demo_motionstreamer.py

import os
import torch
import numpy as np
from models.llama_model import LLaMAHF, LLaMAHFConfig
import models.tae as tae
import options.option_transformer as option_trans
import warnings

import smplx
from utils import bvh, quat
from utils.face_z_align_util import rotation_6d_to_matrix, matrix_to_axis_angle, axis_angle_to_quaternion
from sentence_transformers import SentenceTransformer
warnings.filterwarnings('ignore')

# --- save_motion_as_bvh function is unchanged ---
def save_motion_as_bvh(motion_data, output_path, fps=30):
    print(f"--- Starting direct conversion to BVH: {os.path.basename(output_path)} ---")
    try:
        if isinstance(motion_data, torch.Tensor): motion_data = motion_data.detach().cpu().numpy()
        if motion_data.ndim == 3 and motion_data.shape[0] == 1: motion_data = motion_data.squeeze(0)
        elif motion_data.ndim != 2: raise ValueError(f"Input motion data must be 2D, but got shape {motion_data.shape}")
        njoint = 22; nfrm, _ = motion_data.shape
        rotations_matrix = rotation_6d_to_matrix(torch.from_numpy(motion_data[:, 8+6*njoint : 8+12*njoint]).reshape(nfrm, -1, 6)).numpy()
        global_heading_diff_rot_6d = torch.from_numpy(motion_data[:, 2:8])
        global_heading_diff_rot = rotation_6d_to_matrix(global_heading_diff_rot_6d).numpy()
        global_heading_rot = np.zeros_like(global_heading_diff_rot); global_heading_rot[0] = global_heading_diff_rot[0]
        for i in range(1, nfrm): global_heading_rot[i] = np.matmul(global_heading_diff_rot[i], global_heading_rot[i-1])
        velocities_root_xy = motion_data[:, :2]; height = motion_data[:, 8 : 8+3*njoint].reshape(nfrm, -1, 3)[:, 0, 1]
        inv_global_heading_rot = np.transpose(global_heading_rot, (0, 2, 1)); rotations_matrix[:, 0, ...] = np.matmul(inv_global_heading_rot, rotations_matrix[:, 0, ...])
        velocities_root_xyz = np.zeros((nfrm, 3)); velocities_root_xyz[:, 0] = velocities_root_xy[:, 0]; velocities_root_xyz[:, 2] = velocities_root_xy[:, 1]
        velocities_root_xyz[1:, :] = np.matmul(inv_global_heading_rot[:-1], velocities_root_xyz[1:, :, None]).squeeze(-1)
        root_translation = np.cumsum(velocities_root_xyz, axis=0); root_translation[:, 1] = height
        axis_angle = matrix_to_axis_angle(torch.from_numpy(rotations_matrix)).numpy().reshape(nfrm, -1); poses_24_joints = np.zeros((nfrm, 72)); poses_24_joints[:, :66] = axis_angle
        model = smplx.create(model_path="body_models/human_model_files", model_type="smpl", gender="NEUTRAL"); parents = model.parents.detach().cpu().numpy()
        rest_pose = model().joints.detach().cpu().numpy().squeeze()[:24,:]; offsets = rest_pose - rest_pose[parents]; offsets[0] = np.array([0,0,0])
        rotations_quat = axis_angle_to_quaternion(torch.from_numpy(poses_24_joints.reshape(-1, 24, 3))).numpy(); rotations_euler = np.degrees(quat.to_euler(rotations_quat, order="zyx"))
        positions = np.zeros_like(rotations_quat[..., :3]); positions[:, 0] = root_translation
        joint_names = ["Pelvis", "Left_hip", "Right_hip", "Spine1", "Left_knee", "Right_knee", "Spine2", "Left_ankle", "Right_ankle", "Spine3", "Left_foot", "Right_foot", "Neck", "Left_collar", "Right_collar", "Head", "Left_shoulder", "Right_shoulder", "Left_elbow", "Right_elbow", "Left_wrist", "Right_wrist", "Left_hand", "Right_hand"]
        bvh.save(output_path, {"rotations": rotations_euler, "positions": positions, "offsets": offsets, "parents": parents, "names": joint_names, "order": "zyx", "frametime": 1.0 / fps})
        print(f"✅ BVH file saved successfully to {output_path}")
    except Exception as e:
        print(f"❌ BVH Conversion Failed. Error: {e}"); import traceback; traceback.print_exc()


if __name__ == '__main__':
    comp_device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    args = option_trans.get_args_parser()
    torch.manual_seed(args.seed)

