import tensorflow as tf
import numpy as np
from read_input_data import RID

data_reader = RID()


#################
# training data #
#################
trX = np.array(data_reader.get_data_dict_with_arrays()["data_train"]).astype(np.float32) / 255.0

trY = np.array(data_reader.get_data_dict_with_arrays()["labels_train_coords"]).astype(np.float32)
trY[:, 0] /= 800.0
trY[:, 1] /= 600.0
trY = np.array([trY[:, 0]]).reshape((21, 1))


#############
# test data #
#############
teX = np.array(data_reader.get_data_dict_with_arrays()["data_test"]).astype(np.float32) / 255.0

teY = np.array(data_reader.get_data_dict_with_arrays()["labels_test_coords"]).astype(np.float32)
teY[:, 0] /= 800.0
teY[:, 1] /= 600.0
teY = np.array([teY[:, 0]]).reshape((7, 1))


################
# reshape data #
################
trX = trX.reshape(-1, 600, 800, 1)
teX = teX.reshape(-1, 600, 800, 1)


################################################################################
# weight, bias, conv2d and max_pool methods for initialization and calculation #
################################################################################
def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)


def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)


def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')


def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1], padding='SAME')


##############################
# placeholder for data input #
##############################
x = tf.placeholder("float", shape=[None, 600, 800, 1])
y_train = tf.placeholder("float", shape=[None, 1])

# reshape data
x_image = tf.reshape(x, [-1, 600, 800, 1])


########################
# network architecture #
########################

# conv/maxpool layer
W_conv1 = weight_variable([5, 5, 1, 8])
b_conv1 = bias_variable([8])

h_conv1 = tf.nn.relu((conv2d(x_image, W_conv1) + b_conv1))
h_pool1 = max_pool_2x2(h_conv1)


# conv/maxpool layer
W_conv2 = weight_variable([5, 5, 8, 16])
b_conv2 = bias_variable([16])

h_conv2 = tf.nn.relu((conv2d(h_pool1, W_conv2) + b_conv2))
h_pool2 = max_pool_2x2(h_conv2)


# fully connected layer
W_fc1 = weight_variable([150 * 200 * 16, 100])
b_fc1 = bias_variable([100])

h_pool2_flat = tf.reshape(h_pool2, [-1, 150 * 200 * 16])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)


# fully connected softmax layer
W_fc2 = weight_variable([100, 1])
b_fc2 = bias_variable([1])

y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)


##################
# train and eval #
##################
cross_entropy = -tf.reduce_sum(y_train * tf.log(y_conv))
# cross_entropy = tf.reduce_mean(-(y_train * tf.log(y_conv) + (1 - y_train) * tf.log(1 - y_conv)))

# train_step = tf.train.GradientDescentOptimizer(0.001).minimize(cross_entropy)
train_step = tf.train.AdamOptimizer(0.001).minimize(cross_entropy)
# train_step = tf.train.RMSPropOptimizer(0.001, 0.9).minimize(cross_entropy)

correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_train, 1))

accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))


#################
# session stuff #
#################
sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)



###################
# actual training #
###################
for i in range(1, 1000 + 1):
    # if i % 5 == 0:
    #     train_accuracy = sess.run(accuracy, feed_dict={
    #         x: trX, y_train: trY, keep_prob: 1.0})
    #     print("step {:>3}, training accuracy {}".format(i, train_accuracy))

    # training step
    sess.run(train_step, feed_dict={x: trX, y_train: trY, keep_prob: 0.5})


    #################
    # testing stuff #
    #################

    # testing picture 0
    test_out = sess.run(y_conv, feed_dict={x: [trX[0]],
                                         keep_prob: 1.0})
    print("out[0]: {:>5.5f}  |  out original[0]: {:>5.5f}  |  diff[0]: {:>5.5f}".format(test_out[0][0], trY[0][0], abs(
        test_out[0][0] - trY[0][0])))

    # testing picture 1
    test_out = sess.run(y_conv, feed_dict={x: [trX[1]],
                                         keep_prob: 1.0})
    print("out[1]: {:>5.5f}  |  out original[1]: {:>5.5f}  |  diff[1]: {:>5.5f}".format(test_out[0][0], trY[1][0], abs(
        test_out[0][0] - trY[1][0])))

    # testing picture 2
    test_out = sess.run(y_conv, feed_dict={x: [trX[2]],
                                         keep_prob: 1.0})
    print("out[2]: {:>5.5f}  |  out original[2]: {:>5.5f}  |  diff[2]: {:>5.5f}".format(test_out[0][0], trY[2][0], abs(
        test_out[0][0] - trY[2][0])))

    # testing picture 3
    test_out = sess.run(y_conv, feed_dict={x: [trX[3]],
                                         keep_prob: 1.0})
    print("out[3]: {:>5.5f}  |  out original[3]: {:>5.5f}  |  diff[3]: {:>5.5f}".format(test_out[0][0], trY[3][0], abs(
        test_out[0][0] - trY[3][0])))

    # testing random picture/array
    print(sess.run(y_conv, feed_dict={x: [np.random.rand(600, 800).reshape(600, 800, 1)],
                                         keep_prob: 1.0}))

    # testing array of zeros
    print(sess.run(y_conv, feed_dict={x: [np.zeros([600, 800, 1])],
                                         keep_prob: 1.0}))

    # testing array of ones
    print(sess.run(y_conv, feed_dict={x: [np.ones([600, 800, 1])],
                                         keep_prob: 1.0}))


#################
# test accuracy #
#################
print("\ntesting accuracy...")
print("\ntest accuracy: %g" % sess.run(accuracy, feed_dict={
    x: teX, y_train: teY, keep_prob: 1.0}))