Spaces:
Runtime error
Runtime error
File size: 10,571 Bytes
c81908d 06c5f0c c81908d 27911d6 c81908d 27911d6 c81908d 27911d6 c81908d 27911d6 c81908d 27911d6 c81908d 27911d6 c81908d 27911d6 c81908d 06c5f0c ae26d48 06c5f0c 27911d6 06c5f0c c81908d 06c5f0c ae26d48 06c5f0c be61cf2 06c5f0c ae26d48 06c5f0c be61cf2 7ac3881 a45817e 7060b15 a45817e 7ac3881 7060b15 e96a195 3dcdf92 27911d6 3dcdf92 e96a195 3dcdf92 06c5f0c ae26d48 06c5f0c ae26d48 27911d6 06c5f0c ae26d48 be61cf2 e96a195 be61cf2 7ac3881 a45817e 7060b15 e96a195 3dcdf92 c81908d 06c5f0c 27911d6 c81908d 27911d6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 |
import os
from glob import glob
from subprocess import call
import json
def base():
return {
"slurm":{
"t": 360,
"N": 2,
"n": 8,
},
"model":{
"dataset": "wds",
"seed": 0,
"cross_attention": False,
"num_channels": 3,
"centered": True,
"use_geometric": False,
"beta_min": 0.1,
"beta_max": 20.0,
"num_channels_dae": 128,
"n_mlp": 3,
"ch_mult": [1, 1, 2, 2, 4, 4],
"num_res_blocks": 2,
"attn_resolutions": (16,),
"dropout": 0.0,
"resamp_with_conv": True,
"conditional": True,
"fir": True,
"fir_kernel": [1, 3, 3, 1],
"skip_rescale": True,
"resblock_type": "biggan",
"progressive": "none",
"progressive_input": "residual",
"progressive_combine": "sum",
"embedding_type": "positional",
"fourier_scale": 16.0,
"not_use_tanh": False,
"image_size": 256,
"nz": 100,
"num_timesteps": 4,
"z_emb_dim": 256,
"t_emb_dim": 256,
"text_encoder": "google/t5-v1_1-base",
"masked_mean": True,
"cross_attention_block": "basic",
}
}
def ddgan_cc12m_v2():
cfg = base()
cfg['slurm']['N'] = 2
cfg['slurm']['n'] = 8
return cfg
def ddgan_cc12m_v6():
cfg = base()
cfg['model']['text_encoder'] = "google/t5-v1_1-large"
return cfg
def ddgan_cc12m_v7():
cfg = base()
cfg['model']['classifier_free_guidance_proba'] = 0.2
cfg['slurm']['N'] = 2
cfg['slurm']['n'] = 8
return cfg
def ddgan_cc12m_v8():
cfg = base()
cfg['model']['text_encoder'] = "google/t5-v1_1-large"
cfg['model']['classifier_free_guidance_proba'] = 0.2
return cfg
def ddgan_cc12m_v9():
cfg = base()
cfg['model']['text_encoder'] = "google/t5-v1_1-large"
cfg['model']['classifier_free_guidance_proba'] = 0.2
cfg['model']['num_channels_dae'] = 320
cfg['model']['image_size'] = 64
cfg['model']['batch_size'] = 1
return cfg
def ddgan_cc12m_v11():
cfg = base()
cfg['model']['text_encoder'] = "google/t5-v1_1-large"
cfg['model']['classifier_free_guidance_proba'] = 0.2
cfg['model']['cross_attention'] = True
return cfg
def ddgan_cc12m_v12():
cfg = ddgan_cc12m_v11()
cfg['model']['text_encoder'] = "google/t5-v1_1-xl"
cfg['model']['preprocessing'] = 'random_resized_crop_v1'
return cfg
def ddgan_cc12m_v13():
cfg = ddgan_cc12m_v12()
cfg['model']['discr_type'] = "large_cond_attn"
return cfg
def ddgan_cc12m_v14():
cfg = ddgan_cc12m_v12()
cfg['model']['num_channels_dae'] = 192
return cfg
def ddgan_cc12m_v15():
cfg = ddgan_cc12m_v11()
cfg['model']['mismatch_loss'] = ''
cfg['model']['grad_penalty_cond'] = ''
return cfg
def ddgan_cifar10_cond17():
cfg = base()
cfg['model']['image_size'] = 32
cfg['model']['classifier_free_guidance_proba'] = 0.2
cfg['model']['ch_mult'] = "1 2 2 2"
cfg['model']['cross_attention'] = True
cfg['model']['dataset'] = "cifar10"
cfg['model']['n_mlp'] = 4
return cfg
def ddgan_cifar10_cond18():
