tools/data_process/motionmillion/gen_motion_embedding.py

import sys
sys.path.append('/mnt/shenzhen2cephfs/capybarali/codes/neobot')

import numpy as np
from tqdm import tqdm
import torch, os, random
from mmengine.config import Config
from mmengine.registry import MODELS
import src
import joblib

device = torch.device('cuda')
tmr_cfg_path = 'ckpts/mm_tmr_400e/tmr.py'
tmr_cfg_ckpt_path = 'ckpts/mm_tmr_400e/epoch_400.pth'
mean_path = 'data/motionmillion/motion_272_mean.npy'
std_path = 'data/motionmillion/motion_272_std.npy'

class MotionDataset(torch.utils.data.Dataset):
    def __init__(self, data_root, data_path_file, mean_path, std_path):
        self.data_list = []
        self.data_ids = []
        data_paths = joblib.load(data_path_file)
        for path in data_paths:
            data_id = path.replace('/mnt/shenzhen2cephfs/capybarali/codes/neobot/data/omnihumo/2_gmr_retarget', '')
            path = data_root + data_id.replace('.npz', '')
            self.data_list.append(path)
            self.data_ids.append(data_id)

        self.mean = torch.from_numpy(np.load(mean_path))
        self.std = torch.from_numpy(np.load(std_path))
        self.unit_length = 4
    
    def __len__(self):
        return len(self.data_list)
    
    def __getitem__(self, idx):
        data_path = self.data_list[idx]
        motion = np.load(data_path)

        T = motion.shape[0]

        motion_length = (T // self.unit_length) * self.unit_length
        idx = random.randint(0, T - motion_length)
        motion = torch.from_numpy(motion[idx:idx+motion_length])
        
        motion = (motion - self.mean) / self.std

        return motion, motion_length, self.data_ids[idx]

def collate_fn(batch):
    motion = torch.nn.utils.rnn.pad_sequence([x[0] for x in batch], batch_first=True).float()
    motion_length = torch.tensor([x[1] for x in batch])
    data_id = [x[2] for x in batch]
    return motion, motion_length, data_id

def l2_normalize(x, axis=1, eps=1e-8):
    return x / (np.linalg.norm(x, axis=axis, keepdims=True) + eps)

def kmeans_cosine(X, K, max_iters=100, tol=1e-4, seed=0):
    np.random.seed(seed)

    # 1. L2 normalize data
    X = l2_normalize(X)

    T, C = X.shape

    # 2. random init centroids
    indices = np.random.choice(T, K, replace=False)
    centroids = X[indices]

    for it in range(max_iters):
        # 3. cosine similarity = dot product (since normalized)
        sim = X @ centroids.T              # (T, K)
        labels = np.argmax(sim, axis=1)    # assign by max cosine

        # 4. update centroids
        new_centroids = np.zeros_like(centroids)
        for k in range(K):
            members = X[labels == k]
            if len(members) == 0:
                # empty cluster → reinit
                new_centroids[k] = X[np.random.randint(T)]
            else:
                new_centroids[k] = members.mean(axis=0)

        # 5. normalize centroids
        new_centroids = l2_normalize(new_centroids)

        # 6. convergence check
        shift = np.linalg.norm(centroids - new_centroids)
        centroids = new_centroids

        # ===== 新增：类内相似度统计 =====
        cluster_avg_sim = np.zeros(K)
        cluster_sizes = np.zeros(K, dtype=int)
        for k in range(K):
            members = X[labels == k]
            cluster_sizes[k] = len(members)
            if len(members) == 0:
                cluster_avg_sim[k] = np.nan
            else:
                cluster_avg_sim[k] = (members @ centroids[k]).mean()
        print(cluster_avg_sim.min(), cluster_avg_sim.max(), cluster_avg_sim.mean())
        print(shift)

        if shift < tol:
            break

    return labels, centroids

def main():
    # build model
    tmr_cfg = Config.fromfile(tmr_cfg_path)
    tmr_ckpt = torch.load(tmr_cfg_ckpt_path, weights_only=False, map_location='cpu')
    tmr = MODELS.build(tmr_cfg.model)
    tmr.load_state_dict(tmr_ckpt['state_dict'])
    tmr.to(device).eval()

    # build dataset
    dataset = MotionDataset(data_root='/mnt/shenzhen2cephfs/capybarali/omnihumo/processed_results/omnihumo_272', 
                            data_path_file='data/omnihumo/filtered_gmr_retarget_path.pkl', 
                            mean_path=mean_path, std_path=std_path)
    dataloader = torch.utils.data.DataLoader(dataset, batch_size=256, shuffle=False, collate_fn=collate_fn, 
                                         num_workers=4, drop_last=False)

    all_m_letent = []
    for data in tqdm(dataloader):
        motion, motion_length, data_id = data

        max_len = motion_length.max()
        mask = torch.arange(max_len, device=motion.device).expand(
            motion_length.shape[0], max_len) < motion_length.unsqueeze(1)
        
        with torch.no_grad():
            m_latents = tmr.motion_encoder(motion.to(device), mask=mask.to(device))[:, 0]
        all_m_letent.append(m_latents)
    all_m_letent = torch.concat(all_m_letent, dim=0) # T, C
    labels, centroids = kmeans_cosine(all_m_letent.cpu().numpy(), K=128, max_iters=1000, tol=1e-4, seed=0)
    joblib.dump(labels, 'labels.pkl')

if __name__ == '__main__':
    main()