Spaces:
Running
on
L4
Running
on
L4
File size: 11,376 Bytes
81ecb2b |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.transforms import Compose, Resize, CenterCrop, Normalize, InterpolationMode
import open_clip
from dva.io import load_from_config
def sample_orbit_traj(radius, height, start_theta, end_theta, num_points, world_up=torch.Tensor([0, 1, 0])):
# return [num_points, 3, 4]
angles = torch.rand((num_points, )) * (end_theta - start_theta) + start_theta
return get_pose_on_orbit(radius=radius, height=height, angles=angles, world_up=world_up)
def get_pose_on_orbit(radius, height, angles, world_up=torch.Tensor([0, 1, 0])):
num_points = angles.shape[0]
x = radius * torch.cos(angles)
h = torch.ones((num_points,)) * height
z = radius * torch.sin(angles)
position = torch.stack([x, h, z], dim=-1)
forward = position / torch.norm(position, p=2, dim=-1, keepdim=True)
right = -torch.cross(world_up[None, ...], forward)
right /= torch.norm(right, dim=-1, keepdim=True)
up = torch.cross(forward, right)
up /= torch.norm(up, p=2, dim=-1, keepdim=True)
rotation = torch.stack([right, up, forward], dim=1)
translation = torch.Tensor([0, 0, radius])[None, :, None].repeat(num_points, 1, 1)
return torch.concat([rotation, translation], dim=2)
class DummyImageConditioner(nn.Module):
def __init__(
self,
num_prims,
dim_feat,
prim_shape,
encoder_config,
sample_view=False,
sample_start=torch.pi*0.25,
sample_end=torch.pi*0.75,
):
super().__init__()
self.num_prims = num_prims
self.dim_feat = dim_feat
self.prim_shape = prim_shape
self.sample_view = sample_view
self.sample_start = sample_start
self.sample_end = sample_end
self.encoder = None
@torch.no_grad()
def forward(self, batch, rm, amp, precision_dtype=torch.float32):
return batch['cond']
class ImageConditioner(nn.Module):
def __init__(
self,
num_prims,
dim_feat,
prim_shape,
encoder_config,
sample_view=False,
sample_start=torch.pi*0.25,
sample_end=torch.pi*0.75,
):
super().__init__()
self.num_prims = num_prims
self.dim_feat = dim_feat
self.prim_shape = prim_shape
self.sample_view = sample_view
self.sample_start = sample_start
self.sample_end = sample_end
self.encoder = load_from_config(encoder_config)
def sdf2alpha(self, sdf):
return torch.exp(-(sdf / 0.005) ** 2)
@torch.no_grad()
def forward(self, batch, rm, amp, precision_dtype=torch.float32):
# TODO: replace with real rendering process in primsdf
assert 'input_param' in batch, "No parameters in current batch for rendering image conditions"
prim_volume = batch['input_param']
bs = prim_volume.shape[0]
preds = {}
geo_start_index = 4
geo_end_index = geo_start_index + self.prim_shape ** 3 # non-inclusive
tex_start_index = geo_end_index
tex_end_index = tex_start_index + self.prim_shape ** 3 * 3 # non-inclusive
feat_geo = prim_volume[:, :, geo_start_index: geo_end_index]
feat_tex = prim_volume[:, :, tex_start_index: tex_end_index]
prim_alpha = self.sdf2alpha(feat_geo).reshape(bs, self.num_prims, 1, self.prim_shape, self.prim_shape, self.prim_shape) * 255
prim_rgb = feat_tex.reshape(bs, self.num_prims, 3, self.prim_shape, self.prim_shape, self.prim_shape) * 255
preds['prim_rgba'] = torch.concat([prim_rgb, prim_alpha], dim=2)
pos = prim_volume[:, :, 1:4]
scale = prim_volume[:, :, 0:1]
preds['prim_pos'] = pos.reshape(bs, self.num_prims, 3) * rm.volradius
preds['prim_rot'] = torch.eye(3).to(preds['prim_pos'])[None, None, ...].repeat(bs, self.num_prims, 1, 1)
preds['prim_scale'] = (1 / scale.reshape(bs, self.num_prims, 1).repeat(1, 1, 3))
if not self.sample_view:
preds['Rt'] = torch.Tensor([
[
1.0,
0.0,
0.0,
0.0 * rm.volradius
],
[
0.0,
-1.0,
0.0,
0.0 * rm.volradius
],
[
0.0,
0.0,
-1.0,
5 * rm.volradius
]
]).to(prim_volume)[None, ...].repeat(bs, 1, 1)
else:
preds['Rt'] = sample_orbit_traj(radius=5*rm.volradius, height=0, start_theta=self.sample_start, end_theta=self.sample_end, num_points=bs).to(prim_volume)
preds['K'] = torch.Tensor([
[
2084.9526697685183,
0.0,
512.0
],
[
0.0,
2084.9526697685183,
512.0
],
[
0.0,
0.0,
1.0
]]).to(prim_volume)[None, ...].repeat(bs, 1, 1)
ratio_h = rm.image_height / 1024.
ratio_w = rm.image_width / 1024.
