1、使用placeholder讀記憶體中的資料

最簡單的一種方法是用placeholder，然後以feed_dict將資料給holder的變數，進行傳遞值。如下面程式碼所示：

from __future__ import print_function
import tensorflow as tf
import numpy as np

x1 = tf.placeholder(tf.float32,shape=(3,2))
y1 = tf.placeholder(tf.float32,shape=(2,3))
z1 = tf.matmul(x1,y1)

x2 = tf.placeholder
 
(tf.float32,shape=None)
y2 = tf.placeholder(tf.float32,shape=None)
z2 = x2 + y2

# using feed_dict when placehoder
with tf.Session() as sess:
    z2_value = sess.run(z2,feed_dict={x2:1,y2:2}) 
    print(z2_value)
    rand_x = np.random.rand(3,2)
    rand_y = np.random.rand(2,3)
    z1_value,z2_value = sess.run
 
(
        [z1,z2],                   # run together
        feed_dict={
            x1:rand_x,y1:rand_y,
            x2:1,y2:2
        }
    )
    print(z1_value,z2_value)

2、使用queue讀硬碟中的資料

參考如下的連線，不過感覺佇列讀取方式較為複雜，有了Dataset API後大部分不用此方法。

3、Dataset API

Dataset可以看作是相同型別“元素”的有序列表。在實際使用時，單個“元素”可以是向量，也可以是字串、圖片，甚至是tuple或者dict。

注意下圖的繼承關係

tf.data.TextLineDataset

可以直接從檔案中讀取資料

__init__(
    filenames,
    compression_type=None,
    buffer_size=None
)

程式碼示例：

with tf.Graph().as_default(),tf.Session() as sess:
    # instance a dataset,np.array() => tf.constant => tensorflow
    dataset = tf.data.Dataset.from_tensor_slices(np.array([1,2,3,4,5]))
    # we can also use tf.data.TextLineDataset because this inherit tf.data.Dataset
    # dataset = tf.data.TextLineDataset.from_tensor_slices(np.array([1,2,3,4,5]))
    # return a Iterator over the element of this dataset 
    iterator = dataset.make_one_shot_iterator()
    element = iterator.get_next() # every element is a number
    for i in range(5):
        print(sess.run(element))  # 1,2,3,4,5


##### read data from file
"""
we have a file test.csv:
1,2,0
4,5,1
7,8,2
"""
with tf.Graph().as_default(),tf.Session() as sess:
    dataset = tf.data.TextLineDataset("test.csv")
    iterator = dataset.make_one_shot_iterator()
    element = iterator.get_next() # every element is a vector
    try:
        while True:
            print(sess.run(element))
    except tf.errors.OutOfRangeError:
        print("end!")

##### more complex dataset
"""
1,2,0
4,5,1
7,8,2
the last column is label we create => batch of feature,label
"""

with tf.Graph().as_default(),tf.Session() as sess:
    def to_tensor(line):
        parsed_line = tf.decode_csv(line,[[0.],[0.],[0]]) # => tensor
        #label = parsed_line[-1]
        label =  parsed_line[-1]
        del parsed_line[-1]
        features = parsed_line
        features_names = ['feature_1','feature_2']
        d = dict(zip(features_names,features)),label
        return d

    dataset = tf.data.TextLineDataset("test.csv").map(to_tensor).batch(2)
    iterator = dataset.make_one_shot_iterator()
    batch_features,batch_labels = iterator.get_next()
    try:
        while True:
            batch_fea,batch_lab = sess.run([batch_features,batch_labels])           
            print(batch_fea,batch_lab)
    except tf.errors.OutOfRangeError:
        print("end!")

注意dataloader的使用方式

# create dataloader
dataset = tf.data.Dataset.from_tensor_slices((tfx,tfy)) #reference tf_dataset_basic.py
dataset = dataset.shuffle(buffer_size=1000)
dataset = dataset.batch(32)
dataset = dataset.repeat(5)
iterator = dataset.make_initializable_iterator()

使用dataset具體的一個例子

x = np.random.uniform(-1,1,(1000,1)) 
y = np.power(x,2) + np.random.normal(0,0.1,size=x.shape)
x_train,x_test = np.split(x,[800])
y_train,y_test = np.split(y,[800])
print(
    '\nx_train shape',x_train.shape,
    '\ny_train shape',y_train.shape,
)
"""
plt.scatter(x_train,y_train)
plt.show()
"""

tfx = tf.placeholder(x_train.dtype,x_train.shape)
tfy = tf.placeholder(y_train.dtype,y_train.shape)

# create dataloader
dataset = tf.data.Dataset.from_tensor_slices((tfx,tfy)) #reference tf_dataset_basic.py
dataset = dataset.shuffle(buffer_size=1000)
dataset = dataset.batch(32)
dataset = dataset.repeat(5)
iterator = dataset.make_initializable_iterator()

# built network
batch_x,batch_y = iterator.get_next()  # batch_x:(32,1)
h1 = tf.layers.dense(batch_x,10,tf.nn.relu) # batch_x:(32,10)
out = tf.layers.dense(h1,1) # 32*1
loss = tf.losses.mean_squared_error(batch_y,out)
train = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

with tf.Session() as sess:
    #initializable
    sess.run([iterator.initializer,tf.global_variables_initializer()],
            feed_dict={tfx:x_train,tfy:y_train})
    for step in range(301):
        try:
            _,train_loss = sess.run([train,loss])
            if step % 10 == 0:
                test_loss = sess.run(loss,{batch_x:x_test,batch_y:y_test})
                print('\nsetp:',step,
                    '\ntrain loss:',train_loss,
                    '\ntest loss:',test_loss,
                )
        except tf.errors.OutOfRangeError:
            print("finish!")
            break

完整程式碼在我的github上~

參考資料