File size: 11,755 Bytes
8d7ec14 |
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 |
# Author: Bingxin Ke
# Last modified: 2023-12-11
import logging
from typing import Dict
import numpy as np
import torch
from diffusers import (
DDIMScheduler,
DDPMScheduler,
PNDMScheduler,
DEISMultistepScheduler,
SchedulerMixin,
UNet2DConditionModel,
)
from torch import nn
from torch.nn import Conv2d
from torch.nn.parameter import Parameter
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
from .rgb_encoder import RGBEncoder
from .stacked_depth_AE import StackedDepthAE
class MarigoldPipeline(nn.Module):
"""
Marigold monocular depth estimator.
"""
def __init__(
self,
unet_pretrained_path: Dict, # {path: xxx, subfolder: xxx}
rgb_encoder_pretrained_path: Dict,
depht_ae_pretrained_path: Dict,
noise_scheduler_pretrained_path: Dict,
tokenizer_pretrained_path: Dict,
text_encoder_pretrained_path: Dict,
empty_text_embed=None,
trainable_unet=False,
rgb_latent_scale_factor=0.18215,
depth_latent_scale_factor=0.18215,
noise_scheduler_type=None,
enable_gradient_checkpointing=False,
enable_xformers=True,
) -> None:
super().__init__()
self.rgb_latent_scale_factor = rgb_latent_scale_factor
self.depth_latent_scale_factor = depth_latent_scale_factor
self.device = "cpu"
# ******* Initialize modules *******
# Trainable modules
self.trainable_module_dic: Dict[str, nn.Module] = {}
self.trainable_unet = trainable_unet
# Denoising UNet
self.unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
unet_pretrained_path["path"], subfolder=unet_pretrained_path["subfolder"]
)
logging.info(f"pretrained UNet loaded from: {unet_pretrained_path}")
if 8 != self.unet.config["in_channels"]:
self._replace_unet_conv_in()
logging.warning("Unet conv_in layer is replaced")
if enable_xformers:
self.unet.enable_xformers_memory_efficient_attention()
else:
self.unet.disable_xformers_memory_efficient_attention()
# Image encoder
self.rgb_encoder = RGBEncoder(
pretrained_path=rgb_encoder_pretrained_path["path"],
subfolder=rgb_encoder_pretrained_path["subfolder"],
)
logging.info(
f"pretrained RGBEncoder loaded from: {rgb_encoder_pretrained_path}"
)
self.rgb_encoder.requires_grad_(False)
# Depth encoder-decoder
self.depth_ae = StackedDepthAE(
pretrained_path=depht_ae_pretrained_path["path"],
subfolder=depht_ae_pretrained_path["subfolder"],
)
logging.info(
f"pretrained Depth Autoencoder loaded from: {rgb_encoder_pretrained_path}"
)
# Trainability
# unet
if self.trainable_unet:
self.unet.requires_grad_(True)
self.trainable_module_dic["unet"] = self.unet
logging.debug(f"UNet is set to trainable")
else:
self.unet.requires_grad_(False)
logging.debug(f"UNet is set to frozen")
# Gradient checkpointing
if enable_gradient_checkpointing:
self.unet.enable_gradient_checkpointing()
self.depth_ae.vae.enable_gradient_checkpointing()
# Noise scheduler
if "DDPMScheduler" == noise_scheduler_type:
self.noise_scheduler: SchedulerMixin = DDPMScheduler.from_pretrained(
noise_scheduler_pretrained_path["path"],
subfolder=noise_scheduler_pretrained_path["subfolder"],
)
elif "DDIMScheduler" == noise_scheduler_type:
self.noise_scheduler: SchedulerMixin = DDIMScheduler.from_pretrained(
noise_scheduler_pretrained_path["path"],
subfolder=noise_scheduler_pretrained_path["subfolder"],
)
elif "PNDMScheduler" == noise_scheduler_type:
self.noise_scheduler: SchedulerMixin = PNDMScheduler.from_pretrained(
noise_scheduler_pretrained_path["path"],
subfolder=noise_scheduler_pretrained_path["subfolder"],
)
elif "DEISMultistepScheduler" == noise_scheduler_type:
self.noise_scheduler: SchedulerMixin = DEISMultistepScheduler.from_pretrained(
noise_scheduler_pretrained_path["path"],
subfolder=noise_scheduler_pretrained_path["subfolder"],
)
else:
raise NotImplementedError
# Text embed for empty prompt (always in CPU)
if empty_text_embed is None:
tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(
tokenizer_pretrained_path["path"],
subfolder=tokenizer_pretrained_path["subfolder"],
)
text_encoder: CLIPTextModel = CLIPTextModel.from_pretrained(
text_encoder_pretrained_path["path"],
subfolder=text_encoder_pretrained_path["subfolder"],
)
with torch.no_grad():
self.empty_text_embed = self._encode_text(
"", tokenizer, text_encoder
).detach()#.to(dtype=precision) # [1, 2, 1024]
else:
self.empty_text_embed = empty_text_embed
def from_pretrained(pretrained_path, **kwargs):
return __class__(
unet_pretrained_path={"path": pretrained_path, "subfolder": "unet"},
rgb_encoder_pretrained_path={"path": pretrained_path, "subfolder": "vae"},
