In [1]:

    

import tensorflow as tf
import numpy as np
import math
from matplotlib import pyplot as plt
from tqdm import tqdm_notebook
from IPython.display import clear_output
%matplotlib inline

m = 60
tTr = np.linspace(0, 3*math.pi, m)
om2 = tf.constant(1.0)


    

In [2]:

    

sess = tf.Session()


    

In [3]:

    

#Params
time = tf.placeholder(tf.float32)
A = tf.get_variable("A", initializer = tf.random_uniform(shape=[], minval = -1, maxval = 1))
B = tf.get_variable("B", initializer = tf.random_uniform(shape=[], minval = -1, maxval = 1))
C = tf.get_variable("C", initializer = tf.random_uniform(shape=[], minval = -1, maxval = 1))


#A = tf.get_variable("A", initializer = -1.0)
#B = tf.get_variable("B", initializer = -1.0)
#C = tf.get_variable("C", initializer = 0.0)


    

In [4]:

    

init = tf.global_variables_initializer()
sess.run(init)


    

In [5]:

    

#FORWARD
approximator = A * tf.sin(B*time + C)

#COST
alpha = tf.get_variable("Learning_Rate", initializer = 0.01)
v0 = tf.constant(1.0)
num = tf.constant(m, dtype = tf.float32)
t_0 = tf.constant(0.0)

vel = tf.gradients(approximator, time)
acc = tf.gradients(vel, time)

x0_app = A * tf.sin(B*t_0 + C)
v0_app = A * B * tf.cos(B*t_0 + C)

eq = acc + tf.multiply(approximator, om2)

J = tf.square(tf.reduce_sum(eq))/num + tf.square(x0_app) + tf.square(v0_app - v0)

Ag = tf.gradients(J, A)
Bg = tf.gradients(J, B)
Cg = tf.gradients(J, C)

update_A = tf.assign_sub(A, alpha*tf.reshape(Ag,[])) 
update_B = tf.assign_sub(B, alpha*tf.reshape(Bg,[])) 
update_C = tf.assign_sub(C, alpha*tf.reshape(Cg,[]))


    

In [6]:

    

Err = []
I = []
def grad_descent(N, learn_rate):
    sess.run(tf.assign(alpha, learn_rate))
    for i in tqdm_notebook(range(N)):
        I.append(i)
        j = sess.run(J, {time: tTr})
        Err.append(j)
        
        sess.run(update_A, {time: tTr})
        sess.run(update_B, {time: tTr})
        sess.run(update_C, {time: tTr})
        if i%400==0:
            clear_output()
            print(j)
            
def showGrad():            
    grads = []            
    grads.append(sess.run(Ag, {time: tTr}))
    grads.append(sess.run(Bg, {time: tTr}))
    grads.append(sess.run(Cg, {time: tTr}))
    return grads


    

In [7]:

    

grad_descent(2000, 0.05)


    

    




5.68434e-14

In [8]:

    

print(showGrad())
sess.run(tf.square(v0_app-v0), {time:tTr})

#print(sess.run(approximator, {time:1}))


    

    




[[4.7683716e-07], [4.7683704e-07], [2.1389693e-37]]






    
Out[8]:





5.6843419e-14

In [9]:

    

r=np.sin(tTr)
result = sess.run(approximator, {time: tTr})

plt.plot(tTr, result, 'red')
plt.plot(tTr, r, 'green')
plt.plot(tTr, result - r)


    

    
Out[9]:





[<matplotlib.lines.Line2D at 0x7fba9432ff98>]






    

In [10]:

    

'''
true_init = [tf.assign(A, 1),
tf.assign(B, 1),
tf.assign(C, 0)]
for assignment in true_init:
    sess.run(assignment)


x_res = sess.run(approximator, {time: tTr})

acc_res = sess.run(acc, {time: tTr})

acc_th = -A*B*B*tf.sin(B*time+C)


plt.plot(tTr, x_res)
plt.plot(tTr, acc_res[0])
plt.plot(tTr, sess.run(acc_th, {time: tTr}) - acc_res[0])
'''
print(1)


    

    




1