"""Run the original RAW/EMA validation from an existing RFM checkpoint.""" import argparse from copy import deepcopy import hashlib import json from pathlib import Path import runpy import torch from diffusers import AutoencoderKL from diffusion.model.builder import build_model from scripts.train_rfm import validate def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('--config', help='Python config with resource paths and validation sampling settings') parser.add_argument('--checkpoint', required=True) parser.add_argument('--output', required=True) parser.add_argument('--conditioning-cache') parser.add_argument('--vae-path') args = parser.parse_args() output = Path(args.output) if output.exists() and any(output.iterdir()): raise FileExistsError(f'Use an empty output directory: {output}') checkpoint = torch.load(args.checkpoint, map_location='cpu', weights_only=False, mmap=True) metadata = checkpoint['metadata'] for key, expected in dict(format_version='rfm_v1', prediction_type='flow_velocity', path='linear', time_embedding='999*(1-t)', vae_scaling_factor=.18215).items(): if metadata.get(key) != expected: raise ValueError(f'Incompatible RFM checkpoint: {key}') config = dict(metadata['training_config']) conditioning_hash = metadata['manifest_hashes'][config['conditioning_cache']] if args.config: runtime = runpy.run_path(args.config)['config'] for key in ('model', 'image_size', 'prediction_type', 'path', 'time_embedding', 'out_channels', 'qk_norm'): if key in runtime and runtime[key] != config.get(key): raise ValueError(f'Config differs from checkpoint architecture/convention: {key}') for key in ('conditioning_cache', 'vae_path', 'cfg_scale', 'sampling_method', 'sampling_steps', 'sampling_atol', 'sampling_rtol'): if key in runtime: config[key] = runtime[key] config['work_dir'] = str(output) if args.conditioning_cache: config['conditioning_cache'] = args.conditioning_cache if args.vae_path: config['vae_path'] = args.vae_path if hashlib.sha256(Path(config['conditioning_cache']).read_bytes()).hexdigest() != conditioning_hash: raise ValueError('Conditioning cache differs from the original training cache') text = torch.load(config['conditioning_cache'], map_location='cpu', weights_only=True) if text['seeds'] != config['validation_seeds'] or len(text['prompts']) != 4: raise ValueError('Conditioning prompts/seeds do not match the training configuration') torch.set_num_threads(4) torch.backends.cuda.matmul.allow_tf32 = True model = build_model(config['model'], input_size=64, pred_sigma=False, model_max_length=120, class_dropout_prob=0., qk_norm=config['qk_norm'], use_grad_checkpoint=False, lewei_scale=1.).float() model.load_state_dict(checkpoint['state_dict'], strict=True) ema = deepcopy(model) ema.load_state_dict(checkpoint['state_dict_ema'], strict=True) for network in (model, ema): if not torch.equal(network.y_embedder.y_embedding, text['empty'][0, 0].float()): raise ValueError('Unconditional embedding differs from the training cache') network.cuda().eval().requires_grad_(False) vae = AutoencoderKL.from_pretrained(config['vae_path'], local_files_only=True) vae = vae.cuda().float().eval().requires_grad_(False) if abs(vae.config.scaling_factor - .18215) > 1e-6: raise ValueError('Unexpected VAE scaling factor') step = checkpoint['progress']['optimizer_step'] del checkpoint report = validate(model, ema, vae, text, config, step, torch.device('cuda')) (output / 'validation_config.json').write_text(json.dumps(config, indent=2) + '\n') print(json.dumps(report, indent=2)) if __name__ == '__main__': main()