Spaces:
Sleeping
Sleeping
Download sample.py from wkea/blockdiffusion-api: direct link, hf CLI and curl.
- Browser
- Download file 5.75 kB
-
https://huggingface.co/spaces/wkea/blockdiffusion-api/resolve/main/sample.py
- Command line
-
hf download hf://spaces/wkea/blockdiffusion-api/sample.py
-
curl -L -o sample.py https://huggingface.co/spaces/wkea/blockdiffusion-api/resolve/main/sample.py
5.75 kB
| from __future__ import annotations | |
| import argparse | |
| import os | |
| from typing import Optional | |
| import torch | |
| from blockdiffusion.diffusion.gaussian_diffusion import GaussianDiffusion | |
| from blockdiffusion.diffusion.schedules import get_beta_schedule | |
| from blockdiffusion.models.unet16 import UNet16, UNet16Config | |
| from blockdiffusion.models.text_unet16 import TextCondConfig, TextCondUNet16 | |
| from blockdiffusion.text.simple_tokenizer import SimpleVocab | |
| from blockdiffusion.train.guidance import CFGuidanceModel | |
| from blockdiffusion.utils.checkpoint import load_checkpoint | |
| from blockdiffusion.utils.ema import EMA | |
| from blockdiffusion.utils.image_io import save_tensor_grid | |
| from blockdiffusion.utils.seed import get_device, seed_everything | |
| def parse_args() -> argparse.Namespace: | |
| p = argparse.ArgumentParser(description="BlockDiffusion 16×16 纹理采样入口") | |
| p.add_argument("--ckpt", type=str, required=True, help="训练生成的 checkpoint 路径(.pt)") | |
| p.add_argument("--out_dir", type=str, default="samples", help="输出目录") | |
| p.add_argument("--n", type=int, default=256, help="生成张数") | |
| p.add_argument("--batch", type=int, default=256, help="单次采样 batch(受显存影响)") | |
| p.add_argument("--use_ema", action="store_true", help="使用 EMA 权重采样(推荐)") | |
| p.add_argument("--use_ddim", action="store_true", help="使用 DDIM 快速采样") | |
| p.add_argument("--ddim_steps", type=int, default=50) | |
| p.add_argument("--ddim_eta", type=float, default=0.0) | |
| p.add_argument("--prompt", type=str, default=None, help="提示词(仅条件模型可用)") | |
| p.add_argument("--cfg_scale", type=float, default=3.0, help="CFG scale(仅条件模型可用)") | |
| p.add_argument("--seed", type=int, default=123) | |
| p.add_argument("--device", type=str, default=None) | |
| return p.parse_args() | |
| def main() -> None: | |
| args = parse_args() | |
| seed_everything(args.seed, deterministic=False) | |
| device = get_device(args.device) | |
| ckpt = load_checkpoint(args.ckpt, map_location="cpu") | |
| meta = ckpt.get("meta", {}) | |
| model_type = str(meta.get("model_type", "uncond")) | |
| unet_cfg_dict = meta.get("unet_cfg", None) | |
| if unet_cfg_dict is None: | |
| # 兼容:若旧 checkpoint 未保存配置,则使用默认 16×16 RGB 配置 | |
| unet_cfg = UNet16Config() | |
| timesteps = 1000 | |
| beta_schedule = "cosine" | |
| image_size = 16 | |
| channels = 3 | |
| model_type = "uncond" | |
| else: | |
| unet_cfg = UNet16Config(**unet_cfg_dict) | |
| diff_meta = meta.get("diffusion", {}) | |
| timesteps = int(diff_meta.get("timesteps", 1000)) | |
| beta_schedule = str(diff_meta.get("beta_schedule", "cosine")) | |
| data_meta = meta.get("data", {}) | |
| image_size = int(data_meta.get("image_size", 16)) | |
| channels = int(data_meta.get("channels", unet_cfg.in_channels)) | |
| vocab = None | |
| text_cfg = None | |
| if model_type == "text_cond": | |
| tok_dict = meta.get("tokenizer", None) | |
| text_cfg_dict = meta.get("text_cfg", None) | |
| if tok_dict is None or text_cfg_dict is None: | |
| raise ValueError("该 checkpoint 标记为 text_cond,但缺少 tokenizer/text_cfg") | |
| vocab = SimpleVocab.from_dict(tok_dict) | |
| text_cfg = TextCondConfig(**text_cfg_dict) | |
| model = TextCondUNet16(unet_cfg=unet_cfg, vocab=vocab, text_cfg=text_cfg).to(device) | |
| else: | |
| model = UNet16(unet_cfg).to(device) | |
| model.load_state_dict(ckpt["model"], strict=True) | |
| if args.use_ema and ("ema" in ckpt): | |
| ema = EMA(model, decay=0.9999) | |
| ema.load_state_dict(ckpt["ema"]) | |
| ema.copy_to(model) | |
| betas = get_beta_schedule(beta_schedule, timesteps).to(device) | |
| diffusion = GaussianDiffusion(betas) | |
| os.makedirs(args.out_dir, exist_ok=True) | |
| # 条件采样:如果提供 prompt 且模型支持,则启用 CFG | |
| sampler_model = model | |
| model_kwargs = None | |
| if args.prompt and model_type == "text_cond": | |
| if vocab is None or text_cfg is None: | |
| raise ValueError("内部错误:条件模型缺少 vocab/text_cfg") | |
| tok = torch.tensor([vocab.encode(args.prompt, max_len=text_cfg.max_text_len)], dtype=torch.long, device=device) | |
| m = (tok != vocab.pad_id) | |
| cond_kwargs = {"tokens": tok, "mask": m} | |
| un_tok = torch.full_like(tok, fill_value=vocab.pad_id) | |
| un_m = torch.zeros_like(m) | |
| uncond_kwargs = {"tokens": un_tok, "mask": un_m} | |
| sampler_model = CFGuidanceModel(model, cond_kwargs=cond_kwargs, uncond_kwargs=uncond_kwargs, scale=args.cfg_scale).to(device) | |
| model_kwargs = None | |
| elif args.prompt and model_type != "text_cond": | |
| raise ValueError("当前 checkpoint 是无条件模型,不支持 --prompt;请用带 captions 训练的模型") | |
| all_imgs = [] | |
| remaining = args.n | |
| while remaining > 0: | |
| b = min(args.batch, remaining) | |
| shape = (b, channels, image_size, image_size) | |
| if args.use_ddim: | |
| x = diffusion.ddim_sample_loop( | |
| sampler_model, | |
| shape=shape, | |
| device=device, | |
| steps=args.ddim_steps, | |
| eta=args.ddim_eta, | |
| model_kwargs=model_kwargs, | |
| progress=False, | |
| ) | |
| else: | |
| x = diffusion.p_sample_loop(sampler_model, shape=shape, device=device, model_kwargs=model_kwargs, progress=False) | |
| x = (x.clamp(-1, 1) + 1) * 0.5 | |
| all_imgs.append(x.cpu()) | |
| remaining -= b | |
| imgs = torch.cat(all_imgs, dim=0) | |
| out_path = os.path.join(args.out_dir, "samples.png") | |
| save_tensor_grid(imgs, out_path, nrow=int(max(1, args.n) ** 0.5), padding=1) | |
| if __name__ == "__main__": | |
| main() | |