Spaces:
Sleeping
Sleeping
File size: 8,608 Bytes
ef677f1 9135a28 ef677f1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import random
import numpy as np
from sklearn.linear_model import Ridge
from sklearn.linear_model import LogisticRegression
def to_torch(x):
return torch.from_numpy(x).float()
def to_cuda(x, use_cuda):
if use_cuda:
try:
return x.cuda()
except (RuntimeError, AssertionError) as e:
print(f"Warning: CUDA error: {e}. Falling back to CPU.")
return x
else:
return x
def to_numpy(x):
return x.detach().cpu().numpy()
class VAE(nn.Module):
def __init__(self, input_dim, latent_dim, demo_dim, use_cuda=True):
super(VAE, self).__init__()
self.input_dim = input_dim
self.latent_dim = latent_dim
self.demo_dim = demo_dim
self.use_cuda = use_cuda
self.enc1 = to_cuda(nn.Linear(input_dim, 1000).float(), use_cuda)
self.enc2 = to_cuda(nn.Linear(1000, latent_dim).float(), use_cuda)
self.dec1 = to_cuda(nn.Linear(latent_dim+demo_dim, 1000).float(), use_cuda)
self.dec2 = to_cuda(nn.Linear(1000, input_dim).float(), use_cuda)
def enc(self, x):
x = F.relu(self.enc1(x))
z = self.enc2(x)
return z
def gen(self, n):
return to_cuda(torch.randn(n, self.latent_dim).float(), self.use_cuda)
def dec(self, z, demo):
z = to_cuda(torch.cat([z, demo], dim=1), self.use_cuda)
x = F.relu(self.dec1(z))
x = self.dec2(x)
#x = x.reshape(len(z), 264, 5)
#x = torch.einsum('nac,nbc->nab', x, x)
#a,b = np.triu_indices(264, 1)
#x = x[:,a,b]
return x
def rmse(a, b, mean=torch.mean):
return mean((a-b)**2)**0.5
def latent_loss(z, use_cuda=True):
C = z.T@z
mu = torch.mean(z, dim=0)
tgt1 = to_cuda(torch.eye(z.shape[-1]).float(), use_cuda)*len(z)
tgt2 = to_cuda(torch.zeros(z.shape[-1]).float(), use_cuda)
loss_C = rmse(C, tgt1)
loss_mu = rmse(mu, tgt2)
return loss_C, loss_mu, C, mu
def decor_loss(z, demo, use_cuda=True):
ps = []
losses = []
for di in range(demo.shape[1]):
d = demo[:,di]
d = d - torch.mean(d)
p = torch.einsum('n,nz->z', d, z)
p = p/torch.std(d)
p = p/torch.einsum('nz,nz->z', z, z)
tgt = to_cuda(torch.zeros(z.shape[-1]).float(), use_cuda)
loss = rmse(p, tgt)
losses.append(loss)
ps.append(p)
losses = torch.stack(losses)
return losses, ps
def pretty(x):
return f'{round(float(x), 4)}'
def demo_to_torch(demo, demo_types, pred_stats, use_cuda):
demo_t = []
demo_idx = 0
for d,t,s in zip(demo, demo_types, pred_stats):
if t == 'continuous':
demo_t.append(to_cuda(to_torch(d), use_cuda))
elif t == 'categorical':
for dd in d:
if dd not in s:
print(f'Model not trained with value {dd} for categorical demographic {demo_idx}')
raise Exception('Bad demographic')
for ss in s:
idx = (d == ss).astype('bool')
zeros = torch.zeros(len(d))
zeros[idx] = 1
demo_t.append(to_cuda(zeros, use_cuda))
demo_idx += 1
demo_t = torch.stack(demo_t).permute(1,0)
return demo_t
def train_vae(vae, x, demo, demo_types, nepochs, pperiod, bsize, loss_C_mult, loss_mu_mult, loss_rec_mult, loss_decor_mult, loss_pred_mult, lr, weight_decay, alpha, LR_C, ret_obj):
# Get linear predictors for demographics
pred_w = []
pred_i = []
# Pred stats are mean and std for continuous, and a list of all values for categorical
pred_stats = []
for i,d,t in zip(range(len(demo)), demo, demo_types):
print(f'Fitting auxilliary guidance model for demographic {i} {t}...', end='')
