python - TensorFlow vs. Theano Performance -

in context of neural network research i'm evaluating several approaches on how implement these or library use. i'm comparing tensorflow , theano , i'm struggling getting tenorflow perform well. here simple hello-gradient-benchmark, optimizes scalar multiplication 1 coefficient.

import time  class timer:     def __init__(self, what):       self.what =     def __enter__(self):       self.t1 = time.time()       return self     def __exit__(self,t,v,tb):       t2 = time.time()       print("{0} runs {1:.4f} seconds".format(self.what, t2-self.t1))   def run_tensorflow():     import tensorflow tf     x = tf.placeholder(tf.float32)    y = tf.placeholder(tf.float32)    = tf.variable([1.], tf.float32)     sess = tf.session()    sess.run(tf.global_variables_initializer())     loss = (y-a*x)**2    step = tf.train.gradientdescentoptimizer(0.01).minimize(loss)     def one_step():       sess.run(step, {x:1.,y:0.})     timer('tensorflow') t:       result = [ one_step() n in range(1000) ]   def run_theano():     import theano th     x = th.tensor.dscalar()    y = th.tensor.dscalar()    = th.tensor.dscalar()    l = a*x     loss = (y-l)**2    dloss = th.tensor.grad(loss, a)    dloss_f = th.function([x,y,a], dloss)     = [1.]     def one_step():       a[0] -= 0.01 * dloss_f(1.,0.,a[0])     timer('theano') t:       result = [ one_step() n in range(1000) ]   run_tensorflow() run_theano() 

i'm running program on cpu packages installed via pip. running times 0.36 , 0.043 seconds tensorflow , theano, respectively. see similar performance differences real networks matrix-multiplication overhead should dominate, still tensorflow slower.

i want know if i'm using tensorflow wrongly i'm trying do. should not call run() method within loop?

1. tf , theano designed handling large objects, on order of 1m elements. benchmarking handling of scalars not particularly relevant.

2. this apples-to-oranges comparison: tf, timing both compilation , run time, while in theano, timing run time! because when call theano.function, compilation then. otoh in tf, of work shifted when first call sess.run.

that said, there realistic scenarios when tf slower theano.