Spaces:
Running
on
Zero
Running
on
Zero
File size: 10,477 Bytes
2d1e0bb 180819f 2d1e0bb 180819f 2d1e0bb 108d3f4 180819f 2d1e0bb 4366823 108d3f4 2d1e0bb 180819f 2d1e0bb 180819f 2d1e0bb 108d3f4 2d1e0bb 108d3f4 48c1bdf 108d3f4 48c1bdf 108d3f4 2d1e0bb 3a3ce5e 2d1e0bb 180819f 2d1e0bb 8b9ace1 180819f 212c199 8b9ace1 180819f 8b9ace1 212c199 2d1e0bb 180819f 212c199 2d1e0bb c686e22 2d1e0bb c686e22 2d1e0bb b56f62c 2d1e0bb 40c9ea6 2d1e0bb 8b9ace1 40c9ea6 212c199 8b9ace1 40c9ea6 2d1e0bb 8b9ace1 212c199 40c9ea6 212c199 2d1e0bb c686e22 340cc7b c686e22 2d1e0bb c686e22 2d1e0bb b56f62c 2d1e0bb |
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 |
import gc
import spaces
from safetensors.torch import load_file
from autoregressive.models.gpt_t2i import GPT_models
from tokenizer.tokenizer_image.vq_model import VQ_models
from language.t5 import T5Embedder
import torch
import numpy as np
import PIL
from PIL import Image
from condition.canny import CannyDetector
import time
from autoregressive.models.generate import generate
from condition.midas.depth import MidasDetector
from preprocessor import Preprocessor
models = {
"edge": "checkpoints/edge_base.safetensors",
"depth": "checkpoints/depth_base.safetensors",
}
class Model:
def __init__(self):
self.device = torch.device(
"cuda")
self.base_model_id = ""
self.task_name = ""
self.vq_model = self.load_vq()
self.t5_model = self.load_t5()
# self.gpt_model_edge = self.load_gpt(condition_type='edge')
# self.gpt_model_depth = self.load_gpt(condition_type='depth')
self.gpt_model = self.load_gpt()
self.preprocessor = Preprocessor()
def to(self, device):
self.gpt_model_canny.to('cuda')
def load_vq(self):
vq_model = VQ_models["VQ-16"](codebook_size=16384,
codebook_embed_dim=8)
vq_model.eval()
checkpoint = torch.load(f"checkpoints/vq_ds16_t2i.pt",
map_location="cpu")
vq_model.load_state_dict(checkpoint["model"])
del checkpoint
print("image tokenizer is loaded")
return vq_model
def load_gpt(self, condition_type='edge'):
# gpt_ckpt = models[condition_type]
# precision = torch.bfloat16
precision = torch.float32
latent_size = 512 // 16
gpt_model = GPT_models["GPT-XL"](
block_size=latent_size**2,
cls_token_num=120,
model_type='t2i',
condition_type=condition_type,
adapter_size='base',
).to(device='cpu', dtype=precision)
# model_weight = load_file(gpt_ckpt)
# gpt_model.load_state_dict(model_weight, strict=False)
# gpt_model.eval()
# print("gpt model is loaded")
return gpt_model
def load_gpt_weight(self, condition_type='edge'):
torch.cuda.empty_cache()
gc.collect()
gpt_ckpt = models[condition_type]
model_weight = load_file(gpt_ckpt)
self.gpt_model.load_state_dict(model_weight, strict=False)
self.gpt_model.eval()
torch.cuda.empty_cache()
gc.collect()
# print("gpt model is loaded")
def load_t5(self):
# precision = torch.bfloat16
precision = torch.float32
t5_model = T5Embedder(
device=self.device,
local_cache=True,
cache_dir='checkpoints/flan-t5-xl',
dir_or_name='flan-t5-xl',
torch_dtype=precision,
model_max_length=120,
)
return t5_model
@torch.no_grad()
@spaces.GPU(enable_queue=True)
def process_edge(
self,
image: np.ndarray,
prompt: str,
cfg_scale: float,
temperature: float,
top_k: int,
top_p: int,
seed: int,
low_threshold: int,
high_threshold: int,
control_strength: float,
preprocessor_name: str,
) -> list[PIL.Image.Image]:
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
origin_W, origin_H = image.size
if preprocessor_name == 'Canny':
self.preprocessor.load("Canny")
condition_img = self.preprocessor(
image=image, low_threshold=low_threshold, high_threshold=high_threshold, detect_resolution=512)
elif preprocessor_name == 'Hed':
self.preprocessor.load("HED")
condition_img = self.preprocessor(
image=image,image_resolution=512, detect_resolution=512)
elif preprocessor_name == 'Lineart':
self.preprocessor.load("Lineart")
condition_img = self.preprocessor(
image=image,image_resolution=512, detect_resolution=512)
elif preprocessor_name == 'No preprocess':
condition_img = image
print('get edge')
del self.preprocessor.model
torch.cuda.empty_cache()
condition_img = condition_img.resize((512,512))
W, H = condition_img.size
self.t5_model.model.to('cuda').to(torch.bfloat16)
self.load_gpt_weight('edge')
self.gpt_model.to('cuda').to(torch.bfloat16)
self.vq_model.to('cuda')
condition_img = torch.from_numpy(np.array(condition_img)).unsqueeze(0).permute(0,3,1,2).repeat(1,1,1,1)
condition_img = condition_img.to(self.device)
condition_img = 2*(condition_img/255 - 0.5)
prompts = [prompt] * 1
caption_embs, emb_masks = self.t5_model.get_text_embeddings(prompts)
print(f"processing left-padding...")
