FC perf benchmark

fc_bench.py

import mxnet as mx
import numpy as np
from timeit import default_timer as timer

def get_bench_net(num_hidden=10000):
  data = mx.sym.var('data')
  fc = mx.sym.FullyConnected(data, num_hidden=num_hidden)
  return fc

num_out = 10000
num_example = 1024
net = get_bench_net(num_out)
mod = mx.mod.Module(net)

data = np.random.rand(num_example, 10000)
#label = np.random.rand(num_example, num_out)
nd_iter = mx.io.NDArrayIter(data, batch_size=16)
mod.bind(data_shapes=nd_iter.provide_data, label_shapes=None)
mod.init_params()

start = timer()
for _ in range(100):
  mod.forward_backward()
  mx.nd.waitall()
print('elaped time:', timer() - start)