preds['K'][:, 0:1, :] *= ratio_h
preds['K'][:, 1:2, :] *= ratio_w
rm_preds = rm(
prim_rgba=preds["prim_rgba"],
prim_pos=preds["prim_pos"],
prim_scale=preds["prim_scale"],
prim_rot=preds["prim_rot"],
RT=preds["Rt"],
K=preds["K"],
)
rendered_image = rm_preds['rgba_image'].permute(0, 2, 3, 1)[..., :3].contiguous()
with torch.autocast(device_type='cuda', dtype=precision_dtype, enabled=amp):
results = self.encoder(rendered_image)
return results
class ImageMultiViewConditioner(nn.Module):
def __init__(
self,
num_prims,
dim_feat,
prim_shape,
encoder_config,
sample_view=False,
view_counts=4,
):
super().__init__()
self.num_prims = num_prims
self.dim_feat = dim_feat
self.prim_shape = prim_shape
self.view_counts = view_counts
view_angles = torch.linspace(0.5, 2.5, self.view_counts + 1) * torch.pi
self.view_angles = view_angles[:-1]
self.encoder = load_from_config(encoder_config)
def sdf2alpha(self, sdf):
return torch.exp(-(sdf / 0.005) ** 2)
@torch.no_grad()
def forward(self, batch, rm, amp, precision_dtype=torch.float32):
# TODO: replace with real rendering process in primsdf
assert 'input_param' in batch, "No parameters in current batch for rendering image conditions"
prim_volume = batch['input_param']
bs = prim_volume.shape[0]
preds = {}
geo_start_index = 4
geo_end_index = geo_start_index + self.prim_shape ** 3 # non-inclusive
tex_start_index = geo_end_index
tex_end_index = tex_start_index + self.prim_shape ** 3 * 3 # non-inclusive
feat_geo = prim_volume[:, :, geo_start_index: geo_end_index]
feat_tex = prim_volume[:, :, tex_start_index: tex_end_index]
prim_alpha = self.sdf2alpha(feat_geo).reshape(bs, self.num_prims, 1, self.prim_shape, self.prim_shape, self.prim_shape) * 255
prim_rgb = feat_tex.reshape(bs, self.num_prims, 3, self.prim_shape, self.prim_shape, self.prim_shape) * 255
preds['prim_rgba'] = torch.concat([prim_rgb, prim_alpha], dim=2)
pos = prim_volume[:, :, 1:4]
scale = prim_volume[:, :, 0:1]
preds['prim_pos'] = pos.reshape(bs, self.num_prims, 3) * rm.volradius
preds['prim_rot'] = torch.eye(3).to(preds['prim_pos'])[None, None, ...].repeat(bs, self.num_prims, 1, 1)
preds['prim_scale'] = (1 / scale.reshape(bs, self.num_prims, 1).repeat(1, 1, 3))
preds['K'] = torch.Tensor([
[
2084.9526697685183,
0.0,
512.0
],
[
0.0,
2084.9526697685183,
512.0
],
[
0.0,
0.0,
1.0
]]).to(prim_volume)[None, ...].repeat(bs, 1, 1)
ratio_h = rm.image_height / 1024.
ratio_w = rm.image_width / 1024.
preds['K'][:, 0:1, :] *= ratio_h
preds['K'][:, 1:2, :] *= ratio_w
# we sample view according to view_counts
cond_list = []
for view_ang in self.view_angles:
bs_view_ang = view_ang.repeat(bs,)
preds['Rt'] = get_pose_on_orbit(radius=5*rm.volradius, height=0, angles=bs_view_ang).to(prim_volume)
rm_preds = rm(
prim_rgba=preds["prim_rgba"],
prim_pos=preds["prim_pos"],
prim_scale=preds["prim_scale"],
prim_rot=preds["prim_rot"],
RT=preds["Rt"],
K=preds["K"],
)
rendered_image = rm_preds['rgba_image'].permute(0, 2, 3, 1)[..., :3].contiguous()
with torch.autocast(device_type='cuda', dtype=precision_dtype, enabled=amp):
results = self.encoder(rendered_image)
cond_list.append(results)
final_cond = torch.concat(cond_list, dim=1)
return final_cond
class CLIPImageEncoder(nn.Module):
def __init__(
self,
pretrained_path: str,
model_spec: str = 'ViT-L-14',
):
super().__init__()
self.model, _, _ = open_clip.create_model_and_transforms(model_spec, pretrained=pretrained_path)
self.model_resolution = self.model.visual.image_size
self.preprocess = Compose([
Resize(self.model_resolution, interpolation=InterpolationMode.BICUBIC),
CenterCrop(self.model_resolution),
Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
])
self.model.eval()
# self.tokenizer = open_clip.get_tokenizer(model_spec)
@torch.no_grad()
def forward(self, img):
assert img.shape[-1] == 3
img = img.permute(0, 3, 1, 2) / 255.
image = self.preprocess(img)
image_features = self.model.encode_image(image)
image_features /= image_features.norm(dim=-1, keepdim=True)
return image_features
class CLIPImageTokenEncoder(nn.Module):
def __init__(
self,
pretrained_path: str,
model_spec: str = 'ViT-L-14',
):
super().__init__()
self.model, _, _ = open_clip.create_model_and_transforms(model_spec, pretrained=pretrained_path)
self.model.visual.output_tokens = True
self.model_resolution = self.model.visual.image_size
self.preprocess = Compose([
Resize(self.model_resolution, interpolation=InterpolationMode.BICUBIC),
CenterCrop(self.model_resolution),
Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
])
self.model.eval()
@torch.no_grad()
def forward(self, img):
assert img.shape[-1] == 3
img = img.permute(0, 3, 1, 2) / 255.
image = self.preprocess(img)
_, image_tokens = self.model.encode_image(image)
# [B, T, D] - [B, 256, 1024]
image_tokens /= image_tokens.norm(dim=-1, keepdim=True)
return image_tokens |