Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+123456789.jpg filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,14 @@
+# Jupyter Notebook
+__pycache__/
+*.pyc
+.ipynb_checkpoints/
+*.ipynb_checkpoints/*
+.ipynb_checkpoints/*
+src/samples
+# cache
+cache
+datasets
+test
+wandb
+nohup.out
+

down.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+
+TARGET_DIR="/workspace/d23"
+mkdir -p "$TARGET_DIR"
+
+BASE_URL="https://huggingface.co/datasets/AI-Art-Collab/dtasettar23/resolve/main/d23.tar."
+
+(
+  # Устанавливаем `set -e` внутри subshell, чтобы он завершился при первой ошибке curl
+  set -e
+  # Попробуем от 'a' до 'z' для первого символа суффикса
+  for c1 in {a..z}; do
+    # Попробуем от 'a' до 'z' для второго символа суффикса
+    for c2 in {a..z}; do
+      suffix="${c1}${c2}"
+      url="${BASE_URL}${suffix}"
+      echo "Fetching: $url" >&2
+      # Качаем часть архива. --fail заставит curl завершиться с ошибкой, если файла нет.
+      curl -LsS --fail "$url"
+    done
+  done
+) 2>/dev/null | tar -xv -C "$TARGET_DIR" --wildcards '*.png'
+#    └─ 1 ─┘   └────────── 2 ──────────┘ └─────────── 3 ───────────┘
+
+echo "Extraction of PNG files finished. Check $TARGET_DIR"

test.ipynb

@@ -0,0 +1,192 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "6ca10d55-03ed-4c8b-b32b-8d2f94d77162",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "The config attributes {'block_out_channels': [128, 256, 512, 512, 512], 'force_upcast': False} were passed to AsymmetricAutoencoderKL, but are not expected and will be ignored. Please verify your config.json configuration file.\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "test log-variance: 0.065\n",
+      "Готово\n"
+     ]
+    }
+   ],
+   "source": [
+    "import torch\n",
+    "from PIL import Image\n",
+    "from diffusers import AutoencoderKL,AsymmetricAutoencoderKL\n",
+    "from torchvision.transforms.functional import to_pil_image\n",
+    "import matplotlib.pyplot as plt\n",
+    "import os\n",
+    "from torchvision.transforms import ToTensor, Normalize, CenterCrop\n",
+    "\n",
+    "# путь к вашей картинке\n",
+    "IMG_PATH = \"123456789.jpg\"\n",
+    "OUT_DIR  = \"test\"\n",
+    "device   = \"cuda\"\n",
+    "dtype    = torch.float16      \n",
+    "os.makedirs(OUT_DIR, exist_ok=True)\n",
+    "\n",
+    "# список VAE\n",
+    "VAES = {\n",
+    "    \"test\": \"/workspace/simple_vae2x\",\n",
+    "}\n",
+    "\n",
+    "def load_image(path):\n",
+    "    img = Image.open(path).convert('RGB')\n",
+    "    # обрезаем до кратности 8\n",
+    "    w, h = img.size\n",
+    "    img = CenterCrop((h // 8 * 8, w // 8 * 8))(img)\n",
+    "    tensor = ToTensor()(img).unsqueeze(0)          # [0,1]\n",
+    "    tensor = Normalize(mean=[0.5]*3, std=[0.5]*3)(tensor)  # [-1,1]\n",
+    "    return img, tensor.to(device, dtype=dtype)\n",
+    "\n",
+    "# обратно в PIL\n",
+    "def tensor_to_img(t):\n",
+    "    t = (t * 0.5 + 0.5).clamp(0, 1)\n",
+    "    return to_pil_image(t[0])\n",
+    "\n",
+    "def logvariance(latents):\n",
+    "    \"\"\"Возвращает лог-дисперсию по всем элементам.\"\"\"\n",
+    "    return torch.log(latents.var() + 1e-8).item()\n",
+    "\n",
+    "def plot_latent_distribution(latents, title, save_path):\n",
+    "    \"\"\"Гистограмма + QQ-plot.\"\"\"\n",
+    "    lat = latents.detach().cpu().numpy().flatten()\n",
+    "    plt.figure(figsize=(10, 4))\n",
+    "\n",
+    "    # гистограмма\n",
+    "    plt.subplot(1, 2, 1)\n",
+    "    plt.hist(lat, bins=100, density=True, alpha=0.7, color='steelblue')\n",
+    "    plt.title(f\"{title} histogram\")\n",
+    "    plt.xlabel(\"latent value\")\n",
+    "    plt.ylabel(\"density\")\n",
+    "\n",
+    "    # QQ-plot\n",
+    "    from scipy.stats import probplot\n",
+    "    plt.subplot(1, 2, 2)\n",
+    "    probplot(lat, dist=\"norm\", plot=plt)\n",
+    "    plt.title(f\"{title} QQ-plot\")\n",
+    "\n",
+    "    plt.tight_layout()\n",
+    "    plt.savefig(save_path)\n",
+    "    plt.close()\n",
+    "\n",
+    "for name, repo in VAES.items():\n",
+    "    if name==\"test\":\n",
+    "        vae = AsymmetricAutoencoderKL.from_pretrained(repo, subfolder=\"vae\", torch_dtype=dtype).to(device)\n",
+    "    else:\n",
+    "        vae = AutoencoderKL.from_pretrained(repo, torch_dtype=dtype).to(device)#, subfolder=\"vae\", variant=\"fp16\"\n",
+    "\n",
+    "    cfg   = vae.config\n",
+    "    scale = getattr(cfg, \"scaling_factor\", 1.)\n",
+    "    shift = getattr(cfg, \"shift_factor\", 0.0)\n",
