dim=-1)[-1].item()
     source = source_mapper.get_source(source_idx)
     
     return f"{answer}\n\nŹródło: {source if source else 'Opracowanie własne'}"