new_emb_masks = torch.flip(emb_masks, dims=[-1])
new_caption_embs = []
for idx, (caption_emb,
emb_mask) in enumerate(zip(caption_embs, emb_masks)):
valid_num = int(emb_mask.sum().item())
print(f' prompt {idx} token len: {valid_num}')
new_caption_emb = torch.cat(
[caption_emb[valid_num:], caption_emb[:valid_num]])
new_caption_embs.append(new_caption_emb)
new_caption_embs = torch.stack(new_caption_embs)
c_indices = new_caption_embs * new_emb_masks[:, :, None]
c_emb_masks = new_emb_masks
qzshape = [len(c_indices), 8, H // 16, W // 16]
t1 = time.time()
print(caption_embs.device)
index_sample = generate(
self.gpt_model,
c_indices,
(H // 16) * (W // 16),
c_emb_masks,
condition=condition_img,
cfg_scale=cfg_scale,
temperature=temperature,
top_k=top_k,
top_p=top_p,
sample_logits=True,
control_strength=control_strength,
)
sampling_time = time.time() - t1
print(f"Full sampling takes about {sampling_time:.2f} seconds.")
t2 = time.time()
print(index_sample.shape)
samples = self.vq_model.decode_code(
index_sample, qzshape) # output value is between [-1, 1]
decoder_time = time.time() - t2
print(f"decoder takes about {decoder_time:.2f} seconds.")
# samples = condition_img[0:1]
samples = torch.cat((condition_img[0:1], samples), dim=0)
samples = 255 * (samples * 0.5 + 0.5)
samples = [
Image.fromarray(
sample.permute(1, 2, 0).cpu().detach().numpy().clip(
0, 255).astype(np.uint8)) for sample in samples
]
del condition_img
torch.cuda.empty_cache()
return samples
@torch.no_grad()
@spaces.GPU(enable_queue=True)
def process_depth(
self,
image: np.ndarray,
prompt: str,
cfg_scale: float,
temperature: float,
top_k: int,
top_p: int,
seed: int,
control_strength: float,
preprocessor_name: str
) -> list[PIL.Image.Image]:
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
origin_W, origin_H = image.size
# print(image)
if preprocessor_name == 'depth':
self.preprocessor.load("Depth")
condition_img = self.preprocessor(
image=image,
image_resolution=512,
detect_resolution=512,
)
elif preprocessor_name == 'No preprocess':
condition_img = image
print('get depth')
del self.preprocessor.model
torch.cuda.empty_cache()
condition_img = condition_img.resize((512,512))
W, H = condition_img.size
self.t5_model.model.to(self.device).to(torch.bfloat16)
self.load_gpt_weight('depth')
self.gpt_model.to('cuda').to(torch.bfloat16)
self.vq_model.to(self.device)
condition_img = torch.from_numpy(np.array(condition_img)).unsqueeze(0).permute(0,3,1,2).repeat(1,1,1,1)
condition_img = condition_img.to(self.device)
condition_img = 2*(condition_img/255 - 0.5)
prompts = [prompt] * 1
caption_embs, emb_masks = self.t5_model.get_text_embeddings(prompts)
print(f"processing left-padding...")
new_emb_masks = torch.flip(emb_masks, dims=[-1])
new_caption_embs = []
for idx, (caption_emb,
emb_mask) in enumerate(zip(caption_embs, emb_masks)):
valid_num = int(emb_mask.sum().item())
print(f' prompt {idx} token len: {valid_num}')
new_caption_emb = torch.cat(
[caption_emb[valid_num:], caption_emb[:valid_num]])
new_caption_embs.append(new_caption_emb)
new_caption_embs = torch.stack(new_caption_embs)
c_indices = new_caption_embs * new_emb_masks[:, :, None]
c_emb_masks = new_emb_masks
qzshape = [len(c_indices), 8, H // 16, W // 16]
t1 = time.time()
index_sample = generate(
self.gpt_model,
c_indices,
(H // 16) * (W // 16),
c_emb_masks,
condition=condition_img,
cfg_scale=cfg_scale,
temperature=temperature,
top_k=top_k,
top_p=top_p,
sample_logits=True,
control_strength=control_strength,
)
sampling_time = time.time() - t1
print(f"Full sampling takes about {sampling_time:.2f} seconds.")
t2 = time.time()
print(index_sample.shape)
samples = self.vq_model.decode_code(index_sample, qzshape)
decoder_time = time.time() - t2
print(f"decoder takes about {decoder_time:.2f} seconds.")
condition_img = condition_img.cpu()
samples = samples.cpu()
# samples = condition_img[0:1]
samples = torch.cat((condition_img[0:1], samples), dim=0)
samples = 255 * (samples * 0.5 + 0.5)
samples = [
Image.fromarray(
sample.permute(1, 2, 0).cpu().detach().numpy().clip(0, 255).astype(np.uint8))
for sample in samples
]
del condition_img
torch.cuda.empty_cache()
return samples |