+    "    mean  = getattr(cfg, \"latents_mean\", None)\n",
+    "    std  = getattr(cfg, \"latents_std\",  None)\n",
+    "\n",
+    "    C = 16  # 4 для SDXL\n",
+    "    if mean is not None:\n",
+    "        mean = torch.tensor(mean, device=device, dtype=dtype).view(1, C, 1, 1)\n",
+    "    if std is not None:\n",
+    "        std  = torch.tensor(std,  device=device, dtype=dtype).view(1, C, 1, 1)\n",
+    "    if shift is not None:\n",
+    "        shift = torch.tensor(shift, device=device, dtype=dtype)\n",
+    "    else:\n",
+    "        shift = 0.0 \n",
+    "\n",
+    "    scale = torch.tensor(scale, device=device, dtype=dtype)\n",
+    "\n",
+    "    img, x = load_image(IMG_PATH)\n",
+    "    img.save(os.path.join(OUT_DIR, f\"original.jpg\"))\n",
+    "\n",
+    "    with torch.no_grad():\n",
+    "        # encode\n",
+    "        latents = vae.encode(x).latent_dist.sample().to(dtype)\n",
+    "        if mean is not None and std is not None:\n",
+    "            latents = (latents - mean) / std\n",
+    "        latents = latents * scale + shift\n",
+    "\n",
+    "        lv = logvariance(latents)\n",
+    "        print(f\"{name} log-variance: {lv:.3f}\")\n",
+    "\n",
+    "        # график\n",
+    "        plot_latent_distribution(latents, f\"{name}_latents\",\n",
+    "                                 os.path.join(OUT_DIR, f\"dist_{name}.png\"))\n",
+    "\n",
+    "        # decode\n",
+    "        latents = (latents - shift) / scale\n",
+    "        if mean is not None and std is not None:\n",
+    "            latents = latents * std + mean\n",
+    "        rec = vae.decode(latents).sample\n",
+    "\n",
+    "    tensor_to_img(rec).save(os.path.join(OUT_DIR, f\"decoded_{name}.png\"))\n",
+    "\n",
+    "print(\"Готово\")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "5e930fd3-0aa5-4ed6-beab-e871df009125",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Collecting scipy\n",
+      "  Downloading scipy-1.16.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl.metadata (62 kB)\n",
+      "Requirement already satisfied: numpy<2.6,>=1.25.2 in /usr/local/lib/python3.12/dist-packages (from scipy) (2.1.2)\n",
+      "Downloading scipy-1.16.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl (35.7 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m35.7/35.7 MB\u001b[0m \u001b[31m58.9 MB/s\u001b[0m  \u001b[33m0:00:00\u001b[0mm0:00:01\u001b[0m00:01\u001b[0m\n",
+      "\u001b[?25hInstalling collected packages: scipy\n",
+      "Successfully installed scipy-1.16.2\n"
+     ]
+    }
+   ],
+   "source": [
+    "!pip install scipy"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "72785e98-5dad-48a3-809b-3ab9755ac9db",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

train_vae.py

@@ -0,0 +1,569 @@
+# -*- coding: utf-8 -*-
+import os
+import math
+import re
+import torch
+import numpy as np
+import random
+import gc
+from datetime import datetime
+from pathlib import Path
+
+import torchvision.transforms as transforms
+import torch.nn.functional as F
+from torch.utils.data import DataLoader, Dataset
+from torch.optim.lr_scheduler import LambdaLR
+from diffusers import AutoencoderKL, AsymmetricAutoencoderKL
+# QWEN: импорт класса
+from diffusers import AutoencoderKLQwenImage
+from diffusers import AutoencoderKLWan
+
+from accelerate import Accelerator
+from PIL import Image, UnidentifiedImageError
+from tqdm import tqdm
+import bitsandbytes as bnb
+import wandb
+import lpips   # pip install lpips
+from collections import deque
+
+# --------------------------- Параметры ---------------------------
+ds_path            = "/workspace/d23"
+project            = "vae"
+batch_size         = 4
+base_learning_rate = 6e-6
+min_learning_rate  = 9e-7
+num_epochs         = 50
+sample_interval_share = 10
+use_wandb          = True
+save_model         = True
+use_decay          = True
+optimizer_type     = "adam8bit"
+dtype              = torch.float32
+
+model_resolution   = 256
+high_resolution    = 512 
+limit              = 0
+save_barrier       = 1.03
+warmup_percent     = 0.01
+percentile_clipping = 99
+beta2              = 0.997
+eps                = 1e-8
+clip_grad_norm     = 1.0
+mixed_precision    = "no"
+gradient_accumulation_steps = 4
+generated_folder   = "samples"
+save_as            = "vae"
+num_workers        = 0
+device = None
+
+# --- Режимы обучения ---
+# QWEN: учим только декодер
+train_decoder_only = True
+full_training      = False  # если True — учим весь VAE и добавляем KL (ниже)
+kl_ratio           = 0.00
+
+# Доли лоссов
+loss_ratios = {
