File size: 19,819 Bytes
59da1c6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 |
import matplotlib.pyplot as plt
import pydantic
import time
import numpy as np
from tqdm import tqdm, trange
import torch
from torch import nn
from diffusers import StableDiffusionPipeline
import clip
from dreamsim import dreamsim
from ribs.archives import GridArchive
from ribs.schedulers import Scheduler
from ribs.emitters import GaussianEmitter
import itertools
from ribs.visualize import grid_archive_heatmap
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.cuda.empty_cache()
print("Torch device:", DEVICE)
# Use float16 for GPU, float32 for CPU.
TORCH_DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32
print("Torch dtype:", TORCH_DTYPE)
IMG_WIDTH = 256
IMG_HEIGHT = 256
SD_IN_HEIGHT = 32
SD_IN_WIDTH = 32
SD_CHECKPOINT = "lambdalabs/miniSD-diffusers"
BATCH_SIZE = 4
SD_IN_CHANNELS = 4
SD_IN_SHAPE = (
BATCH_SIZE,
SD_IN_CHANNELS,
SD_IN_HEIGHT,
SD_IN_WIDTH,
)
SDPIPE = StableDiffusionPipeline.from_pretrained(
SD_CHECKPOINT,
torch_dtype=TORCH_DTYPE,
safety_checker=None, # For faster inference.
requires_safety_checker=False,
)
SDPIPE.set_progress_bar_config(disable=True)
SDPIPE = SDPIPE.to(DEVICE)
GRID_SIZE = (20, 20)
SEED = 123
np.random.seed(SEED)
torch.manual_seed(SEED)
# INIT_POP = 200 # Initial population.
# TOTAL_ITRS = 200 # Total number of iterations.
class DivProj(nn.Module):
def __init__(self, input_dim, latent_dim=2):
super().__init__()
self.proj = nn.Sequential(
nn.Linear(in_features=input_dim, out_features=latent_dim),
)
def forward(self, x):
"""Get diversity representations."""
x = self.proj(x)
return x
def calc_dis(self, x1, x2):
"""Calculate diversity distance as (squared) L2 distance."""
x1 = self.forward(x1)
x2 = self.forward(x2)
return torch.sum(torch.square(x1 - x2), -1)
def triplet_delta_dis(self, ref, x1, x2):
"""Calculate delta distance comparing x1 and x2 to ref."""
x1 = self.forward(x1)
x2 = self.forward(x2)
ref = self.forward(ref)
return (torch.sum(torch.square(ref - x1), -1) -
torch.sum(torch.square(ref - x2), -1))
# Triplet loss with margin 0.05.
# The binary preference labels are scaled to y = 1 or -1 for the loss, where y = 1 means x2 is more similar to ref than x1.
loss_fn = lambda y, delta_dis: torch.max(
torch.tensor([0.0]).to(DEVICE), 0.05 - (y * 2 - 1) * delta_dis
).mean()
def fit_div_proj(inputs, dreamsim_features, latent_dim, batch_size=32):
"""Trains the DivProj model on ground-truth labels."""
t = time.time()
model = DivProj(input_dim=inputs.shape[-1], latent_dim=latent_dim)
model.to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
n_pref_data = inputs.shape[0]
ref = inputs[:, 0]
x1 = inputs[:, 1]
x2 = inputs[:, 2]
n_train = int(n_pref_data * 0.75)
n_val = n_pref_data - n_train
# Split data into train and val.
ref_train = ref[:n_train]
x1_train = x1[:n_train]
x2_train = x2[:n_train]
ref_val = ref[n_train:]
x1_val = x1[n_train:]
x2_val = x2[n_train:]
