def train(train_loader, train_loader1, model, criterion, optimizer, var_optimizer, epoch, args, log):
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
rk_losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
train_loader1_iter = iter(train_loader1)
model.train()
end = time.time()
for i, (input, target) in enumerate(train_loader):
data_time.update(time.time() - end)
input = input.cuda(args.gpu, non_blocking=True)
target = target.cuda(args.gpu, non_blocking=True)
input1 = next(train_loader1_iter)
input1 = input1.cuda(args.gpu, non_blocking=True)
bs = input.shape[0]
bs1 = input1.shape[0]
output = model(torch.cat([input, input1.repeat(2, 1, 1, 1)]))
loss = criterion(output[:bs], target)
out1_0 = output[bs:bs+bs1].softmax(-1)
out1_1 = output[bs+bs1:].softmax(-1)
mi1 = ent((out1_0 + out1_1)/2.) - (ent(out1_0) + ent(out1_1))/2.
rank_loss = torch.nn.functional.relu(args.mi_th - mi1).mean()
prec1, prec5 = accuracy(output[:bs], target, topk=(1, 5))
losses.update(loss.detach().item(), bs)
rk_losses.update(rank_loss.detach().item(), bs1)
top1.update(prec1.item(), bs)
top5.update(prec5.item(), bs)
optimizer.zero_grad()
var_optimizer.zero_grad()
(loss+rank_loss*args.alpha).backward()
optimizer.step()
var_optimizer.step()
batch_time.update(time.time() - end)
end = time.time()
if i == len(train_loader) - 1:
print_log(' Epoch: [{:03d}][{:03d}/{:03d}] '
'Time {batch_time.avg:.3f} '
'Data {data_time.avg:.3f} '
'Loss {loss.avg:.4f} '
'RK Loss {rk_loss.avg:.4f} '
'Prec@1 {top1.avg:.3f} '
'Prec@5 {top5.avg:.3f} '.format(
epoch, i, len(train_loader), batch_time=batch_time, rk_loss=rk_losses,
data_time=data_time, loss=losses, top1=top1, top5=top5) + time_string(), log)
return top1.avg, losses.avg
def _calc_pt_plugin(self, pt):
"""Calculate direct entropies and apply PT correction if required """
calc = self.calc
pt_corr = lambda R: (R - 1) / (2 * self.N * np.log(2))
self.H_plugin = {}
if pt: self.H_pt = {}
# compute basic entropies
if 'HX' in calc:
H = ent(self.PX)
self.H_plugin['HX'] = H
if pt:
self.H_pt['HX'] = H + pt_corr(pt_bayescount(self.PX, self.N))
if 'HY' in calc:
H = ent(self.PY)
self.H_plugin['HY'] = H
if pt:
self.H_pt['HY'] = H + pt_corr(pt_bayescount(self.PY, self.N))
if 'HXY' in calc:
H = (self.PY * ent(self.PXY)).sum()
self.H_plugin['HXY'] = H
if pt:
for y in xrange(self.Y_dim):
H += pt_corr(pt_bayescount(self.PXY[:, y], self.Ny[y]))
self.H_pt['HXY'] = H
if 'SiHXi' in calc:
H = ent(self.PXi).sum()
self.H_plugin['SiHXi'] = H
if pt:
for x in xrange(self.X_n):
H += pt_corr(pt_bayescount(self.PXi[:, x], self.N))
self.H_pt['SiHXi'] = H
if 'HiXY' in calc:
H = (self.PY * ent(self.PXiY)).sum()
self.H_plugin['HiXY'] = H
if pt:
for x in xrange(self.X_n):
for y in xrange(self.Y_dim):
H += pt_corr(
pt_bayescount(self.PXiY[:, x, y], self.Ny[y]))
self.H_pt['HiXY'] = H
if 'HiX' in calc:
H = ent(self.PiX)
self.H_plugin['HiX'] = H
if pt:
# no PT correction for HiX
self.H_pt['HiX'] = H
if 'ChiX' in calc:
H = -(self.PX * malog2(
np.ma.array(self.PiX,
copy=False,
mask=(self.PiX <= np.finfo(np.float).eps)))).sum(
axis=0)
self.H_plugin['ChiX'] = H
if pt:
# no PT correction for ChiX
self.H_pt['ChiX'] = H
# for adelman style I(k;spike) (bits/spike)
if 'HXY1' in calc:
if self.Y_m != 2:
raise ValueError, \
"HXY1 calculation only makes sense for spike data, ie Y_m = 2"
H = ent(self.PXY[:, 1])
self.H_plugin['HXY1'] = H
if pt:
self.H_pt['HXY1'] = H + pt_corr(
pt_bayescount(self.PXY[:, 1], self.Ny[1]))
if 'ChiXY1' in calc:
if self.Y_m != 2:
raise ValueError, \
"ChiXY1 calculation only makes sense for spike data, ie Y_m = 2"
H = -np.ma.array(self.PXY[:, 1] * np.log2(self.PX),
copy=False,
mask=(self.PX <= np.finfo(np.float).eps)).sum()
self.H_plugin['ChiXY1'] = H
if pt:
# no PT for ChiXY1
self.H_pt['ChiXY1'] = H
def _calc_pt_plugin(self, pt):
"""Calculate direct entropies and apply PT correction if required """
calc = self.calc
pt_corr = lambda R: (R-1)/(2*self.N*np.log(2))
self.H_plugin = {}
if pt: self.H_pt = {}
# compute basic entropies
if 'HX' in calc:
H = ent(self.PX)
self.H_plugin['HX'] = H
if pt:
self.H_pt['HX'] = H + pt_corr(pt_bayescount(self.PX, self.N))
if 'HY' in calc:
H = ent(self.PY)
self.H_plugin['HY'] = H
if pt:
self.H_pt['HY'] = H + pt_corr(pt_bayescount(self.PY, self.N))
if 'HXY' in calc:
H = (self.PY * ent(self.PXY)).sum()
self.H_plugin['HXY'] = H
if pt:
for y in xrange(self.Y_dim):
H += pt_corr(pt_bayescount(self.PXY[:,y], self.Ny[y]))
self.H_pt['HXY'] = H
if 'SiHXi' in calc:
H = ent(self.PXi).sum()
self.H_plugin['SiHXi'] = H
if pt:
for x in xrange(self.X_n):
H += pt_corr(pt_bayescount(self.PXi[:,x],self.N))
self.H_pt['SiHXi'] = H
if 'HiXY' in calc:
H = (self.PY * ent(self.PXiY)).sum()
self.H_plugin['HiXY'] = H
if pt:
for x in xrange(self.X_n):
for y in xrange(self.Y_dim):
H += pt_corr(pt_bayescount(self.PXiY[:,x,y],self.Ny[y]))
self.H_pt['HiXY'] = H
if 'HiX' in calc:
H = ent(self.PiX)
self.H_plugin['HiX'] = H
if pt:
# no PT correction for HiX
self.H_pt['HiX'] = H
if 'ChiX' in calc:
H = -(self.PX*malog2(np.ma.array(self.PiX,copy=False,
mask=(self.PiX<=np.finfo(np.float).eps)))).sum(axis=0)
self.H_plugin['ChiX'] = H
if pt:
# no PT correction for ChiX
self.H_pt['ChiX'] = H
# for adelman style I(k;spike) (bits/spike)
if 'HXY1' in calc:
if self.Y_m != 2:
raise ValueError, \
"HXY1 calculation only makes sense for spike data, ie Y_m = 2"
H = ent(self.PXY[:,1])
self.H_plugin['HXY1'] = H
if pt:
self.H_pt['HXY1'] = H + pt_corr(pt_bayescount(self.PXY[:,1],self.Ny[1]))
if 'ChiXY1' in calc:
if self.Y_m != 2:
raise ValueError, \
"ChiXY1 calculation only makes sense for spike data, ie Y_m = 2"
H = -np.ma.array(self.PXY[:,1]*np.log2(self.PX),copy=False,
mask=(self.PX<=np.finfo(np.float).eps)).sum()
self.H_plugin['ChiXY1'] = H
if pt:
# no PT for ChiXY1
self.H_pt['ChiXY1'] = H