Spaces:
Runtime error
Runtime error
File size: 15,298 Bytes
c19ca42 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 |
import os
import sys
from collections import namedtuple
from pathlib import Path
import re
import torch
import torch.hub # pylint: disable=ungrouped-imports
from PIL import Image
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
from modules import devices, paths, shared, lowvram, errors
blip_image_eval_size = 384
clip_model_name = 'ViT-L/14'
Category = namedtuple("Category", ["name", "topn", "items"])
re_topn = re.compile(r"\.top(\d+)\.")
def category_types():
return [f.stem for f in Path(shared.interrogator.content_dir).glob('*.txt')]
def download_default_clip_interrogate_categories(content_dir):
shared.log.info("Downloading CLIP categories...")
tmpdir = f"{content_dir}_tmp"
cat_types = ["artists", "flavors", "mediums", "movements"]
try:
os.makedirs(tmpdir, exist_ok=True)
for category_type in cat_types:
torch.hub.download_url_to_file(f"https://raw.githubusercontent.com/pharmapsychotic/clip-interrogator/main/clip_interrogator/data/{category_type}.txt", os.path.join(tmpdir, f"{category_type}.txt"))
os.rename(tmpdir, content_dir)
except Exception as e:
errors.display(e, "downloading default CLIP interrogate categories")
finally:
if os.path.exists(tmpdir):
os.removedirs(tmpdir)
class InterrogateModels:
blip_model = None
clip_model = None
clip_preprocess = None
dtype = None
running_on_cpu = None
def __init__(self, content_dir):
self.loaded_categories = None
self.skip_categories = []
self.content_dir = content_dir
self.running_on_cpu = devices.device_interrogate == torch.device("cpu")
def categories(self):
if not os.path.exists(self.content_dir):
download_default_clip_interrogate_categories(self.content_dir)
if self.loaded_categories is not None and self.skip_categories == shared.opts.interrogate_clip_skip_categories:
return self.loaded_categories
self.loaded_categories = []
if os.path.exists(self.content_dir):
self.skip_categories = shared.opts.interrogate_clip_skip_categories
cat_types = []
for filename in Path(self.content_dir).glob('*.txt'):
cat_types.append(filename.stem)
if filename.stem in self.skip_categories:
continue
m = re_topn.search(filename.stem)
topn = 1 if m is None else int(m.group(1))
with open(filename, "r", encoding="utf8") as file:
lines = [x.strip() for x in file.readlines()]
self.loaded_categories.append(Category(name=filename.stem, topn=topn, items=lines))
return self.loaded_categories
def create_fake_fairscale(self):
class FakeFairscale:
def checkpoint_wrapper(self):
pass
sys.modules["fairscale.nn.checkpoint.checkpoint_activations"] = FakeFairscale
def load_blip_model(self):
self.create_fake_fairscale()
from repositories.blip import models # pylint: disable=unused-import
from repositories.blip.models import blip
import modules.modelloader as modelloader
model_path = os.path.join(paths.models_path, "BLIP")
download_name='model_base_caption_capfilt_large.pth',
shared.log.debug(f'Model interrogate load: type=BLiP model={download_name} path={model_path}')
files = modelloader.load_models(
model_path=model_path,
model_url='https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_base_caption_capfilt_large.pth',
ext_filter=[".pth"],
download_name=download_name,
)
blip_model = blip.blip_decoder(pretrained=files[0], image_size=blip_image_eval_size, vit='base', med_config=os.path.join(paths.paths["BLIP"], "configs", "med_config.json")) # pylint: disable=c-extension-no-member
blip_model.eval()
return blip_model
def load_clip_model(self):
shared.log.debug(f'Model interrogate load: type=CLiP model={clip_model_name} path={shared.opts.clip_models_path}')
import clip
if self.running_on_cpu:
model, preprocess = clip.load(clip_model_name, device="cpu", download_root=shared.opts.clip_models_path)
else:
model, preprocess = clip.load(clip_model_name, download_root=shared.opts.clip_models_path)
model.eval()
model = model.to(devices.device_interrogate)
return model, preprocess
def load(self):
if self.blip_model is None:
self.blip_model = self.load_blip_model()
if not shared.opts.no_half and not self.running_on_cpu:
self.blip_model = self.blip_model.half()
self.blip_model = self.blip_model.to(devices.device_interrogate)
