File size: 12,434 Bytes
44931be
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Open Source Model Licensed under the Apache License Version 2.0
# and Other Licenses of the Third-Party Components therein:
# The below Model in this distribution may have been modified by THL A29 Limited
# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited.

# Copyright (C) 2024 THL A29 Limited, a Tencent company.  All rights reserved.
# The below software and/or models in this distribution may have been
# modified by THL A29 Limited ("Tencent Modifications").
# All Tencent Modifications are Copyright (C) THL A29 Limited.

# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
# except for the third-party components listed below.
# Hunyuan 3D does not impose any additional limitations beyond what is outlined
# in the repsective licenses of these third-party components.
# Users must comply with all terms and conditions of original licenses of these third-party
# components and must ensure that the usage of the third party components adheres to
# all relevant laws and regulations.

# For avoidance of doubts, Hunyuan 3D means the large language models and
# their software and algorithms, including trained model weights, parameters (including
# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
# fine-tuning enabling code and other elements of the foregoing made publicly available
# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.


import os
from typing import Union, List

import numpy as np
import torch
import torch.nn as nn
import yaml

from .attention_blocks import FourierEmbedder, Transformer, CrossAttentionDecoder, PointCrossAttentionEncoder
from .surface_extractors import MCSurfaceExtractor, SurfaceExtractors
from .volume_decoders import VanillaVolumeDecoder, FlashVDMVolumeDecoding, HierarchicalVolumeDecoding
from ...utils import logger, synchronize_timer, smart_load_model


class DiagonalGaussianDistribution(object):
    def __init__(self, parameters: Union[torch.Tensor, List[torch.Tensor]], deterministic=False, feat_dim=1):
        """
        Initialize a diagonal Gaussian distribution with mean and log-variance parameters.

        Args:
            parameters (Union[torch.Tensor, List[torch.Tensor]]): 
                Either a single tensor containing concatenated mean and log-variance along `feat_dim`,
                or a list of two tensors [mean, logvar].
            deterministic (bool, optional): If True, the distribution is deterministic (zero variance). 
                Default is False. feat_dim (int, optional): Dimension along which mean and logvar are 
                concatenated if parameters is a single tensor. Default is 1.
        """
        self.feat_dim = feat_dim
        self.parameters = parameters

        if isinstance(parameters, list):
            self.mean = parameters[0]
            self.logvar = parameters[1]
        else:
            self.mean, self.logvar = torch.chunk(parameters, 2, dim=feat_dim)

        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
        self.deterministic = deterministic
        self.std = torch.exp(0.5 * self.logvar)
        self.var = torch.exp(self.logvar)
        if self.deterministic:
            self.var = self.std = torch.zeros_like(self.mean)

    def sample(self):
        """
        Sample from the diagonal Gaussian distribution.

        Returns:
            torch.Tensor: A sample tensor with the same shape as the mean.
        """
        x = self.mean + self.std * torch.randn_like(self.mean)
        return x

    def kl(self, other=None, dims=(1, 2, 3)):
        """
        Compute the Kullback-Leibler (KL) divergence between this distribution and another.

        If `other` is None, compute KL divergence to a standard normal distribution N(0, I).

        Args:
            other (DiagonalGaussianDistribution, optional): Another diagonal Gaussian distribution.
            dims (tuple, optional): Dimensions along which to compute the mean KL divergence. 
                Default is (1, 2, 3).

        Returns:
            torch.Tensor: The mean KL divergence value.
        """
        if self.deterministic:
            return torch.Tensor([0.])
        else:
            if other is None:
                return 0.5 * torch.mean(torch.pow(self.mean, 2)
                                        + self.var - 1.0 - self.logvar,
                                        dim=dims)
            else:
                return 0.5 * torch.mean(
                    torch.pow(self.mean - other.mean, 2) / other.var
                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
                    dim=dims)

    def nll(self, sample, dims=(1, 2, 3)):
        if self.deterministic:
            return torch.Tensor([0.])
        logtwopi = np.log(2.0 * np.pi)
        return 0.5 * torch.sum(
            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
            dim=dims)

    def mode(self):
        return self.mean


class VectsetVAE(nn.Module):

    @classmethod
    @synchronize_timer('VectsetVAE Model Loading')
    def from_single_file(
        cls,
        ckpt_path,
        config_path,
        device='cuda',
        dtype=torch.float16,
        use_safetensors=None,
        **kwargs,
    ):
        # load config
        with open(config_path, 'r') as f:
            config = yaml.safe_load(f)

        # load ckpt
        if use_safetensors:
            ckpt_path = ckpt_path.replace('.ckpt', '.safetensors')
        if not os.path.exists(ckpt_path):
            raise FileNotFoundError(f"Model file {ckpt_path} not found")

        logger.info(f"Loading model from {ckpt_path}")
        if use_safetensors:
            import safetensors.torch
            ckpt = safetensors.torch.load_file(ckpt_path, device='cpu')
        else:
            ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=True)

        model_kwargs = config['params']
        model_kwargs.update(kwargs)

        model = cls(**model_kwargs)
        model.load_state_dict(ckpt)
        model.to(device=device, dtype=dtype)
        return model

