Function super_resolution

generate_samples.py:223–244

(model, args, raw_text, seq, output_path="./samples")

            generate_images_once(model, args, raw_text, seq, num=args.batch_size, output_path=output_path)

def super_resolution(model, args, raw_text, seq, output_path="./samples"):
    tokenizer = get_tokenizer()
    model.eval()
    if not os.path.exists(output_path):
        os.makedirs(output_path)
    with torch.no_grad():
        start_time = time.time()
        output_tokens_list = magnify(model, tokenizer, seq[-32**2:], seq[:-32**2], args)

        print("\nTaken time {:.2f}\n".format(time.time() - start_time), flush=True)
        print("\nContext:", raw_text, flush=True)
        output_file_prefix = raw_text.replace('/', '')[:20]
        output_file = os.path.join(output_path, f"{output_file_prefix}-{datetime.now().strftime('%m-%d-%H-%M-%S')}.jpg")
        imgs = []
        if args.debug:
            imgs.append(torch.nn.functional.interpolate(tokenizer.img_tokenizer.DecodeIds(seq[-32**2:]), size=(512, 512)))
        for seq in output_tokens_list:
            decoded_txts, decoded_imgs = tokenizer.DecodeIds(seq.tolist())
            imgs.extend(decoded_imgs)
        imgs = torch.cat(imgs, dim=0)
        print("\nSave to: ", output_file, flush=True)
        save_image(imgs, output_file, normalize=True)

def post_selection(model, args, raw_text, seq, output_path):
    tokenizer = get_tokenizer()