if self.clip_model is None:
self.clip_model, self.clip_preprocess = self.load_clip_model()
if not shared.opts.no_half and not self.running_on_cpu:
self.clip_model = self.clip_model.half()
self.clip_model = self.clip_model.to(devices.device_interrogate)
self.dtype = next(self.clip_model.parameters()).dtype
def send_clip_to_ram(self):
if not shared.opts.interrogate_keep_models_in_memory:
if self.clip_model is not None:
self.clip_model = self.clip_model.to(devices.cpu)
def send_blip_to_ram(self):
if not shared.opts.interrogate_keep_models_in_memory:
if self.blip_model is not None:
self.blip_model = self.blip_model.to(devices.cpu)
def unload(self):
self.send_clip_to_ram()
self.send_blip_to_ram()
devices.torch_gc()
def rank(self, image_features, text_array, top_count=1):
import clip
devices.torch_gc()
if shared.opts.interrogate_clip_dict_limit != 0:
text_array = text_array[0:int(shared.opts.interrogate_clip_dict_limit)]
top_count = min(top_count, len(text_array))
text_tokens = clip.tokenize(list(text_array), truncate=True).to(devices.device_interrogate)
text_features = self.clip_model.encode_text(text_tokens).type(self.dtype)
text_features /= text_features.norm(dim=-1, keepdim=True)
similarity = torch.zeros((1, len(text_array))).to(devices.device_interrogate)
for i in range(image_features.shape[0]):
similarity += (100.0 * image_features[i].unsqueeze(0) @ text_features.T).softmax(dim=-1)
similarity /= image_features.shape[0]
top_probs, top_labels = similarity.cpu().topk(top_count, dim=-1)
return [(text_array[top_labels[0][i].numpy()], (top_probs[0][i].numpy()*100)) for i in range(top_count)]
def generate_caption(self, pil_image):
gpu_image = transforms.Compose([
transforms.Resize((blip_image_eval_size, blip_image_eval_size), interpolation=InterpolationMode.BICUBIC),
transforms.ToTensor(),
transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
])(pil_image).unsqueeze(0).type(self.dtype).to(devices.device_interrogate)
with devices.inference_context():
caption = self.blip_model.generate(gpu_image, sample=False, num_beams=shared.opts.interrogate_clip_num_beams, min_length=shared.opts.interrogate_clip_min_length, max_length=shared.opts.interrogate_clip_max_length)
return caption[0]
def interrogate(self, pil_image):
res = ""
shared.state.begin('interrogate')
try:
if shared.backend == shared.Backend.ORIGINAL and (shared.cmd_opts.lowvram or shared.cmd_opts.medvram):
lowvram.send_everything_to_cpu()
devices.torch_gc()
self.load()
if isinstance(pil_image, list):
pil_image = pil_image[0]
if isinstance(pil_image, dict) and 'name' in pil_image:
pil_image = Image.open(pil_image['name'])
pil_image = pil_image.convert("RGB")
caption = self.generate_caption(pil_image)
self.send_blip_to_ram()
devices.torch_gc()
res = caption
clip_image = self.clip_preprocess(pil_image).unsqueeze(0).type(self.dtype).to(devices.device_interrogate)
with devices.inference_context(), devices.autocast():
image_features = self.clip_model.encode_image(clip_image).type(self.dtype)
image_features /= image_features.norm(dim=-1, keepdim=True)
for _name, topn, items in self.categories():
matches = self.rank(image_features, items, top_count=topn)
for match, score in matches:
if shared.opts.interrogate_return_ranks:
res += f", ({match}:{score/100:.3f})"
else:
res += f", {match}"
except Exception as e:
errors.display(e, 'interrogate')
res += "<error>"
self.unload()
shared.state.end()
return res
# --------- interrrogate ui
ci = None
low_vram = False
class BatchWriter:
def __init__(self, folder):
self.folder = folder
self.csv, self.file = None, None
def add(self, file, prompt):
txt_file = os.path.splitext(file)[0] + ".txt"
with open(os.path.join(self.folder, txt_file), 'w', encoding='utf-8') as f:
f.write(prompt)
def close(self):
if self.file is not None:
self.file.close()
def get_clip_models():
import open_clip
return ['/'.join(x) for x in open_clip.list_pretrained()]
def load_interrogator(model):
from clip_interrogator import Config, Interrogator
global ci # pylint: disable=global-statement
if ci is None:
config = Config(device=devices.get_optimal_device(), cache_path=os.path.join(paths.models_path, 'Interrogator'), clip_model_name=model, quiet=True)
if low_vram:
config.apply_low_vram_defaults()
