def _train(self, net, tr, va):
restart = (
"rmse", # loss op name
0.5 * (((va['T'] - va['T'].mean())**2).mean()**0.5), # threshold
10 # epoch
)
lr = exp_lr(1e-3, 1e-4, 3000, floor=1e-4)
net.fit(tr, va, 25, 4000, lr=lr, silent=True, restart_check=restart)
return net
def test(estimator, data=None, pretrained=False):
# test("dohe", "data/benzene.npz")
preproc, (tr, va, te) = load_simple(
filename=data,
sizes=[50e3, 2e3, 10e3],
verbose=False
)
if re.match(r'^doh[oe]\d*$', estimator):
p100 = re.match(r'\D+(\d+)', estimator)
p = int(p100.group(1))/100 if p100 is not None else 0.8
print(estimator,'with p', p)
net = honet if 'doho' in estimator else henet
print(net, "(selected dropout variance estimator class)")
nn = net(preproc.scaling_params, tr[0], len(tr), mu_max=10, keep_prob=p, length_scale=20, n_predictions=50, tf_config=None, name=estimator)
elif estimator == 'density':
nn = densnet(preproc.scaling_params, tr[0], mu_max=10, tf_config=None, name=estimator)
elif estimator == 'density_full':
nn = fulldensnet(preproc.scaling_params, tr[0], mu_max=10, tf_config=None, name=estimator)
elif estimator == 'density_re':
nn = redensnet(preproc.scaling_params, tr[0], mu_max=10, tf_config=None, name=estimator)
else:
raise ValueError('estimator not defined: {}'.format(estimator))
if pretrained:
nn.load()
else:
# lr = cycle_lr(1e-3, 1e-5, 5, 200*len(tr)/25, decay="exp")
lr = exp_lr(1e-3, 1e-4, 3000, floor=1e-4)
# lr = 1e-3
tstart = time.time()
restart = ("rmse", 1.5, 100)
nn.fit(tr, va, 25, 4000, lr=lr, silent=True, stop_at={"rmse":0.05}, restart_check=restart)
# nn.fit(tr, va, 25, 3000, lr=lr, silent=True, stop_at={"rmse":0.05})
# nn.fit(tr, va, 250, 100, lr=lr, plot_loss='rmse', silent=False)
# nn.fit(tr, va, 25, 10/60, lr=lr, plot_loss='rmse', silent=False)
# nn.fit(tr, va, 25, 30/60, lr=lr, plot_loss='rmse', silent=False)
# nn.fit(tr, va, 25, 10, lr=lr, plot_loss='rmse', silent=False)
sec = time.time() - tstart
print("fit in {} hours".format(sec/60/60))
nn.save()
pr = nn.predict(va)
print(((va['E']-pr['E'])**2).mean()**0.5, "va rmse")
pr = nn.predict(te)
print(((te['E']-pr['E'])**2).mean()**0.5, "te rmse")
pp.hist(pr["U"], bins=100)
pp.title("U {}".format(estimator))
pp.savefig(os.path.join("figs", estimator + ".jpg"))
nn.close()
def _train(self, net, tr, va):
restart = (
"rmse", # loss op name
0.5 * (((va['T'] - va['T'].mean())**2).mean()**0.5), # threshold
30 # epoch
)
stop = {'rmse': 0.06}
lr = exp_lr(1e-3, 1e-4, 3000, floor=1e-4)
net.fit(tr,
va,
25,
12000,
lr=lr,
silent=True,
restart_check=restart,
stop_at=stop)
return net
def _train(self, net, tr, va):
lr = exp_lr(1e-3, 1e-5, 1000, floor=5e-5)
net.fit(tr, va, 200, 30, lr=lr, silent=True)
return net
def _train(self, net, tr, va):
lr = exp_lr(1e-3, 1e-4, 3000, floor=1e-4)
net.fit(tr, va, 25, 4000, lr=lr, silent=True)
return net