+    "lpips": 0.75,
+    "edge":  0.05,
+    "mse":   0.10,
+    "mae":   0.10,
+    "kl":    0.00,  # активируем при full_training=True
+}
+median_coeff_steps = 256
+
+resize_long_side = 1280  # ресайз длинной стороны исходных картинок
+
+# QWEN: конфиг загрузки модели
+vae_kind      = "kl"  # "qwen" или "kl" (обычный)
+
+Path(generated_folder).mkdir(parents=True, exist_ok=True)
+
+accelerator = Accelerator(
+    mixed_precision=mixed_precision,
+    gradient_accumulation_steps=gradient_accumulation_steps
+)
+device = accelerator.device
+
+# reproducibility
+seed = int(datetime.now().strftime("%Y%m%d"))
+torch.manual_seed(seed); np.random.seed(seed); random.seed(seed)
+torch.backends.cudnn.benchmark = False
+
+# --------------------------- WandB ---------------------------
+if use_wandb and accelerator.is_main_process:
+    wandb.init(project=project, config={
+        "batch_size": batch_size,
+        "base_learning_rate": base_learning_rate,
+        "num_epochs": num_epochs,
+        "optimizer_type": optimizer_type,
+        "model_resolution": model_resolution,
+        "high_resolution": high_resolution,
+        "gradient_accumulation_steps": gradient_accumulation_steps,
+        "train_decoder_only": train_decoder_only,
+        "full_training": full_training,
+        "kl_ratio": kl_ratio,
+        "vae_kind": vae_kind,
+    })
+
+# --------------------------- VAE ---------------------------
+def get_core_model(model):
+    m = model
+    # если модель уже обёрнута torch.compile
+    if hasattr(m, "_orig_mod"):
+        m = m._orig_mod
+    return m
+
+def is_video_vae(model) -> bool:
+    # WAN/Qwen — это видео-VAEs
+    if vae_kind in ("wan", "qwen"):
+        return True
+    # fallback по структуре (если понадобится)
+    try:
+        core = get_core_model(model)
+        enc = getattr(core, "encoder", None)
+        conv_in = getattr(enc, "conv_in", None)
+        w = getattr(conv_in, "weight", None)
+        if isinstance(w, torch.nn.Parameter):
+            return w.ndim == 5
+    except Exception:
+        pass
+    return False
+
+# загрузка
+if vae_kind == "qwen":
+    vae = AutoencoderKLQwenImage.from_pretrained("Qwen/Qwen-Image", subfolder="vae")
+else:
+    if vae_kind == "wan":
+        vae = AutoencoderKLWan.from_pretrained(project)
+    else:
+        # старое поведение (пример)
+        if model_resolution==high_resolution:
+            vae = AutoencoderKL.from_pretrained(project)
+        else:
+            vae = AsymmetricAutoencoderKL.from_pretrained(project)
+
+vae = vae.to(dtype)
+
+# torch.compile (опционально)
+if hasattr(torch, "compile"):
+    try:
+        vae = torch.compile(vae)
+    except Exception as e:
+        print(f"[WARN] torch.compile failed: {e}")
+
+# --------------------------- Freeze/Unfreeze ---------------------------
+core = get_core_model(vae)
+
+for p in core.parameters():
+    p.requires_grad = False
+
+unfrozen_param_names = []
+
+if full_training and not train_decoder_only:
+    for name, p in core.named_parameters():
+        p.requires_grad = True
+        unfrozen_param_names.append(name)
+        loss_ratios["kl"] = float(kl_ratio)
+        trainable_module = core
+else:
+# учим только декодер + post_quant_conv на "ядре" модели
+    if hasattr(core, "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
+
+print(f"[INFO] Разморожено параметров: {len(unfrozen_param_names)}. Первые 200 имён:")
+for nm in unfrozen_param_names[:200]:
+    print(" ", nm)
+
+# --------------------------- Датасет ---------------------------
+class PngFolderDataset(Dataset):
+    def __init__(self, root_dir, min_exts=('.png',), resolution=1024, limit=0):
+        self.root_dir = root_dir
+        self.resolution = resolution
+        self.paths = []
+        for root, _, files in os.walk(root_dir):
+            for fname in files:
+                if fname.lower().endswith(tuple(ext.lower() for ext in min_exts)):
+                    self.paths.append(os.path.join(root, fname))
+        if limit:
+            self.paths = self.paths[:limit]
+        valid = []
+        for p in self.paths:
+            try:
+                with Image.open(p) as im:
+                    im.verify()
+                valid.append(p)
+            except (OSError, UnidentifiedImageError):
+                continue
+        self.paths = valid
+        if len(self.paths) == 0:
+            raise RuntimeError(f"No valid PNG images found under {root_dir}")
+        random.shuffle(self.paths)
+
+    def __len__(self):
+        return len(self.paths)
+
+    def __getitem__(self, idx):