shared.log.info(f'Interrogate load: config={config}')
ci = Interrogator(config)
elif model != ci.config.clip_model_name:
ci.config.clip_model_name = model
shared.log.info(f'Interrogate load: config={ci.config}')
ci.load_clip_model()
def unload_clip_model():
if ci is not None:
shared.log.debug('Interrogate offload')
ci.caption_model = ci.caption_model.to(devices.cpu)
ci.clip_model = ci.clip_model.to(devices.cpu)
ci.caption_offloaded = True
ci.clip_offloaded = True
devices.torch_gc()
def interrogate(image, mode, caption=None):
shared.log.info(f'Interrogate: image={image} mode={mode} config={ci.config}')
if mode == 'best':
prompt = ci.interrogate(image, caption=caption)
elif mode == 'caption':
prompt = ci.generate_caption(image) if caption is None else caption
elif mode == 'classic':
prompt = ci.interrogate_classic(image, caption=caption)
elif mode == 'fast':
prompt = ci.interrogate_fast(image, caption=caption)
elif mode == 'negative':
prompt = ci.interrogate_negative(image)
else:
raise RuntimeError(f"Unknown mode {mode}")
return prompt
def interrogate_image(image, model, mode):
shared.state.begin()
shared.state.job = 'interrogate'
try:
if shared.backend == shared.Backend.ORIGINAL and (shared.cmd_opts.lowvram or shared.cmd_opts.medvram):
lowvram.send_everything_to_cpu()
devices.torch_gc()
load_interrogator(model)
image = image.convert('RGB')
shared.log.info(f'Interrogate: image={image} mode={mode} config={ci.config}')
prompt = interrogate(image, mode)
except Exception as e:
prompt = f"Exception {type(e)}"
shared.log.error(f'Interrogate: {e}')
shared.state.end()
return prompt
def interrogate_batch(batch_files, batch_folder, batch_str, model, mode, write):
files = []
if batch_files is not None:
files += [f.name for f in batch_files]
if batch_folder is not None:
files += [f.name for f in batch_folder]
if batch_str is not None and len(batch_str) > 0 and os.path.exists(batch_str) and os.path.isdir(batch_str):
files += [os.path.join(batch_str, f) for f in os.listdir(batch_str) if f.lower().endswith(('.png', '.jpg', '.jpeg', '.webp'))]
if len(files) == 0:
shared.log.error('Interrogate batch no images')
return ''
shared.state.begin()
shared.state.job = 'batch interrogate'
prompts = []
try:
if shared.backend == shared.Backend.ORIGINAL and (shared.cmd_opts.lowvram or shared.cmd_opts.medvram):
lowvram.send_everything_to_cpu()
devices.torch_gc()
load_interrogator(model)
shared.log.info(f'Interrogate batch: images={len(files)} mode={mode} config={ci.config}')
captions = []
# first pass: generate captions
for file in files:
caption = ""
try:
if shared.state.interrupted:
break
image = Image.open(file).convert('RGB')
caption = ci.generate_caption(image)
except Exception as e:
shared.log.error(f'Interrogate caption: {e}')
finally:
captions.append(caption)
# second pass: interrogate
if write:
writer = BatchWriter(os.path.dirname(files[0]))
for idx, file in enumerate(files):
try:
if shared.state.interrupted:
break
image = Image.open(file).convert('RGB')
prompt = interrogate(image, mode, caption=captions[idx])
prompts.append(prompt)
if write:
writer.add(file, prompt)
except OSError as e:
shared.log.error(f'Interrogate batch: {e}')
if write:
writer.close()
ci.config.quiet = False
unload_clip_model()
except Exception as e:
shared.log.error(f'Interrogate batch: {e}')
shared.state.end()
return '\n\n'.join(prompts)
def analyze_image(image, model):
load_interrogator(model)
image = image.convert('RGB')
image_features = ci.image_to_features(image)
top_mediums = ci.mediums.rank(image_features, 5)
top_artists = ci.artists.rank(image_features, 5)
top_movements = ci.movements.rank(image_features, 5)
top_trendings = ci.trendings.rank(image_features, 5)
top_flavors = ci.flavors.rank(image_features, 5)
medium_ranks = dict(zip(top_mediums, ci.similarities(image_features, top_mediums)))
artist_ranks = dict(zip(top_artists, ci.similarities(image_features, top_artists)))
movement_ranks = dict(zip(top_movements, ci.similarities(image_features, top_movements)))
trending_ranks = dict(zip(top_trendings, ci.similarities(image_features, top_trendings)))
flavor_ranks = dict(zip(top_flavors, ci.similarities(image_features, top_flavors)))
return medium_ranks, artist_ranks, movement_ranks, trending_ranks, flavor_ranks
|