-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathcifar10_combine.py
167 lines (126 loc) · 5.77 KB
/
cifar10_combine.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
import tensorflow as tf
import math
from tqdm import tqdm
from tensorflow.python.platform import flags
from torch.utils.data import DataLoader, Dataset
from models import ResNetModel, CelebAModel
from utils import ReplayBuffer, GaussianBlur
import os.path as osp
import numpy as np
from logger import TensorBoardOutputFormat
from scipy.misc import imsave
from torchvision import transforms
import os
from itertools import product
from PIL import Image
import torch
flags.DEFINE_integer('batch_size', 256, 'Size of inputs')
flags.DEFINE_integer('data_workers', 4, 'Number of workers to do things')
flags.DEFINE_string('logdir', 'cachedir', 'directory for logging')
flags.DEFINE_string('savedir', 'cachedir', 'location where log of experiments will be stored')
flags.DEFINE_integer('num_filters', 64, 'number of filters for conv nets -- 32 for miniimagenet, 64 for omniglot.')
flags.DEFINE_float('step_lr', 10.0, 'size of gradient descent size')
flags.DEFINE_bool('cclass', True, 'not cclass')
flags.DEFINE_bool('proj_cclass', False, 'use for backwards compatibility reasons')
flags.DEFINE_bool('spec_norm', True, 'Whether to use spectral normalization on weights')
flags.DEFINE_bool('use_bias', True, 'Whether to use bias in convolution')
flags.DEFINE_bool('use_attention', False, 'Whether to use self attention in network')
flags.DEFINE_integer('num_steps', 200, 'number of steps to optimize the label')
flags.DEFINE_string('task', 'negation_figure', 'conceptcombine, combination_figure, negation_figure, or_figure, negation_eval')
flags.DEFINE_bool('eval', False, 'Whether to quantitively evaluate models')
flags.DEFINE_bool('latent_energy', False, 'latent energy in model')
flags.DEFINE_bool('proj_latent', False, 'Projection of latents')
# Whether to train for gentest
flags.DEFINE_bool('train', False, 'whether to train on generalization into multiple different predictions')
FLAGS = flags.FLAGS
def conceptcombine(model_list, select_idx):
n = 64
labels = []
for six in select_idx:
label_ix = np.eye(10)[six]
label_batch = np.tile(label_ix[None, :], (n, 1))
label = torch.Tensor(label_batch).cuda()
labels.append(label)
im = torch.rand(n, 3, 32, 32).cuda()
im_noise = torch.randn_like(im).detach()
def get_color_distortion(s=1.0):
# s is the strength of color distortion.
color_jitter = transforms.ColorJitter(0.8*s, 0.8*s, 0.8*s, 0.4*s)
rnd_color_jitter = transforms.RandomApply([color_jitter], p=0.8)
rnd_gray = transforms.RandomGrayscale(p=0.2)
color_distort = transforms.Compose([
rnd_color_jitter,
rnd_gray])
return color_distort
color_transform = get_color_distortion()
im_size = 32
transform = transforms.Compose([transforms.RandomResizedCrop(im_size, scale=(0.02, 1.0)), transforms.RandomHorizontalFlip(), color_transform, GaussianBlur(kernel_size=5), transforms.ToTensor()])
# First get good initializations for sampling
for i in range(10):
for i in range(20):
im_noise.normal_()
im = im + 0.001 * im_noise
# im.requires_grad = True
im.requires_grad_(requires_grad=True)
energy = 0
for model, label in zip(model_list, labels):
energy = model.forward(im, label) + energy
# print("step: ", i, energy.mean())
im_grad = torch.autograd.grad([energy.sum()], [im])[0]
im = im - FLAGS.step_lr * im_grad
im = im.detach()
im = torch.clamp(im, 0, 1)
im = im.detach().cpu().numpy().transpose((0, 2, 3, 1))
im = (im * 255).astype(np.uint8)
ims = []
for i in range(im.shape[0]):
im_i = np.array(transform(Image.fromarray(np.array(im[i]))))
ims.append(im_i)
im = torch.Tensor(np.array(ims)).cuda()
# Then refine the images
for i in range(FLAGS.num_steps):
im_noise.normal_()
im = im + 0.001 * im_noise
# im.requires_grad = True
im.requires_grad_(requires_grad=True)
energy = 0
for model, label in zip(model_list, labels):
energy = model.forward(im, label) + energy
print("step: ", i, energy.mean())
im_grad = torch.autograd.grad([energy.sum()], [im])[0]
im = im - FLAGS.step_lr * im_grad
im = im.detach()
im = torch.clamp(im, 0, 1)
output = im.detach().cpu().numpy()
output = output.transpose((0, 2, 3, 1))
output = output.reshape((-1, 8, 32, 32, 3)).transpose((0, 2, 1, 3, 4)).reshape((-1, 32 * 8, 3))
imsave("debug.png", output)
def combine_main(models, resume_iters, select_idx):
model_list = []
for model, resume_iter in zip(models, resume_iters):
model_path = osp.join("cachedir", model, "model_{}.pth".format(resume_iter))
checkpoint = torch.load(model_path)
FLAGS_model = checkpoint['FLAGS']
model_base = ResNetModel(FLAGS_model)
model_base.load_state_dict(checkpoint['ema_model_state_dict_0'])
model_base = model_base.cuda()
model_list.append(model_base)
conceptcombine(model_list, select_idx)
if __name__ == "__main__":
models_orig = ['cifar10_cond_807', 'cifar10_cond_807']
resume_iters_orig = ["15000", "15000"]
models = []
resume_iters = []
select_idx = []
##################################
# Settings for the composition_figure
models = models + [models_orig[1]]
resume_iters = resume_iters + [resume_iters_orig[1]]
select_idx = select_idx + [6]
models = models + [models_orig[0]]
resume_iters = resume_iters + [resume_iters_orig[0]]
select_idx = select_idx + [9]
FLAGS.step_lr = FLAGS.step_lr / len(models)
# List of 4 attributes that might be good
# Young -> Female -> Smiling -> Wavy
combine_main(models, resume_iters, select_idx)