+        p = self.paths[idx % len(self.paths)]
+        with Image.open(p) as img:
+            img = img.convert("RGB")
+            if not resize_long_side or resize_long_side <= 0:
+                return img
+            w, h = img.size
+            long = max(w, h)
+            if long <= resize_long_side:
+                return img
+            scale = resize_long_side / float(long)
+            new_w = int(round(w * scale))
+            new_h = int(round(h * scale))
+            return img.resize((new_w, new_h), Image.LANCZOS)
+
+def random_crop(img, sz):
+    w, h = img.size
+    if w < sz or h < sz:
+        img = img.resize((max(sz, w), max(sz, h)), Image.LANCZOS)
+    x = random.randint(0, max(1, img.width - sz))
+    y = random.randint(0, max(1, img.height - sz))
+    return img.crop((x, y, x + sz, y + sz))
+
+tfm = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
+])
+
+dataset = PngFolderDataset(ds_path, min_exts=('.png',), resolution=high_resolution, limit=limit)
+if len(dataset) < batch_size:
+    raise RuntimeError(f"Not enough valid images ({len(dataset)}) to form a batch of size {batch_size}")
+
+def collate_fn(batch):
+    imgs = []
+    for img in batch:
+        img = random_crop(img, high_resolution)
+        imgs.append(tfm(img))
+    return torch.stack(imgs)
+
+dataloader = DataLoader(
+    dataset,
+    batch_size=batch_size,
+    shuffle=True,
+    collate_fn=collate_fn,
+    num_workers=num_workers,
+    pin_memory=True,
+    drop_last=True
+)
+
+# --------------------------- Оптимизатор ---------------------------
+def get_param_groups(module, weight_decay=0.001):
+    no_decay = ["bias", "LayerNorm.weight", "layer_norm.weight", "ln_1.weight", "ln_f.weight"]
+    decay_params, no_decay_params = [], []
+    for n, p in vae.named_parameters():  # глобально по vae, с фильтром requires_grad
+        if not p.requires_grad:
+            continue
+        if any(nd in n for nd in no_decay):
+            no_decay_params.append(p)
+        else:
+            decay_params.append(p)
+    return [
+        {"params": decay_params, "weight_decay": weight_decay},
+        {"params": no_decay_params, "weight_decay": 0.0},
+    ]
+
+def get_param_groups(module, weight_decay=0.001):
+    no_decay_tokens = ("bias", "norm", "rms", "layernorm")
+    decay_params, no_decay_params = [], []
+    for n, p in module.named_parameters():
+        if not p.requires_grad:
+            continue
+        n_l = n.lower()
+        if any(t in n_l for t in no_decay_tokens):
+            no_decay_params.append(p)
+        else:
+            decay_params.append(p)
+    return [
+        {"params": decay_params, "weight_decay": weight_decay},
+        {"params": no_decay_params, "weight_decay": 0.0},
+    ]
+
+def create_optimizer(name, param_groups):
+    if name == "adam8bit":
+        return bnb.optim.AdamW8bit(param_groups, lr=base_learning_rate, betas=(0.9, beta2), eps=eps)
+    raise ValueError(name)
+
+param_groups = get_param_groups(get_core_model(vae), weight_decay=0.001)
+optimizer = create_optimizer(optimizer_type, param_groups)
+
+# --------------------------- LR schedule ---------------------------
+batches_per_epoch = len(dataloader)
+steps_per_epoch = int(math.ceil(batches_per_epoch / float(gradient_accumulation_steps)))
+total_steps = steps_per_epoch * num_epochs
+
+def lr_lambda(step):
+    if not use_decay:
+        return 1.0
+    x = float(step) / float(max(1, total_steps))
+    warmup = float(warmup_percent)
+    min_ratio = float(min_learning_rate) / float(base_learning_rate)
+    if x < warmup:
+        return min_ratio + (1.0 - min_ratio) * (x / warmup)
+    decay_ratio = (x - warmup) / (1.0 - warmup)
+    return min_ratio + 0.5 * (1.0 - min_ratio) * (1.0 + math.cos(math.pi * decay_ratio))
+
+scheduler = LambdaLR(optimizer, lr_lambda)
+
+# Подготовка
+dataloader, vae, optimizer, scheduler = accelerator.prepare(dataloader, vae, optimizer, scheduler)
+trainable_params = [p for p in vae.parameters() if p.requires_grad]
+
+# --------------------------- LPIPS и вспомогательные ---------------------------
+_lpips_net = None
+def _get_lpips():
+    global _lpips_net
+    if _lpips_net is None:
+        _lpips_net = lpips.LPIPS(net='vgg', verbose=False).eval().to(accelerator.device).eval()
+    return _lpips_net
+
+_sobel_kx = torch.tensor([[[[-1,0,1],[-2,0,2],[-1,0,1]]]], dtype=torch.float32)
+_sobel_ky = torch.tensor([[[[-1,-2,-1],[0,0,0],[1,2,1]]]], dtype=torch.float32)
+def sobel_edges(x: torch.Tensor) -> torch.Tensor:
+    C = x.shape[1]