cfg = ddgan_cifar10_cond17()
cfg['model']['text_encoder'] = "google/t5-v1_1-xl"
return cfg
def ddgan_cifar10_cond19():
cfg = ddgan_cifar10_cond17()
cfg['model']['discr_type'] = 'small_cond_attn'
cfg['model']['mismatch_loss'] = ''
cfg['model']['grad_penalty_cond'] = ''
return cfg
def ddgan_laion_aesthetic_v1():
cfg = ddgan_cc12m_v11()
cfg['model']['dataset_root'] = '"/p/scratch/ccstdl/cherti1/LAION-aesthetic/output/{00000..05038}.tar"'
return cfg
def ddgan_laion_aesthetic_v2():
cfg = ddgan_laion_aesthetic_v1()
cfg['model']['discr_type'] = "large_cond_attn"
return cfg
def ddgan_laion_aesthetic_v3():
cfg = ddgan_laion_aesthetic_v1()
cfg['model']['text_encoder'] = "google/t5-v1_1-xl"
cfg['model']['mismatch_loss'] = ''
cfg['model']['grad_penalty_cond'] = ''
return cfg
def ddgan_laion_aesthetic_v4():
cfg = ddgan_laion_aesthetic_v1()
cfg['model']['text_encoder'] = "openclip/ViT-L-14-336/openai"
return cfg
def ddgan_laion_aesthetic_v5():
cfg = ddgan_laion_aesthetic_v1()
cfg['model']['mismatch_loss'] = ''
cfg['model']['grad_penalty_cond'] = ''
return cfg
def ddgan_laion2b_v1():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['mismatch_loss'] = ''
cfg['model']['grad_penalty_cond'] = ''
cfg['model']['num_channels_dae'] = 224
cfg['model']['batch_size'] = 2
cfg['model']['discr_type'] = "large_cond_attn"
cfg['model']['preprocessing'] = 'random_resized_crop_v1'
return cfg
def ddgan_laion_aesthetic_v6():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['no_lr_decay'] = ''
return cfg
def ddgan_laion_aesthetic_v7():
cfg = ddgan_laion_aesthetic_v6()
cfg['model']['r1_gamma'] = 5
return cfg
def ddgan_laion_aesthetic_v8():
cfg = ddgan_laion_aesthetic_v6()
cfg['model']['num_timesteps'] = 8
return cfg
def ddgan_laion_aesthetic_v9():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['num_channels_dae'] = 384
return cfg
def ddgan_sd_v1():
cfg = ddgan_laion_aesthetic_v3()
return cfg
def ddgan_sd_v2():
cfg = ddgan_laion_aesthetic_v3()
return cfg
def ddgan_sd_v3():
cfg = ddgan_laion_aesthetic_v3()
return cfg
def ddgan_sd_v4():
cfg = ddgan_laion_aesthetic_v3()
return cfg
def ddgan_sd_v5():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['num_timesteps'] = 8
return cfg
def ddgan_sd_v6():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['num_channels_dae'] = 192
return cfg
def ddgan_sd_v7():
cfg = ddgan_laion_aesthetic_v3()
return cfg
def ddgan_sd_v8():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['image_size'] = 512
return cfg
def ddgan_laion_aesthetic_v12():
cfg = ddgan_laion_aesthetic_v3()
return cfg
def ddgan_laion_aesthetic_v13():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['text_encoder'] = "openclip/ViT-H-14/laion2b_s32b_b79k"
return cfg
def ddgan_laion_aesthetic_v14():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['text_encoder'] = "openclip/ViT-H-14/laion2b_s32b_b79k"
return cfg
def ddgan_sd_v9():
cfg = ddgan_laion_aesthetic_v3()
cfg['model']['text_encoder'] = "openclip/ViT-H-14/laion2b_s32b_b79k"
return cfg
def ddgan_sd_v10():
cfg = ddgan_sd_v9()
cfg['model']['num_timesteps'] = 2
return cfg
def ddgan_laion2b_v2():
cfg = ddgan_sd_v9()
return cfg
def ddgan_ddb_v1():
cfg = ddgan_sd_v10()
return cfg
def ddgan_sd_v11():
cfg = ddgan_sd_v10()
cfg['model']['image_size'] = 512
return cfg
