forked from tnwei/vqgan-clip-app
-
Notifications
You must be signed in to change notification settings - Fork 0
/
logic.py
405 lines (345 loc) · 13.8 KB
/
logic.py
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
from typing import Optional, List, Tuple
from PIL import Image
import argparse
import clip
from vqgan_utils import (
load_vqgan_model,
MakeCutouts,
parse_prompt,
resize_image,
Prompt,
synth,
checkin,
TVLoss,
)
import torch
from torchvision.transforms import functional as TF
import torch.nn as nn
from torch.nn import functional as F
from torch import optim
from torchvision import transforms
import cv2
import numpy as np
import kornia.augmentation as K
class Run:
"""
Subclass this to house your own implementation of CLIP-based image generation
models within the UI
"""
def __init__(self):
"""
Set up the run's config here
"""
pass
def load_model(self):
"""
Load models here. Separated this from __init__ to allow loading model state
from a previous run
"""
pass
def model_init(self):
"""
Continue run setup, for items that require the models to be in=place.
Call once after load_model
"""
pass
def iterate(self):
"""
Place iteration logic here. Outputs results for human consumption at
every step.
"""
pass
class VQGANCLIPRun(Run):
def __init__(
# Inputs
self,
text_input: str = "the first day of the waters",
vqgan_ckpt: str = "vqgan_imagenet_f16_16384",
num_steps: int = 300,
image_x: int = 300,
image_y: int = 300,
init_image: Optional[Image.Image] = None,
image_prompts: List[Image.Image] = [],
continue_prev_run: bool = False,
seed: Optional[int] = None,
mse_weight=0.5,
mse_weight_decay=0.1,
mse_weight_decay_steps=50,
tv_loss_weight=1e-3,
use_cutout_augmentations: bool = True,
# use_augs: bool = True,
# noise_fac: float = 0.1,
# use_noise: Optional[float] = None,
# mse_withzeros=True,
## **kwargs, # Use this to receive Streamlit objects ## Call from main UI
use_scrolling_zooming: bool = False,
translation_x: int = 0,
translation_y: int = 0,
rotation_angle: float = 0,
zoom_factor: float = 1,
transform_interval: int = 10,
device: Optional[torch.device] = None,
) -> None:
super().__init__()
self.text_input = text_input
self.vqgan_ckpt = vqgan_ckpt
self.num_steps = num_steps
self.image_x = image_x
self.image_y = image_y
self.init_image = init_image
self.image_prompts = image_prompts
self.continue_prev_run = continue_prev_run
self.seed = seed
# Setup ------------------------------------------------------------------------------
# Split text by "|" symbol
texts = [phrase.strip() for phrase in text_input.split("|")]
if texts == [""]:
texts = []
# Leaving most of this untouched
self.args = argparse.Namespace(
prompts=texts,
image_prompts=image_prompts,
noise_prompt_seeds=[],
noise_prompt_weights=[],
size=[int(image_x), int(image_y)],
init_image=init_image,
init_weight=mse_weight,
# clip.available_models()
# ['RN50', 'RN101', 'RN50x4', 'ViT-B/32']
# Visual Transformer seems to be the smallest
clip_model="ViT-B/32",
vqgan_config=f"assets/{vqgan_ckpt}.yaml",
vqgan_checkpoint=f"assets/{vqgan_ckpt}.ckpt",
step_size=0.05,
cutn=64,
cut_pow=1.0,
display_freq=50,
seed=seed,
)
if device is None:
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
else:
self.device = device
print("Using device:", device)
self.iterate_counter = 0
# self.use_augs = use_augs
# self.noise_fac = noise_fac
# self.use_noise = use_noise
# self.mse_withzeros = mse_withzeros
self.init_mse_weight = mse_weight
self.mse_weight = mse_weight
self.mse_weight_decay = mse_weight_decay
self.mse_weight_decay_steps = mse_weight_decay_steps
