File size: 8,277 Bytes
c64dfa4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cadd88e
c64dfa4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cadd88e
c64dfa4
 
 
 
cadd88e
c64dfa4
 
cadd88e
c64dfa4
 
 
cadd88e
c64dfa4
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
import os
import time
import random
import datetime
import os.path as osp
from functools import partial

import tqdm
from omegaconf import OmegaConf

import torch
import gradio as gr

from mld.config import get_module_config
from mld.data.get_data import get_dataset
from mld.models.modeltype.mld import MLD
from mld.utils.utils import set_seed
from mld.data.humanml.utils.plot_script import plot_3d_motion

os.environ["TOKENIZERS_PARALLELISM"] = "false"

WEBSITE = """
<div class="embed_hidden">
<h1 style='text-align: center'> MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model </h1>
<h2 style='text-align: center'>
<a href="https://github.com/Dai-Wenxun/" target="_blank"><nobr>Wenxun Dai</nobr><sup>1</sup></a> &emsp;
<a href="https://lhchen.top/" target="_blank"><nobr>Ling-Hao Chen</nobr></a><sup>1</sup> &emsp;
<a href="https://wangjingbo1219.github.io/" target="_blank"><nobr>Jingbo Wang</nobr></a><sup>2</sup> &emsp;
<a href="https://moonsliu.github.io/" target="_blank"><nobr>Jinpeng Liu</nobr></a><sup>1</sup> &emsp;
<a href="https://daibo.info/" target="_blank"><nobr>Bo Dai</nobr></a><sup>2</sup> &emsp;
<a href="https://andytang15.github.io/" target="_blank"><nobr>Yansong Tang</nobr></a><sup>1</sup>
</h2>
<h2 style='text-align: center'>
<nobr><sup>1</sup>Tsinghua University</nobr> &emsp;
<nobr><sup>2</sup>Shanghai AI Laboratory</nobr>
</h2>
</div>
"""

WEBSITE_bottom = """
<div class="embed_hidden">
<p>
Space adapted from <a href="https://huggingface.co/spaces/Mathux/TMR" target="_blank">TMR</a> 
and <a href="https://huggingface.co/spaces/MeYourHint/MoMask" target="_blank">MoMask</a>.
</p>
</div>
"""

EXAMPLES = [
    "a person does a jump",
    "a person waves both arms in the air.",
    "The person takes 4 steps backwards.",
    "this person bends forward as if to bow.",
    "The person was pushed but did not fall.",
    "a man walks forward in a snake like pattern.",
    "a man paces back and forth along the same line.",
    "with arms out to the sides a person walks forward",
    "A man bends down and picks something up with his right hand.",
    "The man walked forward, spun right on one foot and walked back to his original position.",
    "a person slightly bent over with right hand pressing against the air walks forward slowly"
]

if not os.path.exists("./experiments_t2m/"):
    os.system("bash prepare/download_pretrained_models.sh")
if not os.path.exists('./deps/glove/'):
    os.system("bash prepare/download_glove.sh")
if not os.path.exists('./deps/sentence-t5-large/'):
    os.system("bash prepare/prepare_t5.sh")
if not os.path.exists('./deps/t2m/'):
    os.system("bash prepare/download_t2m_evaluators.sh")
if not os.path.exists('./datasets/humanml3d/'):
    os.system("bash prepare/prepare_tiny_humanml3d.sh")

DEFAULT_TEXT = "A person is "
MAX_VIDEOS = 8
NUM_ROWS = 2
NUM_COLS = MAX_VIDEOS // NUM_ROWS
EXAMPLES_PER_PAGE = 12
T2M_CFG = "./configs_v1/motionlcm_t2m.yaml"

device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
print("device: ", device)

cfg = OmegaConf.load(T2M_CFG)
cfg_root = os.path.dirname(T2M_CFG)
cfg_model = get_module_config(cfg.model, cfg.model.target, cfg_root)
cfg = OmegaConf.merge(cfg, cfg_model)
set_seed(cfg.SEED_VALUE)

name_time_str = osp.join(cfg.NAME, datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S"))
cfg.output_dir = osp.join(cfg.TEST_FOLDER, name_time_str)
vis_dir = osp.join(cfg.output_dir, 'samples')
os.makedirs(cfg.output_dir, exist_ok=False)
os.makedirs(vis_dir, exist_ok=False)

state_dict = torch.load(cfg.TEST.CHECKPOINTS, map_location="cpu")["state_dict"]
print("Loading checkpoints from {}".format(cfg.TEST.CHECKPOINTS))

is_lcm = False
lcm_key = 'denoiser.time_embedding.cond_proj.weight'  # unique key for CFG
if lcm_key in state_dict:
    is_lcm = True
    time_cond_proj_dim = state_dict[lcm_key].shape[1]
    cfg.model.denoiser.params.time_cond_proj_dim = time_cond_proj_dim
print(f'Is LCM: {is_lcm}')

dataset = get_dataset(cfg)
model = MLD(cfg, dataset)
model.to(device)
model.eval()
model.requires_grad_(False)
model.load_state_dict(state_dict)

