File size: 14,282 Bytes
823807d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
from functools import partial
import os

import torch
import numpy as np
import gradio as gr
import random
import shutil

print(f"Is CUDA available: {torch.cuda.is_available()}")
print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")

import os
from os.path import join as pjoin

import torch.nn.functional as F

from models.mask_transformer.transformer import MaskTransformer, ResidualTransformer
from models.vq.model import RVQVAE, LengthEstimator

from options.hgdemo_option import EvalT2MOptions
from utils.get_opt import get_opt

from utils.fixseed import fixseed
from visualization.joints2bvh import Joint2BVHConvertor
from torch.distributions.categorical import Categorical

from utils.motion_process import recover_from_ric
from utils.plot_script import plot_3d_motion

from utils.paramUtil import t2m_kinematic_chain

from gen_t2m import load_vq_model, load_res_model, load_trans_model, load_len_estimator

clip_version = 'ViT-B/32'

WEBSITE = """
<div class="embed_hidden">
<h1 style='text-align: center'> MoMask: Generative Masked Modeling of 3D Human Motions </h1>
<h2 style='text-align: center'>
<a href="https://ericguo5513.github.io" target="_blank"><nobr>Chuan Guo*</nobr></a> &emsp;
<a href="https://yxmu.foo/" target="_blank"><nobr>Yuxuan Mu*</nobr></a> &emsp;
<a href="https://scholar.google.com/citations?user=w4e-j9sAAAAJ&hl=en" target="_blank"><nobr>Muhammad Gohar Javed*</nobr></a> &emsp;
<a href="https://sites.google.com/site/senwang1312home/" target="_blank"><nobr>Sen Wang</nobr></a> &emsp;
<a href="https://www.ece.ualberta.ca/~lcheng5/" target="_blank"><nobr>Li Cheng</nobr></a>
</h2>
<h2 style='text-align: center'>
<nobr>arXiv 2023</nobr>
</h2>
<h3 style="text-align:center;">
<a target="_blank" href="https://arxiv.org/abs/2312.00063"> <button type="button" class="btn btn-primary btn-lg"> Paper </button></a> &ensp;
<a target="_blank" href="https://github.com/EricGuo5513/momask-codes"> <button type="button" class="btn btn-primary btn-lg"> Code </button></a> &ensp;
<a target="_blank" href="https://ericguo5513.github.io/momask/"> <button type="button" class="btn btn-primary btn-lg"> Webpage </button></a> &ensp;
<a target="_blank" href="https://ericguo5513.github.io/source_files/momask_2023_bib.txt"> <button type="button" class="btn btn-primary btn-lg"> BibTex </button></a>
</h3>
<h3> Description </h3>
<p>
πŸ”₯πŸ”₯πŸ”₯ This space presents an interactive demo for <a href='https://ericguo5513.github.io/momask/' target='_blank'><b>MoMask</b></a>, a method for text-to-motion generation!!! It generates human motions (skeletal animations) based on your descriptions. To gain a better understanding of our work, you could try the provided examples first. πŸ”₯πŸ”₯πŸ”₯
</p>
<p>
πŸš€πŸš€πŸš€ In addition, we provide a link to download the generated human skeletal motion in <b>BVH</b> file format, compatible with CG software such as Blender!!! πŸš€πŸš€πŸš€
</p>
<p>
😁😁😁 If you find this demo interesting, we would appreciate your star on our <a href="https://github.com/EricGuo5513/momask-codes" target="_blank">github</a>. More details could be found on our <a href='https://ericguo5513.github.io/momask/' target='_blank'>webpage</a>. 🫢🫢🫢
</p>
<p>
If you have any issues on this space or feature requests, we warmly welcome you to contact us through our <a href="https://github.com/EricGuo5513/momask-codes/issues" target="_blank">github repo</a> or <a href="mailto:[email protected]?subject =[MoMask]Feedback&body = Message">email</a>.
</p>
</div>
"""
WEBSITE_bottom = """
<p>
We thanks <a href="https://huggingface.co./spaces/Mathux/TMR" target="_blank">TMR</a> for this cool space template.
</p>
</div>
"""

EXAMPLES = [
   "A person is running on a treadmill.", "The person takes 4 steps backwards.", 
   "A person jumps up and then lands.", "The person was pushed but did not fall.", 
   "The person does a salsa dance.", "A figure streches it hands and arms above its head.",
   "This person kicks with his right leg then jabs several times.",
   "A person stands for few seconds and picks up his arms and shakes them.",
   "A person walks in a clockwise circle and stops where he began.",
   "A man bends down and picks something up with his right hand.",
   "A person walks with a limp, their left leg gets injured.",
   "A person repeatedly blocks their face with their right arm.",
   "The person holds his left foot with his left hand, puts his right foot up and left hand up too.",
   "A person stands, crosses left leg in front of the right, lowering themselves until they are sitting, both hands on the floor before standing and uncrossing legs.",
   "The man walked forward, spun right on one foot and walked back to his original position.",
   "A man is walking forward then steps over an object then continues walking forward.",
]

