-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathlatent_distances.py
146 lines (122 loc) · 5.3 KB
/
latent_distances.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
import matplotlib
import matplotlib.pyplot as plt
import os
import utilities
from perturbation_learning import cvae, perturbations, datasets
import torch
import torch.nn.functional as F
import torchvision.utils
from torchvision.utils import save_image
import argparse
parser = argparse.ArgumentParser(
description='Latent space calculator',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('name', type=str)
args = parser.parse_args()
name = args.name
print(f"loading model and dataloader for {name}")
config_dict = utilities.get_config(f'configs/{name}.json')
config = utilities.config_to_namedtuple(config_dict)
train_loader, test_loader, val_loader = datasets.loaders[config.dataset.type](config)
model = cvae.models[config.model.type](config)
model.cuda()
h = perturbations.hs[config.perturbation.test_type](config.perturbation)
model.load_state_dict(torch.load(f'experiments/{name}/checkpoints/checkpoint_latest.pth')['model_state_dict'])
model.eval();
if False:
print("plotting pictures")
with torch.no_grad():
for batch_idx, batch in enumerate(test_loader):
data = batch[0]
hdata = h(batch)
data = data.to(config.device)
hdata = hdata.to(config.device)
output = model(data, hdata)
hsample = model.sample(data)
n = min(data.size(0), 8)
recon_hbatch = output[0]
if 'cifar10' in name:
pass
else:
data, hdata, recon_hbatch = [F.interpolate(t, size=(125,187), mode="bilinear") for t in [data,hdata,recon_hbatch]]
hcomparison = torch.cat([
data[:n],
hdata[:n],
recon_hbatch.view(*hdata.size())[:n]])
save_image(hcomparison.cpu(),
os.path.join('figures', f'hreconstruction.png'), nrow=n)
save_image(hsample[:min(64,config.eval.batch_size)],
os.path.join('figures', f'hsample.png'))
repeat_hsample = torch.cat([model.sample(data)[:8].unsqueeze(1) for i in range(8)],dim=1)
repeat_hsample = repeat_hsample.view(-1,*hdata.size()[1:])
save_image(repeat_hsample[:min(64,config.eval.batch_size)],
os.path.join('figures', f'repeat_hsample.png'))
break
if False:
print("making interpolation")
for batch in test_loader:
data = batch[0]
break
hdata = h(batch)
data = data.to(config.device)
hdata = hdata.to(config.device)
output = model(data, hdata)
recon_hbatch = output[0]
n = 4
nperturb = 12
l = [(data[i*nperturb:(i+1)*nperturb],
hdata[i*nperturb:(i+1)*nperturb],
recon_hbatch.view(*hdata.size())[i*nperturb:(i+1)*nperturb]) for i in range(n)]
hcomparison = torch.cat([inner for outer in l for inner in outer])
z = model.recognition(data, hdata)[0][[0,4,6,10]]
nrow = 10
latent_dim = config.model.latent_dim
if isinstance(latent_dim, list):
latent_dim = sum(latent_dim)
def interpolate(z, n=6):
z1, z2, z3, z4 = z
z12 = torch.cat([(t*z1 + (1-t)*z2).unsqueeze(0) for t in torch.linspace(0,1,n)])
z34 = torch.cat([(t*z3 + (1-t)*z4).unsqueeze(0) for t in torch.linspace(0,1,n)])
z1234 = torch.cat([(t*z12 + (1-t)*z34).unsqueeze(0) for t in torch.linspace(0,1,n)])
return z1234
zbatch = interpolate(z, n=nrow).view(-1,latent_dim)
Xbatch = data[:1].repeat(nrow*nrow,1,1,1)
img_interpolation = model.generator(Xbatch, zbatch)
fig, ax = plt.subplots(figsize=(12,12))
ax.imshow(torchvision.utils.make_grid(img_interpolation, nrow=nrow).cpu().detach().permute(1,2,0).numpy())
fig.savefig(f'figures/latent_interpolation_{name}.png')
print("loading all dataset loader")
config_dict = utilities.get_config(f'configs/{name}.json')
if 'cifar10' in name:
config_dict['dataset']['type'] = 'cifar10c_all'
elif 'mi_unet' in name:
config_dict['dataset']['mode'] = 'all'
config_dict['eval']['batch_size']=32
config = utilities.config_to_namedtuple(config_dict)
train_loader, test_loader, val_loader = datasets.loaders[config.dataset.type](config)
loops = 5 if 'mnist' in name else 0
print("making distribution")
norms = []
with torch.no_grad():
for _ in range(loops):
for batch in val_loader:
data = batch[0]
hdata = h(batch)
data = data.to(config.device)
hdata = hdata.to(config.device)
mu_prior, logvar_prior = model.prior(data)
mu_recog, logvar_recog = model.recognition(data,hdata)
std = torch.exp(0.5*logvar_prior)
norms.append(((mu_recog - mu_prior)/std).view(data.size(0),-1).norm(p=2,dim=1))
import numpy as np
import scipy.stats as stats
norms_np = torch.cat(norms).cpu().numpy()
print(f'Max {norms_np.max()} Mean {norms_np.mean()} Std {norms_np.std()}')
p = np.percentile(norms_np, [25,50,75,99.9])
print(f'25% {p[0]} 50% {p[1]} 75% {p[2]} 99% {p[3]}')
x = np.linspace(0,25,100)
density = stats.gaussian_kde(norms_np)
fig, ax = plt.subplots(figsize=(4,4))
ax.plot(x, density(x))
fig.savefig(f'figures/latent_distribution_{name}.png')
np.save(f"data/latent_norms_{name}.npy", norms_np)