self.use_cutout_augmentations = use_cutout_augmentations
# For TV loss
self.tv_loss_weight = tv_loss_weight
self.use_scrolling_zooming = use_scrolling_zooming
self.translation_x = translation_x
self.translation_y = translation_y
self.rotation_angle = rotation_angle
self.zoom_factor = zoom_factor
self.transform_interval = transform_interval
def load_model(
self, prev_model: nn.Module = None, prev_perceptor: nn.Module = None
) -> Optional[Tuple[nn.Module, nn.Module]]:
if self.continue_prev_run is True:
self.model = prev_model
self.perceptor = prev_perceptor
return None
else:
self.model = load_vqgan_model(
self.args.vqgan_config, self.args.vqgan_checkpoint
).to(self.device)
self.perceptor = (
clip.load(self.args.clip_model, jit=False)[0]
.eval()
.requires_grad_(False)
.to(self.device)
)
return self.model, self.perceptor
def model_init(self, init_image: Image.Image = None) -> None:
cut_size = self.perceptor.visual.input_resolution
e_dim = self.model.quantize.e_dim
f = 2 ** (self.model.decoder.num_resolutions - 1)
if self.use_cutout_augmentations:
noise_fac = 0.1
augs = nn.Sequential(
K.RandomHorizontalFlip(p=0.5),
K.RandomSharpness(0.3, p=0.4),
K.RandomAffine(degrees=30, translate=0.1, p=0.8, padding_mode="border"),
K.RandomPerspective(0.2, p=0.4),
K.ColorJitter(hue=0.01, saturation=0.01, p=0.7),
)
else:
noise_fac = None
augs = None
self.make_cutouts = MakeCutouts(
cut_size,
self.args.cutn,
cut_pow=self.args.cut_pow,
noise_fac=noise_fac,
augs=augs,
)
n_toks = self.model.quantize.n_e
toksX, toksY = self.args.size[0] // f, self.args.size[1] // f
sideX, sideY = toksX * f, toksY * f
self.z_min = self.model.quantize.embedding.weight.min(dim=0).values[
None, :, None, None
]
self.z_max = self.model.quantize.embedding.weight.max(dim=0).values[
None, :, None, None
]
if self.seed is not None:
torch.manual_seed(self.seed)
else:
self.seed = torch.seed() # Trigger a seed, retrieve the utilized seed
# Initialization order: continue_prev_im, init_image, then only random init
if init_image is not None:
init_image = init_image.resize((sideX, sideY), Image.LANCZOS)
self.z, *_ = self.model.encode(
TF.to_tensor(init_image).to(self.device).unsqueeze(0) * 2 - 1
)
elif self.args.init_image:
pil_image = self.args.init_image
pil_image = pil_image.resize((sideX, sideY), Image.LANCZOS)
self.z, *_ = self.model.encode(
TF.to_tensor(pil_image).to(self.device).unsqueeze(0) * 2 - 1
)
else:
one_hot = F.one_hot(
torch.randint(n_toks, [toksY * toksX], device=self.device), n_toks
).float()
self.z = one_hot @ self.model.quantize.embedding.weight
self.z = self.z.view([-1, toksY, toksX, e_dim]).permute(0, 3, 1, 2)
self.z_orig = self.z.clone()
self.z.requires_grad_(True)
self.opt = optim.Adam([self.z], lr=self.args.step_size)
self.normalize = transforms.Normalize(
mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711],
)
self.pMs = []
for prompt in self.args.prompts:
txt, weight, stop = parse_prompt(prompt)
embed = self.perceptor.encode_text(
clip.tokenize(txt).to(self.device)
).float()
self.pMs.append(Prompt(embed, weight, stop).to(self.device))
for uploaded_image in self.args.image_prompts:
# path, weight, stop = parse_prompt(prompt)
# img = resize_image(Image.open(fetch(path)).convert("RGB"), (sideX, sideY))
img = resize_image(uploaded_image.convert("RGB"), (sideX, sideY))
batch = self.make_cutouts(TF.to_tensor(img).unsqueeze(0).to(self.device))
embed = self.perceptor.encode_image(self.normalize(batch)).float()
self.pMs.append(Prompt(embed, weight, stop).to(self.device))
for seed, weight in zip(
self.args.noise_prompt_seeds, self.args.noise_prompt_weights
