Upload folder using huggingface_hub

test.ipynb

@@ -2,7 +2,7 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 7,
+   "execution_count": 1,
    "id": "6ca10d55-03ed-4c8b-b32b-8d2f94d77162",
    "metadata": {},
    "outputs": [

train_vae.py

@@ -29,9 +29,9 @@ from collections import deque
 # --------------------------- Параметры ---------------------------
 ds_path            = "/workspace/d23"
 project            = "vae"
-batch_size         = 4
-base_learning_rate = 6e-6
-min_learning_rate  = 9e-7
+batch_size         = 2
+base_learning_rate = 2e-6
+min_learning_rate  = 7e-7
 num_epochs         = 25
 sample_interval_share = 10
 use_wandb          = True
@@ -40,8 +40,8 @@ use_decay          = True
 optimizer_type     = "adam8bit"
 dtype              = torch.float32
 
-model_resolution   = 256
-high_resolution    = 512 
+model_resolution   = 384
+high_resolution    = 768 
 limit              = 0
 save_barrier       = 1.3
 warmup_percent     = 0.01
@@ -50,7 +50,7 @@ beta2              = 0.997
 eps                = 1e-8
 clip_grad_norm     = 1.0
 mixed_precision    = "no"
-gradient_accumulation_steps = 4
+gradient_accumulation_steps = 8
 generated_folder   = "samples"
 save_as            = "vae"
 num_workers        = 0
@@ -167,16 +167,24 @@ if full_training and not train_decoder_only:
         loss_ratios["kl"] = float(kl_ratio)
         trainable_module = core
 else:
-# учим только декодер + post_quant_conv на "ядре" модели
+    # учим только 0-й блок декодера + post_quant_conv
     if hasattr(core, "decoder"):
-        for name, p in core.decoder.named_parameters():
-            p.requires_grad = True
-            unfrozen_param_names.append(f"decoder.{name}")
+        # --- только 0-й up_block ---
+        if hasattr(core.decoder, "up_blocks") and len(core.decoder.up_blocks) > 0:
+            for name, p in core.decoder.up_blocks[0].named_parameters():
+                p.requires_grad = True
+                unfrozen_param_names.append(f"decoder.up_blocks.0.{name}")
+        else:
+            print("[WARN] Decoder has no up_blocks — fallback to full decoder")
+            for name, p in core.decoder.named_parameters():
+                p.requires_grad = True
+                unfrozen_param_names.append(f"decoder.{name}")
     if hasattr(core, "post_quant_conv"):
         for name, p in core.post_quant_conv.named_parameters():
             p.requires_grad = True
             unfrozen_param_names.append(f"post_quant_conv.{name}")
-            trainable_module = core.decoder if hasattr(core, "decoder") else core
+    trainable_module = core.decoder if hasattr(core, "decoder") else core
+
 
 print(f"[INFO] Разморожено параметров: {len(unfrozen_param_names)}. Первые 200 имён:")
 for nm in unfrozen_param_names[:200]:

vae/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:71b36395ec5c24620ff6943eb10111bcb05e0af4c0ad35fde61020f148be115e
+oid sha256:da1a8a728f24c5841cd0be986c90096d63800f8403bfba4b0e8c25c82c521bbf
 size 444559412