+    kx = _sobel_kx.to(x.device, x.dtype).repeat(C, 1, 1, 1)
+    ky = _sobel_ky.to(x.device, x.dtype).repeat(C, 1, 1, 1)
+    gx = F.conv2d(x, kx, padding=1, groups=C)
+    gy = F.conv2d(x, ky, padding=1, groups=C)
+    return torch.sqrt(gx * gx + gy * gy + 1e-12)
+
+class MedianLossNormalizer:
+    def __init__(self, desired_ratios: dict, window_steps: int):
+        s = sum(desired_ratios.values())
+        self.ratios = {k: (v / s) if s > 0 else 0.0 for k, v in desired_ratios.items()}
+        self.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
+        self.window = window_steps
+
+    def update_and_total(self, abs_losses: dict):
+        for k, v in abs_losses.items():
+            if k in self.buffers:
+                self.buffers[k].append(float(v.detach().abs().cpu()))
+        meds = {k: (np.median(self.buffers[k]) if len(self.buffers[k]) > 0 else 1.0) for k in self.buffers}
+        coeffs = {k: (self.ratios[k] / max(meds[k], 1e-12)) for k in self.ratios}
+        total = sum(coeffs[k] * abs_losses[k] for k in abs_losses if k in coeffs)
+        return total, coeffs, meds
+
+if full_training and not train_decoder_only:
+    loss_ratios["kl"] = float(kl_ratio)
+normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
+
+# --------------------------- Сэмплы ---------------------------
+@torch.no_grad()
+def get_fixed_samples(n=3):
+    idx = random.sample(range(len(dataset)), min(n, len(dataset)))
+    pil_imgs = [dataset[i] for i in idx]
+    tensors = []
+    for img in pil_imgs:
+        img = random_crop(img, high_resolution)
+        tensors.append(tfm(img))
+    return torch.stack(tensors).to(accelerator.device, dtype)
+
+fixed_samples = get_fixed_samples()
+
+@torch.no_grad()
+def _to_pil_uint8(img_tensor: torch.Tensor) -> Image.Image:
+    arr = ((img_tensor.float().clamp(-1, 1) + 1.0) * 127.5).clamp(0, 255).byte().cpu().numpy().transpose(1, 2, 0)
+    return Image.fromarray(arr)
+
+@torch.no_grad()
+def generate_and_save_samples(step=None):
+    try:
+        temp_vae = accelerator.unwrap_model(vae).eval()
+        lpips_net = _get_lpips()
+        with torch.no_grad():
+            orig_high = fixed_samples
+            orig_low = F.interpolate(orig_high, size=(model_resolution, model_resolution), mode="bilinear", align_corners=False)
+            model_dtype = next(temp_vae.parameters()).dtype
+            orig_low = orig_low.to(dtype=model_dtype)
+
+            # QWEN: добавляем T=1 на encode/decode и снимаем при сравнении
+            if is_video_vae(temp_vae):
+                x_in = orig_low.unsqueeze(2)           # [B,3,1,H,W]
+                enc = temp_vae.encode(x_in)
+                latents_mean = enc.latent_dist.mean
+                dec = temp_vae.decode(latents_mean).sample  # [B,3,1,H,W]
+                rec = dec.squeeze(2)                   # [B,3,H,W]
+            else:
+                enc = temp_vae.encode(orig_low)
+                latents_mean = enc.latent_dist.mean
+                rec = temp_vae.decode(latents_mean).sample
+
+        if rec.shape[-2:] != orig_high.shape[-2:]:
+            rec = F.interpolate(rec, size=orig_high.shape[-2:], mode="bilinear", align_corners=False)
+
+        first_real = _to_pil_uint8(orig_high[0])
+        first_dec  = _to_pil_uint8(rec[0])
+        first_real.save(f"{generated_folder}/sample_real.jpg", quality=95)
+        first_dec.save(f"{generated_folder}/sample_decoded.jpg", quality=95)
+
+        for i in range(rec.shape[0]):
+            _to_pil_uint8(rec[i]).save(f"{generated_folder}/sample_{i}.jpg", quality=95)
+
+        lpips_scores = []
+        for i in range(rec.shape[0]):
+            orig_full = orig_high[i:i+1].to(torch.float32)
+            rec_full  = rec[i:i+1].to(torch.float32)
+            if rec_full.shape[-2:] != orig_full.shape[-2:]:
+                rec_full = F.interpolate(rec_full, size=orig_full.shape[-2:], mode="bilinear", align_corners=False)
+            lpips_val = lpips_net(orig_full, rec_full).item()
+            lpips_scores.append(lpips_val)
+        avg_lpips = float(np.mean(lpips_scores))
+
+        if use_wandb and accelerator.is_main_process:
+            wandb.log({"lpips_mean": avg_lpips}, step=step)
+            wandb.log({
+                "sample/real": wandb.Image(first_real, caption="real"),
+                "sample/decoded": wandb.Image(first_dec, caption="decoded"),
+            }, step=step)
+    finally:
+        gc.collect()
+        torch.cuda.empty_cache()
+
+if accelerator.is_main_process and save_model:
+    print("Генерация сэмплов до старта обучения...")