# Split DreamSim features into train and val.
ref_dreamsim_features = dreamsim_features[:, 0]
x1_dreamsim_features = dreamsim_features[:, 1]
x2_dreamsim_features = dreamsim_features[:, 2]
ref_gt_train = ref_dreamsim_features[:n_train]
x1_gt_train = x1_dreamsim_features[:n_train]
x2_gt_train = x2_dreamsim_features[:n_train]
ref_gt_val = ref_dreamsim_features[n_train:]
x1_gt_val = x1_dreamsim_features[n_train:]
x2_gt_val = x2_dreamsim_features[n_train:]
val_acc = []
n_iters_per_epoch = max((n_train) // batch_size, 1)
for epoch in range(200):
for _ in range(n_iters_per_epoch):
optimizer.zero_grad()
idx = np.random.choice(n_train, batch_size)
batch_ref = ref_train[idx].float()
batch1 = x1_train[idx].float()
batch2 = x2_train[idx].float()
# Get delta distance from model.
delta_dis = model.triplet_delta_dis(batch_ref, batch1, batch2)
# Get preference labels from DreamSim features.
gt_dis = torch.nn.functional.cosine_similarity(
ref_gt_train[idx], x2_gt_train[idx], dim=-1
) - torch.nn.functional.cosine_similarity(
ref_gt_train[idx], x1_gt_train[idx], dim=-1
)
gt = (gt_dis > 0).to(TORCH_DTYPE) # if distance from the two sims are greater than 0, convert gt to torch_type
loss = loss_fn(gt, delta_dis)
loss.backward()
optimizer.step()
# Validate.
n_correct = 0
n_total = 0
with torch.no_grad():
idx = np.arange(n_val)
batch_ref = ref_val[idx].float()
batch1 = x1_val[idx].float()
batch2 = x2_val[idx].float()
delta_dis = model.triplet_delta_dis(batch_ref, batch1, batch2)
pred = delta_dis > 0
gt_dis = torch.nn.functional.cosine_similarity(
ref_gt_val[idx], x2_gt_val[idx], dim=-1
) - torch.nn.functional.cosine_similarity(
ref_gt_val[idx], x1_gt_val[idx], dim=-1
)
gt = gt_dis > 0
n_correct += (pred == gt).sum().item()
n_total += len(idx)
acc = n_correct / n_total
val_acc.append(acc)
# Early stopping if val_acc does not improve for 10 epochs.
if epoch > 10 and np.mean(val_acc[-10:]) < np.mean(val_acc[-11:-1]):
break
print(
f"{np.round(time.time()- t, 1)}s ({epoch+1} epochs) | DivProj (n={n_pref_data}) fitted with val acc.: {acc}"
)
return model.to(TORCH_DTYPE), acc
def compute_diversity_measures(clip_features, diversity_model):
with torch.no_grad():
measures = diversity_model(clip_features).detach().cpu().numpy()
return measures
def tensor_to_list(tensor):
sols = tensor.detach().cpu().numpy().astype(np.float32)
return sols.reshape(sols.shape[0], -1)
def list_to_tensor(list_):
sols = np.array(list_).reshape(
len(list_), 4, SD_IN_HEIGHT, SD_IN_WIDTH
) # Hard-coded for now.
return torch.tensor(sols, dtype=TORCH_DTYPE, device=DEVICE)
def create_scheduler(
sols,
objs,
clip_features,
diversity_model,
seed=None,
):
measures = compute_diversity_measures(clip_features, diversity_model)
archive_bounds = np.array(
[np.quantile(measures, 0.01, axis=0), np.quantile(measures, 0.99, axis=0)]
).T
sols = tensor_to_list(sols)
# Set up archive.
archive = GridArchive(
solution_dim=len(sols[0]), dims=GRID_SIZE, ranges=archive_bounds, seed=SEED
)
# Add initial solutions to the archive.
archive.add(sols, objs, measures)
# Set up the GaussianEmitter.
emitters = [
GaussianEmitter(
archive=archive,
sigma=0.1,
initial_solutions=archive.sample_elites(BATCH_SIZE)["solution"],
batch_size=BATCH_SIZE,
seed=SEED,
)
]
