2511

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b9b5e44ec1e34e89b0a6024bf233b486c425325b537b42577bad7cc9da3c7a9
-size 5180264
+oid sha256:7754efea243246c21b73b743ea55055cff9ed385f22d119ee489931185366cf1
+size 8316949

train.py

@@ -447,35 +447,22 @@ def get_negative_embedding(neg_prompt="", batch_size=1):
     # Загрузка моделей (если ещё не загружены)
     if not hasattr(get_negative_embedding, "tokenizer"):
         get_negative_embedding.tokenizer = AutoTokenizer.from_pretrained(
-            "Qwen/Qwen3-Embedding-0.6B", padding_side="left"
+            "Qwen/Qwen3-0.6B"
         )
         get_negative_embedding.text_model = AutoModel.from_pretrained(
-            "Qwen/Qwen3-Embedding-0.6B"
+            "Qwen/Qwen3-0.6B"
         ).to(device).eval()
 
-    # Вспомогательная функция для пулинга
-    def last_token_pool(last_hidden_states, attention_mask):
-        left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
-        if left_padding:
-            return last_hidden_states[:, -1]
-        else:
-            sequence_lengths = attention_mask.sum(dim=1) - 1
-            batch_size = last_hidden_states.shape[0]
-            return last_hidden_states[
-                torch.arange(batch_size, device=last_hidden_states.device),
-                sequence_lengths
-            ]
-
     # Вычисление эмбеддинга
-    def encode_texts(texts, max_length=512):
+    def encode_texts(texts, max_length=150):
         with torch.inference_mode():
             toks = get_negative_embedding.tokenizer(
                 texts, return_tensors="pt", padding="max_length", truncation=True, max_length=max_length
             ).to(device)
-            outs = get_negative_embedding.text_model(**toks)
-            emb = last_token_pool(outs.last_hidden_state, toks["attention_mask"])
-            emb = emb.unsqueeze(1)  # Добавляем размерность sequence_length
-            return emb
+
+            outs = get_negative_embedding.text_model(**toks, output_hidden_states=True)
+            hidden_states = outs.hidden_states[-1]  # [B, L, D]
+            return hidden_states
 
     # Возвращаем эмбеддинг
     if not neg_prompt:

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4fbf975ae1854d56d63011e6062054fe4abbd45ca011f66013bf0264f657d890
+oid sha256:c093c51d6be0a372c91657f1abdce28a61577134ef09eac531479ca384cef933
 size 6184944280