+    generate_and_save_samples(0)
+
+accelerator.wait_for_everyone()
+
+# --------------------------- Тренировка ---------------------------
+progress = tqdm(total=total_steps, disable=not accelerator.is_local_main_process)
+global_step = 0
+min_loss = float("inf")
+sample_interval = max(1, total_steps // max(1, sample_interval_share * num_epochs))
+
+for epoch in range(num_epochs):
+    vae.train()
+    batch_losses, batch_grads = [], []
+    track_losses = {k: [] for k in loss_ratios.keys()}
+
+    for imgs in dataloader:
+        with accelerator.accumulate(vae):
+            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)
+            else:
+                imgs_low = imgs
+
+            model_dtype = next(vae.parameters()).dtype
+            imgs_low_model = imgs_low.to(dtype=model_dtype) if imgs_low.dtype != model_dtype else imgs_low
+
+            # QWEN: encode/decode с T=1
+            if is_video_vae(vae):
+                x_in = imgs_low_model.unsqueeze(2)             # [B,3,1,H,W]
+                enc = vae.encode(x_in)
+                latents = enc.latent_dist.mean if train_decoder_only else enc.latent_dist.sample()
+                dec = vae.decode(latents).sample               # [B,3,1,H,W]
+                rec = dec.squeeze(2)                           # [B,3,H,W]
+            else:
+                enc = vae.encode(imgs_low_model)
+                latents = enc.latent_dist.mean if train_decoder_only else enc.latent_dist.sample()
+                rec = vae.decode(latents).sample
+
+            if rec.shape[-2:] != imgs.shape[-2:]:
+                rec = F.interpolate(rec, size=imgs.shape[-2:], mode="bilinear", align_corners=False)
+
+            rec_f32 = rec.to(torch.float32)
+            imgs_f32 = imgs.to(torch.float32)
+
+            abs_losses = {
+                "mae":   F.l1_loss(rec_f32, imgs_f32),
+                "mse":   F.mse_loss(rec_f32, imgs_f32),
+                "lpips": _get_lpips()(rec_f32, imgs_f32).mean(),
+                "edge":  F.l1_loss(sobel_edges(rec_f32), sobel_edges(imgs_f32)),
+            }
+
+            if full_training and not train_decoder_only:
+                mean   = enc.latent_dist.mean
+                logvar = enc.latent_dist.logvar
+                kl = -0.5 * torch.mean(1 + logvar - mean.pow(2) - logvar.exp())
+                abs_losses["kl"] = kl
+            else:
+                abs_losses["kl"] = torch.tensor(0.0, device=accelerator.device, dtype=torch.float32)
+
+            total_loss, coeffs, meds = normalizer.update_and_total(abs_losses)
+
+            if torch.isnan(total_loss) or torch.isinf(total_loss):
+                raise RuntimeError("NaN/Inf loss")
+
+            accelerator.backward(total_loss)
+
+            grad_norm = torch.tensor(0.0, device=accelerator.device)
+            if accelerator.sync_gradients:
+                grad_norm = accelerator.clip_grad_norm_(trainable_params, clip_grad_norm)
+                optimizer.step()
+                scheduler.step()
+                optimizer.zero_grad(set_to_none=True)
+                global_step += 1
+                progress.update(1)
+
+            if accelerator.is_main_process:
+                try:
+                    current_lr = optimizer.param_groups[0]["lr"]
+                except Exception:
+                    current_lr = scheduler.get_last_lr()[0]
+
+                batch_losses.append(total_loss.detach().item())
+                batch_grads.append(float(grad_norm.detach().cpu().item()) if isinstance(grad_norm, torch.Tensor) else float(grad_norm))
+                for k, v in abs_losses.items():
+                    track_losses[k].append(float(v.detach().item()))
+
+                if use_wandb and accelerator.sync_gradients:
+                    log_dict = {
+                        "total_loss": float(total_loss.detach().item()),
+                        "learning_rate": current_lr,
+                        "epoch": epoch,
+                        "grad_norm": batch_grads[-1],
+                    }
+                    for k, v in abs_losses.items():
+                        log_dict[f"loss_{k}"] = float(v.detach().item())
+                    for k in coeffs:
+                        log_dict[f"coeff_{k}"] = float(coeffs[k])
+                        log_dict[f"median_{k}"] = float(meds[k])
+                    wandb.log(log_dict, step=global_step)
+
+            if global_step > 0 and global_step % sample_interval == 0:
+                if accelerator.is_main_process:
+                    generate_and_save_samples(global_step)
+                accelerator.wait_for_everyone()
+
+                n_micro = sample_interval * gradient_accumulation_steps
+                avg_loss = float(np.mean(batch_losses[-n_micro:])) if len(batch_losses) >= n_micro else float(np.mean(batch_losses)) if batch_losses else float("nan")
+                avg_grad = float(np.mean(batch_grads[-n_micro:])) if len(batch_grads) >= 1 else float(np.mean(batch_grads)) if batch_grads else 0.0
+
+                if accelerator.is_main_process:
+                    print(f"Epoch {epoch} step {global_step} loss: {avg_loss:.6f}, grad_norm: {avg_grad:.6f}, lr: {current_lr:.9f}")
+                    if save_model and avg_loss < min_loss * save_barrier:
+                        min_loss = avg_loss
+                        accelerator.unwrap_model(vae).save_pretrained(save_as)
+                    if use_wandb:
+                        wandb.log({"interm_loss": avg_loss, "interm_grad": avg_grad}, step=global_step)
+
+    if accelerator.is_main_process:
+        epoch_avg = float(np.mean(batch_losses)) if batch_losses else float("nan")
+        print(f"Epoch {epoch} done, avg loss {epoch_avg:.6f}")
+        if use_wandb:
+            wandb.log({"epoch_loss": epoch_avg, "epoch": epoch + 1}, step=global_step)
+
+# --------------------------- Финальное сохранение ---------------------------
+if accelerator.is_main_process:
+    print("Training finished – saving final model")
+    if save_model:
+        accelerator.unwrap_model(vae).save_pretrained(save_as)
+
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+print("Готово!")

