File size: 12,619 Bytes
491eded |
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 |
from dataclasses import dataclass
import os
import sys
import torch
import trimesh
from torch import nn
from transformers import AutoModelForCausalLM
from transformers.generation.logits_process import LogitsProcessorList
from einops import rearrange
from modules.bbox_gen.models.image_encoder import DINOv2ImageEncoder
from modules.bbox_gen.config import parse_structured
from modules.bbox_gen.models.bboxopt import BBoxOPT, BBoxOPTConfig
from modules.bbox_gen.utils.bbox_tokenizer import BoundsTokenizerDiag
from modules.bbox_gen.models.bbox_gen_models import GroupEmbedding, MultiModalProjector, MeshDecodeLogitsProcessor, SparseStructureEncoder
current_dir = os.path.dirname(os.path.abspath(__file__))
modules_dir = os.path.dirname(os.path.dirname(current_dir))
partfield_dir = os.path.join(modules_dir, 'PartField')
if partfield_dir not in sys.path:
sys.path.insert(0, partfield_dir)
import importlib.util
from partfield.config import default_argument_parser, setup
class BboxGen(nn.Module):
@dataclass
class Config:
# encoder config
encoder_dim_feat: int = 3
encoder_dim: int = 64
encoder_heads: int = 4
encoder_token_num: int = 256
encoder_qkv_bias: bool = False
encoder_use_ln_post: bool = True
encoder_use_checkpoint: bool = False
encoder_num_embed_freqs: int = 8
encoder_embed_include_pi: bool = False
encoder_init_scale: float = 0.25
encoder_random_fps: bool = True
encoder_learnable_query: bool = False
encoder_layers: int = 4
group_embedding_dim: int = 64
# decoder config
vocab_size: int = 518
decoder_hidden_size: int = 1536
decoder_num_hidden_layers: int = 24
decoder_ffn_dim: int = 6144
decoder_heads: int = 16
decoder_use_flash_attention: bool = True
decoder_gradient_checkpointing: bool = True
# data config
bins: int = 64
BOS_id: int = 64
EOS_id: int = 65
PAD_id: int = 66
max_length: int = 2187 # bos + 50x2x3 + 1374 + 512
voxel_token_length: int = 1886
voxel_token_placeholder: int = -1
# tokenizer config
max_group_size: int = 50
# voxel encoder
partfield_encoder_path: str = ""
cfg: Config
def __init__(self, cfg):
super().__init__()
self.cfg = parse_structured(self.Config, cfg)
self.image_encoder = DINOv2ImageEncoder(
model_name="facebook/dinov2-with-registers-large",
)
self.image_projector = MultiModalProjector(
in_features=(1024 + self.cfg.group_embedding_dim),
out_features=self.cfg.decoder_hidden_size,
)
self.group_embedding = GroupEmbedding(
max_group_size=self.cfg.max_group_size,
hidden_size=self.cfg.group_embedding_dim,
)
self.decoder_config = BBoxOPTConfig(
vocab_size=self.cfg.vocab_size,
hidden_size=self.cfg.decoder_hidden_size,
num_hidden_layers=self.cfg.decoder_num_hidden_layers,
ffn_dim=self.cfg.decoder_ffn_dim,
max_position_embeddings=self.cfg.max_length,
num_attention_heads=self.cfg.decoder_heads,
pad_token_id=self.cfg.PAD_id,
bos_token_id=self.cfg.BOS_id,
eos_token_id=self.cfg.EOS_id,
use_cache=True,
init_std=0.02,
)
if self.cfg.decoder_use_flash_attention:
self.decoder: BBoxOPT = AutoModelForCausalLM.from_config(
self.decoder_config,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2"
)
else:
self.decoder: BBoxOPT = AutoModelForCausalLM.from_config(
self.decoder_config,
)
if self.cfg.decoder_gradient_checkpointing:
self.decoder.gradient_checkpointing_enable()
self.logits_processor = LogitsProcessorList()
self.logits_processor.append(MeshDecodeLogitsProcessor(
bins=self.cfg.bins,
BOS_id=self.cfg.BOS_id,
EOS_id=self.cfg.EOS_id,
