2512

dataset.py

@@ -21,14 +21,14 @@ from datetime import timedelta
 dtype = torch.float32
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 batch_size = 5
-min_size = 384 #320 #192 #256 #192
-max_size = 768 #640 #384 #256 #384
-step = 32 #64
+min_size = 192 #384 #320 #192 #256 #192
+max_size = 320 #768 #640 #384 #256 #384
+step = 64 #64
 empty_share = 0.0
 limit = 0
 # Основная процедура обработки
-folder_path = "/workspace/alchemist" #alchemist"
-save_path = "/workspace/sdxs/datasets/alchemist" #"alchemist"
+folder_path = "/workspace/butterfly" #alchemist"
+save_path = "/workspace/sdxs/datasets/butterfly" #"alchemist"
 os.makedirs(save_path, exist_ok=True)
 
 # Функция для очистки CUDA памяти

src/sdxs_create.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cd9557eaff4978dc3758d8632f07a8f6a9406b1ce5033529d8e0cb389336a372
-size 37957
+oid sha256:f7e7a0f7dec0f14e09690f5c4cf83505b42361436f3eec7a0edfb3c6ccf3bf91
+size 37958

src/sdxs_create_05.ipynb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c170b5c6e418e9d7e94bf475aff653bde77a1b852db94b8254beaeef8a36d9ce
+size 37749

train.py

@@ -29,10 +29,10 @@ from transformers import AutoTokenizer, AutoModel
 # --------------------------- Параметры ---------------------------
 ds_path = "/workspace/sdxs/datasets/768"
 project = "unet"
-batch_size = 36
-base_learning_rate = 2.7e-5
-min_learning_rate = 1e-5 #2.7e-5
-num_epochs = 20
+batch_size = 16
+base_learning_rate = 2e-5 #2.7e-5
+min_learning_rate = 2e-6 #2.7e-5
+num_epochs = 10
 sample_interval_share = 5
 max_length = 192
 use_wandb = True
@@ -651,7 +651,7 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
     
     for step, (latents, embeddings, attention_mask) in enumerate(dataloader):
         with accelerator.accumulate(unet):
-            if save_model == False and step == 5 :
+            if save_model == False and epoch == 0 and step == 5 :
                 used_gb = torch.cuda.max_memory_allocated() / 1024**3
                 print(f"Шаг {step}: {used_gb:.2f} GB")
                 
@@ -772,8 +772,8 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
 
 if accelerator.is_main_process:
     print("Обучение завершено! Сохраняем финальную модель...")
-    if save_model:
-        save_checkpoint(unet,"fp16")
+    #if save_model:
+    save_checkpoint(unet,"fp16")
     if use_comet_ml:
         comet_experiment.end()
 accelerator.free_memory()

unet/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:80410b994cba2961422e6bade67a2a6a41abaf90e6dd89c0226402f45ef113cb
-size 1892
+oid sha256:fb10c12478625b355068dbfce29e8db3c8a52d03e10508452e4feb5404545584
+size 1876

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:87a0d7fc2710a7f65ba6038642507109b1d5d2a22040903c5e54d01627566ca4
-size 3722280432
+oid sha256:5bd1bc9f4967f9060624d6d5536ed84fe051cd0203f5b26b083249093bd696d2
+size 7444321360