forked from yael-vinker/CLIPasso
-
Notifications
You must be signed in to change notification settings - Fork 0
/
painterly_rendering.py
182 lines (157 loc) · 6.69 KB
/
painterly_rendering.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
import warnings
warnings.filterwarnings('ignore')
warnings.simplefilter('ignore')
import argparse
import math
import os
import sys
import time
import traceback
import numpy as np
import PIL
import torch
import torch.nn as nn
import torch.nn.functional as F
import wandb
from PIL import Image
from torchvision import models, transforms
from tqdm import tqdm
import config
import sketch_utils as utils
from models.loss import Loss
from models.painter_params import Painter, PainterOptimizer
def load_renderer(args, target_im=None, mask=None):
renderer = Painter(num_strokes=args.num_paths, args=args,
num_segments=args.num_segments,
imsize=args.image_scale,
device=args.device,
target_im=target_im,
mask=mask)
renderer = renderer.to(args.device)
return renderer
def get_target(args):
target = Image.open(args.target)
if target.mode == "RGBA":
# Create a white rgba background
new_image = Image.new("RGBA", target.size, "WHITE")
# Paste the image on the background.
new_image.paste(target, (0, 0), target)
target = new_image
target = target.convert("RGB")
masked_im, mask = utils.get_mask_u2net(args, target)
if args.mask_object:
target = masked_im
if args.fix_scale:
target = utils.fix_image_scale(target)
transforms_ = []
if target.size[0] != target.size[1]:
transforms_.append(transforms.Resize(
(args.image_scale, args.image_scale), interpolation=PIL.Image.BICUBIC))
else:
transforms_.append(transforms.Resize(
args.image_scale, interpolation=PIL.Image.BICUBIC))
transforms_.append(transforms.CenterCrop(args.image_scale))
transforms_.append(transforms.ToTensor())
data_transforms = transforms.Compose(transforms_)
target_ = data_transforms(target).unsqueeze(0).to(args.device)
return target_, mask
def main(args):
loss_func = Loss(args)
inputs, mask = get_target(args)
utils.log_input(args.use_wandb, 0, inputs, args.output_dir)
renderer = load_renderer(args, inputs, mask)
optimizer = PainterOptimizer(args, renderer)
counter = 0
configs_to_save = {"loss_eval": []}
best_loss, best_fc_loss = 100, 100
best_iter, best_iter_fc = 0, 0
min_delta = 1e-5
terminate = False
renderer.set_random_noise(0)
img = renderer.init_image(stage=0)
optimizer.init_optimizers()
for epoch in tqdm(range(args.num_iter)):
renderer.set_random_noise(epoch)
if args.lr_scheduler:
optimizer.update_lr(counter)
start = time.time()
optimizer.zero_grad_()
sketches = renderer.get_image().to(args.device)
losses_dict = loss_func(sketches, inputs.detach(
), renderer.get_color_parameters(), renderer, counter, optimizer)
loss = sum(list(losses_dict.values()))
loss.backward()
optimizer.step_()
if epoch % args.save_interval == 0:
utils.plot_batch(inputs, sketches, f"{args.output_dir}/jpg_logs", counter,
use_wandb=args.use_wandb, title=f"iter{epoch}.jpg")
renderer.save_svg(
f"{args.output_dir}/svg_logs", f"svg_iter{epoch}")
if epoch % args.eval_interval == 0:
with torch.no_grad():
losses_dict_eval = loss_func(sketches, inputs, renderer.get_color_parameters(
), renderer.get_points_parans(), counter, optimizer, mode="eval")
loss_eval = sum(list(losses_dict_eval.values()))
configs_to_save["loss_eval"].append(loss_eval.item())
for k in losses_dict_eval.keys():
if k not in configs_to_save.keys():
configs_to_save[k] = []
configs_to_save[k].append(losses_dict_eval[k].item())
if args.clip_fc_loss_weight:
if losses_dict_eval["fc"].item() < best_fc_loss:
best_fc_loss = losses_dict_eval["fc"].item(
) / args.clip_fc_loss_weight
best_iter_fc = epoch
# print(
# f"eval iter[{epoch}/{args.num_iter}] loss[{loss.item()}] time[{time.time() - start}]")
cur_delta = loss_eval.item() - best_loss
if abs(cur_delta) > min_delta:
if cur_delta < 0:
best_loss = loss_eval.item()
best_iter = epoch
terminate = False
utils.plot_batch(
inputs, sketches, args.output_dir, counter, use_wandb=args.use_wandb, title="best_iter.jpg")
renderer.save_svg(args.output_dir, "best_iter")
if args.use_wandb:
wandb.run.summary["best_loss"] = best_loss
wandb.run.summary["best_loss_fc"] = best_fc_loss
wandb_dict = {"delta": cur_delta,
"loss_eval": loss_eval.item()}
for k in losses_dict_eval.keys():
wandb_dict[k + "_eval"] = losses_dict_eval[k].item()
wandb.log(wandb_dict, step=counter)
if abs(cur_delta) <= min_delta:
if terminate:
break
terminate = True
if counter == 0 and args.attention_init:
utils.plot_atten(renderer.get_attn(), renderer.get_thresh(), inputs, renderer.get_inds(),
args.use_wandb, "{}/{}.jpg".format(
args.output_dir, "attention_map"),
args.saliency_model, args.display_logs)
if args.use_wandb:
wandb_dict = {"loss": loss.item(), "lr": optimizer.get_lr()}
for k in losses_dict.keys():
wandb_dict[k] = losses_dict[k].item()
wandb.log(wandb_dict, step=counter)
counter += 1
renderer.save_svg(args.output_dir, "final_svg")
path_svg = os.path.join(args.output_dir, "best_iter.svg")
utils.log_sketch_summary_final(
path_svg, args.use_wandb, args.device, best_iter, best_loss, "best total")
return configs_to_save
if __name__ == "__main__":
args = config.parse_arguments()
final_config = vars(args)
try:
configs_to_save = main(args)
except BaseException as err:
print(f"Unexpected error occurred:\n {err}")
print(traceback.format_exc())
sys.exit(1)
for k in configs_to_save.keys():
final_config[k] = configs_to_save[k]
np.save(f"{args.output_dir}/config.npy", final_config)
if args.use_wandb:
wandb.finish()