v3

src/dataset_combine.py

@@ -65,4 +65,4 @@ def combine_datasets(main_dataset_path, datasets_to_add):
     
     return combined
 
-combine_datasets("/workspace/sdxs3d/datasets/mjnj", ["/workspace/animesfw384"])
+combine_datasets("/workspace/sdxs3d/datasets/batch0_640", ["/workspace/sdxs3d/datasets/esh640"])

src/dataset_combine_zstd.py

@@ -0,0 +1,47 @@
+import os
+import shutil
+from datasets import load_from_disk, concatenate_datasets
+
+def combine_datasets(main_dataset_path, datasets_to_add):
+    try:
+        main_dataset = load_from_disk(main_dataset_path)
+        print(f"Загружен основной датасет: {main_dataset_path} ({len(main_dataset)} записей)")
+    except Exception as e:
+        print(f"Ошибка загрузки основного датасета: {e}")
+        return None
+
+    all_datasets = [main_dataset]
+
+    for path in datasets_to_add:
+        try:
+            ds = load_from_disk(path)
+            all_datasets.append(ds)
+            print(f"Добавлен датасет: {path} ({len(ds)} записей)")
+        except Exception as e:
+            print(f"Ошибка загрузки датасета {path}: {e}")
+
+    combined = concatenate_datasets(all_datasets)
+    print(f"Объединение {len(all_datasets)} датасетов...")
+
+    temp_dir = f"{main_dataset_path}_temp"
+    if os.path.exists(temp_dir):
+        shutil.rmtree(temp_dir)
+
+    try:
+        print(f"Сохранение в сжатом формате во временную директорию {temp_dir}...")
+        combined.save_to_disk(temp_dir)#, compression="zstd")  # <-- Сжатие
+
+        print(f"Обновление основного датасета...")
+        if os.path.exists(main_dataset_path):
+            shutil.rmtree(main_dataset_path)
+        shutil.copytree(temp_dir, main_dataset_path)
+        shutil.rmtree(temp_dir)
+
+        print(f"✅ Объединенный датасет ({len(combined)} записей) успешно сохранен в: {main_dataset_path}")
+    except Exception as e:
+        print(f"Ошибка при сохранении датасета: {e}")
+        print(f"Временные данные сохранены в: {temp_dir}")
+
+    return combined
+
+combine_datasets("/workspace/sdxs3d/datasets/esh640", ["/workspace/sdxs3d/datasets/batch0_640_temp"])

src/dataset_from_folder_qwen.py

@@ -27,8 +27,8 @@ step = 64 #64
 empty_share = 0.05
 limit = 0
 # Основная процедура обработки
-folder_path = "/workspace/sdxs3d/datasets/eshooshoo_all" #alchemist"
-save_path = "/workspace/sdxs3d/datasets/esh640" #"alchemist"
+folder_path = "/workspace/sdxs3d/datasets/batch1" #alchemist"
+save_path = "/workspace/sdxs3d/datasets/batch1_640" #"alchemist"
 os.makedirs(save_path, exist_ok=True)
 
 # Функция для очистки CUDA памяти

train.py

@@ -26,16 +26,16 @@ import torch.nn.functional as F
 from collections import deque
 
 # --------------------------- Параметры ---------------------------
-ds_path = "/workspace/sdxs3d/datasets/esh640"
+ds_path = "/workspace/sdxs3d/datasets/batch0_640_temp"
 project = "unet"
-batch_size = 64
+batch_size = 32
 base_learning_rate = 6e-5
 min_learning_rate = 1e-5
-num_epochs = 80
+num_epochs = 10
 # samples/save per epoch
-sample_interval_share = 20
-use_wandb = True
-use_comet_ml = False
+sample_interval_share = 2
+use_wandb = False
+use_comet_ml = True
 save_model = True
 use_decay = True
 fbp = False # fused backward pass
@@ -52,7 +52,7 @@ torch.backends.cudnn.allow_tf32 = True
 torch.backends.cuda.enable_mem_efficient_sdp(False)
 dtype = torch.float32
 save_barrier = 1.006
-warmup_percent = 0.01
+warmup_percent = 0.005
 percentile_clipping = 99 # 8bit optim
 betta2 = 0.99
 eps = 1e-8
@@ -480,7 +480,8 @@ def get_negative_embedding(neg_prompt="", batch_size=1):
     # Возвращаем эмбеддинг
     if not neg_prompt:
         hidden_dim = 1024  # Размерность эмбеддинга Qwen3-Embedding-0.6B
-        return torch.zeros((batch_size, 1, hidden_dim), dtype=dtype, device=device)
+        seq_len = 150
+        return torch.zeros((batch_size, seq_len, hidden_dim), dtype=dtype, device=device)
 
     uncond_emb = encode_texts([neg_prompt]).to(dtype=dtype, device=device)
     uncond_emb = uncond_emb.repeat(batch_size, 1, 1)  # Добавляем батч
@@ -533,10 +534,19 @@ def generate_and_save_samples(fixed_samples_cpu,empty_embeddings, step):
                 # guidance: удваиваем батч
                 if guidance_scale != 1:
                     latent_model_input = torch.cat([latents, latents], dim=0)
-                    text_embeddings_batch = torch.cat([empty_embeddings, sample_text_embeddings], dim=0)
+                
+                    # empty_embeddings: [1, 1, hidden_dim] → повторяем по seq_len и batch
+                    seq_len = sample_text_embeddings.shape[1]
+                    hidden_dim = sample_text_embeddings.shape[2]
+                    empty_embeddings_exp = empty_embeddings.expand(-1, seq_len, hidden_dim)  # [1, seq_len, hidden_dim]
+                    empty_embeddings_exp = empty_embeddings_exp.repeat(sample_text_embeddings.shape[0], 1, 1)  # [batch, seq_len, hidden_dim]
+                
+                    text_embeddings_batch = torch.cat([empty_embeddings_exp, sample_text_embeddings], dim=0)
                 else:
                     latent_model_input = latents
                     text_embeddings_batch = sample_text_embeddings
+
+
         
                 # предсказание потока (velocity)
                 model_out = original_model(latent_model_input, t, encoder_hidden_states=text_embeddings_batch)

unet/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:afc06beff07034f0ce9f671c83222e7f78eedc3b3ce93293143accdebef1b111
-size 1887
+oid sha256:fac879374be0c1cd3ffcbfa188e72f50576f25326aada5da077b0fc7af442ad4
+size 1874

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9d933d318f2d42b37c31065a09c14ee0c03ec05a10d672667743a089d396086b
-size 3092571208
+oid sha256:74563fd4bef0e9f811d7ed6571b6064707beafa8f6ea7f4eeae0dbdfe819f902
+size 6184944280