Human Motion Diffusion Model 使用笔记

目录

依赖项：

依赖项目：

验证数据集

体验网址：

推理代码整理，一键运行

渲染mesh

可视化代码：

---

GitHub - GuyTevet/motion-diffusion-model: The official PyTorch implementation of the paper "Human Motion Diffusion Model"

依赖项：

spacy

依赖项目：

GitHub - EricGuo5513/HumanML3D: HumanML3D: A large and diverse 3d human motion-language dataset.

验证数据集

HumanML3D/HumanML3D at main · EricGuo5513/HumanML3D · GitHub

第1步，把目录 HumanML3D/HumanML3D拷贝到dataset目录中，

第2步，把HumanML3D数据集中的texts.zip解压到目录texts

体验网址：

daanelson/motion_diffusion_model – Run with an API on Replicate

推理代码整理，一键运行

# This code is based on https://github.com/openai/guided-diffusion
"""
Generate a large batch of image samples from a model and save them as a large
numpy array. This can be used to produce samples for FID evaluation.
"""
from utils.fixseed import fixseed
import os
import numpy as np
import torch
from utils.parser_util import generate_args
from utils.model_util import create_model_and_diffusion, load_model_wo_clip
from utils import dist_util
from model.cfg_sampler import ClassifierFreeSampleModel
from data_loaders.get_data import get_dataset_loader
from data_loaders.humanml.scripts.motion_process import recover_from_ric
import data_loaders.humanml.utils.paramUtil as paramUtil
from data_loaders.humanml.utils.plot_script import plot_3d_motion
import shutil
from data_loaders.tensors import collate


def main():
    args = generate_args()
    fixseed(args.seed)
    args.text_prompt=r''
    args.input_text=r'assets/example_text_prompts.txt'
    args.model_path = r'humanml_trans_enc_512/model000200000.pt'
    out_path = args.output_dir
    name = os.path.basename(os.path.dirname(args.model_path))
    niter = os.path.basename(args.model_path).replace('model', '').replace('.pt', '')
    max_frames = 196 if args.dataset in ['kit', 'humanml'] else 60
    fps = 12.5 if args.dataset == 'kit' else 20
    n_frames = min(max_frames, int(args.motion_length*fps))
    is_using_data = not any([args.input_text, args.text_prompt, args.action_file, args.action_name])
    dist_util.setup_dist(args.device)
    if out_path == '':
        out_path = os.path.join(os.path.dirname(args.model_path),
                                'samples_{}_{}_seed{}'.format(name, niter, args.seed))
        if args.text_prompt != '':
            out_path += '_' + args.text_prompt.replace(' ', '_').replace('.', '')
        elif args.input_text != '':
            out_path += '_' + os.path.basename(args.input_text).replace('.txt', '').replace(' ', '_').replace('.', '')

    # this block must be called BEFORE the dataset is loaded
    if args.text_prompt != '':
        texts = [args.text_prompt]
        args.num_samples = 1
    elif args.input_text != '':
        assert os.path.exists(args.input_text)
        with open(args.input_text, 'r') as fr:
            texts = fr.readlines()
        texts = [s.replace('\n', '') for s in texts]
        args.num_samples = len(texts)
    elif args.action_name:
        action_text = [args.action_name]
        args.num_samples = 1
    elif args.action_file != '':
        assert os.path.exists(args.action_file)
        with open(args.action_file, 'r') as fr:
            action_text = fr.readlines()
        action_text = [s.replace('\n', '') for s in action_text]
        args.num_samples = len(action_text)

    assert args.num_samples <= args.batch_size, \
        f'Please either increase batch_size({args.batch_size}) or reduce num_samples({args.num_samples})'
    # So why do we need this check? In order to protect GPU from a memory overload in the following line.
    # If your GPU can handle batch size larger then default, you can specify it through --batch_size flag.
    # If it doesn't, and you still want to sample more prompts, run this script with different seeds
    # (specify through the --seed flag)
    args.batch_size = args.num_samples  # Sampling a single batch from the testset, with exactly args.num_samples

    print('Loading dataset...')
    data = load_dataset(args, max_frames, n_frames)
    total_num_samples = args.num_samples * args.num_repetitions

    print("Creating model and diffusion...")
    model, diffusion = create_model_and_diffusion(args, data)

    print(f"Loading checkpoints from [{args.model_path}]...")
    state_dict = torch.load(args.model_path, map_location='cpu')
    load_model_wo_clip(model, state_dict)

    if args.guidance_param != 1:
        model = ClassifierFreeSampleModel(model)   # wrapping model with the classifier-free sampler
    model.to(dist_util.dev())
    model.eval()  # disable random masking