# Return the archive and scheduler.
return archive, Scheduler(archive, emitters)
def plot_archive(archive):
plt.figure(figsize=(6, 4.5))
grid_archive_heatmap(archive, vmin=0, vmax=100)
plt.xlabel("Diversity Metric 1")
plt.ylabel("Diversity Metric 2")
return plt
def run_qdhf(prompt:str, init_pop: int=200, total_itrs: int=200):
INIT_POP = init_pop
TOTAL_ITRS = total_itrs
# This tutorial uses ViT-B/32, you may use other checkpoints depending on your resources and need.
CLIP_MODEL, CLIP_PREPROCESS = clip.load("ViT-B/32", device=DEVICE)
CLIP_MODEL.eval()
for p in CLIP_MODEL.parameters():
p.requires_grad_(False)
def compute_clip_scores(imgs, text, return_clip_features=False):
"""Computes CLIP scores for a batch of images and a given text prompt."""
img_tensor = torch.stack([CLIP_PREPROCESS(img) for img in imgs]).to(DEVICE)
tokenized_text = clip.tokenize([text]).to(DEVICE)
img_logits, _text_logits = CLIP_MODEL(img_tensor, tokenized_text)
img_logits = img_logits.detach().cpu().numpy().astype(np.float32)[:, 0]
img_logits = 1 / img_logits * 100
# Remap the objective from minimizing [0, 10] to maximizing [0, 100]
img_logits = (10.0 - img_logits) * 10.0
if return_clip_features:
clip_features = CLIP_MODEL.encode_image(img_tensor).to(TORCH_DTYPE)
return img_logits, clip_features
else:
return img_logits
DREAMSIM_MODEL, DREAMSIM_PREPROCESS = dreamsim(
pretrained=True, dreamsim_type="open_clip_vitb32", device=DEVICE
)
def evaluate_lsi(
latents,
prompt,
return_features=False,
diversity_model=None,
):
"""Evaluates the objective of LSI for a batch of latents and a given text prompt."""
images = SDPIPE(
prompt,
num_images_per_prompt=latents.shape[0],
latents=latents,
# num_inference_steps=1, # For testing.
).images
objs, clip_features = compute_clip_scores(
images,
prompt,
return_clip_features=True,
)
images = torch.cat([DREAMSIM_PREPROCESS(img) for img in images]).to(DEVICE)
dreamsim_features = DREAMSIM_MODEL.embed(images)
if diversity_model is not None:
measures = compute_diversity_measures(clip_features, diversity_model)
else:
measures = None
if return_features:
return objs, measures, clip_features, dreamsim_features
else:
return objs, measures
update_schedule = [1, 21, 51, 101] # Iterations on which to update the archive.
n_pref_data = 1000 # Number of preferences used in each update.
archive = None
best = 0.0
for itr in trange(1, TOTAL_ITRS + 1):
# Update archive and scheduler if needed.
if itr in update_schedule:
if archive is None:
tqdm.write("Initializing archive and diversity projection.")
all_sols = []
all_clip_features = []
all_dreamsim_features = []
all_objs = []
# Sample random solutions and get judgment on similarity.
n_batches = INIT_POP // BATCH_SIZE
for _ in range(n_batches):
sols = torch.randn(SD_IN_SHAPE, device=DEVICE, dtype=TORCH_DTYPE)
objs, _, clip_features, dreamsim_features = evaluate_lsi(
sols, prompt, return_features=True
)
all_sols.append(sols)
all_clip_features.append(clip_features)
all_dreamsim_features.append(dreamsim_features)
all_objs.append(objs)
all_sols = torch.concat(all_sols, dim=0)
all_clip_features = torch.concat(all_clip_features, dim=0)
all_dreamsim_features = torch.concat(all_dreamsim_features, dim=0)
all_objs = np.concatenate(all_objs, axis=0)