):
gen = torch.Generator().manual_seed(seed)
embed = torch.empty([1, self.perceptor.visual.output_dim]).normal_(
generator=gen
)
self.pMs.append(Prompt(embed, weight).to(self.device))
def _ascend_txt(self) -> List:
out = synth(self.model, self.z)
iii = self.perceptor.encode_image(
self.normalize(self.make_cutouts(out))
).float()
result = {}
if self.args.init_weight:
result["mse_loss"] = F.mse_loss(self.z, self.z_orig) * self.mse_weight / 2
# MSE regularization scheduler
with torch.no_grad():
# if not the first step
# and is time for step change
# and both weight decay steps and magnitude are nonzero
# and MSE isn't zero already
if (
self.iterate_counter > 0
and self.iterate_counter % self.mse_weight_decay_steps == 0
and self.mse_weight_decay != 0
and self.mse_weight_decay_steps != 0
and self.mse_weight != 0
):
self.mse_weight = self.mse_weight - self.mse_weight_decay
# Don't allow changing sign
# Basically, caps MSE at zero if decreasing from positive
# But, also prevents MSE from becoming positive if -MSE intended
if self.init_mse_weight > 0:
self.mse_weight = max(self.mse_weight, 0)
else:
self.mse_weight = min(self.mse_weight, 0)
print(f"updated mse weight: {self.mse_weight}")
tv_loss_fn = TVLoss()
result["tv_loss"] = tv_loss_fn(self.z) * self.tv_loss_weight
for count, prompt in enumerate(self.pMs):
result[f"prompt_loss_{count}"] = prompt(iii)
return result
def iterate(self) -> Tuple[List[float], Image.Image]:
if not self.use_scrolling_zooming:
# Forward prop
self.opt.zero_grad()
losses = self._ascend_txt()
# Grab an image
im: Image.Image = checkin(self.model, self.z)
# Backprop
loss = sum([j for i, j in losses.items()])
loss.backward()
self.opt.step()
with torch.no_grad():
self.z.copy_(self.z.maximum(self.z_min).minimum(self.z_max))
# Advance iteration counter
self.iterate_counter += 1
print(
f"Step {self.iterate_counter} losses: {[(i, j.item()) for i, j in losses.items()]}"
)
# Output stuff useful for humans
return [(i, j.item()) for i, j in losses.items()], im
else:
# Grab current image
im_before_transform: Image.Image = checkin(self.model, self.z)
# Convert for use in OpenCV
imarr = np.array(im_before_transform)
imarr = cv2.cvtColor(imarr, cv2.COLOR_RGB2BGR)
translation = np.float32(
[[1, 0, self.translation_x], [0, 1, self.translation_y]]
)
imcenter = (imarr.shape[1] // 2, imarr.shape[0] // 2)
rotation = cv2.getRotationMatrix2D(
imcenter, angle=self.rotation_angle, scale=self.zoom_factor
)
trans_mat = np.vstack([translation, [0, 0, 1]])
rot_mat = np.vstack([rotation, [0, 0, 1]])
transformation_matrix = np.matmul(rot_mat, trans_mat)
outarr = cv2.warpPerspective(
imarr,
transformation_matrix,
(imarr.shape[1], imarr.shape[0]),
borderMode=cv2.BORDER_WRAP,
)
transformed_im = Image.fromarray(cv2.cvtColor(outarr, cv2.COLOR_BGR2RGB))
# Encode as z, reinit
self.z, *_ = self.model.encode(
TF.to_tensor(transformed_im).to(self.device).unsqueeze(0) * 2 - 1
)
self.z.requires_grad_(True)
self.opt = optim.Adam([self.z], lr=self.args.step_size)
for _ in range(self.transform_interval):
# Forward prop
self.opt.zero_grad()
losses = self._ascend_txt()
# Grab an image
im: Image.Image = checkin(self.model, self.z)
# Backprop
loss = sum([j for i, j in losses.items()])
loss.backward()
self.opt.step()
with torch.no_grad():
self.z.copy_(self.z.maximum(self.z_min).minimum(self.z_max))
# Advance iteration counter
self.iterate_counter += 1
print(
f"Step {self.iterate_counter} losses: {[(i, j.item()) for i, j in losses.items()]}"
)
# Output stuff useful for humans
return [(i, j.item()) for i, j in losses.items()], im