

In [1]:

    
import tensorflow as tf



In [2]:

    
from pycalphad import Database, Model, variables as v
from pycalphad.codegen.sympydiff_utils import build_functions
from sympy import lambdify
import numpy as np

dbf = Database('Al-Cu-Zr_Zhou.tdb')
mod = Model(dbf, ['AL', 'CU', 'ZR'], 'LIQUID')



In [3]:

    
mod.variables









    Out[3]:





[T, LIQUID0AL, LIQUID0CU, LIQUID0ZR]

Expression Building



In [4]:

    
%time bfr = build_functions(mod.GM, mod.variables, include_grad=True)









    



Wall time: 377 ms



In [5]:

    
%%timeit
tf_func = lambdify(mod.variables, mod.GM, modules='tensorflow')
tf_grads = lambdify(mod.variables, [mod.GM.diff(x) for x in mod.variables], modules='tensorflow')









    



632 ms ± 20.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)



In [6]:

    
tf_func = lambdify(mod.variables, mod.GM, modules='tensorflow')
tf_grads = lambdify(mod.variables, [mod.GM.diff(x) for x in mod.variables], modules='tensorflow')



In [7]:

    
%%timeit
func_xla = tf.function(experimental_compile=True)(tf_func)
grad_xla = tf.function(experimental_compile=True)(tf_grads)









    



93 µs ± 3.83 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)



In [8]:

    
func_xla = tf.function(experimental_compile=True)(tf_func)
grad_xla = tf.function(experimental_compile=True)(tf_grads)

Function Evaluation



In [9]:

    
out = np.array([0.])
%timeit bfr.func.unsafe_real(np.array([300., 0.3, 0.3, 0.4]), out)
print(out)









    



3.94 µs ± 189 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
[-41178.9428136]



In [10]:

    
out = np.zeros(4)
%timeit bfr.grad.unsafe_real(np.array([300., 0.3, 0.3, 0.4]), out)
print(out)









    



4.27 µs ± 84.5 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
[-3.39000632e+01 -6.67993032e+04 -2.12186404e+04 -3.12432907e+04]



In [11]:

    
%timeit func_xla(300.0, 0.3, 0.3, 0.4)









    



202 µs ± 76.5 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)



In [12]:

    
%timeit grad_xla(300.0, 0.3, 0.3, 0.4)









    



299 µs ± 71.8 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)



In [13]:

    
temps = tf.constant(np.linspace(300., 2000., num=1000, dtype=np.float32))
%timeit grad_xla(temps, 0.3, 0.3, 0.4)









    



480 µs ± 76.4 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)



In [ ]: