forked from nshepperd/jax-guided-diffusion
-
Notifications
You must be signed in to change notification settings - Fork 7
/
execute.py
339 lines (285 loc) · 12.2 KB
/
execute.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
# Based on Katherine Crowson's CLIP guided diffusion notebook
# (https://colab.research.google.com/drive/12a_Wrfi2_gwwAuN3VvMTwVMz9TfqctNj)
# including a port of https://github.com/crowsonkb/guided-diffusion to jax.
import math
import io
import sys
import time
import os
import functools
from functools import partial
from PIL import Image
import requests
import numpy as np
import jax
import jax.numpy as jnp
import jaxtorch
from jaxtorch import PRNG, Context, ParamState, Module
from tqdm import tqdm
sys.path.append('./CLIP_JAX')
import clip_jax
from lib.script_util import create_model_and_diffusion, model_and_diffusion_defaults
from lib import util
from lib.util import pil_from_tensor, pil_to_tensor
# Define necessary functions
def fetch(url_or_path):
if str(url_or_path).startswith('http://') or str(url_or_path).startswith('https://'):
r = requests.get(url_or_path)
r.raise_for_status()
fd = io.BytesIO()
fd.write(r.content)
fd.seek(0)
return fd
return open(url_or_path, 'rb')
class MakeCutouts(object):
def __init__(self, cut_size, cutn, cut_pow=1.):
self.cut_size = cut_size
self.cutn = cutn
self.cut_pow = cut_pow
def key(self):
return (self.cut_size,self.cutn,self.cut_pow)
def __hash__(self):
return hash(self.key())
def __eq__(self, other):
if isinstance(other, MakeCutouts):
return type(self) is type(other) and self.key() == other.key()
return NotImplemented
def __call__(self, input, key):
[b, c, h, w] = input.shape
rng = PRNG(key)
max_size = min(h, w)
min_size = min(h, w, self.cut_size)
cut_us = jax.random.uniform(rng.split(), shape=[self.cutn])**self.cut_pow
sizes = (min_size + cut_us * (max_size - min_size + 1)).astype(jnp.int32).clamp(min_size, max_size)
offsets_x = jax.random.randint(rng.split(), [self.cutn], 0, w - sizes + 1)
offsets_y = jax.random.randint(rng.split(), [self.cutn], 0, h - sizes + 1)
cutouts = util.cutouts_images(input, offsets_x, offsets_y, sizes)
cutouts = cutouts.rearrange('b n c h w -> (n b) c h w')
return cutouts
class StaticCutouts(MakeCutouts):
def __init__(self, cut_size, cutn, size):
self.cut_size = cut_size
self.cutn = cutn
self.size = size
def key(self):
return (self.cut_size,self.cutn,self.size)
def __call__(self, input, key):
[b, c, h, w] = input.shape
rng = PRNG(key)
sizes = jnp.array([self.size]*self.cutn).astype(jnp.int32)
offsets_x = jax.random.randint(rng.split(), [self.cutn], 0, w - sizes + 1)
offsets_y = jax.random.randint(rng.split(), [self.cutn], 0, h - sizes + 1)
cutouts = util.cutouts_images(input, offsets_x, offsets_y, sizes)
cutouts = cutouts.rearrange('b n c h w -> (n b) c h w')
return cutouts
def Normalize(mean, std):
mean = jnp.array(mean).reshape(3,1,1)
std = jnp.array(std).reshape(3,1,1)
def forward(image):
return (image - mean) / std
return forward
def norm1(x):
"""Normalize to the unit sphere."""
return x / x.square().sum(axis=-1, keepdims=True).sqrt()
def spherical_dist_loss(x, y):
x = norm1(x)
y = norm1(y)
return (x - y).square().sum(axis=-1).sqrt().div(2).arcsin().square().mul(2)
def cborfile(path):
with fetch(path) as fp:
return jaxtorch.cbor.load(fp)
def tv_loss(input):
"""L2 total variation loss, as in Mahendran et al."""