depht_ae_pretrained_path={"path": pretrained_path, "subfolder": "vae"},
noise_scheduler_pretrained_path={
"path": pretrained_path,
"subfolder": "scheduler",
},
tokenizer_pretrained_path={
"path": pretrained_path,
"subfolder": "tokenizer",
},
text_encoder_pretrained_path={
"path": pretrained_path,
"subfolder": "text_encoder",
},
**kwargs,
)
def _replace_unet_conv_in(self):
# Replace the first layer to accept 8 in_channels. Only applied when loading pretrained SD U-Net
_weight = self.unet.conv_in.weight.clone() # [320, 4, 3, 3]
_bias = self.unet.conv_in.bias.clone() # [320]
_weight = _weight.repeat((1, 2, 1, 1)) # Keep selected channel(s)
# half the activation magnitude
_weight *= 0.5
_bias *= 0.5
# new conv_in channel
_n_convin_out_channel = self.unet.conv_in.out_channels
_new_conv_in = Conv2d(
8, _n_convin_out_channel, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)
)
_new_conv_in.weight = Parameter(_weight)
_new_conv_in.bias = Parameter(_bias)
self.unet.conv_in = _new_conv_in
# replace config
self.unet.config["in_channels"] = 8
return
def to(self, device):
self.rgb_encoder.to(device)
self.depth_ae.to(device)
self.unet.to(device)
self.empty_text_embed = self.empty_text_embed.to(device)
self.device = device
return self
def forward(
self,
rgb_in,
num_inference_steps: int = 50,
num_output_inter_results: int = 0,
show_pbar=False,
init_depth_latent=None,
return_depth_latent=False,
):
device = rgb_in.device
precision = self.unet.dtype
# Set timesteps
self.noise_scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.noise_scheduler.timesteps # [T]
# Encode image
rgb_latent = self.encode_rgb(rgb_in)
# Initial depth map (noise)
if init_depth_latent is not None:
init_depth_latent = init_depth_latent.to(dtype=precision)
assert (
init_depth_latent.shape == rgb_latent.shape
), "initial depth latent should be the size of [B, 4, H/8, W/8]"
depth_latent = init_depth_latent
depth_latent = torch.randn(rgb_latent.shape, device=device, dtype=precision)
else:
depth_latent = torch.randn(rgb_latent.shape, device=device) # [B, 4, h, w]
# Expand text embeding for batch
batch_empty_text_embed = self.empty_text_embed.repeat(
(rgb_latent.shape[0], 1, 1)
).to(device=device, dtype=precision) # [B, 2, 1024]
# Export intermediate denoising steps
if num_output_inter_results > 0:
depth_latent_ls = []
inter_steps = []
_idx = (
-1
* (
np.arange(0, num_output_inter_results)
* num_inference_steps
/ num_output_inter_results
)
.round()
.astype(int)
- 1
)
steps_to_output = timesteps[_idx]
# Denoising loop
if show_pbar:
iterable = tqdm(enumerate(timesteps), total=len(timesteps), leave=False, desc="denoising")
else:
iterable = enumerate(timesteps)
for i, t in iterable:
unet_input = torch.cat(
[rgb_latent, depth_latent], dim=1
) # this order is important
unet_input = unet_input.to(dtype=precision)
# predict the noise residual
noise_pred = self.unet(
unet_input, t, encoder_hidden_states=batch_empty_text_embed
).sample # [B, 4, h, w]
# compute the previous noisy sample x_t -> x_t-1
depth_latent = self.noise_scheduler.step(
noise_pred, t, depth_latent
).prev_sample.to(dtype=precision)
if num_output_inter_results > 0 and t in steps_to_output:
depth_latent_ls.append(depth_latent.detach().clone())
#depth_latent_ls = depth_latent_ls.to(dtype=precision)
inter_steps.append(t - 1)
# Decode depth latent
if num_output_inter_results > 0:
assert 0 in inter_steps
depth = [self.decode_depth(lat) for lat in depth_latent_ls]
if return_depth_latent:
return depth, inter_steps, depth_latent_ls
else:
return depth, inter_steps
else:
depth = self.decode_depth(depth_latent)
if return_depth_latent:
return depth, depth_latent
else:
return depth
def encode_rgb(self, rgb_in):
rgb_latent = self.rgb_encoder(rgb_in) # [B, 4, h, w]
rgb_latent = rgb_latent * self.rgb_latent_scale_factor
return rgb_latent
def encode_depth(self, depth_in):
depth_latent = self.depth_ae.encode(depth_in)
depth_latent = depth_latent * self.depth_latent_scale_factor
return depth_latent
def decode_depth(self, depth_latent):
#depth_latent = depth_latent.to(dtype=torch.float16)
depth_latent = depth_latent / self.depth_latent_scale_factor
depth = self.depth_ae.decode(depth_latent) # [B, 1, H, W]
return depth
@staticmethod
def _encode_text(prompt, tokenizer, text_encoder):
text_inputs = tokenizer(
prompt,
padding="do_not_pad",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids.to(text_encoder.device)
text_embed = text_encoder(text_input_ids)[0]
return text_embed
|