Spaces:
Running
on
Zero
Running
on
Zero
File size: 10,652 Bytes
178f950 |
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 |
from typing import *
import torch
import torch.nn as nn
import numpy as np
from transformers import CLIPTextModel, AutoTokenizer
# import open3d as o3d
from .base import Pipeline
from . import samplers
from ..modules import sparse as sp
class TrellisTextTo3DPipeline(Pipeline):
"""
Pipeline for inferring Trellis text-to-3D models.
Args:
models (dict[str, nn.Module]): The models to use in the pipeline.
sparse_structure_sampler (samplers.Sampler): The sampler for the sparse structure.
slat_sampler (samplers.Sampler): The sampler for the structured latent.
slat_normalization (dict): The normalization parameters for the structured latent.
text_cond_model (str): The name of the text conditioning model.
"""
def __init__(
self,
models: dict[str, nn.Module] = None,
sparse_structure_sampler: samplers.Sampler = None,
slat_sampler: samplers.Sampler = None,
slat_normalization: dict = None,
text_cond_model: str = None,
):
if models is None:
return
super().__init__(models)
self.sparse_structure_sampler = sparse_structure_sampler
self.slat_sampler = slat_sampler
self.sparse_structure_sampler_params = {}
self.slat_sampler_params = {}
self.slat_normalization = slat_normalization
self._init_text_cond_model(text_cond_model)
@staticmethod
def from_pretrained(path: str) -> "TrellisTextTo3DPipeline":
"""
Load a pretrained model.
Args:
path (str): The path to the model. Can be either local path or a Hugging Face repository.
"""
pipeline = super(TrellisTextTo3DPipeline, TrellisTextTo3DPipeline).from_pretrained(path)
new_pipeline = TrellisTextTo3DPipeline()
new_pipeline.__dict__ = pipeline.__dict__
args = pipeline._pretrained_args
new_pipeline.sparse_structure_sampler = getattr(samplers, args['sparse_structure_sampler']['name'])(**args['sparse_structure_sampler']['args'])
new_pipeline.sparse_structure_sampler_params = args['sparse_structure_sampler']['params']
new_pipeline.slat_sampler = getattr(samplers, args['slat_sampler']['name'])(**args['slat_sampler']['args'])
new_pipeline.slat_sampler_params = args['slat_sampler']['params']
new_pipeline.slat_normalization = args['slat_normalization']
new_pipeline._init_text_cond_model(args['text_cond_model'])
return new_pipeline
def _init_text_cond_model(self, name: str):
"""
Initialize the text conditioning model.
"""
# load model
model = CLIPTextModel.from_pretrained(name)
tokenizer = AutoTokenizer.from_pretrained(name)
model.eval()
model = model.cuda()
self.text_cond_model = {
'model': model,
'tokenizer': tokenizer,
}
self.text_cond_model['null_cond'] = self.encode_text([''])
@torch.no_grad()
def encode_text(self, text: List[str]) -> torch.Tensor:
"""
Encode the text.
"""
assert isinstance(text, list) and all(isinstance(t, str) for t in text), "text must be a list of strings"
encoding = self.text_cond_model['tokenizer'](text, max_length=77, padding='max_length', truncation=True, return_tensors='pt')
tokens = encoding['input_ids'].cuda()
embeddings = self.text_cond_model['model'](input_ids=tokens).last_hidden_state
return embeddings
def get_cond(self, prompt: List[str]) -> dict:
"""
Get the conditioning information for the model.
Args:
prompt (List[str]): The text prompt.
Returns:
dict: The conditioning information
"""
cond = self.encode_text(prompt)
neg_cond = self.text_cond_model['null_cond']
return {
'cond': cond,
'neg_cond': neg_cond,
}
def sample_sparse_structure(
self,
cond: dict,
num_samples: int = 1,
sampler_params: dict = {},
) -> torch.Tensor:
"""
Sample sparse structures with the given conditioning.
Args:
cond (dict): The conditioning information.
num_samples (int): The number of samples to generate.
sampler_params (dict): Additional parameters for the sampler.
"""
# Sample occupancy latent
flow_model = self.models['sparse_structure_flow_model']
reso = flow_model.resolution
noise = torch.randn(num_samples, flow_model.in_channels, reso, reso, reso).to(self.device)
sampler_params = {**self.sparse_structure_sampler_params, **sampler_params}
z_s = self.sparse_structure_sampler.sample(
flow_model,
noise,
**cond,
**sampler_params,
verbose=True
).samples
# Decode occupancy latent
decoder = self.models['sparse_structure_decoder']
coords = torch.argwhere(decoder(z_s)>0)[:, [0, 2, 3, 4]].int()
return coords
def decode_slat(
self,
slat: sp.SparseTensor,
formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
) -> dict:
"""
Decode the structured latent.