    if is_using_data:
        iterator = iter(data)
        _, model_kwargs = next(iterator)
    else:
        collate_args = [{'inp': torch.zeros(n_frames), 'tokens': None, 'lengths': n_frames}] * args.num_samples
        is_t2m = any([args.input_text, args.text_prompt])
        if is_t2m:
            # t2m
            collate_args = [dict(arg, text=txt) for arg, txt in zip(collate_args, texts)]
        else:
            # a2m
            action = data.dataset.action_name_to_action(action_text)
            collate_args = [dict(arg, action=one_action, action_text=one_action_text) for
                            arg, one_action, one_action_text in zip(collate_args, action, action_text)]
        _, model_kwargs = collate(collate_args)

    all_motions = []
    all_lengths = []
    all_text = []

    for rep_i in range(args.num_repetitions):
        print(f'### Sampling [repetitions #{rep_i}]')

        # add CFG scale to batch
        if args.guidance_param != 1:
            model_kwargs['y']['scale'] = torch.ones(args.batch_size, device=dist_util.dev()) * args.guidance_param

        sample_fn = diffusion.p_sample_loop

        sample = sample_fn(
            model,
            # (args.batch_size, model.njoints, model.nfeats, n_frames),  # BUG FIX - this one caused a mismatch between training and inference
            (args.batch_size, model.njoints, model.nfeats, max_frames),  # BUG FIX
            clip_denoised=False,
            model_kwargs=model_kwargs,
            skip_timesteps=0,  # 0 is the default value - i.e. don't skip any step
            init_image=None,
            progress=True,
            dump_steps=None,
            noise=None,
            const_noise=False,
        )

        # Recover XYZ *positions* from HumanML3D vector representation
        if model.data_rep == 'hml_vec':
            n_joints = 22 if sample.shape[1] == 263 else 21
            sample = data.dataset.t2m_dataset.inv_transform(sample.cpu().permute(0, 2, 3, 1)).float()
            sample = recover_from_ric(sample, n_joints)
            sample = sample.view(-1, *sample.shape[2:]).permute(0, 2, 3, 1)

        rot2xyz_pose_rep = 'xyz' if model.data_rep in ['xyz', 'hml_vec'] else model.data_rep
        rot2xyz_mask = None if rot2xyz_pose_rep == 'xyz' else model_kwargs['y']['mask'].reshape(args.batch_size, n_frames).bool()
        sample = model.rot2xyz(x=sample, mask=rot2xyz_mask, pose_rep=rot2xyz_pose_rep, glob=True, translation=True,
                               jointstype='smpl', vertstrans=True, betas=None, beta=0, glob_rot=None,
                               get_rotations_back=False)

        if args.unconstrained:
            all_text += ['unconstrained'] * args.num_samples
        else:
            text_key = 'text' if 'text' in model_kwargs['y'] else 'action_text'
            all_text += model_kwargs['y'][text_key]

        all_motions.append(sample.cpu().numpy())
        all_lengths.append(model_kwargs['y']['lengths'].cpu().numpy())

        print(f"created {len(all_motions) * args.batch_size} samples")


    all_motions = np.concatenate(all_motions, axis=0)
    all_motions = all_motions[:total_num_samples]  # [bs, njoints, 6, seqlen]
    all_text = all_text[:total_num_samples]
    all_lengths = np.concatenate(all_lengths, axis=0)[:total_num_samples]

    if os.path.exists(out_path):
        shutil.rmtree(out_path)
    os.makedirs(out_path)

    npy_path = os.path.join(out_path, 'results.npy')
    print(f"saving results file to [{npy_path}]")
    np.save(npy_path,
            {'motion': all_motions, 'text': all_text, 'lengths': all_lengths,
             'num_samples': args.num_samples, 'num_repetitions': args.num_repetitions})
    with open(npy_path.replace('.npy', '.txt'), 'w') as fw:
        fw.write('\n'.join(all_text))
    with open(npy_path.replace('.npy', '_len.txt'), 'w') as fw:
        fw.write('\n'.join([str(l) for l in all_lengths]))

    print(f"saving visualizations to [{out_path}]...")
    skeleton = paramUtil.kit_kinematic_chain if args.dataset == 'kit' else paramUtil.t2m_kinematic_chain

    sample_files = []
    num_samples_in_out_file = 7

    sample_print_template, row_print_template, all_print_template, \
    sample_file_template, row_file_template, all_file_template = construct_template_variables(args.unconstrained)