# Initialize the diversity projection model.
div_proj_data = []
div_proj_labels = []
for _ in range(n_pref_data):
idx = np.random.choice(all_sols.shape[0], 3)
div_proj_data.append(all_clip_features[idx])
div_proj_labels.append(all_dreamsim_features[idx])
div_proj_data = torch.concat(div_proj_data, dim=0)
div_proj_labels = torch.concat(div_proj_labels, dim=0)
div_proj_data = div_proj_data.reshape(n_pref_data, 3, -1)
div_proj_label = div_proj_labels.reshape(n_pref_data, 3, -1)
diversity_model, div_proj_acc = fit_div_proj(
div_proj_data,
div_proj_label,
latent_dim=2,
)
else:
tqdm.write("Updating archive and diversity projection.")
# Get all the current solutions and collect feedback.
all_sols = list_to_tensor(archive.data("solution"))
n_batches = np.ceil(len(all_sols) / BATCH_SIZE).astype(int)
all_clip_features = []
all_dreamsim_features = []
all_objs = []
for i in range(n_batches):
sols = all_sols[i * BATCH_SIZE : (i + 1) * BATCH_SIZE]
objs, _, clip_features, dreamsim_features = evaluate_lsi(
sols, prompt, return_features=True
)
all_clip_features.append(clip_features)
all_dreamsim_features.append(dreamsim_features)
all_objs.append(objs)
all_clip_features = torch.concat(
all_clip_features, dim=0
) # n_pref_data * 3, dim
all_dreamsim_features = torch.concat(all_dreamsim_features, dim=0)
all_objs = np.concatenate(all_objs, axis=0)
# Update the diversity projection model.
additional_features = []
additional_labels = []
for _ in range(n_pref_data):
idx = np.random.choice(all_sols.shape[0], 3)
additional_features.append(all_clip_features[idx])
additional_labels.append(all_dreamsim_features[idx])
additional_features = torch.concat(additional_features, dim=0)
additional_labels = torch.concat(additional_labels, dim=0)
additional_div_proj_data = additional_features.reshape(n_pref_data, 3, -1)
additional_div_proj_label = additional_labels.reshape(n_pref_data, 3, -1)
div_proj_data = torch.concat(
(div_proj_data, additional_div_proj_data), axis=0
)
div_proj_label = torch.concat(
(div_proj_label, additional_div_proj_label), axis=0
)
diversity_model, div_proj_acc = fit_div_proj(
div_proj_data,
div_proj_label,
latent_dim=2,
)
archive, scheduler = create_scheduler(
all_sols,
all_objs,
all_clip_features,
diversity_model,
seed=SEED,
)
# Primary QD loop.
sols = scheduler.ask()
sols = list_to_tensor(sols)
objs, measures, clip_features, dreamsim_features = evaluate_lsi(
sols, prompt, return_features=True, diversity_model=diversity_model
)
best = max(best, max(objs))
scheduler.tell(objs, measures)
# This can be used as a flag to save on the final iteration, but note that
# we do not save results in this tutorial.
final_itr = itr == TOTAL_ITRS
# Update the summary statistics for the archive.
qd_score, coverage = archive.stats.norm_qd_score, archive.stats.coverage
tqdm.write(f"QD score: {np.round(qd_score, 2)} Coverage: {coverage * 100}")
plt = plot_archive(archive)
yield archive, plt
plt = plot_archive(archive)
return archive, plt
def many_pictures(archive, prompt:str):
# Modify this to determine how many images to plot along each dimension.
img_freq = (
4, # Number of columns of images.
4, # Number of rows of images.