Args:
slat (sp.SparseTensor): The structured latent.
formats (List[str]): The formats to decode the structured latent to.
Returns:
dict: The decoded structured latent.
"""
ret = {}
if 'mesh' in formats:
ret['mesh'] = self.models['slat_decoder_mesh'](slat)
if 'gaussian' in formats:
ret['gaussian'] = self.models['slat_decoder_gs'](slat)
if 'radiance_field' in formats:
ret['radiance_field'] = self.models['slat_decoder_rf'](slat)
return ret
def sample_slat(
self,
cond: dict,
coords: torch.Tensor,
sampler_params: dict = {},
) -> sp.SparseTensor:
"""
Sample structured latent with the given conditioning.
Args:
cond (dict): The conditioning information.
coords (torch.Tensor): The coordinates of the sparse structure.
sampler_params (dict): Additional parameters for the sampler.
"""
# Sample structured latent
flow_model = self.models['slat_flow_model']
noise = sp.SparseTensor(
feats=torch.randn(coords.shape[0], flow_model.in_channels).to(self.device),
coords=coords,
)
sampler_params = {**self.slat_sampler_params, **sampler_params}
slat = self.slat_sampler.sample(
flow_model,
noise,
**cond,
**sampler_params,
verbose=True
).samples
std = torch.tensor(self.slat_normalization['std'])[None].to(slat.device)
mean = torch.tensor(self.slat_normalization['mean'])[None].to(slat.device)
slat = slat * std + mean
return slat
@torch.no_grad()
def run(
self,
prompt: str,
num_samples: int = 1,
seed: int = 42,
sparse_structure_sampler_params: dict = {},
slat_sampler_params: dict = {},
formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
) -> dict:
"""
Run the pipeline.
Args:
prompt (str): The text prompt.
num_samples (int): The number of samples to generate.
seed (int): The random seed.
sparse_structure_sampler_params (dict): Additional parameters for the sparse structure sampler.
slat_sampler_params (dict): Additional parameters for the structured latent sampler.
formats (List[str]): The formats to decode the structured latent to.
"""
cond = self.get_cond([prompt])
torch.manual_seed(seed)
coords = self.sample_sparse_structure(cond, num_samples, sparse_structure_sampler_params)
slat = self.sample_slat(cond, coords, slat_sampler_params)
return self.decode_slat(slat, formats)
'''
def voxelize(self, mesh: o3d.geometry.TriangleMesh) -> torch.Tensor:
"""
Voxelize a mesh.
Args:
mesh (o3d.geometry.TriangleMesh): The mesh to voxelize.
sha256 (str): The SHA256 hash of the mesh.
output_dir (str): The output directory.
"""
vertices = np.asarray(mesh.vertices)
aabb = np.stack([vertices.min(0), vertices.max(0)])
center = (aabb[0] + aabb[1]) / 2
scale = (aabb[1] - aabb[0]).max()
vertices = (vertices - center) / scale
vertices = np.clip(vertices, -0.5 + 1e-6, 0.5 - 1e-6)
mesh.vertices = o3d.utility.Vector3dVector(vertices)
voxel_grid = o3d.geometry.VoxelGrid.create_from_triangle_mesh_within_bounds(mesh, voxel_size=1/64, min_bound=(-0.5, -0.5, -0.5), max_bound=(0.5, 0.5, 0.5))
vertices = np.array([voxel.grid_index for voxel in voxel_grid.get_voxels()])
return torch.tensor(vertices).int().cuda()
@torch.no_grad()
def run_variant(
self,
mesh: o3d.geometry.TriangleMesh,
prompt: str,
num_samples: int = 1,
seed: int = 42,
slat_sampler_params: dict = {},
formats: List[str] = ['mesh', 'gaussian', 'radiance_field'],
) -> dict:
"""
Run the pipeline for making variants of an asset.
Args:
mesh (o3d.geometry.TriangleMesh): The base mesh.
prompt (str): The text prompt.
num_samples (int): The number of samples to generate.
seed (int): The random seed
slat_sampler_params (dict): Additional parameters for the structured latent sampler.
formats (List[str]): The formats to decode the structured latent to.
"""
cond = self.get_cond([prompt])
coords = self.voxelize(mesh)
coords = torch.cat([
torch.arange(num_samples).repeat_interleave(coords.shape[0], 0)[:, None].int().cuda(),
coords.repeat(num_samples, 1)
], 1)
torch.manual_seed(seed)
slat = self.sample_slat(cond, coords, slat_sampler_params)
return self.decode_slat(slat, formats)
''' |