# input = jnp.pad(input, ((0,0), (0,0), (0,1), (0,1)), mode='edge')
# x_diff = input[..., :-1, 1:] - input[..., :-1, :-1]
# y_diff = input[..., 1:, :-1] - input[..., :-1, :-1]
# return (x_diff**2 + y_diff**2).mean([1, 2, 3])
x_diff = input[..., :, 1:] - input[..., :, :-1]
y_diff = input[..., 1:, :] - input[..., :-1, :]
return x_diff.square().mean([1,2,3]) + y_diff.square().mean([1,2,3])
def downscale2d(image, f):
[c, n, h, w] = image.shape
return jax.image.resize(image, [c, n, h//f, w//f], method='linear')
def rms(x):
return x.square().mean().sqrt()
# Model settings
model_config = model_and_diffusion_defaults()
model_config.update({
'attention_resolutions': '32, 16, 8',
'class_cond': False,
'diffusion_steps': 1000,
'rescale_timesteps': True,
'timestep_respacing': '1000',
'image_size': 512,
'learn_sigma': True,
'noise_schedule': 'linear',
'num_channels': 256,
'num_head_channels': 64,
'num_res_blocks': 2,
'resblock_updown': True,
# 'use_fp16': True,
'use_scale_shift_norm': True,
})
# Load models
model, diffusion = create_model_and_diffusion(**model_config)
model_params = ParamState(model.labeled_parameters_())
model_params.initialize(jax.random.PRNGKey(0))
print('Loading state dict...')
# with open('256x256_diffusion_uncond.cbor', 'rb') as fp:
with open('512x512_diffusion_uncond_finetune_008100.cbor', 'rb') as fp:
jax_state_dict = jaxtorch.cbor.load(fp)
model.load_state_dict(model_params, jax_state_dict)
def exec_model(model_params, x, timesteps, y=None):
cx = Context(model_params, jax.random.PRNGKey(0))
return model(cx, x, timesteps, y=y)
exec_model_jit = functools.partial(jax.jit(exec_model), model_params)
def base_cond_fn(x, t, y, text_embed, style_embed, cur_t, key, model_params, clip_params, clip_guidance_scale, style_guidance_scale, tv_scale, sat_scale, make_cutouts, make_cutouts_style):
rng = PRNG(key)
n = x.shape[0]
def denoise(x):
my_t = jnp.ones([n], dtype=jnp.int32) * cur_t
out = diffusion.p_mean_variance(functools.partial(exec_model,model_params),
x, my_t,
clip_denoised=False,
model_kwargs={'y': y})
fac = diffusion.sqrt_one_minus_alphas_cumprod[cur_t]
x_in = out['pred_xstart'] * fac + x * (1 - fac)
return x_in
(x_in, backward) = jax.vjp(denoise, x)
def main_clip_loss(x_in, key):
clip_in = normalize(make_cutouts(x_in.add(1).div(2), key))
image_embeds = emb_image(clip_in, clip_params).reshape([make_cutouts.cutn, n, 512])
# Method 1. Average the clip embeds, then compute great circle distance.
# losses = spherical_dist_loss(image_embeds.mean(0), text_embed)
# Method 2. Compute great circle losses for clip embeds, then average.
losses = spherical_dist_loss(image_embeds, text_embed).mean(0)
return losses.sum() * clip_guidance_scale
# Scan method, should reduce jit times...
num_cuts = 4
keys = jnp.stack([rng.split() for _ in range(num_cuts)])
main_clip_grad = jax.lax.scan(lambda total, key: (total + jax.grad(main_clip_loss)(x_in, key), key),
jnp.zeros_like(x_in),
keys)[0] / num_cuts
if style_embed is not None:
def style_loss(x_in, key):
clip_in = normalize(make_cutouts_style(x_in.add(1).div(2), key))
image_embeds = emb_image(clip_in, clip_params).reshape([make_cutouts_style.cutn, n, 512])
style_losses = spherical_dist_loss(image_embeds, style_embed).mean(0)
return style_losses.sum() * style_guidance_scale
main_clip_grad += jax.grad(style_loss)(x_in, rng.split())
def sum_tv_loss(x_in, f=None):
if f is not None:
x_in = downscale2d(x_in, f)
return tv_loss(x_in).sum() * tv_scale
tv_grad_512 = jax.grad(sum_tv_loss)(x_in)
tv_grad_256 = jax.grad(partial(sum_tv_loss,f=2))(x_in)
tv_grad_128 = jax.grad(partial(sum_tv_loss,f=4))(x_in)
main_clip_grad += tv_grad_512 + tv_grad_256 + tv_grad_128
def saturation_loss(x_in):
return jnp.abs(x_in - x_in.clamp(min=-1,max=1)).mean()
sat_grad = sat_scale * jax.grad(saturation_loss)(x_in)
main_clip_grad += sat_grad
return (-backward(main_clip_grad)[0], tv_grad_512, tv_grad_256, tv_grad_128, sat_grad)
base_cond_fn = jax.jit(base_cond_fn, static_argnames=['make_cutouts', 'make_cutouts_style'])
print('Loading CLIP model...')
