trans_test.py

import chainer
from chainer import cuda
import cupy
import numpy
import time

ls = numpy.random.randint(1, 100, size=1000)
ls = numpy.sort(ls)[::-1]
unit = 100
xs = [cupy.empty((l, unit), 'f') for l in ls]

cuda.to_cpu(xs[0])

begin = time.time()
for i in range(10):
    chainer.functions.transpose_sequence(xs)
cuda.to_cpu(xs[0])
end = time.time()

print(end - begin)