import tensorflow as tf
import numpy as np
import math

np.random.seed(10)

#program to take 3 features, multiply them together to produce a 4th row

numRows = 4
numCols = 200

def randomTrainingData():
    lower = -10
    upper = 10
    mat = np.random.rand(numRows,numCols) * (upper - lower) + lower

    for col in range(numCols):
        prod = 1
        for f in mat[:-1,col]:
            prod *= f
        mat[-1,col] = prod

    return mat


def nn(input):
    pass

def main():
    trainingData = randomTrainingData()
    X = trainingData[:-1,:]
    Y = trainingData[-1,:].reshape(1,numCols)
    inputPlaceholder = tf.placeholder(tf.float32,
                                      shape=(numRows-1, numCols))

    hiddenUnits = 100

    ################################################

    weights1 = tf.Variable(
        tf.random_uniform([hiddenUnits, numRows-1],-.1, .1),
        name = 'weights1')
    biases1 = tf.Variable(tf.zeros([hiddenUnits,1]),
                          name = 'biases1')
    hidden = tf.nn.sigmoid(tf.matmul(weights1,inputPlaceholder) + biases1)

    numOutputs = 1

    weights2 = tf.Variable(
        tf.random_uniform([numOutputs, hiddenUnits],-.1,.1),
        name = 'weights2')
    biases2 = tf.Variable(tf.zeros([numOutputs]), name='biases2')
    outputs = tf.matmul(weights2, hidden) + biases2

    ########################################################

    targetPlaceholder = tf.placeholder(tf.float32, shape=(1,numCols))
    learningRate = 1e-5
    optimizer = tf.train.GradientDescentOptimizer(learningRate)
#    loss = tf.reduce_mean(tf.square(outputs-targetPlaceholder))
    loss = tf.square(outputs-targetPlaceholder)
    tr = optimizer.minimize(loss)

    ###############################################

    init = tf.initialize_all_variables()

    with tf.Session() as sess:
        dd={inputPlaceholder:X, targetPlaceholder:Y}
        dd2={inputPlaceholder:X}
        sess.run(init,feed_dict = dd)

        for i in range(100000):
            sess.run(tr, feed_dict=dd)


            if (i) % 5000 == 0:
                print "iteration:",i
                ll = sess.run(loss, feed_dict = dd)
                print np.mean(ll[0])
                oo = sess.run(outputs,feed_dict=dd2)
                print oo[0,0],Y[0,0]


if __name__ == "__main__":
    main()