    for sample_i in range(args.num_samples):
        rep_files = []
        for rep_i in range(args.num_repetitions):
            caption = all_text[rep_i*args.batch_size + sample_i]
            length = all_lengths[rep_i*args.batch_size + sample_i]
            motion = all_motions[rep_i*args.batch_size + sample_i].transpose(2, 0, 1)[:length]
            save_file = sample_file_template.format(sample_i, rep_i)
            print(sample_print_template.format(caption, sample_i, rep_i, save_file))
            animation_save_path = os.path.join(out_path, save_file)
            plot_3d_motion(animation_save_path, skeleton, motion, dataset=args.dataset, title=caption, fps=fps)
            # Credit for visualization: https://github.com/EricGuo5513/text-to-motion
            rep_files.append(animation_save_path)

        sample_files = save_multiple_samples(args, out_path,
                                               row_print_template, all_print_template, row_file_template, all_file_template,
                                               caption, num_samples_in_out_file, rep_files, sample_files, sample_i)

    abs_path = os.path.abspath(out_path)
    print(f'[Done] Results are at [{abs_path}]')


def save_multiple_samples(args, out_path, row_print_template, all_print_template, row_file_template, all_file_template,
                          caption, num_samples_in_out_file, rep_files, sample_files, sample_i):
    all_rep_save_file = row_file_template.format(sample_i)
    all_rep_save_path = os.path.join(out_path, all_rep_save_file)
    ffmpeg_rep_files = [f' -i {f} ' for f in rep_files]
    hstack_args = f' -filter_complex hstack=inputs={args.num_repetitions}' if args.num_repetitions > 1 else ''
    ffmpeg_rep_cmd = f'ffmpeg -y -loglevel warning ' + ''.join(ffmpeg_rep_files) + f'{hstack_args} {all_rep_save_path}'
    os.system(ffmpeg_rep_cmd)
    print(row_print_template.format(caption, sample_i, all_rep_save_file))
    sample_files.append(all_rep_save_path)
    if (sample_i + 1) % num_samples_in_out_file == 0 or sample_i + 1 == args.num_samples:
        # all_sample_save_file =  f'samples_{(sample_i - len(sample_files) + 1):02d}_to_{sample_i:02d}.mp4'
        all_sample_save_file = all_file_template.format(sample_i - len(sample_files) + 1, sample_i)
        all_sample_save_path = os.path.join(out_path, all_sample_save_file)
        print(all_print_template.format(sample_i - len(sample_files) + 1, sample_i, all_sample_save_file))
        ffmpeg_rep_files = [f' -i {f} ' for f in sample_files]
        vstack_args = f' -filter_complex vstack=inputs={len(sample_files)}' if len(sample_files) > 1 else ''
        ffmpeg_rep_cmd = f'ffmpeg -y -loglevel warning ' + ''.join(
            ffmpeg_rep_files) + f'{vstack_args} {all_sample_save_path}'
        os.system(ffmpeg_rep_cmd)
        sample_files = []
    return sample_files


def construct_template_variables(unconstrained):
    row_file_template = 'sample{:02d}.mp4'
    all_file_template = 'samples_{:02d}_to_{:02d}.mp4'
    if unconstrained:
        sample_file_template = 'row{:02d}_col{:02d}.mp4'
        sample_print_template = '[{} row #{:02d} column #{:02d} | -> {}]'
        row_file_template = row_file_template.replace('sample', 'row')
        row_print_template = '[{} row #{:02d} | all columns | -> {}]'
        all_file_template = all_file_template.replace('samples', 'rows')
        all_print_template = '[rows {:02d} to {:02d} | -> {}]'
    else:
        sample_file_template = 'sample{:02d}_rep{:02d}.mp4'
        sample_print_template = '["{}" ({:02d}) | Rep #{:02d} | -> {}]'
        row_print_template = '[ "{}" ({:02d}) | all repetitions | -> {}]'
        all_print_template = '[samples {:02d} to {:02d} | all repetitions | -> {}]'

    return sample_print_template, row_print_template, all_print_template, \
           sample_file_template, row_file_template, all_file_template


def load_dataset(args, max_frames, n_frames):
    data = get_dataset_loader(name=args.dataset,
                              batch_size=args.batch_size,
                              num_frames=max_frames,
                              split='test',
                              hml_mode='text_only')
    if args.dataset in ['kit', 'humanml']:
        data.dataset.t2m_dataset.fixed_length = n_frames
    return data


if __name__ == "__main__":
    main()

渲染mesh

import argparse
import os
from visualize import vis_utils
import shutil
from tqdm import tqdm