def ddgan_ddb_v2():
cfg = ddgan_ddb_v1()
cfg['model']['num_timesteps'] = 1
return cfg
def ddgan_ddb_v3():
cfg = ddgan_ddb_v1()
cfg['model']['num_channels_dae'] = 192
cfg['model']['num_timesteps'] = 2
return cfg
def ddgan_ddb_v4():
cfg = ddgan_ddb_v1()
cfg['model']['num_channels_dae'] = 256
cfg['model']['num_timesteps'] = 2
return cfg
def ddgan_ddb_v5():
cfg = ddgan_ddb_v2()
return cfg
def ddgan_ddb_v6():
cfg = ddgan_ddb_v3()
return cfg
def ddgan_ddb_v7():
cfg = ddgan_ddb_v1()
return cfg
def ddgan_ddb_v9():
cfg = ddgan_ddb_v3()
cfg['model']['attn_resolutions'] = [4, 8, 16, 32]
return cfg
def ddgan_laion_aesthetic_v15():
cfg = ddgan_ddb_v3()
return cfg
def ddgan_ddb_v10():
cfg = ddgan_ddb_v9()
return cfg
def ddgan_ddb_v11():
cfg = ddgan_ddb_v3()
cfg['model']['text_encoder'] = "openclip/ViT-g-14/laion2B-s12B-b42K"
return cfg
def ddgan_ddb_v12():
cfg = ddgan_ddb_v3()
cfg['model']['text_encoder'] = "openclip/ViT-bigG-14/laion2b_s39b_b160k"
return cfg
models = [
ddgan_cifar10_cond17, # cifar10, cross attn for discr
ddgan_cifar10_cond18, # cifar10, xl encoder
ddgan_cifar10_cond19, # cifar10, xl encoder
ddgan_cc12m_v2, # baseline (no large text encoder, no classifier guidance)
ddgan_cc12m_v6, # like v2 but using large T5 text encoder
ddgan_cc12m_v7, # like v2 but with classifier guidance
ddgan_cc12m_v8, # like v6 but classifier guidance
ddgan_cc12m_v9, # ~1B model but 64x64 resolution
ddgan_cc12m_v11, # large text encoder + cross attention + classifier free guidance
ddgan_cc12m_v12, # T5-XL + cross attention + classifier free guidance + random_resized_crop_v1
ddgan_cc12m_v13, # T5-XL + cross attention + classifier free guidance + random_resized_crop_v1 + cond attn
ddgan_cc12m_v14, # T5-XL + cross attention + classifier free guidance + random_resized_crop_v1 + 300M model
ddgan_cc12m_v15, # fine-tune v11 with --mismatch_loss and --grad_penalty_cond
ddgan_laion_aesthetic_v1, # like ddgan_cc12m_v11 but fine-tuned on laion aesthetic
ddgan_laion_aesthetic_v2, # like ddgan_laion_aesthetic_v1 but trained from scratch with the new cross attn discr
ddgan_laion_aesthetic_v3, # like ddgan_laion_aesthetic_v1 but trained from scratch with T5-XL (continue from 23aug with mismatch and grad penalty and random_resized_crop_v1)
ddgan_laion_aesthetic_v4, # like ddgan_laion_aesthetic_v1 but trained from scratch with OpenAI's ClipEncoder
ddgan_laion_aesthetic_v5, # fine-tune ddgan_laion_aesthetic_v1 with mismatch and cond grad penalty losses
ddgan_laion_aesthetic_v6, # like v3 but without lr decay
ddgan_laion_aesthetic_v7, # like v6 but with r1 gamma of 5 instead of 1, trying to constrain the discr more.
ddgan_laion_aesthetic_v8, # like v6 but with 8 timesteps
ddgan_laion_aesthetic_v9,
ddgan_laion_aesthetic_v12,
ddgan_laion_aesthetic_v13,
ddgan_laion_aesthetic_v14,
ddgan_laion_aesthetic_v15,
ddgan_laion2b_v1,
ddgan_sd_v1,
ddgan_sd_v2,
ddgan_sd_v3,
ddgan_sd_v4,
ddgan_sd_v5,
ddgan_sd_v6,
ddgan_sd_v7,
ddgan_sd_v8,
ddgan_sd_v9,
ddgan_sd_v10,
ddgan_sd_v11,
ddgan_laion2b_v2,
ddgan_ddb_v1,
ddgan_ddb_v2,
ddgan_ddb_v3,
ddgan_ddb_v4,
ddgan_ddb_v5,
ddgan_ddb_v6,
ddgan_ddb_v7,
ddgan_ddb_v9,
ddgan_ddb_v10,
ddgan_ddb_v11,
ddgan_ddb_v12,
]
def get_model_config(model_name):
for model in models:
if model.__name__ == model_name:
return model()['model'] |