Anonymous-123/ImageNet-Editing

1"""2Like image_sample.py, but use a noisy image classifier to guide the sampling3process towards more realistic images.4"""5 6import argparse7import os8 9import numpy as np10import torch as th11import torch.distributed as dist12import torch.nn.functional as F13 14from guided_diffusion import dist_util, logger15from guided_diffusion.script_util import (16    NUM_CLASSES,17    model_and_diffusion_defaults,18    classifier_defaults,19    create_model_and_diffusion,20    create_classifier,21    add_dict_to_argparser,22    args_to_dict,23)24 25 26def main():27    args = create_argparser().parse_args()28 29    dist_util.setup_dist()30    logger.configure()31 32    logger.log("creating model and diffusion...")33    model, diffusion = create_model_and_diffusion(34        **args_to_dict(args, model_and_diffusion_defaults().keys())35    )36    model.load_state_dict(37        dist_util.load_state_dict(args.model_path, map_location="cpu")38    )39    model.to(dist_util.dev())40    if args.use_fp16:41        model.convert_to_fp16()42    model.eval()43 44    logger.log("loading classifier...")45    classifier = create_classifier(**args_to_dict(args, classifier_defaults().keys()))46    classifier.load_state_dict(47        dist_util.load_state_dict(args.classifier_path, map_location="cpu")48    )49    classifier.to(dist_util.dev())50    if args.classifier_use_fp16:51        classifier.convert_to_fp16()52    classifier.eval()53 54    def cond_fn(x, t, y=None):55        assert y is not None56        with th.enable_grad():57            x_in = x.detach().requires_grad_(True)58            logits = classifier(x_in, t)59            log_probs = F.log_softmax(logits, dim=-1)60            selected = log_probs[range(len(logits)), y.view(-1)]61            return th.autograd.grad(selected.sum(), x_in)[0] * args.classifier_scale62 63    def model_fn(x, t, y=None):64        assert y is not None65        return model(x, t, y if args.class_cond else None)66 67    logger.log("sampling...")68    all_images = []69    all_labels = []70    while len(all_images) * args.batch_size < args.num_samples:71        model_kwargs = {}72        classes = th.randint(73            low=0, high=NUM_CLASSES, size=(args.batch_size,), device=dist_util.dev()74        )75        model_kwargs["y"] = classes76        sample_fn = (77            diffusion.p_sample_loop if not args.use_ddim else diffusion.ddim_sample_loop78        )79        sample = sample_fn(80            model_fn,81            (args.batch_size, 3, args.image_size, args.image_size),82            clip_denoised=args.clip_denoised,83            model_kwargs=model_kwargs,84            cond_fn=cond_fn,85            device=dist_util.dev(),86        )87        sample = ((sample + 1) * 127.5).clamp(0, 255).to(th.uint8)88        sample = sample.permute(0, 2, 3, 1)89        sample = sample.contiguous()90 91        gathered_samples = [th.zeros_like(sample) for _ in range(dist.get_world_size())]92        dist.all_gather(gathered_samples, sample)  # gather not supported with NCCL93        all_images.extend([sample.cpu().numpy() for sample in gathered_samples])94        gathered_labels = [th.zeros_like(classes) for _ in range(dist.get_world_size())]95        dist.all_gather(gathered_labels, classes)96        all_labels.extend([labels.cpu().numpy() for labels in gathered_labels])97        logger.log(f"created {len(all_images) * args.batch_size} samples")98 99    arr = np.concatenate(all_images, axis=0)100    arr = arr[: args.num_samples]101    label_arr = np.concatenate(all_labels, axis=0)102    label_arr = label_arr[: args.num_samples]103    if dist.get_rank() == 0:104        shape_str = "x".join([str(x) for x in arr.shape])105        out_path = os.path.join(logger.get_dir(), f"samples_{shape_str}.npz")106        logger.log(f"saving to {out_path}")107        np.savez(out_path, arr, label_arr)108 109    dist.barrier()110    logger.log("sampling complete")111 112 113def create_argparser():114    defaults = dict(115        clip_denoised=True,116        num_samples=10000,117        batch_size=16,118        use_ddim=False,119        model_path="",120        classifier_path="",121        classifier_scale=1.0,122    )123    defaults.update(model_and_diffusion_defaults())124    defaults.update(classifier_defaults())125    parser = argparse.ArgumentParser()126    add_dict_to_argparser(parser, defaults)127    return parser128 129 130if __name__ == "__main__":131    main()132