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--input_path", type=str, default=r"E:\project\202404\motion-diffusion-model-main\humanml_trans_enc_512\samples_humanml_trans_enc_512_000200000_seed10_example_text_prompts\sample00_rep00.mp4", help='stick figure mp4 file to be rendered.')
    parser.add_argument("--cuda", type=bool, default=True, help='')
    parser.add_argument("--device", type=int, default=0, help='')
    params = parser.parse_args()

    assert params.input_path.endswith('.mp4')
    parsed_name = os.path.basename(params.input_path).replace('.mp4', '').replace('sample', '').replace('rep', '')
    sample_i, rep_i = [int(e) for e in parsed_name.split('_')]
    npy_path = os.path.join(os.path.dirname(params.input_path), 'results.npy')
    out_npy_path = params.input_path.replace('.mp4', '_smpl_params.npy')
    assert os.path.exists(npy_path)
    results_dir = params.input_path.replace('.mp4', '_obj')
    if os.path.exists(results_dir):
        shutil.rmtree(results_dir)
    os.makedirs(results_dir)

    npy2obj = vis_utils.npy2obj(npy_path, sample_i, rep_i,
                                device=params.device, cuda=params.cuda)

    print('Saving obj files to [{}]'.format(os.path.abspath(results_dir)))
    for frame_i in tqdm(range(npy2obj.real_num_frames)):
        npy2obj.save_obj(os.path.join(results_dir, 'frame{:03d}.obj'.format(frame_i)), frame_i)

    print('Saving SMPL params to [{}]'.format(os.path.abspath(out_npy_path)))
    npy2obj.save_npy(out_npy_path)

可视化代码：

自带的可视化代码windows系统不能显示，经过改进实现可以显示的代码：
 

import math
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from matplotlib.animation import FuncAnimation, FFMpegFileWriter
from mpl_toolkits.mplot3d.art3d import Poly3DCollection
import mpl_toolkits.mplot3d.axes3d as p3
# import cv2
from textwrap import wrap

def list_cut_average(ll, intervals):
    if intervals == 1:
        return ll

    bins = math.ceil(len(ll) * 1.0 / intervals)
    ll_new = []
    for i in range(bins):
        l_low = intervals * i
        l_high = l_low + intervals
        l_high = l_high if l_high < len(ll) else len(ll)
        ll_new.append(np.mean(ll[l_low:l_high]))
    return ll_new

def plot_3d_motion(save_path, kinematic_tree, joints, title, dataset, figsize=(3, 3), fps=120, radius=3, vis_mode='default', gt_frames=[]):
    fig = plt.figure(figsize=figsize)
    ax = fig.add_subplot(111, projection='3d')
    plt.tight_layout()

    title = '\n'.join(wrap(title, 20))

    colors_blue = ["#4D84AA", "#5B9965", "#61CEB9", "#34C1E2", "#80B79A"]  # GT color
    colors_orange = ["#DD5A37", "#D69E00", "#B75A39", "#FF6D00", "#DDB50E"]  # Generation color
    colors = colors_orange
    def init():
        ax.set_xlim3d([-radius / 2, radius / 2])
        ax.set_ylim3d([0, radius])
        ax.set_zlim3d([-radius / 3., radius * 2 / 3.])
        fig.suptitle(title, fontsize=10)
        ax.grid(b=False)
        return fig,

    def update(index):
        while len(ax.lines) > 0:
            ax.lines[0].remove()
        while len(ax.collections) > 0:
            ax.collections[0].remove()

        ax.view_init(elev=120, azim=-90)
        ax.dist = 7.5

        # Draw motion
        used_colors = colors_blue if index in gt_frames else colors
        for i, (chain, color) in enumerate(zip(kinematic_tree, used_colors)):
            linewidth = 4.0 if i < 5 else 2.0
            ax.plot3D(data[index, chain, 0], data[index, chain, 1], data[index, chain, 2], linewidth=linewidth, color=color)

        plt.axis('off')
        ax.set_xticklabels([])
        ax.set_yticklabels([])
        ax.set_zticklabels([])

        return fig,

    data = joints.copy().reshape(len(joints), -1, 3)
    frame_number = data.shape[0]

    ani = FuncAnimation(fig, update, frames=frame_number, interval=1000 / fps, repeat=False, init_func=init)
    ani.save(save_path, fps=fps)

    # plt.show()
    plt.close()

if __name__ == '__main__':
    save_path='0415.mp4'
    dataset='humanml'
    save_path='0415.mp4'
    title='title'

    kinematic_tree=[[0, 2, 5, 8, 11], [0, 1, 4, 7, 10], [0, 3, 6, 9, 12, 15], [9, 14, 17, 19, 21], [9, 13, 16, 18, 20]]

    npz_data = np.load(r"E:\04151.npz", allow_pickle=True)
    joints = npz_data['joints_3d'].item()['data']
    joints/=20
    plot_3d_motion(save_path, kinematic_tree, joints, title, dataset, figsize=(3, 3), fps=120, radius=3, vis_mode='default', gt_frames=[])