forked from sportsracer48/pytti
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ImageGuide.py
166 lines (144 loc) · 6.32 KB
/
ImageGuide.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
from torch import optim, nn
from pytti.Notebook import tqdm
from pytti import *
import pandas as pd
import math
from labellines import labelLines
from scipy.signal import savgol_filter
def unpack_dict(D, n=2):
ds = [{k: V[i] for k, V in D.items()} for i in range(n)]
return tuple(ds)
def smooth_dataframe(df, window_size):
"""Applies a moving average filter to the columns of df."""
smoothed_df = pd.DataFrame(index=df.index, columns=df.columns)
for key in df.columns:
smoothed_df[key] = savgol_filter(df[key], window_size, 2, mode='nearest')
return smoothed_df
class DirectImageGuide():
"""
Image guide that uses an optimizer and torch autograd to optimize an image representation.
Based on the BigGan+CLIP algorithm by advadnoun (https://twitter.com/advadnoun).
"""
def __init__(self, image_rep, embedder, optimizer=None, lr=None, **optimizer_params):
self.image_rep = image_rep
self.embedder = embedder
if lr is None:
lr = image_rep.lr
optimizer_params['lr'] = lr
self.optimizer_params = optimizer_params
if optimizer is None:
self.optimizer = optim.Adam(image_rep.parameters(), **optimizer_params)
else:
self.optimizer = optimizer
self.dataframe = []
def run_steps(self, n_steps, prompts, interp_prompts, loss_augs, stop=-math.inf, interp_steps=0, i_offset=0, skipped_steps=0):
"""Runs the optimizer."""
for i in tqdm(range(n_steps)):
self.update(i + i_offset, i + skipped_steps)
losses = self.train(i + skipped_steps, prompts, interp_prompts, loss_augs, interp_steps=interp_steps)
if losses['TOTAL'] <= stop:
break
return i + 1
def set_optim(self, opt=None):
if opt is not None:
self.optimizer = opt
else:
self.optimizer = optim.Adam(self.image_rep.parameters(), **self.optimizer_params)
def clear_dataframe(self):
self.dataframe = []
def plot_losses(self, axs):
def plot_dataframe(df, ax, legend=False):
keys = df.columns.tolist()
keys.sort(reverse=True, key=lambda k: df[k].iloc[-1])
ax.clear()
df[keys].plot(ax=ax, legend=legend)
if legend:
ax.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
ax.tick_params(labelbottom=True, labeltop=False, labelleft=True, labelright=False,
bottom=True, top=False, left=True, right=False)
labelLines(ax.get_lines(), align=False)
return
dfs = self.dataframe[:]
if dfs:
dfs[0] = smooth_dataframe(dfs[0], 17)
for i, (df, ax) in enumerate(zip(dfs, axs)):
if len(df.index) < 2:
return False
if not df.empty:
plot_dataframe(df, ax, legend=(i == 0))
ax.set_ylabel('Loss')
ax.set_xlabel('Step')
return True
def update(self, i, stage_i):
"""Update hook called every step."""
pass
def train(self, i, prompts, interp_prompts, loss_augs, interp_steps=0, save_loss=True):
"""Performs a training step."""
self.optimizer.zero_grad()
z = self.image_rep.decode_training_tensor()
# Precompute formatted inputs to avoid redundant computations
if self.embedder is not None:
image_embeds, offsets, sizes = self.embedder(self.image_rep, input=z)
# Cache formatted inputs for prompts and interpolation prompts
all_prompts = prompts + interp_prompts
formatted_inputs = {}
for prompt in set(all_prompts):
formatted_inputs[prompt] = {
'embeds': format_input(image_embeds, self.embedder, prompt),
'offsets': format_input(offsets, self.embedder, prompt),
'sizes': format_input(sizes, self.embedder, prompt)
}
else:
formatted_inputs = {}
# Cache formatted inputs for loss augmentations
formatted_z = {}
for aug in loss_augs:
formatted_z[aug] = format_input(z, self.image_rep, aug)
# Compute interpolation factor
if i < interp_steps:
t = i / interp_steps
interp_losses = [prompt(formatted_inputs[prompt]['embeds'],
formatted_inputs[prompt]['offsets'],
formatted_inputs[prompt]['sizes'])[0] * (1 - t) for prompt in interp_prompts]
else:
t = 1
interp_losses = [0]
# Compute prompt losses
prompt_losses = {}
for prompt in prompts:
loss = prompt(formatted_inputs[prompt]['embeds'],
formatted_inputs[prompt]['offsets'],
formatted_inputs[prompt]['sizes'])
# Scale loss by interpolation factor
loss[0].mul_(t)
prompt_losses[prompt] = loss
# Compute augmentation losses
aug_losses = {}
for aug in loss_augs:
loss = aug(formatted_z[aug], self.image_rep)
aug_losses[aug] = loss
# Compute image losses
image_augs = self.image_rep.image_loss()
image_losses = {}
for aug in image_augs:
loss = aug(self.image_rep)
image_losses[aug] = loss
# Aggregate losses
total_loss = sum(loss[0] for loss in prompt_losses.values()) + \
sum(loss[0] for loss in aug_losses.values()) + \
sum(loss[0] for loss in image_losses.values()) + \
sum(interp_losses)
# Prepare loss tracking
if save_loss:
loss_dict = {'TOTAL': float(total_loss)}
loss_dict.update({str(k): float(v[0]) for k, v in prompt_losses.items()})
loss_dict.update({str(k): float(v[0]) for k, v in aug_losses.items()})
loss_dict.update({str(k): float(v[0]) for k, v in image_losses.items()})
if not self.dataframe:
self.dataframe = [pd.DataFrame(loss_dict, index=[i])]
else:
self.dataframe[0] = pd.concat([self.dataframe[0], pd.DataFrame(loss_dict, index=[i])])
total_loss.backward()
self.optimizer.step()
self.image_rep.update()
return {'TOTAL': float(total_loss)}