Upload folder using huggingface_hub

config.json

@@ -1,7 +1,7 @@
 {
   "_class_name": "AsymmetricAutoencoderKL",
   "_diffusers_version": "0.36.0",
-  "_name_or_path": "vae9",
+  "_name_or_path": "vae10",
   "act_fn": "silu",
   "block_out_channels": [
     128,

diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:17ca100e6990c559a994e7589fb1e6d8bec5e76d3ea70ac5e4009238cbaf9f26
+oid sha256:adb25d8783707b9a6267acb396f44a2966fd6e50dbe23ff6ed142020385a178b
 size 427466716

train_vae_fdl.py

@@ -29,11 +29,11 @@ from collections import deque
 
 # --------------------------- Параметры ---------------------------
 ds_path            = "/workspace/d23"
-project            = "vae9"
+project            = "vae10"
 batch_size         = 1
 base_learning_rate = 6e-6
 min_learning_rate  = 7e-7
-num_epochs         = 4
+num_epochs         = 2
 sample_interval_share = 25
 use_wandb          = True
 save_model         = True
@@ -41,8 +41,8 @@ use_decay          = True
 optimizer_type     = "adam8bit"
 dtype              = torch.float32
 
-model_resolution   = 384 #288
-high_resolution    = 768 #576 
+model_resolution   = 512 #288
+high_resolution    = 1024 #576 
 limit              = 0
 save_barrier       = 1.3
 warmup_percent     = 0.005
@@ -498,7 +498,7 @@ for epoch in range(num_epochs):
             imgs = imgs.to(accelerator.device)
 
             if high_resolution != model_resolution:
-                imgs_low = F.interpolate(imgs, size=(model_resolution, model_resolution), mode="bilinear", align_corners=False)
+                imgs_low = F.interpolate(imgs, size=(model_resolution, model_resolution),mode="area") # mode="bilinear", align_corners=False)
             else:
                 imgs_low = imgs