if t == 'continuous':
pred_stats.append([np.mean(d), np.std(d)])
reg = Ridge(alpha=alpha).fit(x, d)
reg_w = to_cuda(to_torch(reg.coef_), vae.use_cuda)
reg_i = reg.intercept_
pred_w.append(reg_w)
pred_i.append(reg_i)
elif t == 'categorical':
pred_stats.append(sorted(list(set(list(d)))))
reg = LogisticRegression(C=LR_C).fit(x, d)
# Binary
if len(reg.coef_) == 1:
reg_w = to_cuda(to_torch(reg.coef_[0]), vae.use_cuda)
reg_i = reg.intercept_[0]
pred_w.append(-reg_w)
pred_i.append(-reg_i)
pred_w.append(reg_w)
pred_i.append(reg_i)
# Categorical
else:
for i in range(len(reg.coef_)):
reg_w = to_cuda(to_torch(reg.coef_[i]), vae.use_cuda)
reg_i = reg.intercept_[i]
pred_w.append(reg_w)
pred_i.append(reg_i)
else:
print(f'demographic type "{t}" not "continuous" or "categorical"')
raise Exception('Bad demographic type')
print(' done')
ret_obj.pred_stats = pred_stats
# Convert input to pytorch
print('Converting input to pytorch')
x = to_cuda(to_torch(x), vae.use_cuda)
# Convert demographics to pytorch
print('Converting demographics to pytorch')
demo_t = demo_to_torch(demo, demo_types, pred_stats, vae.use_cuda)
# Training loop
print('Beginning VAE training')
ce = nn.CrossEntropyLoss()
optim = torch.optim.Adam(vae.parameters(), lr=lr, weight_decay=weight_decay)
for e in range(nepochs):
for bs in range(0,len(x),bsize):
xb = x[bs:(bs+bsize)]
db = demo_t[bs:(bs+bsize)]
optim.zero_grad()
# Reconstruct
z = vae.enc(xb)
y = vae.dec(z, db)
loss_C, loss_mu, _, _ = latent_loss(z, vae.use_cuda)
loss_decor, _ = decor_loss(z, db, vae.use_cuda)
loss_decor = sum(loss_decor)
loss_rec = rmse(xb, y)
# Sample demographics
demo_gen = []
for s,t in zip(pred_stats, demo_types):
if t == 'continuous':
mu = s[0]
std = s[1]
dd = torch.randn(100).float()
dd = dd*std+mu
dd = to_cuda(dd, vae.use_cuda)
demo_gen.append(dd)
elif t == 'categorical':
idx = random.randint(0, len(s)-1)
for i in range(len(s)):
if idx == i:
dd = torch.ones(100).float()
else:
dd = torch.zeros(100).float()
dd = to_cuda(dd, vae.use_cuda)
demo_gen.append(dd)
demo_gen = torch.stack(demo_gen).permute(1,0)
# Generate
z = vae.gen(100)
y = vae.dec(z, demo_gen)
# Regressor/classifier guidance loss
losses_pred = []
idcs = []
dg_idx = 0
for s,t in zip(pred_stats, demo_types):
if t == 'continuous':
yy = y@pred_w[dg_idx]+pred_i[dg_idx]
loss = rmse(demo_gen[:,dg_idx], yy)
losses_pred.append(loss)
idcs.append(float(demo_gen[0,dg_idx]))
dg_idx += 1
elif t == 'categorical':
loss = 0
for i in range(len(s)):
yy = y@pred_w[dg_idx]+pred_i[dg_idx]
loss += ce(torch.stack([-yy, yy], dim=1), demo_gen[:,dg_idx].long())
idcs.append(int(demo_gen[0,dg_idx]))
dg_idx += 1
losses_pred.append(loss)
total_loss = loss_C_mult*loss_C + loss_mu_mult*loss_mu + loss_rec_mult*loss_rec + loss_decor_mult*loss_decor + loss_pred_mult*sum(losses_pred)
total_loss.backward()
optim.step()
if e%pperiod == 0 or e == nepochs-1:
print(f'Epoch {e} ', end='')
print(f'ReconLoss {pretty(loss_rec)} ', end='')
print(f'CovarianceLoss {pretty(loss_C)} ', end='')
print(f'MeanLoss {pretty(loss_mu)} ', end='')
print(f'DecorLoss {pretty(loss_decor)} ', end='')
losses_pred = [pretty(loss) for loss in losses_pred]
print(f'GuidanceTargets {idcs} GuidanceLosses {losses_pred} ', end='')
print()
print('Training complete.')
|