Spaces:
Paused
Paused
File size: 8,636 Bytes
9d7cf7f | 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 | from dataclasses import asdict, dataclass
from omegaconf import OmegaConf
from scipy.spatial import cKDTree # type: ignore
from torch import nn, Tensor
from typing import Dict, List
import math
import numpy as np
import random
import torch
import torch.nn.functional as F
from src.rig_package.info.asset import Asset
from .spec import ModelSpec, ModelInput, VaeInput
from .skin_vae.autoencoders import SkinFSQCVAEModel
try:
from flash_attn_interface import flash_attn_func # type: ignore
except Exception as e:
from flash_attn.flash_attn_interface import flash_attn_func as _flash_attn_func
def flash_attn_func(*args, **kwargs):
res = _flash_attn_func(*args, **kwargs)
return res, None
class Perceiver(nn.Module):
def __init__(self, channels, out_tokens, num_heads=8):
super().__init__()
self.q_vec = nn.Parameter(torch.randn(out_tokens // num_heads, num_heads, channels) * 0.02)
self.num_heads = num_heads
self.head_dim = channels // num_heads
self.k_proj = nn.Linear(channels, channels)
self.v_proj = nn.Linear(channels, channels)
self.out_proj = nn.Linear(channels, channels)
def forward(self, x: Tensor) -> Tensor:
B, N, C = x.shape
k = self.k_proj(x) # [B, N, C]
v = self.v_proj(x) # [B, N, C]
q_repeated = self.q_vec.repeat(B, 1, 1, 1)
q = q_repeated.view(B, -1, self.num_heads, self.head_dim).type(torch.bfloat16)
k = k.view(B, -1, self.num_heads, self.head_dim)
v = v.view(B, -1, self.num_heads, self.head_dim)
hidden_states, _ = flash_attn_func(q, k, v)
hidden_states = hidden_states.view(B, -1, self.num_heads * self.head_dim) # type: ignore
hidden_states = self.out_proj(hidden_states)
return hidden_states
class SkinVAEModel(ModelSpec):
def __init__(self, model_config, transform_config, tokenizer_config=None):
super().__init__(model_config, transform_config, tokenizer_config)
cfg = self.model_config
self.cond_tokens = cfg['sample']['cond_tokens']
self.compress_tokens = cfg['sample']['compress_tokens']
self.sample_tokens = cfg['sample']['sample_tokens']
self.only_dense = cfg['sample'].get('only_dense', False)
self.model_type = cfg.get('type', 'fsqc')
if self.model_type == 'fsqc':
self.model = SkinFSQCVAEModel(**cfg['model'], sample_tokens=self.sample_tokens)
else:
raise NotImplementedError()
if self.sample_tokens != self.compress_tokens:
self.down_perceiver = Perceiver(self.model.latent_channels, self.compress_tokens)
if self.sample_tokens != self.compress_tokens:
self.up_perceiver = Perceiver(self.model.latent_channels, self.sample_tokens)
def compile_model(self):
self.model.compile_model()
@property
def vocab_size(self) -> int:
return self.model.FSQ.codebook_size
@property
def latent_channels(self) -> int:
return self.model.latent_channels
def encode(self, vae_input: VaeInput, num_tokens: int=4, j: int=0, full: bool=False, encode_repeat: int=4, return_cond: bool=True):
raise NotImplementedError()
def decode(self, z: Tensor, sampled_cond: Tensor, cond_tokens: Tensor, full: bool=False, encode_repeat: int=4) -> Tensor:
assert z.shape[0] == sampled_cond.shape[0] == cond_tokens.shape[0]
if full:
l = z.shape[0]
s = []
for i in range(0, l, encode_repeat):
t = min(l,i+encode_repeat)
if self.sample_tokens != self.compress_tokens:
_z = self.up_perceiver(z[i:t])
else:
_z = z[i:t]
logits = self.model._decode(z=_z, cond=cond_tokens[i:t], sampled_points=sampled_cond[i:t])
s.append(logits)
return torch.cat(s, dim=0)
else:
if self.sample_tokens != self.compress_tokens:
z = self.up_perceiver(z)
logits = self.model._decode(z=z, cond=cond_tokens, sampled_points=sampled_cond)
return logits
def get_loss_dict(
self,
skin_pred: Tensor,
skin_gt: Tensor,
) -> Dict[str, Tensor]:
raise NotImplementedError()
def get_input(self, batch: Dict) -> VaeInput:
vertices: Tensor = batch['vertices'].float() # (B, N, 3)
normals: Tensor = batch['normals'].float() # (B, N, 3)
uniform_skin: List[Tensor] = batch['uniform_skin'] # [(N, J)]
dense_skin: List[Tensor] = batch['dense_skin'] # [(J, skin_samples)]
dense_vertices: List[Tensor] = batch['dense_vertices'] # [(J, skin_samples, 3)]
dense_normals: List[Tensor] = batch['dense_normals'] # [(J, skin_samples, 3)]
dense_indices: List[List[int]] = batch['dense_indices'] # [List[J]]
B = vertices.shape[0]
uniform_cond = torch.cat([vertices, normals], dim=-1).float()
dense_cond = []
for i in range(B):
dense_cond.append(torch.cat([dense_vertices[i], dense_normals[i]], dim=-1).float())
uniform_skin = [s.float() for s in uniform_skin]
dense_skin = [s.float() for s in dense_skin]
return VaeInput(
dense_cond=dense_cond,
dense_skin=dense_skin,
dense_indices=dense_indices,
uniform_cond=uniform_cond,
uniform_skin=uniform_skin,
)
@torch.autocast(device_type='cuda', dtype=torch.bfloat16)
def training_step(self, batch: Dict) -> Dict:
raise NotImplementedError()
def process_fn(self, batch: List[ModelInput], is_train: bool = True) -> List[Dict]:
res = []
for b in batch:
asset = b.asset
assert asset is not None
assert asset.sampled_vertex_groups is not None
assert 'skin' in asset.sampled_vertex_groups
assert asset.meta is not None
assert 'dense_indices' in asset.meta
assert 'dense_skin' in asset.meta
assert 'dense_vertices' in asset.meta
assert 'dense_normals' in asset.meta
_d = {
'vertices': asset.sampled_vertices,
'normals': b.asset.sampled_normals,
'non': {
'uniform_skin': asset.sampled_vertex_groups['skin'],
'num_bones': asset.J,
'skin_samples': asset.skin_samples,
'dense_indices': asset.meta['dense_indices'],
'dense_skin': asset.meta['dense_skin'],
'dense_vertices': asset.meta['dense_vertices'],
'dense_normals': asset.meta['dense_normals'],
}
}
res.append(_d)
return res
def forward(self, batch: Dict) -> Dict:
return self.training_step(batch=batch)
@torch.autocast('cuda', dtype=torch.bfloat16)
def predict_step(self, batch: Dict) -> Dict:
vertices: Tensor = batch['vertices'].float() # (B, N, 3)
num_bones: List[int] = batch['num_bones']
B = vertices.shape[0]
N = vertices.shape[1]
vae_input = self.get_input(batch=batch)
num_tokens = 4
z, cond_tokens, indices, _ = self.encode(vae_input=vae_input, num_tokens=num_tokens, full=True, encode_repeat=8)
assert cond_tokens is not None
z = self.model.FSQ.indices_to_codes(indices).reshape(z.shape)
_skin_pred = self.decode(z=z, sampled_cond=vae_input.get_flatten_uniform_cond(), cond_tokens=cond_tokens[vae_input.get_flatten_indices()], full=True, encode_repeat=8)
_skin_pred = _skin_pred.squeeze(-1)
tot = 0
results = []
for i in range(B):
asset: Asset = batch['model_input'][i].asset.copy()
skin_pred = torch.zeros((N, num_bones[i]), dtype=vertices.dtype, device=vertices.device)
for j in range(vae_input.get_len(i=i)):
skin_pred[:, vae_input.true_j(i=i, j=j)] = _skin_pred[tot]
tot += 1
sampled_vertices = vertices[i].detach().float().cpu().numpy()
tree = cKDTree(sampled_vertices)
distances, indices = tree.query(asset.vertices)
sampled_skin = skin_pred.detach().float().cpu().numpy()[indices]
asset.skin = sampled_skin
results.append(asset)
return {
'results': results,
} |