-
Notifications
You must be signed in to change notification settings - Fork 2
/
dalle.py
205 lines (158 loc) · 5.64 KB
/
dalle.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
import io
import os
import argparse
import numpy as np
import torch
import torchvision
import torchvision.transforms as T
import torchvision.transforms.functional as TF
import torch.nn.functional as F
import clip
import PIL
from dall_e import map_pixels, unmap_pixels, load_model
parser = argparse.ArgumentParser()
parser.add_argument(
'--output_path',
type=str,
default='./generations',
help='',
)
parser.add_argument(
'--prompt',
type=str,
default='A delicious avocado',
help='',
)
parser.add_argument(
'--lr',
type=float,
default=3e-1,
help='',
)
parser.add_argument(
'--img_save_freq',
type=int,
default=5,
help='',
)
args = parser.parse_args()
output_path = args.output_path
prompt = args.prompt
lr = args.lr
img_save_freq = args.img_save_freq
output_dir = os.path.join(output_path, f'"{prompt}"')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print("USING ", device)
target_img_size = 256
final_img_size = 512
def preprocess(img):
min_img_dim = min(img.size)
if min_img_dim < target_img_size:
raise ValueError(f'min dim for img {min_img_dim} < {target_img_size}')
img_ratio = target_img_size / min_img_dim
min_img_dim = (round(img_ratio * img.size[1]),
round(img_ratio * img.size[0]))
img = TF.resize(img, min_img_dim, interpolation=PIL.Image.LANCZOS)
img = TF.center_crop(img, output_size=2 * [target_img_size])
img = torch.unsqueeze(T.ToTensor()(img), 0)
return map_pixels(img)
def compute_clip_loss(img, text):
img = clip_transform(img)
img = torch.nn.functional.upsample_bilinear(img, (224, 224))
img_logits = clip_model.encode_image(img)
tokenized_text = clip.tokenize([text]).to(device).detach().clone()
text_logits = clip_model.encode_text(tokenized_text)
loss = 10 * -torch.cosine_similarity(text_logits, img_logits).mean()
return loss
def get_stacked_random_crops(img, num_random_crops=64):
img_size = [img.shape[2], img.shape[3]]
crop_list = []
for _ in range(num_random_crops):
crop_size_y = int(img_size[0] * torch.zeros(1, ).uniform_(.75, .95))
crop_size_x = int(img_size[1] * torch.zeros(1, ).uniform_(.75, .95))
y_offset = torch.randint(0, img_size[0] - crop_size_y, ())
x_offset = torch.randint(0, img_size[1] - crop_size_x, ())
crop = img[:, :, y_offset:y_offset + crop_size_y,
x_offset:x_offset + crop_size_x]
crop = torch.nn.functional.upsample_bilinear(crop, (224, 224))
# crop = torch.nn.functional.interpolate(
# crop,
# (224, 224),
# mode='bilinear',
# )
crop_list.append(crop)
img = torch.cat(crop_list, axis=0)
return img
clip_model, clip_preprocess = clip.load("ViT-B/32", device=device)
clip_model.eval()
clip_transform = torchvision.transforms.Compose([
# clip_preprocess.transforms[2],
clip_preprocess.transforms[4],
])
dec = load_model("https://cdn.openai.com/dall-e/decoder.pkl", device)
dec.eval()
scale_x = 1
scale_y = 1
z_logits = torch.rand((1, 8192, 64 * scale_y, int(64 * scale_x))).cuda()
z_logits = torch.nn.Parameter(z_logits, requires_grad=True)
optimizer = torch.optim.Adam(
params=[z_logits],
lr=lr,
betas=(0.9, 0.999),
)
final_x_rec = torch.zeros(
[1, 3, final_img_size * scale_y, int(final_img_size * scale_x)])
counter = 0
rec_steps = 3
x_rec_merged = torch.zeros([1, 3, 512, 512])
while True:
final_x_rec = final_x_rec.detach().clone()
x_rec_merged = x_rec_merged.detach().clone()
for s_y in range(scale_y):
for s_x in np.linspace(0, scale_x - 1, rec_steps):
z_logits_part = z_logits[:, :, int(64 * s_y):int(64 * (s_y + 1)),
int(64 * s_x):int(64 * (s_x + 1))]
z = torch.nn.functional.gumbel_softmax(
z_logits_part.permute(0, 2, 3, 1).reshape(1, 64**2, 8192),
hard=False,
dim=1,
).view(1, 8192, 64, 64)
x_stats = dec(z).float()
x_rec = unmap_pixels(torch.sigmoid(x_stats[:, :3]))
# x_rec_stacked = get_stacked_random_crops(
# x_rec,
# num_random_crops=16,
# )
step_img_size = int(final_img_size / rec_steps)
y_init_img_part = step_img_size * s_y
x_init_img_part = int(step_img_size * s_x)
y_final_img_part = y_init_img_part + step_img_size * (s_y + 1)
x_final_img_part = x_init_img_part + step_img_size * int(s_x + 1)
x_rec_part = x_rec[:, :, 0:(step_img_size * (s_y + 1)),
0:(step_img_size * int(s_x + 1))]
x_rec_merged[:, :, y_init_img_part:y_final_img_part,
x_init_img_part:
x_final_img_part] = x_rec_part
final_x_rec[:, :,
int(512 * s_y):int(512 * (s_y + 1)),
int(512 * s_x):int(512 * (s_x + 1))] = x_rec_merged
final_x_rec_stacked = get_stacked_random_crops(
final_x_rec,
num_random_crops=64,
)
loss = compute_clip_loss(final_x_rec_stacked, prompt)
# final_loss = compute_clip_loss(final_x_rec_stacked, prompt)
# final_loss = 0
# loss = (part_loss + final_loss)/2
print(loss)
# print(z_logits[0, 0, 0])
loss /= rec_steps
optimizer.zero_grad()
loss.backward()
optimizer.step()
counter += 1
if counter % img_save_freq == 0:
x_rec = T.ToPILImage(mode='RGB')(final_x_rec[0])
x_rec.save(f"{output_dir}/{counter}.png")