Pytorch slowdown test

import torch
import time
from torch.autograd import Variable
import numpy as np

"""
This program gets slower and slower by time. No noticeable change in memory
consumption seen on nvidia-smi.
"""

usecuda = 1

dtype = torch.FloatTensor
N=25
if usecuda:
  N = 100*N
  dtype = torch.cuda.FloatTensor

xdim = (100,50,2)

x0 = torch.randn(*xdim).type(dtype)
z0 = torch.randn(N, xdim[0]).type(dtype)

x = Variable(x0, requires_grad=True)
z = Variable(z0, requires_grad=False)

def updateZ(z, b):
  z -= (0.1*b.lt(0.1).sum(2).type(dtype) + 0.01)
  z *= 0.999
  return z

t0 = time.clock()
tsum = 0
tsump = 1
i = 0
while True:
  b = (x.repeat(N,1,1,1).view(N, -1, 2)).pow(2).sum(2).sqrt()
  c = z.view(N, xdim[0], 1).expand(N, xdim[0], xdim[1])*(1/b.view(N, xdim[0], xdim[1]).pow(0.5))
  loss = c.sum()
  loss.backward()
  z = updateZ(z, b.view(N, -1 , xdim[1]))
  x.data -= 0.0000000001*x.grad.data
  x.grad.data.zero_()
  t1 = time.clock()
  tsum += t1 - t0
  if i%100 == 99:
    print("%.4f +%.1f%%" % (tsum, tsum*100/tsump - 100))
    tsump = tsum
    tsum = 0
  i = i + 1
  t0 = t1