Delete train_style_lora.py

train_style_lora.py

@@ -1,157 +0,0 @@
-#!/usr/bin/env python3
-"""Real, resumable SDXL LoRA trainer for Hugging Face ZeroGPU Spaces.
-
-ZeroGPU allocates the GPU per function call, so training is intentionally split
-into small resumable chunks. Each call resumes from weights/checkpoints/.
-The final REAL LoRA is weights/offex3dstyle.safetensors.
-"""
-from __future__ import annotations
-
-import json
-import os
-import re
-import shutil
-import subprocess
-import sys
-import urllib.request
-from pathlib import Path
-
-import spaces  # import before torch/diffusers in a ZeroGPU Space
-
-ROOT = Path(__file__).resolve().parent
-DATA = ROOT / "training_data"
-OUT = ROOT / "weights"
-CKPT = OUT / "checkpoints"
-TRAIN_SCRIPT = ROOT / "_hf_train_dreambooth_lora_sdxl.py"
-FINAL = OUT / "offex3dstyle.safetensors"
-STATE = OUT / "training_state.json"
-
-MODEL = os.getenv("MODEL_ID", "stabilityai/stable-diffusion-xl-base-1.0")
-HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HF_HUB_TOKEN", "")
-TRAIN_SCRIPT_URL = os.getenv(
-    "DIFFUSERS_TRAIN_SCRIPT_URL",
-    "https://raw.githubusercontent.com/huggingface/diffusers/main/"
-    "examples/dreambooth/train_dreambooth_lora_sdxl.py",
-)
-
-# Keep each ZeroGPU request bounded. Run Train again to continue the next chunk.
-CHUNK_STEPS = int(os.getenv("LORA_CHUNK_STEPS", "40"))
-TOTAL_STEPS = int(os.getenv("MAX_TRAIN_STEPS", "800"))
-RESOLUTION = int(os.getenv("TRAIN_RESOLUTION", "1024"))
-LR = float(os.getenv("LEARNING_RATE", "1e-4"))
-RANK = int(os.getenv("LORA_RANK", "16"))
-BATCH = int(os.getenv("TRAIN_BATCH_SIZE", "1"))
-ACCUM = int(os.getenv("GRAD_ACCUM", "1"))
-SEED = int(os.getenv("TRAIN_SEED", "42"))
-INSTANCE = os.getenv("LORA_TRIGGER", "offex3dstyle")
-
-OUT.mkdir(exist_ok=True)
-CKPT.mkdir(exist_ok=True)
-
-
-def images():
-    return sorted(p for p in DATA.iterdir() if p.suffix.lower() in {".png", ".jpg", ".jpeg", ".webp"})
-
-
-def latest_checkpoint():
-    found = []
-    for p in CKPT.glob("checkpoint-*"):
-        m = re.fullmatch(r"checkpoint-(\d+)", p.name)
-        if m and p.is_dir():
-            found.append((int(m.group(1)), p))
-    return str(max(found)[1]) if found else None
-
-
-def current_step():
-    p = latest_checkpoint()
-    if not p:
-        return 0
-    m = re.search(r"checkpoint-(\d+)$", p)
-    return int(m.group(1)) if m else 0
-
-
-def write_state(status, step, message=""):
-    STATE.write_text(json.dumps({
-        "status": status,
-        "step": step,
-        "total_steps": TOTAL_STEPS,
-        "images": len(images()),
-        "message": message,
-        "final": str(FINAL) if FINAL.exists() else None,
-    }, indent=2), encoding="utf-8")
-
-
-def download_trainer():
-    if not TRAIN_SCRIPT.exists():
-        print("Downloading official Diffusers SDXL LoRA trainer...")
-        urllib.request.urlretrieve(TRAIN_SCRIPT_URL, TRAIN_SCRIPT)
-
-
-def build_command(target_steps, resume):
-    cmd = [
-        sys.executable, str(TRAIN_SCRIPT),
-        "--pretrained_model_name_or_path", MODEL,
-        "--instance_data_dir", str(DATA),
-        "--output_dir", str(OUT),
-        "--instance_prompt", f"a {INSTANCE} stylized 3D character",
-        "--resolution", str(RESOLUTION),
-        "--train_batch_size", str(BATCH),
-        "--gradient_accumulation_steps", str(ACCUM),
-        "--gradient_checkpointing",
-        "--learning_rate", str(LR),
-        "--lr_scheduler", "constant",
-        "--lr_warmup_steps", "0",
-        "--max_train_steps", str(target_steps),
-        "--checkpointing_steps", str(max(10, min(40, CHUNK_STEPS))),
-        "--seed", str(SEED),
-        "--rank", str(RANK),
-        "--mixed_precision", "fp16",
-        "--use_8bit_adam",
-        "--cache_latents",
-    ]
-    if HF_TOKEN:
-        cmd += ["--token", HF_TOKEN]
-    if resume:
-        cmd += ["--resume_from_checkpoint", resume]
-    return cmd
-
-
-@spaces.GPU(duration=60, size="large")
-def train_chunk():
-    """Run one real training chunk on ZeroGPU and return a status message."""
-    imgs = images()
-    if len(imgs) < 5:
-        return f"❌ training_data में कम से कम 5 images चाहिए; मिली {len(imgs)}।"
-    if FINAL.exists():
-        return f"✅ LoRA पहले से मौजूद है: {FINAL}"
-
-    download_trainer()
-    start = current_step()
-    if start >= TOTAL_STEPS:
-        return f"ℹ️ {start}/{TOTAL_STEPS} steps पूरे हैं; final export verify करें।"
-
-    target = min(start + CHUNK_STEPS, TOTAL_STEPS)
-    resume = latest_checkpoint()
-    write_state("training", start, f"Running {start} -> {target}")
-    print(f"REAL SDXL LoRA chunk: {start} -> {target}; resume={resume}")
-
-    try:
-        subprocess.run(build_command(target, resume), check=True)
-    except subprocess.CalledProcessError as e:
-        write_state("error", start, f"trainer exited with code {e.returncode}")
-        return f"❌ Training failed (exit {e.returncode}). Logs में error देखें।"
-
-    exported = OUT / "pytorch_lora_weights.safetensors"
-    if target >= TOTAL_STEPS and exported.exists():
-        shutil.copy2(exported, FINAL)
-        write_state("complete", target, "Final LoRA exported.")
-        return f"🎉 Training complete: {FINAL.name}"
-
-    step = current_step()
-    write_state("paused", step, "Chunk complete; run Train again to continue.")
-    return f"✅ Chunk complete: {step}/{TOTAL_STEPS}. Train फिर दबाओ।"
-
-
-if __name__ == "__main__":
-    print("This trainer is intended to be called from the Space UI via train_chunk().")
-    print("It creates a REAL safetensors LoRA only after actual CUDA training completes.")