import torch import os class ZeroVelocityBaseline(torch.nn.Module): def __init__(self, **kwargs): super(ZeroVelocityBaseline, self).__init__(**kwargs) def forward(self, x: torch.Tensor, ph: int = 1) -> torch.Tensor: last_frame = x[..., -1, :, :].unsqueeze(-3).clone() B, T, J, D = last_frame.shape last_frame = last_frame.broadcast_to((B, ph, J, D)) return last_frame def load_model_config_exp(checkpoint_path): """ Load the config file and the experiment folder from the checkpoint_path exp_folder: is a absolute or relative path cfg: dict of configuration """ exp_folder = checkpoint_path return {}, exp_folder def get_eval_out_folder(exp_folder, checkpoint_path, data_split, cfg): """ Create folder where to store the evaluation results """ stats_folder = os.path.join( exp_folder, f"eval_seed{cfg['seed']}", data_split) os.makedirs(stats_folder, exist_ok=True) return stats_folder def prepare_model(config, skeleton, silent=False, **kwargs): """ This function must take as input a config_file and at least a second arg (skeleton) and return a laoded model on a device (cpu or cuda) model, device = prepare_model(config, skeleton, **kwargs) """ for i in range(torch.cuda.device_count()): if not silent: print(f"> GPU {i} ready: {torch.cuda.get_device_name(i)}") device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.set_grad_enabled(False) baseline_name = config['method_name'] alg_baseline = globals()[baseline_name]() alg_baseline = alg_baseline.to(device) alg_baseline.eval() return alg_baseline, device def get_prediction(obs, model, num_samples=50, **kwargs): """ Generate num_samples predictions from model and input obs. return a dict or anything that will be used by process_evaluation_pair """ pred = model(obs, ph=kwargs['pred_length']) pred = pred.unsqueeze(1) # [batch, 1, t, n_joints, 3] return pred def process_evaluation_pair(dataset, target, pred_dict, **kwargs): """ Process the target and the prediction and return them in the right format for metrics computation """ skeleton = dataset.skeleton pred, mm_gt, obs = pred_dict['pred'], pred_dict['mm_gt'], pred_dict['obs'] batch_size, n_samples, seq_length, num_joints, features = pred.shape target = skeleton.transform_to_metric_space(target) pred = skeleton.transform_to_metric_space(pred) obs = skeleton.transform_to_metric_space(obs) mm_gt = [skeleton.transform_to_metric_space( gt) for gt in mm_gt] if mm_gt is not None else None # batch_size, n_samples, n_diffusion_steps, seq_length, num_joints, features = pred.shape assert features == 3 and list(target.shape) == [ batch_size, seq_length, num_joints, features] # return dummy altent prediction return target, pred, mm_gt, obs