FPS = eval(f"cfg.DATASET.{cfg.DATASET.NAME.upper()}.FRAME_RATE")


@torch.no_grad()
def generate(text_, motion_len_):
    batch = {"text": [text_] * MAX_VIDEOS, "length": [motion_len_] * MAX_VIDEOS}

    s = time.time()
    joints = model(batch)[0]
    runtime_infer = round(time.time() - s, 3)

    s = time.time()
    path = []
    for i in tqdm.tqdm(range(len(joints))):
        uid = random.randrange(999999999)
        video_path = osp.join(vis_dir, f"sample_{uid}.mp4")
        plot_3d_motion(video_path, joints[i].detach().cpu().numpy(), '', fps=FPS)
        path.append(video_path)
    runtime_draw = round(time.time() - s, 3)

    runtime_info = f'Inference {len(joints)} motions, Runtime (Inference): {runtime_infer}s, ' \
                   f'Runtime (Draw Skeleton): {runtime_draw}s, device: {device} '

    return path, runtime_info


def generate_component(generate_function, text_, motion_len_, num_inference_steps_, guidance_scale_):
    if text_ == DEFAULT_TEXT or text_ == "" or text_ is None:
        return [None] * MAX_VIDEOS + ["Please modify the default text prompt."]

    model.cfg.model.scheduler.num_inference_steps = num_inference_steps_
    model.guidance_scale = guidance_scale_
    motion_len_ = max(36, min(int(float(motion_len_) * FPS), 196))
    paths, info = generate_function(text_, motion_len_)
    paths = paths + [None] * (MAX_VIDEOS - len(paths))
    return paths + [info]


theme = gr.themes.Default(primary_hue="purple", secondary_hue="gray")
generate_and_show = partial(generate_component, generate)

with gr.Blocks(theme=theme) as demo:
    gr.HTML(WEBSITE)
    videos = []

    with gr.Row():
        with gr.Column(scale=3):
            text = gr.Textbox(
                show_label=True,
                label="Text prompt",
                value=DEFAULT_TEXT,
            )

            with gr.Row():
                with gr.Column(scale=1):
                    motion_len = gr.Slider(
                        minimum=1.8,
                        maximum=9.8,
                        step=0.2,
                        value=5.0,
                        label="Motion length",
                        info="Motion duration in seconds: [1.8s, 9.8s] (FPS = 20)."
                    )

                with gr.Column(scale=1):
                    num_inference_steps = gr.Slider(
                        minimum=1,
                        maximum=4,
                        step=1,
                        value=1,
                        label="Inference steps",
                        info="Number of inference steps.",
                    )

            cfg = gr.Slider(
                minimum=1,
                maximum=15,
                step=0.5,
                value=7.5,
                label="CFG",
                info="Classifier-free diffusion guidance.",
            )

            gen_btn = gr.Button("Generate", variant="primary")
            clear = gr.Button("Clear", variant="secondary")

            results = gr.Textbox(show_label=True,
                                 label='Inference info (runtime and device)',
                                 info='Real-time inference cannot be achieved using the free CPU. Local GPU deployment is recommended.',
                                 interactive=False)

        with gr.Column(scale=2):
            examples = gr.Examples(
                examples=EXAMPLES,
                inputs=[text],
                examples_per_page=EXAMPLES_PER_PAGE)

    for i in range(NUM_ROWS):
        with gr.Row():
            for j in range(NUM_COLS):
                video = gr.Video(autoplay=True, loop=True)
                videos.append(video)

    # gr.HTML(WEBSITE_bottom)

    gen_btn.click(
        fn=generate_and_show,
        inputs=[text, motion_len, num_inference_steps, cfg],
        outputs=videos + [results],
    )
    text.submit(
        fn=generate_and_show,
        inputs=[text, motion_len, num_inference_steps, cfg],
        outputs=videos + [results],
    )


    def clear_videos():
        return [None] * MAX_VIDEOS + [DEFAULT_TEXT] + [None]


    clear.click(fn=clear_videos, outputs=videos + [text] + [results])

demo.launch()