PAD_id=self.cfg.PAD_id,
vertices_num=2,
))
self.tokenizer = BoundsTokenizerDiag(
bins=self.cfg.bins,
BOS_id=self.cfg.BOS_id,
EOS_id=self.cfg.EOS_id,
PAD_id=self.cfg.PAD_id,
)
self._load_partfield_encoder()
self.partfield_voxel_encoder = SparseStructureEncoder(
in_channels=451,
channels=[448, 448, 448, 1024],
latent_channels=448,
num_res_blocks=1,
num_res_blocks_middle=1,
norm_type="layer",
)
def _load_partfield_encoder(self):
# Load PartField encoder
model_spec = importlib.util.spec_from_file_location(
"partfield.partfield_encoder",
os.path.join(partfield_dir, "partfield", "partfield_encoder.py")
)
model_module = importlib.util.module_from_spec(model_spec)
model_spec.loader.exec_module(model_module)
Model = model_module.Model
parser = default_argument_parser()
args = []
args.extend(["-c", os.path.join(partfield_dir, "configs/final/demo.yaml")])
args.append("--opts")
args.extend(["continue_ckpt", self.cfg.partfield_encoder_path])
parsed_args = parser.parse_args(args)
cfg = setup(parsed_args, freeze=False)
self.partfield_encoder = Model(cfg)
self.partfield_encoder.eval()
weights = torch.load(self.cfg.partfield_encoder_path)["state_dict"]
self.partfield_encoder.load_state_dict(weights)
for param in self.partfield_encoder.parameters():
param.requires_grad = False
print("PartField encoder loaded")
def _prepare_lm_inputs(self, voxel_token, input_ids):
inputs_embeds = torch.zeros(input_ids.shape[0], input_ids.shape[1], self.cfg.decoder_hidden_size, device=input_ids.device, dtype=voxel_token.dtype)
voxel_token_mask = (input_ids == self.cfg.voxel_token_placeholder)
inputs_embeds[voxel_token_mask] = voxel_token.view(-1, self.cfg.decoder_hidden_size)
inputs_embeds[~voxel_token_mask] = self.decoder.get_input_embeddings()(input_ids[~voxel_token_mask]).to(dtype=inputs_embeds.dtype)
attention_mask = (input_ids != self.cfg.PAD_id)
return inputs_embeds, attention_mask.long()
def forward(self, batch):
image_latents = self.image_encoder(batch['images'])
masks = batch['masks']
masks_emb = self.group_embedding(masks)
masks_emb = rearrange(masks_emb, 'b c h w -> b (h w) c') # B x Q x C
group_emb = torch.zeros((image_latents.shape[0], image_latents.shape[1], masks_emb.shape[2]), device=image_latents.device, dtype=image_latents.dtype)
group_emb[:, :masks_emb.shape[1], :] = masks_emb
image_latents = torch.cat([image_latents, group_emb], dim=-1)
image_latents = self.image_projector(image_latents)
points = batch['points'][..., :3]
rot_matrix = torch.tensor([[1, 0, 0], [0, 0, -1], [0, 1, 0]], device=points.device, dtype=points.dtype)
rot_points = torch.matmul(points, rot_matrix)
rot_points = rot_points * (2 * 0.9) # from (-0.5, 0.5) to (-1, 1)
partfield_feat = self.partfield_encoder.encode(rot_points)
feat_volume = torch.zeros((points.shape[0], 448, 64, 64, 64), device=partfield_feat.device, dtype=partfield_feat.dtype)
whole_voxel_index = batch['whole_voxel_index'] # (b, m, 3)
batch_size, num_points = whole_voxel_index.shape[0], whole_voxel_index.shape[1]
batch_indices = torch.arange(batch_size, device=whole_voxel_index.device).unsqueeze(1).expand(-1, num_points) # (b, m)
batch_flat = batch_indices.flatten() # (b*m,)
x_flat = whole_voxel_index[..., 0].flatten() # (b*m,)
y_flat = whole_voxel_index[..., 1].flatten() # (b*m,)
z_flat = whole_voxel_index[..., 2].flatten() # (b*m,)
partfield_feat_flat = partfield_feat.reshape(-1, 448) # (b*m, 448)
feat_volume[batch_flat, :, x_flat, y_flat, z_flat] = partfield_feat_flat
