blockdiffusion-api / sample_cpu.py
wkea's picture
Deploy Space
c388140 verified
Raw History Blame Contribute Delete
7.57 kB
from __future__ import annotations
"""
BlockDiffusion:CPU 推理/采样脚本(强制使用 CPU)
设计目标:
- 即使机器有 GPU,也只使用 CPU(避免 CUDA 初始化与显存占用)
- 复用项目现有的 ckpt 元信息/模型构建/EMA/CFG/扩散采样逻辑
注意:
- CPU 采样会明显更慢,建议配合 --use_ddim 与较小的 --ddim_steps / --batch
"""
import argparse
import os
from typing import Optional
# 必须在 import torch 之前设置,否则可能已经触发 CUDA 初始化
os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
import torch
from blockdiffusion.diffusion.gaussian_diffusion import GaussianDiffusion
from blockdiffusion.diffusion.schedules import get_beta_schedule
from blockdiffusion.models.text_unet16 import TextCondConfig, TextCondUNet16
from blockdiffusion.models.unet16 import UNet16, UNet16Config
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 seed_everything
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description="BlockDiffusion CPU 采样入口(16×16 纹理)")
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_cpu", help="输出目录")
p.add_argument("--out_name", type=str, default="samples.png", help="输出文件名(建议 .png)")
p.add_argument("--n", type=int, default=16, help="生成张数(CPU 建议较小)")
p.add_argument("--batch", type=int, default=16, help="单次采样 batch(CPU 建议较小)")
p.add_argument("--threads", type=int, default=0, help="torch CPU 线程数(0 表示不设置)")
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, help="DDIM steps(越小越快)")
p.add_argument("--ddim_eta", type=float, default=0.0, help="DDIM eta(0=确定性)")
p.add_argument("--prompt", type=str, default=None, help="提示词(仅 text_cond 模型可用)")
p.add_argument("--cfg_scale", type=float, default=3.0, help="CFG scale(仅 text_cond 模型可用)")
p.add_argument("--seed", type=int, default=123, help="随机种子")
return p.parse_args()
def _set_cpu_threads(threads: int) -> None:
"""
可选设置 CPU 线程数,便于在不同机器上控制推理占用。
"""
threads = int(threads)
if threads <= 0:
return
try:
torch.set_num_threads(threads)
except Exception:
# 某些环境可能限制 set_num_threads;这里静默忽略即可
pass
try:
# inter-op 线程在部分版本/平台上可能不可设置
torch.set_num_interop_threads(threads)
except Exception:
pass
@torch.no_grad()
def main() -> None:
args = parse_args()
_set_cpu_threads(args.threads)
seed_everything(args.seed, deterministic=False)
# 强制 CPU
device = torch.device("cpu")
ckpt = load_checkpoint(args.ckpt, map_location="cpu")
meta = ckpt.get("meta", {}) or {}
model_type = str(meta.get("model_type", "uncond"))
unet_cfg_dict = meta.get("unet_cfg", None)
if unet_cfg_dict is None:
# 兼容:旧 checkpoint 未保存配置
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", {}) or {}
timesteps = int(diff_meta.get("timesteps", 1000))
beta_schedule = str(diff_meta.get("beta_schedule", "cosine"))
data_meta = meta.get("data", {}) or {}
image_size = int(data_meta.get("image_size", 16))
channels = int(data_meta.get("channels", unet_cfg.in_channels))
vocab: Optional[SimpleVocab] = None
text_cfg: Optional[TextCondConfig] = 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)
model.eval()
if args.use_ema and ("ema" in ckpt):
ema = EMA(model, decay=0.9999)
ema.load_state_dict(ckpt["ema"])
ema.copy_to(model)
model.eval()
betas = get_beta_schedule(beta_schedule, timesteps).to(device)
diffusion = GaussianDiffusion(betas)
# 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=float(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")
os.makedirs(args.out_dir, exist_ok=True)
out_path = os.path.join(args.out_dir, args.out_name)
all_imgs = []
remaining = int(args.n)
while remaining > 0:
b = min(int(args.batch), remaining)
shape = (b, int(channels), int(image_size), int(image_size))
if args.use_ddim:
x = diffusion.ddim_sample_loop(
sampler_model,
shape=shape,
device=device,
steps=int(args.ddim_steps),
eta=float(args.ddim_eta),
model_kwargs=model_kwargs,
progress=True,
)
else:
x = diffusion.p_sample_loop(
sampler_model,
shape=shape,
device=device,
model_kwargs=model_kwargs,
progress=True,
)
x = (x.clamp(-1, 1) + 1) * 0.5
all_imgs.append(x.cpu())
remaining -= b
imgs = torch.cat(all_imgs, dim=0)
nrow = int(max(1, int(args.n)) ** 0.5)
save_tensor_grid(imgs, out_path, nrow=nrow, padding=1)
if __name__ == "__main__":
main()