# Show closest text in the training


# css to make videos look nice
# var(--block-border-color); TODO
CSS = """
.generate_video {
    position: relative;
    margin-left: auto;
    margin-right: auto;
    box-shadow: var(--block-shadow);
    border-width: var(--block-border-width);
    border-color: #000000;
    border-radius: var(--block-radius);
    background: var(--block-background-fill);
    width: 25%;
    line-height: var(--line-sm);
}
}
"""


DEFAULT_TEXT = "A person is "


if not os.path.exists("checkpoints/t2m"):
    os.system("bash prepare/download_models_demo.sh")

##########################
######Preparing demo######
##########################
parser = EvalT2MOptions()
opt = parser.parse()
fixseed(opt.seed)
opt.device = torch.device("cpu" if opt.gpu_id == -1 else "cuda:" + str(opt.gpu_id))
dim_pose = 263
root_dir = pjoin(opt.checkpoints_dir, opt.dataset_name, opt.name)
model_dir = pjoin(root_dir, 'model')
model_opt_path = pjoin(root_dir, 'opt.txt')
model_opt = get_opt(model_opt_path, device=opt.device)

######Loading RVQ######
vq_opt_path = pjoin(opt.checkpoints_dir, opt.dataset_name, model_opt.vq_name, 'opt.txt')
vq_opt = get_opt(vq_opt_path, device=opt.device)
vq_opt.dim_pose = dim_pose
vq_model, vq_opt = load_vq_model(vq_opt)

model_opt.num_tokens = vq_opt.nb_code
model_opt.num_quantizers = vq_opt.num_quantizers
model_opt.code_dim = vq_opt.code_dim

######Loading R-Transformer######
res_opt_path = pjoin(opt.checkpoints_dir, opt.dataset_name, opt.res_name, 'opt.txt')
res_opt = get_opt(res_opt_path, device=opt.device)
res_model = load_res_model(res_opt, vq_opt, opt)

assert res_opt.vq_name == model_opt.vq_name

######Loading M-Transformer######
t2m_transformer = load_trans_model(model_opt, opt, 'latest.tar')

#####Loading Length Predictor#####
length_estimator = load_len_estimator(model_opt)

t2m_transformer.eval()
vq_model.eval()
res_model.eval()
length_estimator.eval()

res_model.to(opt.device)
t2m_transformer.to(opt.device)
vq_model.to(opt.device)
length_estimator.to(opt.device)

opt.nb_joints = 22
mean = np.load(pjoin(opt.checkpoints_dir, opt.dataset_name, model_opt.vq_name, 'meta', 'mean.npy'))
std = np.load(pjoin(opt.checkpoints_dir, opt.dataset_name, model_opt.vq_name, 'meta', 'std.npy'))
def inv_transform(data):
    return data * std + mean

kinematic_chain = t2m_kinematic_chain
converter = Joint2BVHConvertor()
cached_dir = './cached'
uid = 12138
animation_path = pjoin(cached_dir, f'{uid}')
os.makedirs(animation_path, exist_ok=True)

@torch.no_grad()
def generate(
    text, uid, motion_length=0, use_ik=True, seed=10107, repeat_times=1,
):
    # fixseed(seed)
    print(text)
    prompt_list = []
    length_list = []
    est_length = False
    prompt_list.append(text)
    if motion_length == 0:
        est_length = True
    else:
        length_list.append(motion_length)

    if est_length:
        print("Since no motion length are specified, we will use estimated motion lengthes!!")
        text_embedding = t2m_transformer.encode_text(prompt_list)
        pred_dis = length_estimator(text_embedding)
        probs = F.softmax(pred_dis, dim=-1)  # (b, ntoken)
        token_lens = Categorical(probs).sample()  # (b, seqlen)
    else:
        token_lens = torch.LongTensor(length_list) // 4
        token_lens = token_lens.to(opt.device).long()