transfer_simplevae.ipynb

@@ -0,0 +1,240 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "c15deb04-94a0-4073-a174-adcd22af10b8",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "✅ Создана новая модель: <class 'diffusers.models.autoencoders.autoencoder_asym_kl.AsymmetricAutoencoderKL'>\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "The config attributes {'block_out_channels': [128, 256, 512, 512, 512], 'force_upcast': False} were passed to AsymmetricAutoencoderKL, but are not expected and will be ignored. Please verify your config.json configuration file.\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "--- Перенос весов ---\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100%|██████████| 248/248 [00:00<00:00, 142199.23it/s]\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "✅ Перенос завершён.\n",
+      "Статистика:\n",
+      "  перенесено: 142\n",
+      "  дублировано: 26\n",
+      "  сдвинуто: 106\n",
+      "  пропущено: 0\n",
+      "\n",
+      "Неперенесённые ключи (первые 20):\n",
+      "   decoder.condition_encoder.layers.0.weight\n",
+      "   decoder.condition_encoder.layers.0.bias\n",
+      "   decoder.condition_encoder.layers.1.weight\n",
+      "   decoder.condition_encoder.layers.1.bias\n",
+      "   decoder.condition_encoder.layers.2.weight\n",
+      "   decoder.condition_encoder.layers.2.bias\n",
+      "   decoder.condition_encoder.layers.3.weight\n",
+      "   decoder.condition_encoder.layers.3.bias\n",
+      "   decoder.condition_encoder.layers.4.weight\n",
+      "   decoder.condition_encoder.layers.4.bias\n"
+     ]
+    }
+   ],
+   "source": [
+    "from diffusers.models import AsymmetricAutoencoderKL, AutoencoderKL\n",
+    "import torch\n",
+    "from tqdm import tqdm\n",
+    "\n",
+    "# ---- Конфиг новой модели ----\n",
+    "config = {\n",
+    "    \"_class_name\": \"AsymmetricAutoencoderKL\",\n",
+    "    \"act_fn\": \"silu\",\n",
+    "    \"in_channels\": 3,\n",
+    "    \"out_channels\": 3,\n",
+    "    \"scaling_factor\": 1.0,\n",
+    "    \"norm_num_groups\": 32,\n",
+    "    \"down_block_out_channels\": [128, 256, 512, 512],\n",
+    "    \"down_block_types\": [\n",
+    "        \"DownEncoderBlock2D\",\n",
+    "        \"DownEncoderBlock2D\",\n",
+    "        \"DownEncoderBlock2D\",\n",
+    "        \"DownEncoderBlock2D\",\n",
+    "    ],\n",
+    "    \"latent_channels\": 16,\n",
+    "    \"up_block_out_channels\": [128, 256, 512, 512, 512],  # +1 блок\n",
+    "    \"up_block_types\": [\n",
+    "        \"UpDecoderBlock2D\",\n",
+    "        \"UpDecoderBlock2D\",\n",
+    "        \"UpDecoderBlock2D\",\n",
+    "        \"UpDecoderBlock2D\",\n",
+    "        \"UpDecoderBlock2D\",\n",
+    "    ],\n",
+    "}\n",
+    "\n",
+    "# ---- Создание пустой асимметричной модели ----\n",
+    "vae = AsymmetricAutoencoderKL(\n",
+    "    act_fn=config[\"act_fn\"],\n",
+    "    down_block_out_channels=config[\"down_block_out_channels\"],\n",
+    "    down_block_types=config[\"down_block_types\"],\n",
+    "    latent_channels=config[\"latent_channels\"],\n",
+    "    up_block_out_channels=config[\"up_block_out_channels\"],\n",
+    "    up_block_types=config[\"up_block_types\"],\n",
+    "    in_channels=config[\"in_channels\"],\n",
+    "    out_channels=config[\"out_channels\"],\n",
+    "    scaling_factor=config[\"scaling_factor\"],\n",
+    "    norm_num_groups=config[\"norm_num_groups\"],\n",
+    "    layers_per_down_block=2,\n",
+    "    layers_per_up_block = 2,\n",
+    "    sample_size = 1024\n",
+    ")\n",
+    "\n",
+    "vae.save_pretrained(\"asymmetric_vae_empty\")\n",
+    "print(\"✅ Создана новая модель:\", type(vae))\n",
+    "\n",
+    "# ---- Функция переноса весов ----\n",
+    "def transfer_weights_with_duplication(old_path, new_path, save_path=\"asymmetric_vae\", device=\"cuda\", dtype=torch.float16):\n",
+    "    old_vae = AutoencoderKL.from_pretrained(old_path,subfolder=\"vae\").to(device, dtype=dtype)\n",
+    "    new_vae = AsymmetricAutoencoderKL.from_pretrained(new_path).to(device, dtype=dtype)\n",
+    "\n",
+    "    old_sd = old_vae.state_dict()\n",
+    "    new_sd = new_vae.state_dict()\n",
+    "\n",
+    "    transferred_keys = set()\n",
+    "    transfer_stats = {\"перенесено\": 0, \"дублировано\": 0, \"сдвинуто\": 0, \"пропущено\": 0}\n",
+    "\n",
+    "    print(\"\\n--- Перенос весов ---\")\n",
+    "\n",
