forked from stephenportillo/SDSS-VAE
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainVAE.py
167 lines (146 loc) · 6.33 KB
/
trainVAE.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
import numpy as np
from astroML.datasets import sdss_corrected_spectra
import torch.utils.data
from torch import optim
from InfoVAE import InfoVAE
import os
tag = '64k_2par_20190703'
os.makedirs(tag)
data = np.load('spec64k.npz')
rawspec = sdss_corrected_spectra.reconstruct_spectra(data)
lam = sdss_corrected_spectra.compute_wavelengths(data)
spec = rawspec / data['norms'][:,None] # normalize spectra with PCA norm
spec_err = data['spec_err'] / data['norms'][:,None]
lowspecerr = (data['spec_err'] < 0.25) # some errors ~ 0 or are negative?
spec_weig = 1./(spec_err*spec_err + 1./(2e6)) # soften largest weights
meanspec = np.mean(spec, axis=0)
mask = data['mask']
spec_weig[mask] = 0 # if spectrum masked, act as if err = inf
spec_weig[lowspecerr] = 0
print('median weight %0.2e' % (np.median(spec_weig[spec_weig > 0])))
print('maximum weight %0.2e' % (np.amax(spec_weig)))
# only train and validate on spectroscopically confirmed galaxies/QSOs
spec_cln = data['lineindex_cln']
specmask = (spec_cln > 1) & (spec_cln < 7)
# deterministic outlier inclusion in training set
np.random.seed(20190605)
outlierfrac = 0.000
noutlier = min(int(spec.shape[0] * outlierfrac), np.sum(specmask == False))
outlieridx = np.random.choice(np.where(specmask == False)[0], size=noutlier, replace=False)
specmasktrain = (spec_cln > 1) & (spec_cln < 7)
specmasktrain[outlieridx] = True
# deterministic class exclusion in training set
excludefrac = 0.000
excludeclas = 4
nexclude = int(np.sum(spec_cln == excludeclas) * excludefrac)
excludeidx = np.random.choice(np.where(spec_cln == excludeclas)[0], size=nexclude, replace=False)
specmasktrain[excludeidx] = False
trainfrac = 0.75
ntrain = int(spec.shape[0] * trainfrac)
nvalid = spec.shape[0] - ntrain
nfeat = spec.shape[1]
np.random.seed(20190425) # make validation set deterministic
permutation = np.random.permutation(spec.shape[0])
np.random.seed()
trainidx = permutation[0:ntrain]
trainidx = trainidx[np.where(specmasktrain[trainidx])[0]]
valididx = permutation[-1-nvalid:-1]
valididx = valididx[np.where(specmask[valididx])[0]]
trainspec = spec[trainidx,:]
validspec = spec[valididx,:]
trainmask = mask[trainidx,:].astype(np.float32)
validmask = mask[valididx,:].astype(np.float32)
trainspecweig = spec_weig[trainidx,:]
validspecweig = spec_weig[valididx,:]
np.savez(tag+'/datasplit.npz', trainidx=trainidx, valididx=valididx)
torch.set_num_threads(4)
batch_size=64
dataloader = torch.utils.data.DataLoader(
torch.utils.data.TensorDataset(torch.tensor(trainspec-meanspec, dtype=torch.float32), torch.tensor(trainspecweig, dtype=torch.float32)),
batch_size=batch_size,
shuffle=True)
valdloader = torch.utils.data.DataLoader(
torch.utils.data.TensorDataset(torch.tensor(validspec-meanspec, dtype=torch.float32), torch.tensor(validspecweig, dtype=torch.float32)),
batch_size=nvalid)
def train(): #model, optimizer, epoch, min_valid_loss, badepochs
model.train()
train_loss = 0
train_logL = 0
for batch_idx, data in enumerate(dataloader):
spectrum, weig = data
optimizer.zero_grad()
loss, logL, KLD, MMD = model.loss(spectrum, weig, epoch)
loss.backward()
train_loss += loss.item()
train_logL += logL.item()
optimizer.step()
train_loss /= len(dataloader.dataset)
with torch.no_grad():
model.eval()
valid_loss = 0
valid_logL = 0
valid_KLD = 0
valid_MMD = 0
for valddata in valdloader:
spectrum, weig = valddata
loss, logL, KLD, MMD = model.loss(spectrum, weig, epoch)
valid_loss += loss.item()
valid_logL += logL.item()
valid_KLD += KLD.item()
valid_MMD += MMD.item()
valid_loss /= len(valdloader.dataset)
valid_logL /= -len(valdloader.dataset)
valid_KLD /= len(valdloader.dataset)
valid_MMD /= len(valdloader.dataset)
return valid_loss, valid_logL, valid_KLD, valid_MMD
class EarlyStopper:
def __init__(self, precision=1e-3, patience=10):
self.precision = precision
self.patience = patience
self.badepochs = 0
self.min_valid_loss = float('inf')
def step(self, valid_loss):
if valid_loss < self.min_valid_loss*(1-self.precision):
self.badepochs = 0
self.min_valid_loss = valid_loss
else:
self.badepochs += 1
return not (self.badepochs == self.patience)
epochs = 200
log_interval = 10
mdl_ncode = 2
n_config = 100
mdl_MSE = np.zeros(n_config)
mdl_KLD = np.zeros(n_config)
mdl_MMD = np.zeros(n_config)
for config in range(n_config):
alpha = 0
lambd = np.exp(np.random.uniform(0, np.log(1e5)))
dropout = 0#0.9*np.random.uniform()
dfac = 1./(1.-dropout)
nhidden = int(np.ceil(np.exp(np.random.uniform(np.log(dfac*mdl_ncode+1), np.log(dfac*2*nfeat)))))
nhidden2 = int(np.ceil(np.exp(np.random.uniform(np.log(dfac*mdl_ncode+1), np.log(nhidden)))))
print('config %i, alpha = %0.1f, lambda = %0.1f, dropout = %0.2f; 2 hidden layers with %i, %i nodes' % (config, alpha, lambd, dropout, nhidden, nhidden2))
model = InfoVAE(alpha=alpha, lambd=lambd, nfeat=nfeat, nhidden=nhidden, nhidden2=nhidden2, ncode=mdl_ncode, dropout=dropout)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', verbose=True, patience=5)
stopper = EarlyStopper(patience=10)
for epoch in range(1, epochs + 1):
valid_loss, valid_logL, valid_KLD, valid_MMD = train()
if epoch % log_interval == 0:
print('====> Epoch: {} VALIDATION Loss: {:.2e} logL: {:.2e} KL: {:.2e} MMD: {:.2e}'.format(
epoch, valid_loss, valid_logL, valid_KLD, valid_MMD))
scheduler.step(valid_loss)
if (not stopper.step(valid_loss)) or (epoch == epochs):
print('Stopping')
print('====> Epoch: {} VALIDATION Loss: {:.2e} logL: {:.2e} KL: {:.2e} MMD: {:.2e}'.format(
epoch, valid_loss, valid_logL, valid_KLD, valid_MMD))
model.MSE = -valid_logL
model.KLD = valid_KLD
model.MMD = valid_MMD
mdl_MSE[config] = model.MSE
mdl_KLD[config] = model.KLD
mdl_MMD[config] = model.MMD
torch.save(model, tag+'/%04i.pth' % config)
break
np.savez(tag+'/metrics.npz', MSE=mdl_MSE, KLD=mdl_KLD, MMD=mdl_MMD)