    @classmethod
    def from_pretrained(
        cls,
        model_path,
        device='cuda',
        dtype=torch.float16,
        use_safetensors=False,
        variant='fp16',
        subfolder='hunyuan3d-vae-v2-1',
        **kwargs,
    ):
        config_path, ckpt_path = smart_load_model(
            model_path,
            subfolder=subfolder,
            use_safetensors=use_safetensors,
            variant=variant
        )

        return cls.from_single_file(
            ckpt_path,
            config_path,
            device=device,
            dtype=dtype,
            use_safetensors=use_safetensors,
            **kwargs
        )
        
    def init_from_ckpt(self, path, ignore_keys=()):
        state_dict = torch.load(path, map_location="cpu")
        state_dict = state_dict.get("state_dict", state_dict)
        keys = list(state_dict.keys())
        for k in keys:
            for ik in ignore_keys:
                if k.startswith(ik):
                    print("Deleting key {} from state_dict.".format(k))
                    del state_dict[k]
        missing, unexpected = self.load_state_dict(state_dict, strict=False)
        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
        if len(missing) > 0:
            print(f"Missing Keys: {missing}")
            print(f"Unexpected Keys: {unexpected}")

    def __init__(
        self,
        volume_decoder=None,
        surface_extractor=None
    ):
        super().__init__()
        if volume_decoder is None:
            volume_decoder = VanillaVolumeDecoder()
        if surface_extractor is None:
            surface_extractor = MCSurfaceExtractor()
        self.volume_decoder = volume_decoder
        self.surface_extractor = surface_extractor

    def latents2mesh(self, latents: torch.FloatTensor, **kwargs):
        with synchronize_timer('Volume decoding'):
            grid_logits = self.volume_decoder(latents, self.geo_decoder, **kwargs)
        with synchronize_timer('Surface extraction'):
            outputs = self.surface_extractor(grid_logits, **kwargs)
        return outputs

    def enable_flashvdm_decoder(
        self,
        enabled: bool = True,
        adaptive_kv_selection=True,
        topk_mode='mean',
        mc_algo='dmc',
    ):
        if enabled:
            if adaptive_kv_selection:
                self.volume_decoder = FlashVDMVolumeDecoding(topk_mode)
            else:
                self.volume_decoder = HierarchicalVolumeDecoding()
            if mc_algo not in SurfaceExtractors.keys():
                raise ValueError(f'Unsupported mc_algo {mc_algo}, available:{list(SurfaceExtractors.keys())}')
            self.surface_extractor = SurfaceExtractors[mc_algo]()
        else:
            self.volume_decoder = VanillaVolumeDecoder()
            self.surface_extractor = MCSurfaceExtractor()


class ShapeVAE(VectsetVAE):
    def __init__(
        self,
        *,
        num_latents: int,
        embed_dim: int,
        width: int,
        heads: int,
        num_decoder_layers: int,
        num_encoder_layers: int = 8,
        pc_size: int = 5120,
        pc_sharpedge_size: int = 5120,
        point_feats: int = 3,
        downsample_ratio: int = 20,
        geo_decoder_downsample_ratio: int = 1,
        geo_decoder_mlp_expand_ratio: int = 4,
        geo_decoder_ln_post: bool = True,
        num_freqs: int = 8,
        include_pi: bool = True,
        qkv_bias: bool = True,
        qk_norm: bool = False,
        label_type: str = "binary",
        drop_path_rate: float = 0.0,
        scale_factor: float = 1.0,
        use_ln_post: bool = True,
        ckpt_path = None
    ):
        super().__init__()
        self.geo_decoder_ln_post = geo_decoder_ln_post
        self.downsample_ratio = downsample_ratio

        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)

        self.encoder = PointCrossAttentionEncoder(
            fourier_embedder=self.fourier_embedder,
            num_latents=num_latents,
            downsample_ratio=self.downsample_ratio,
            pc_size=pc_size,
            pc_sharpedge_size=pc_sharpedge_size,
            point_feats=point_feats,
            width=width,
            heads=heads,
            layers=num_encoder_layers,
            qkv_bias=qkv_bias,
            use_ln_post=use_ln_post,
            qk_norm=qk_norm
        )

        self.pre_kl = nn.Linear(width, embed_dim * 2)
        self.post_kl = nn.Linear(embed_dim, width)

        self.transformer = Transformer(
            n_ctx=num_latents,
            width=width,
            layers=num_decoder_layers,
            heads=heads,
            qkv_bias=qkv_bias,
            qk_norm=qk_norm,
            drop_path_rate=drop_path_rate
        )

        self.geo_decoder = CrossAttentionDecoder(
            fourier_embedder=self.fourier_embedder,
            out_channels=1,
            num_latents=num_latents,
            mlp_expand_ratio=geo_decoder_mlp_expand_ratio,
            downsample_ratio=geo_decoder_downsample_ratio,
            enable_ln_post=self.geo_decoder_ln_post,
            width=width // geo_decoder_downsample_ratio,
            heads=heads // geo_decoder_downsample_ratio,
            qkv_bias=qkv_bias,
            qk_norm=qk_norm,
            label_type=label_type,
        )

        self.scale_factor = scale_factor
        self.latent_shape = (num_latents, embed_dim)

        if ckpt_path is not None:
            self.init_from_ckpt(ckpt_path)

    def forward(self, latents):
        latents = self.post_kl(latents)
        latents = self.transformer(latents)
        return latents

    def encode(self, surface, sample_posterior=True):
        pc, feats = surface[:, :, :3], surface[:, :, 3:]
        latents, _ = self.encoder(pc, feats)
        # print(latents.shape, self.pre_kl.weight.shape)
        moments = self.pre_kl(latents)
        posterior = DiagonalGaussianDistribution(moments, feat_dim=-1)
        if sample_posterior:
            latents = posterior.sample()
        else:
            latents = posterior.mode()
        return latents

    def decode(self, latents):
        latents = self.post_kl(latents)
        latents = self.transformer(latents)
        return latents