xyz_volume = torch.zeros((points.shape[0], 3, 64, 64, 64), device=points.device, dtype=points.dtype)
xyz_volume[batch_flat, :, x_flat, y_flat, z_flat] = points.reshape(-1, 3)
feat_volume = torch.cat([feat_volume, xyz_volume], dim=1)
feat_volume = self.partfield_voxel_encoder(feat_volume)
feat_volume = rearrange(feat_volume, 'b c x y z -> b (x y z) c')
voxel_token = torch.cat([image_latents, feat_volume], dim=1) # B x N x D
input_ids = batch['input_ids']
inputs_embeds, attention_mask = self._prepare_lm_inputs(voxel_token, input_ids)
output = self.decoder(
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
return_dict=True,
)
return {
"logits": output.logits,
}
def gen_mesh_from_bounds(self, bounds, random_color):
bboxes = []
for j in range(bounds.shape[0]):
bbox = trimesh.primitives.Box(bounds=bounds[j])
color = random_color[j]
bbox.visual.vertex_colors = color
bboxes.append(bbox)
mesh = trimesh.Scene(bboxes)
return mesh
def generate(self, batch):
image_latents = self.image_encoder(batch['images'])
masks = batch['masks']
masks_emb = self.group_embedding(masks)
masks_emb = rearrange(masks_emb, 'b c h w -> b (h w) c') # B x Q x C
group_emb = torch.zeros((image_latents.shape[0], image_latents.shape[1], masks_emb.shape[2]), device=image_latents.device, dtype=image_latents.dtype)
group_emb[:, :masks_emb.shape[1], :] = masks_emb
image_latents = torch.cat([image_latents, group_emb], dim=-1)
image_latents = self.image_projector(image_latents)
points = batch['points'][..., :3]
rot_matrix = torch.tensor([[1, 0, 0], [0, 0, -1], [0, 1, 0]], device=points.device, dtype=points.dtype)
rot_points = torch.matmul(points, rot_matrix)
rot_points = rot_points * (2 * 0.9) # from (-0.5, 0.5) to (-1, 1)
partfield_feat = self.partfield_encoder.encode(rot_points)
feat_volume = torch.zeros((points.shape[0], 448, 64, 64, 64), device=partfield_feat.device, dtype=partfield_feat.dtype)
whole_voxel_index = batch['whole_voxel_index'] # (b, m, 3)
batch_size, num_points = whole_voxel_index.shape[0], whole_voxel_index.shape[1]
batch_indices = torch.arange(batch_size, device=whole_voxel_index.device).unsqueeze(1).expand(-1, num_points) # (b, m)
batch_flat = batch_indices.flatten() # (b*m,)
x_flat = whole_voxel_index[..., 0].flatten() # (b*m,)
y_flat = whole_voxel_index[..., 1].flatten() # (b*m,)
z_flat = whole_voxel_index[..., 2].flatten() # (b*m,)
partfield_feat_flat = partfield_feat.reshape(-1, 448) # (b*m, 448)
feat_volume[batch_flat, :, x_flat, y_flat, z_flat] = partfield_feat_flat
xyz_volume = torch.zeros((points.shape[0], 3, 64, 64, 64), device=points.device, dtype=points.dtype)
xyz_volume[batch_flat, :, x_flat, y_flat, z_flat] = points.reshape(-1, 3)
feat_volume = torch.cat([feat_volume, xyz_volume], dim=1)
feat_volume = self.partfield_voxel_encoder(feat_volume)
feat_volume = rearrange(feat_volume, 'b c x y z -> b (x y z) c')
voxel_token = torch.cat([image_latents, feat_volume], dim=1) # B x N x D
meshes = []
mesh_names = []
bboxes = []
output = self.decoder.generate(
inputs_embeds=voxel_token,
max_new_tokens=self.cfg.max_length - voxel_token.shape[1],
logits_processor=self.logits_processor,
do_sample=True,
top_k=5,
top_p=0.95,
temperature=0.5,
use_cache=True,
)
for i in range(output.shape[0]):
bounds = self.tokenizer.decode(output[i].detach().cpu().numpy(), coord_rg=(-0.5, 0.5))
# mesh = self.gen_mesh_from_bounds(bounds, batch['random_color'][i])
# meshes.append(mesh)
mesh_names.append("topk=5")
bboxes.append(bounds)
return {
# 'meshes': meshes,
'mesh_names': mesh_names,
'bboxes': bboxes,
}
|