)
# List of images.
imgs = []
# Convert archive to a df with solutions available.
df = archive.data(return_type="pandas")
# Compute the min and max measures for which solutions were found.
measure_bounds = np.array(
[
(df["measures_0"].min(), df["measures_0"].max()),
(df["measures_1"].min(), df["measures_1"].max()),
]
)
archive_bounds = np.array(
[archive.boundaries[0][[0, -1]], archive.boundaries[1][[0, -1]]]
)
delta_measures_0 = (archive_bounds[0][1] - archive_bounds[0][0]) / img_freq[0]
delta_measures_1 = (archive_bounds[1][1] - archive_bounds[1][0]) / img_freq[1]
for col, row in itertools.product(range(img_freq[1]), range(img_freq[0])):
# Compute bounds of a box in measure space.
measures_0_low = archive_bounds[0][0] + delta_measures_0 * row
measures_0_high = archive_bounds[0][0] + delta_measures_0 * (row + 1)
measures_1_low = archive_bounds[1][0] + delta_measures_1 * col
measures_1_high = archive_bounds[1][0] + delta_measures_1 * (col + 1)
if row == 0:
measures_0_low = measure_bounds[0][0]
if col == 0:
measures_1_low = measure_bounds[1][0]
if row == img_freq[0] - 1:
measures_0_high = measure_bounds[0][1]
if col == img_freq[1] - 1:
measures_0_high = measure_bounds[1][1]
# Query for a solution with measures within this box.
query_string = (
f"{measures_0_low} <= measures_0 & measures_0 <= {measures_0_high} & "
f"{measures_1_low} <= measures_1 & measures_1 <= {measures_1_high}"
)
df_box = df.query(query_string)
if not df_box.empty:
# Randomly sample a solution from the box.
# Stable Diffusion solutions have SD_IN_CHANNELS * SD_IN_HEIGHT * SD_IN_WIDTH
# dimensions, so the final solution col is solution_(x-1).
sol = (
df_box.loc[
:,
"solution_0" : "solution_{}".format(
SD_IN_CHANNELS * SD_IN_HEIGHT * SD_IN_WIDTH - 1
),
]
.sample(n=1)
.iloc[0]
)
# Convert the latent vector solution to an image.
latents = torch.tensor(sol.to_numpy()).reshape(
(1, SD_IN_CHANNELS, SD_IN_HEIGHT, SD_IN_WIDTH)
)
latents = latents.to(TORCH_DTYPE).to(DEVICE)
img = SDPIPE(
prompt,
num_images_per_prompt=1,
latents=latents,
# num_inference_steps=1, # For testing.
).images[0]
img = torch.from_numpy(np.array(img)).permute(2, 0, 1) / 255.0
imgs.append(img)
else:
imgs.append(torch.zeros((3, IMG_HEIGHT, IMG_WIDTH)))
from torchvision.utils import make_grid
def create_archive_tick_labels(measure_range, num_ticks):
delta = (measure_range[1] - measure_range[0]) / num_ticks
ticklabels = [round(delta * p + measure_range[0], 3) for p in range(num_ticks + 1)]
return ticklabels
plt.figure(figsize=(img_freq[0] * 2, img_freq[0] * 2))
img_grid = make_grid(imgs, nrow=img_freq[0], padding=0)
img_grid = np.transpose(img_grid.cpu().numpy(), (1, 2, 0))
plt.imshow(img_grid)
plt.xlabel("")
num_x_ticks = img_freq[0]
x_ticklabels = create_archive_tick_labels(measure_bounds[0], num_x_ticks)
x_tick_range = img_grid.shape[1]
x_ticks = np.arange(0, x_tick_range + 1e-9, step=x_tick_range / num_x_ticks)
plt.xticks(x_ticks, x_ticklabels)
plt.ylabel("")
num_y_ticks = img_freq[1]
y_ticklabels = create_archive_tick_labels(measure_bounds[1], num_y_ticks)
y_ticklabels.reverse()
y_tick_range = img_grid.shape[0]
y_ticks = np.arange(0, y_tick_range + 1e-9, step=y_tick_range / num_y_ticks)
plt.yticks(y_ticks, y_ticklabels)
plt.tight_layout()
return plt
|