import tensorflow as tf
import os
import pickle
import numpy as np

CIFAR_DIR = "C:/Users/tianr/Desktop/cifar-10-batches-py"
print(os.listdir(CIFAR_DIR))

def load_data(filename):
    """read data from data file"""
    with open(filename,'rb') as f:
        data = pickle.load(f,encoding='bytes')
        return data[b'data'],data[b'labels']

# API：tesorflow.Dataset
class CifarData:
    def __init__(self,filenames,need_shuffle):
        all_data = []
        all_labels = []
        for filename in filenames:
            data,labels = load_data(filename)
            all_data.append(data)
            all_labels.append(labels)
        self._data = np.vstack(all_data)
        self._data = self._data / 127.5 - 1# 归一化将像素缩放到[-1,1]
        self._labels = np.hstack(all_labels)
        print(self._data.shape)
        print(self._labels.shape)
        
        self._num_examples = self._data.shape[0]
        self._need_shuffle = need_shuffle
        self._indicator = 0 #定义数据起始位置
        if self._need_shuffle:
            self._shuffle_data()
            
    def _shuffle_data(self):
        # [0,1,2,3,4,5] -> [3,1,2,5,0,4]
        p = np.random.permutation(self._num_examples)
        self._data = self._data[p]
        self._labels = self._labels[p]
        
    def next_batch(self,batch_size):
        """return batch_size examples as a batch"""
        end_indicator = self._indicator + batch_size
        if end_indicator > self._num_examples:
            if self._need_shuffle:
                self._shuffle_data()
                end_indicator = batch_size
            else:
                raise Exception("have no more examples")
        if end_indicator > self._num_examples:
            raise Exception("batch size is larger than all examples")
        batch_data = self._data[self._indicator: end_indicator]
        batch_labels = self._labels[self._indicator: end_indicator]
        self._indicator = end_indicator
        return batch_data,batch_labels
    
train_filenames = [os.path.join(CIFAR_DIR,'data_batch_%d' % i) for i in range(1,6)]
test_filenames = [os.path.join(CIFAR_DIR,'test_batch')]

train_data = CifarData(train_filenames,True)
test_data = CifarData(test_filenames,False)

x = tf.placeholder(tf.float32, [None, 3072])
y = tf.placeholder(tf.int64, [None])
# [None], eg: [0,5,6,3]
x_image = tf.reshape(x, [-1, 3, 32, 32])
# 32*32
x_image = tf.transpose(x_image, perm=[0, 2, 3, 1])

# conv1: 神经元图， feature_map, 输出图像
conv1 = tf.layers.conv2d(x_image,
                         32, # output channel number
                         (3,3), # kernel size
                         padding = 'same',
                         activation = tf.nn.relu,
                         name = 'conv1')

# 16 * 16
pooling1 = tf.layers.max_pooling2d(conv1,
                                   (2, 2), # kernel size
                                   (2, 2), # stride
                                   name = 'pool1')


conv2 = tf.layers.conv2d(pooling1,
                         32, # output channel number
                         (3,3), # kernel size
                         padding = 'same',
                         activation = tf.nn.relu,
                         name = 'conv2')

# 8 * 8
pooling2 = tf.layers.max_pooling2d(conv2,
                                   (2, 2), # kernel size
                                   (2, 2), # stride
                                   name = 'pool2')

conv3 = tf.layers.conv2d(pooling2,
                         32, # output channel number
                         (3,3), # kernel size
                         padding = 'same',
                         activation = tf.nn.relu,
                         name = 'conv3')

# 4 * 4 * 32
pooling3 = tf.layers.max_pooling2d(conv3,
                                   (2, 2), # kernel size
                                   (2, 2), # stride
                                   name = 'pool3')
# [None, 4 * 4 * 32]
flatten = tf.layers.flatten(pooling3)
y_ = tf.layers.dense(flatten, 10)

loss = tf.losses.sparse_softmax_cross_entropy(labels=y, logits=y_)
# y_ -> sofmax
# y -> one_hot
# loss = ylogy_

# indices
predict = tf.argmax(y_, 1)
# [1,0,1,1,1,0,0,0]
correct_prediction = tf.equal(predict, y)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float64))

with tf.name_scope('train_op'):
    train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)

    init = tf.global_variables_initializer()
batch_size = 20
train_steps = 10000
test_steps = 100

with tf.Session() as sess:
    sess.run(init)
    for i in range(train_steps):
        batch_data, batch_labels = train_data.next_batch(batch_size)
        loss_val, acc_val, _ = sess.run(
            [loss, accuracy, train_op],
            feed_dict={
                x: batch_data,
                y: batch_labels})
        if (i+1) % 50 == 0:
            print('[Train] Step: %d, loss: %4.5f, acc: %4.5f' % (i+1, loss_val, acc_val))
        if (i+1) % 500 == 0:
            test_data = CifarData(test_filenames, False)
            all_test_acc_val = []
            for j in range(test_steps):
                test_batch_data, test_batch_labels \
                    = test_data.next_batch(batch_size)
                test_acc_val = sess.run(
                    [accuracy],
                    feed_dict = {
                        x: test_batch_data, 
                        y: test_batch_labels
                    })
                all_test_acc_val.append(test_acc_val)
            test_acc = np.mean(all_test_acc_val)
            print('[Test ] Step: %d, acc: %4.5f' % (i+1, test_acc))