    # --- Load Models ---
    print("Loading models for MotionStreamer...")
    t5_model = SentenceTransformer('sentencet5-xxl/')
    t5_model.eval()
    for p in t5_model.parameters():
        p.requires_grad = False

    print("Loading Causal TAE (t2m_babel) checkpoint...")
    tae_net = tae.Causal_HumanTAE(
        hidden_size=1024, down_t=2, stride_t=2, depth=3, dilation_growth_rate=3,
        latent_dim=16, clip_range=[-30, 20]
    )
    tae_ckpt = torch.load('Causal_TAE_t2m_babel/net_last.pth', map_location='cpu')
    tae_net.load_state_dict(tae_ckpt['net'], strict=True)
    tae_net.eval()
    tae_net.to(comp_device)

    config = LLaMAHFConfig.from_name('Normal_size')
    config.block_size = 78
    trans_encoder = LLaMAHF(config, args.num_diffusion_head_layers, args.latent_dim, comp_device)

    # --- THIS IS THE FIX ---
    print("Loading your trained MotionStreamer checkpoint from 'motionstreamer_model/latest.pth'...")
    # Make sure this path is correct relative to where you run the script
    checkpoint_path = 'motionstreamer_model/latest.pth'
    trans_ckpt = torch.load(checkpoint_path, map_location='cpu')

    # Create a new state dict without the 'module.' prefix
    unwrapped_state_dict = {}
    for key, value in trans_ckpt['trans'].items():
        if key.startswith('module.'):
            # Strip the 'module.' prefix
            unwrapped_state_dict[key[len('module.'):]] = value
        else:
            # Keep keys that don't have the prefix (just in case)
            unwrapped_state_dict[key] = value

    # Load the unwrapped state dict
    trans_encoder.load_state_dict(unwrapped_state_dict, strict=True)
    print("Successfully loaded unwrapped checkpoint.")
    # --- END FIX ---

    trans_encoder.eval()
    trans_encoder.to(comp_device)

    # --- Rest of the script is unchanged ---
    print("Loading mean/std from BABEL dataset...")
    mean = np.load('babel_272/t2m_babel_mean_std/Mean.npy')
    std = np.load('babel_272/t2m_babel_mean_std/Std.npy')

    motion_history = torch.empty(0, 16).to(comp_device)
    cfg_scale = 10.0

    print(f"Generating motion for text: '{args.text}' with CFG scale: {cfg_scale}")
    with torch.no_grad():
        # Use the new two-forward sampling method to match training
        _, motion_latents = trans_encoder.sample_for_eval_CFG_babel_inference_two_forward(
            B_text=args.text,
            A_motion=motion_history,
            tokenizer='t5-xxl',
            clip_model=t5_model,
            device=comp_device,
            cfg=cfg_scale,
            length=240,
            temperature=1.3
        )

        print("Decoding latents to full motion...")
        motion_seqs = tae_net.forward_decoder(motion_latents)

    motion = motion_seqs.detach().cpu().numpy()
    motion_denormalized = motion * std + mean

    output_dir = 'demo_output_streamer'
    if not os.path.exists(output_dir): os.makedirs(output_dir)

    output_bvh_path = os.path.join(output_dir, f'{args.text.replace(" ", "_")}_cfg{cfg_scale}.bvh')
    save_motion_as_bvh(motion_denormalized, output_bvh_path, fps=30)