    m_length = token_lens * 4
    captions = prompt_list
    datas = []
    for r in range(repeat_times):
        mids = t2m_transformer.generate(captions, token_lens,
                                        timesteps=opt.time_steps,
                                        cond_scale=opt.cond_scale,
                                        temperature=opt.temperature,
                                        topk_filter_thres=opt.topkr,
                                        gsample=opt.gumbel_sample)
        mids = res_model.generate(mids, captions, token_lens, temperature=1, cond_scale=5)
        pred_motions = vq_model.forward_decoder(mids)
        pred_motions = pred_motions.detach().cpu().numpy()
        data = inv_transform(pred_motions)
        ruid = random.randrange(99999)
        for k, (caption, joint_data)  in enumerate(zip(captions, data)):
            animation_path = pjoin(cached_dir, f'{uid}')
            os.makedirs(animation_path, exist_ok=True)
            joint_data = joint_data[:m_length[k]]
            joint = recover_from_ric(torch.from_numpy(joint_data).float(), 22).numpy()
            bvh_path = pjoin(animation_path, "sample_repeat%d.bvh" % (r))
            save_path = pjoin(animation_path, "sample_repeat%d_%d.mp4"%(r, ruid))
            if use_ik:
                print("Using IK")
                _, joint = converter.convert(joint, filename=bvh_path, iterations=100)
            else:
                _, joint = converter.convert(joint, filename=bvh_path, iterations=100, foot_ik=False)
            plot_3d_motion(save_path, kinematic_chain, joint, title=caption, fps=20)
            np.save(pjoin(animation_path, "sample_repeat%d.npy"%(r)), joint)
        data_unit = {
            "url": pjoin(animation_path, "sample_repeat%d_%d.mp4"%(r, ruid))
            }
        datas.append(data_unit)

    return datas


# HTML component
def get_video_html(data, video_id, width=700, height=700):
    url = data["url"]
    # class="wrap default svelte-gjihhp hide"
    # <div class="contour_video" style="position: absolute; padding: 10px;">
    # width="{width}" height="{height}"
    video_html = f"""
<h2 style='text-align: center'>
<a href="file/{pjoin(animation_path, "sample_repeat0.bvh")}" download="sample.bvh"><b>BVH Download</b></a>
</h2>
<video class="generate_video" width="{width}" height="{height}" style="center" preload="auto" muted playsinline onpause="this.load()"
autoplay loop disablepictureinpicture id="{video_id}">
  <source src="file/{url}" type="video/mp4">
  Your browser does not support the video tag.
</video>
"""
    return video_html

def generate_component(generate_function, text, motion_len='0', postprocess='IK'):
    if text == DEFAULT_TEXT or text == "" or text is None:
        return [None for _ in range(1)]
    # uid = random.randrange(99999)
    try:
        motion_len = max(0, min(int(float(motion_len) * 20), 196))
    except:
        motion_len = 0
    use_ik = postprocess == 'IK'
    datas = generate_function(text, uid, motion_len, use_ik)
    htmls = [get_video_html(data, idx) for idx, data in enumerate(datas)]
    return htmls


# LOADING

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

with gr.Blocks(css=CSS, theme=theme) as demo:
    gr.Markdown(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.Textbox(
                        show_label=True,
                        label="Motion length (<10s)",
                        value=0,
                        info="Specify the motion length; 0 to use the default auto-setting.",
                    )
                with gr.Column(scale=1):
                    use_ik = gr.Radio(
                        ["Raw", "IK"],
                        label="Post-processing",
                        value="IK",
                        info="Use basic inverse kinematic (IK) for foot contact locking",
                    )
            gen_btn = gr.Button("Generate", variant="primary")
            clear = gr.Button("Clear", variant="secondary")
            gr.Markdown(
                        f"""
                            
                        """
                    )

        with gr.Column(scale=2):

            def generate_example(text):
                return generate_and_show(text)

            examples = gr.Examples(
                examples=[[x, None, None] for x in EXAMPLES],
                inputs=[text],
                examples_per_page=10,
                run_on_click=False,
                cache_examples=False,
                fn=generate_example,
                outputs=[],
            )

    i = -1
    # should indent
    for _ in range(1):
        with gr.Row():
            for _ in range(1):
                i += 1
                video = gr.HTML()
                videos.append(video)
    gr.Markdown(WEBSITE_bottom)
    # connect the examples to the output
    # a bit hacky
    examples.outputs = videos

    def load_example(example_id):
        processed_example = examples.non_none_processed_examples[example_id]
        return gr.utils.resolve_singleton(processed_example)

    examples.dataset.click(
        load_example,
        inputs=[examples.dataset],
        outputs=examples.inputs_with_examples,  # type: ignore
        show_progress=False,
        postprocess=False,
        queue=False,
    ).then(fn=generate_example, inputs=examples.inputs, outputs=videos)

    gen_btn.click(
        fn=generate_and_show,
        inputs=[text, motion_len, use_ik],
        outputs=videos,
    )
    text.submit(
        fn=generate_and_show,
        inputs=[text, motion_len, use_ik],
        outputs=videos,
    )

    def clear_videos():
        return [None for x in range(1)] + [DEFAULT_TEXT]

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

demo.launch(share=True)