+    "    for k, v in tqdm(old_sd.items()):\n",
+    "        # === Копирование энкодера ===\n",
+    "        if \"encoder\" in k or \"quant_conv\" in k or \"post_quant_conv\" in k:\n",
+    "            if k in new_sd and new_sd[k].shape == v.shape:\n",
+    "                new_sd[k] = v.clone()\n",
+    "                transferred_keys.add(k)\n",
+    "                transfer_stats[\"перенесено\"] += 1\n",
+    "            continue\n",
+    "\n",
+    "        # === Перенос декодера ===\n",
+    "        if \"decoder.up_blocks\" in k:\n",
+    "            parts = k.split(\".\")\n",
+    "            idx = int(parts[2])\n",
+    "\n",
+    "            # сдвигаем индекс на +1 (так как добавлен новый блок в начало)\n",
+    "            new_idx = idx + 1\n",
+    "            new_k = \".\".join([parts[0], parts[1], str(new_idx)] + parts[3:])\n",
+    "            if new_k in new_sd and new_sd[new_k].shape == v.shape:\n",
+    "                new_sd[new_k] = v.clone()\n",
+    "                transferred_keys.add(new_k)\n",
+    "                transfer_stats[\"сдвинуто\"] += 1\n",
+    "            continue\n",
+    "\n",
+    "        # === Перенос прочих совпадающих ключей ===\n",
+    "        if k in new_sd and new_sd[k].shape == v.shape:\n",
+    "            new_sd[k] = v.clone()\n",
+    "            transferred_keys.add(k)\n",
+    "            transfer_stats[\"перенесено\"] += 1\n",
+    "\n",
+    "    # === Дублирование весов старого 512→512 блока в новый ===\n",
+    "    ref_prefix = \"decoder.up_blocks.1\"  # старый первый up-блок (512→512)\n",
+    "    new_prefix = \"decoder.up_blocks.0\"  # новый добавленный блок\n",
+    "    for k, v in old_sd.items():\n",
+    "        if k.startswith(ref_prefix):\n",
+    "            new_k = k.replace(ref_prefix, new_prefix)\n",
+    "            if new_k in new_sd and new_sd[new_k].shape == v.shape:\n",
+    "                new_sd[new_k] = v.clone()\n",
+    "                transferred_keys.add(new_k)\n",
+    "                transfer_stats[\"дублировано\"] += 1\n",
+    "\n",
+    "    # === Загрузка и сохранение ===\n",
+    "    new_vae.load_state_dict(new_sd, strict=False)\n",
+    "    new_vae.save_pretrained(save_path)\n",
+    "\n",
+    "    print(\"\\n✅ Перенос завершён.\")\n",
+    "    print(\"Статистика:\")\n",
+    "    for k, v in transfer_stats.items():\n",
+    "        print(f\"  {k}: {v}\")\n",
+    "\n",
+    "    missing = [k for k in new_sd.keys() if k not in transferred_keys]\n",
+    "    if missing:\n",
+    "        print(\"\\nНеперенесённые ключи (первые 20):\")\n",
+    "        for k in missing[:20]:\n",
+    "            print(\"  \", k)\n",
+    "\n",
+    "# ---- Запуск ----\n",
+    "transfer_weights_with_duplication(\"AiArtLab/simplevae\", \"asymmetric_vae_empty\", save_path=\"vae\")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "65653a65-e7c2-4b67-bc17-62c21cfd1db8",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Collecting hf_transfer\n",
+      "  Downloading hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (1.7 kB)\n",
+      "Downloading hf_transfer-0.1.9-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (3.6 MB)\n",
+      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m3.6/3.6 MB\u001b[0m \u001b[31m34.5 MB/s\u001b[0m  \u001b[33m0:00:00\u001b[0m\n",
+      "\u001b[?25hInstalling collected packages: hf_transfer\n",
+      "Successfully installed hf_transfer-0.1.9\n"
+     ]
+    }
+   ],
+   "source": [
+    "!pip install hf_transfer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "59fcafb9-6d89-49b4-8362-b4891f591687",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

vae/config.json

@@ -0,0 +1,48 @@
+{
+  "_class_name": "AsymmetricAutoencoderKL",
+  "_diffusers_version": "0.35.1",
+  "_name_or_path": "vae",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    512
+  ],
+  "down_block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": false,
+  "in_channels": 3,
+  "latent_channels": 16,
+  "layers_per_down_block": 2,
+  "layers_per_up_block": 2,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "sample_size": 1024,
+  "scaling_factor": 1.0,
+  "up_block_out_channels": [
+    128,
+    256,
+    512,
+    512,
+    512
+  ],
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ]
+}

vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8aef462535483be8d283418a62726921310dd8adcd60ee8418ebea5836316627
+size 444559412