-
Notifications
You must be signed in to change notification settings - Fork 14
/
edit_cli_single_multi_turn.py
209 lines (182 loc) · 8.85 KB
/
edit_cli_single_multi_turn.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
from __future__ import annotations
import math
import os.path
import random
import shutil
import sys
from argparse import ArgumentParser
from tqdm import tqdm
import einops
import k_diffusion as K
import numpy as np
import torch
import torch.nn as nn
from einops import rearrange
from omegaconf import OmegaConf
from PIL import Image, ImageOps
from torch import autocast
import json
import os
sys.path.append("./stable_diffusion")
from stable_diffusion.ldm.util import instantiate_from_config
class CFGDenoiser(nn.Module):
def __init__(self, model):
super().__init__()
self.inner_model = model
def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
cfg_cond = {
"c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
"c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
}
out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)
def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
print(f"Loading model from {ckpt}")
print(ckpt)
pl_sd = torch.load(ckpt, map_location="cpu")
if "global_step" in pl_sd:
print(f"Global Step: {pl_sd['global_step']}")
sd = pl_sd["state_dict"]
if vae_ckpt is not None:
print(f"Loading VAE from {vae_ckpt}")
vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
sd = {
k: vae_sd[k[len("first_s tage_model.") :]] if k.startswith("first_stage_model.") else v
for k, v in sd.items()
}
model = instantiate_from_config(config.model)
m, u = model.load_state_dict(sd, strict=False)
if len(m) > 0 and verbose:
print("missing keys:")
print(m)
if len(u) > 0 and verbose:
print("unexpected keys:")
print(u)
return model
def main():
parser = ArgumentParser()
parser.add_argument("--resolution", default=512, type=int)
parser.add_argument("--steps", default=100, type=int)
parser.add_argument("--config", default="configs/generate.yaml", type=str) # default 2.1
parser.add_argument("--ckpt", default="checkpoints/MagicBrush-epoch-52-step-4999.ckpt", type=str) # default 2.1
parser.add_argument("--vae-ckpt", default=None, type=str)
parser.add_argument("--input_path", required=True, type=str, default='./dev/images')
parser.add_argument("--output_path", required=True, type=str, default='./DevMagicBrushOutput')
parser.add_argument("--cfg-text", default=7.5, type=float)
parser.add_argument("--cfg-image", default=1.5, type=float)
parser.add_argument("--debug", action="store_true")
parser.add_argument("--skip_iter", action="store_true")
parser.add_argument("--seed", type=int)
args = parser.parse_args()
# with open('your_input_path_dir/series_instructions.json') as fp:
# with open(os.path.join(args.input_path + '-meta', 'series_instructions.json')) as fp:
with open(os.path.join(args.input_path, '..', 'edit_sessions.json')) as fp:
data_json = json.load(fp)
if not os.path.exists(args.output_path):
os.makedirs(args.output_path)
print(1)
config = OmegaConf.load(args.config)
model = load_model_from_config(config, args.ckpt, args.vae_ckpt)
model.eval().cuda()
model_wrap = K.external.CompVisDenoiser(model)
model_wrap_cfg = CFGDenoiser(model_wrap)
null_token = model.get_learned_conditioning([""])
seed = random.randint(0, 100000) if args.seed is None else args.seed
def edit_image(input_path, output_path, edict_instruction):
input_image = Image.open(input_path).convert("RGB")
width, height = input_image.size
factor = args.resolution / max(width, height)
factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
width = int((width * factor) // 64) * 64
height = int((height * factor) // 64) * 64
input_image = ImageOps.fit(input_image, (width, height), method=Image.Resampling.LANCZOS)
with torch.no_grad(), autocast("cuda"), model.ema_scope():
cond = {}
cond["c_crossattn"] = [model.get_learned_conditioning([edict_instruction])]
input_image = 2 * torch.tensor(np.array(input_image)).float() / 255 - 1
input_image = rearrange(input_image, "h w c -> 1 c h w").to(model.device)
cond["c_concat"] = [model.encode_first_stage(input_image).mode()]
uncond = {}
uncond["c_crossattn"] = [null_token]
uncond["c_concat"] = [torch.zeros_like(cond["c_concat"][0])]
sigmas = model_wrap.get_sigmas(args.steps)
extra_args = {
"cond": cond,
"uncond": uncond,
"text_cfg_scale": args.cfg_text,
"image_cfg_scale": args.cfg_image,
}
torch.manual_seed(seed)
z = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
z = K.sampling.sample_euler_ancestral(model_wrap_cfg, z, sigmas, extra_args=extra_args)
x = model.decode_first_stage(z)
x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0)
x = 255.0 * rearrange(x, "1 c h w -> h w c")
edited_image = Image.fromarray(x.type(torch.uint8).cpu().numpy())
edited_image.save(output_path)
if args.debug:
data_json = data_json[:3]
# iterative edit
if args.skip_iter:
print("Skip Iterative (Mutli-Turn) Editing......")
else:
print("Iterative (Mutli-Turn) Editing......")
# for idx, datas in enumerate(data_json):
for image_id, datas in tqdm(data_json.items()):
for turn_id, data in enumerate(datas):
print("data", data)
image_name = data['input']
image_dir = image_name.split('-')[0]
# 139306-input.png
# 139306-output1.png
# image_dir is 139306
if turn_id == 0: # first enter
image_path = os.path.join(args.input_path, image_dir, image_name)
else:
image_path = save_output_img_path
edit_instruction = data['instruction']
save_output_dir_path = os.path.join(args.output_path, image_dir)
if not os.path.exists(save_output_dir_path):
os.makedirs(save_output_dir_path)
if turn_id == 0:
save_output_img_path = os.path.join(save_output_dir_path, image_dir+'_1.png')
else:
save_output_img_path = os.path.join(save_output_dir_path, image_dir+'_iter_' +str(turn_id + 1)+'.png')
if os.path.exists(save_output_img_path):
print('already generated, skip')
continue
print('image_name', image_name)
print('image_path', image_path)
print('save_output_img_path', save_output_img_path)
print('edit_instruction', edit_instruction)
edit_image(image_path, save_output_img_path, edit_instruction)
print("Independent (Single-Turn) Editing......")
# for idx, datas in enumerate(data_json):
for image_id, datas in tqdm(data_json.items()):
for turn_id, data in enumerate(datas):
image_name = data['input']
image_dir = image_name.split('-')[0]
image_path = os.path.join(args.input_path, image_dir, image_name)
edit_instruction = data['instruction']
save_outut_dir_path = os.path.join(args.output_path, image_dir)
if not os.path.exists(save_outut_dir_path):
os.makedirs(save_outut_dir_path)
if turn_id == 0:
save_output_img_path = os.path.join(save_outut_dir_path, image_dir+'_1.png')
# if os.path.exists(save_output_img_path): # already generated in iterative editing
# print('already generated in iterative editing')
# continue
else:
save_output_img_path = os.path.join(save_outut_dir_path, image_dir+'_inde_' +str(turn_id + 1)+'.png')
if os.path.exists(save_output_img_path): # already generated in iterative (multi-turn) editing.
print('already generated, skip')
continue
print('image_name', image_name)
print('image_path', image_path)
print('save_output_img_path', save_output_img_path)
print('edit_instruction', edit_instruction)
edit_image(image_path, save_output_img_path, edit_instruction)
if __name__ == "__main__":
main()