FoundationVision/LlamaGen

1import time2import argparse3import torch4from torchvision.utils import save_image5 6from tokenizer.tokenizer_image.vq_model import VQ_models7from serve.gpt_model import GPT_models8from serve.llm import LLM 9from vllm import SamplingParams10 11 12def main(args):13    # Setup PyTorch:14    torch.manual_seed(args.seed)15    torch.backends.cudnn.deterministic = True16    torch.backends.cudnn.benchmark = False17    torch.set_grad_enabled(False)18    device = "cuda" if torch.cuda.is_available() else "cpu"19 20    # create and load model21    vq_model = VQ_models[args.vq_model](22        codebook_size=args.codebook_size,23        codebook_embed_dim=args.codebook_embed_dim)24    vq_model.to(device)25    vq_model.eval()26    checkpoint = torch.load(args.vq_ckpt, map_location="cpu")27    vq_model.load_state_dict(checkpoint["model"])28    del checkpoint29    print(f"image tokenizer is loaded")30 31    # Labels to condition the model with (feel free to change):32    class_labels = [207, 360, 387, 974, 88, 979, 417, 279]33    latent_size = args.image_size // args.downsample_size34    qzshape = [len(class_labels), args.codebook_embed_dim, latent_size, latent_size]35    prompt_token_ids = [[cind] for cind in class_labels]36    if args.cfg_scale > 1.0:37        prompt_token_ids.extend([[args.num_classes] for _ in range(len(prompt_token_ids))])38    # Create an LLM.39    llm = LLM(40        args=args, 41        model='autoregressive/serve/fake_json/{}.json'.format(args.gpt_model), 42        gpu_memory_utilization=0.9, 43        skip_tokenizer_init=True)44    print(f"gpt model is loaded")45 46    # Create a sampling params object.47    sampling_params = SamplingParams(48        temperature=args.temperature, top_p=args.top_p, top_k=args.top_k, 49        max_tokens=latent_size ** 2)50 51    # Generate texts from the prompts. The output is a list of RequestOutput objects52    # that contain the prompt, generated text, and other information.53    t1 = time.time()54    outputs = llm.generate(55        prompt_token_ids=prompt_token_ids,56        sampling_params=sampling_params,57        use_tqdm=False)58    sampling_time = time.time() - t159    print(f"gpt sampling takes about {sampling_time:.2f} seconds.") 60 61    # decode to image62    index_sample = torch.tensor([output.outputs[0].token_ids for output in outputs], device=device)63    if args.cfg_scale > 1.0:64        index_sample = index_sample[:len(class_labels)]65    t2 = time.time()66    samples = vq_model.decode_code(index_sample, qzshape) # output value is between [-1, 1]67    decoder_time = time.time() - t268    print(f"decoder takes about {decoder_time:.2f} seconds.")69 70    # Save and display images:71    save_image(samples, "sample_{}.png".format(args.gpt_type), nrow=4, normalize=True, value_range=(-1, 1))72    print(f"image is saved to sample_{args.gpt_type}.png")73 74 75if __name__ == '__main__':76    parser = argparse.ArgumentParser()77    parser.add_argument("--gpt-model", type=str, choices=list(GPT_models.keys()), default="GPT-B")78    parser.add_argument("--gpt-ckpt", type=str, required=True, help="ckpt path for gpt model")79    parser.add_argument("--gpt-type", type=str, choices=['c2i', 't2i'], default="c2i", help="class-conditional or text-conditional")80    parser.add_argument("--from-fsdp", action='store_true')81    parser.add_argument("--cls-token-num", type=int, default=1, help="max token number of condition input")82    parser.add_argument("--precision", type=str, default='bf16', choices=["none", "fp16", "bf16"])83    parser.add_argument("--compile", action='store_true', default=False)84    parser.add_argument("--vq-model", type=str, choices=list(VQ_models.keys()), default="VQ-16")85    parser.add_argument("--vq-ckpt", type=str, required=True, help="ckpt path for vq model")86    parser.add_argument("--codebook-size", type=int, default=16384, help="codebook size for vector quantization")87    parser.add_argument("--codebook-embed-dim", type=int, default=8, help="codebook dimension for vector quantization")88    parser.add_argument("--image-size", type=int, choices=[256, 384, 512], default=384)89    parser.add_argument("--downsample-size", type=int, choices=[8, 16], default=16)90    parser.add_argument("--num-classes", type=int, default=1000)91    parser.add_argument("--cfg-scale", type=float, default=4.0)92    parser.add_argument("--seed", type=int, default=0)93    parser.add_argument("--top-k", type=int, default=2000,help="top-k value to sample with")94    parser.add_argument("--temperature", type=float, default=1.0, help="temperature value to sample with")95    parser.add_argument("--top-p", type=float, default=1.0, help="top-p value to sample with")96    args = parser.parse_args()97    main(args)98