image_fn, text_fn, clip_params, _ = clip_jax.load('ViT-B/32') #, "cpu")
clip_size = 224
normalize = Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711])
def txt(prompt):
"""Returns normalized embedding."""
text = clip_jax.tokenize([prompt])
text_embed = text_fn(clip_params, text)
return norm1(text_embed.reshape(512))
def emb_image(image, clip_params=None):
return norm1(image_fn(clip_params, image))
title = ['sigil of the knight of time. trending on ArtStation']
prompt = [txt(t) for t in title]
style_embed = norm1(jnp.array(cborfile('data/openimages_512x_png_embed224.cbor'))) - norm1(jnp.array(cborfile('data/imagenet_512x_jpg_embed224.cbor')))
batch_size = 1
clip_guidance_scale = 2000
style_guidance_scale = 300
tv_scale = 150
sat_scale = 150
cutn = 32 # effective cutn is 4x this because we do 4 iterations in base_cond_fn
cut_pow = 0.5
style_cutn = 32
n_batches = 4
init_image = None
skip_timesteps = 0
seed = 1
# Actually do the run
print('Starting run...')
def run():
rng = PRNG(jax.random.PRNGKey(seed))
text_embed = prompt
init = None
# if init_image is not None:
# init = Image.open(fetch(init_image)).convert('RGB')
# init = init.resize((model_config['image_size'], model_config['image_size']), Image.LANCZOS)
# init = TF.to_tensor(init).to(device).unsqueeze(0).mul(2).sub(1)
cur_t = None
make_cutouts = MakeCutouts(clip_size, cutn, cut_pow=cut_pow)
make_cutouts_style = StaticCutouts(clip_size, style_cutn, size=224)
def cond_fn(x, t, y=None):
# Triggers recompilation if cutout parameters have changed (cutn or cut_pow).
grad = base_cond_fn(x, jnp.array(t), y,
text_embed=text_embed,
style_embed=style_embed,
cur_t=jnp.array(cur_t),
key=rng.split(),
model_params=model_params,
clip_params=clip_params,
clip_guidance_scale = clip_guidance_scale,
style_guidance_scale = style_guidance_scale,
tv_scale = tv_scale,
sat_scale = sat_scale,
make_cutouts=make_cutouts,
make_cutouts_style=make_cutouts_style)
(grad, tv1, tv2, tv4, sat) = grad
if int(t)%10 == 0:
print(t, rms(tv1), rms(tv2), rms(tv4), rms(sat))
magnitude = grad.square().mean().sqrt()
grad = grad / magnitude * magnitude.clamp(max=0.1)
return grad
for i in range(n_batches):
if type(prompt) is list:
text_embed = prompt[i % len(prompt)]
this_title = title[i % len(prompt)]
else:
text_embed = prompt
this_title = title
cur_t = diffusion.num_timesteps - skip_timesteps - 1
samples = diffusion.p_sample_loop_progressive(
exec_model_jit,
(batch_size, 3, model_config['image_size'], model_config['image_size']),
rng=rng,
clip_denoised=False,
model_kwargs={},
cond_fn=cond_fn,
progress=tqdm,
skip_timesteps=skip_timesteps,
init_image=init,
)
for j, sample in enumerate(samples):
cur_t -= 1
if j % 100 == 0 or cur_t == -1:
print()
for k, image in enumerate(sample['pred_xstart']):
filename = f'progress_{i * batch_size + k:05}.png'
# print(k, type(image).mro())
# For some reason this comes out as a numpy array. Huh?
image = pil_from_tensor(jnp.array(image).add(1).div(2))
image.save(filename)
print(f'Wrote {filename}')
# for k in range(batch_size):
# filename = f'progress_{i * batch_size + k:05}.png'
# timestring = time.strftime('%Y%m%d%H%M%S')
# os.makedirs('samples', exist_ok=True)
# dname = f'samples/{timestring}_{k}_{title}.png'
# with open(filename, 'rb') as fp:
# data = fp.read()
# with open(dname, 'wb') as fp:
# fp.write(data)
# files.download(dname)
run()