1. 程式人生 > >tf.nn.dynamic_rnn返回值詳解

tf.nn.dynamic_rnn返回值詳解

函式原型

tf.nn.dynamic_rnn(
    cell,
    inputs,
    sequence_length=None,
    initial_state=None,
    dtype=None,
    parallel_iterations=None,
    swap_memory=False,
    time_major=False,
    scope=None
)


例項講解:

import tensorflow as tf
import numpy as np
 
n_steps = 2
n_inputs = 3
n_neurons = 5
 
X = tf.placeholder(tf.float32, [None, n_steps, n_inputs])
basic_cell = tf.contrib.rnn.BasicRNNCell(num_units=n_neurons)
 
seq_length = tf.placeholder(tf.int32, [None])
outputs, states = tf.nn.dynamic_rnn(basic_cell, X, dtype=tf.float32,
                                    sequence_length=seq_length)
 
init = tf.global_variables_initializer()
 
X_batch = np.array([
        # step 0     step 1
        [[0, 1, 2], [9, 8, 7]], # instance 1
        [[3, 4, 5], [0, 0, 0]], # instance 2 (padded with zero vectors)
        [[6, 7, 8], [6, 5, 4]], # instance 3
        [[9, 0, 1], [3, 2, 1]], # instance 4
    ])
seq_length_batch = np.array([2, 1, 2, 2])
 
with tf.Session() as sess:
    init.run()
    outputs_val, states_val = sess.run(
        [outputs, states], feed_dict={X: X_batch, seq_length: seq_length_batch})
    print("outputs_val.shape:", outputs_val.shape, "states_val.shape:", states_val.shape)
    print("outputs_val:", outputs_val, "states_val:", states_val)


log info:

outputs_val.shape: (4, 2, 5) states_val.shape: (4, 5)
outputs_val: 
[[[ 0.53073734 -0.61281306 -0.5437517   0.7320347  -0.6109526 ]
  [ 0.99996936  0.99990636 -0.9867181   0.99726075 -0.99999976]]
 
 [[ 0.9931584   0.5877845  -0.9100412   0.988892   -0.9982337 ]
  [ 0.          0.          0.          0.          0.        ]]
 
 [[ 0.99992317  0.96815354 -0.985101    0.9995968  -0.9999936 ]
  [ 0.99948144  0.9998127  -0.57493806  0.91015154 -0.99998355]]
 
 [[ 0.99999255  0.9998929   0.26732785  0.36024097 -0.99991137]
  [ 0.98875254  0.9922327   0.6505734   0.4732064  -0.9957567 ]]] 
states_val:
 [[ 0.99996936  0.99990636 -0.9867181   0.99726075 -0.99999976]
 [ 0.9931584   0.5877845  -0.9100412   0.988892   -0.9982337 ]
 [ 0.99948144  0.9998127  -0.57493806  0.91015154 -0.99998355]
 [ 0.98875254  0.9922327   0.6505734   0.4732064  -0.9957567 ]]


首先輸入X是一個 [batch_size,step,input_size] = [4,2,3] 的tensor,注意我們這裡呼叫的是BasicRNNCell,只有一層迴圈網路,outputs是最後一層每個step的輸出,它的結構是[batch_size,step,n_neurons] = [4,2,5],states是每一層的最後那個step的輸出,由於本例中,我們的迴圈網路只有一個隱藏層,所以它就代表這一層的最後那個step的輸出,因此它和step的大小是沒有關係的,我們的X有4個樣本組成,輸出神經元大小n_neurons是5,因此states的結構就是[batch_size,n_neurons] = [4,5],最後我們觀察資料,states的每條資料正好就是outputs的最後一個step的輸出。

下面我們繼續講解多個隱藏層的情況,這裡是三個隱藏層,注意我們這裡仍然是呼叫BasicRNNCell

import tensorflow as tf
import numpy as np
 
n_steps = 2
n_inputs = 3
n_neurons = 5
n_layers = 3
 
X = tf.placeholder(tf.float32, [None, n_steps, n_inputs])
seq_length = tf.placeholder(tf.int32, [None])
 
layers = [tf.contrib.rnn.BasicRNNCell(num_units=n_neurons,
                                      activation=tf.nn.relu)
          for layer in range(n_layers)]
multi_layer_cell = tf.contrib.rnn.MultiRNNCell(layers)
outputs, states = tf.nn.dynamic_rnn(multi_layer_cell, X, dtype=tf.float32, sequence_length=seq_length)
 
init = tf.global_variables_initializer()
 
X_batch = np.array([
        # step 0     step 1
        [[0, 1, 2], [9, 8, 7]], # instance 1
        [[3, 4, 5], [0, 0, 0]], # instance 2 (padded with zero vectors)
        [[6, 7, 8], [6, 5, 4]], # instance 3
        [[9, 0, 1], [3, 2, 1]], # instance 4
    ])
 
seq_length_batch = np.array([2, 1, 2, 2])
 
with tf.Session() as sess:
    init.run()
    outputs_val, states_val = sess.run(
        [outputs, states], feed_dict={X: X_batch, seq_length: seq_length_batch})
    print("outputs_val.shape:", outputs, "states_val.shape:", states)
    print("outputs_val:", outputs_val, "states_val:", states_val)


log info:

outputs_val.shape: 
Tensor("rnn/transpose_1:0", shape=(?, 2, 5), dtype=float32) 
 
states_val.shape: 
(<tf.Tensor 'rnn/while/Exit_3:0' shape=(?, 5) dtype=float32>, 
 <tf.Tensor 'rnn/while/Exit_4:0' shape=(?, 5) dtype=float32>, 
 <tf.Tensor 'rnn/while/Exit_5:0' shape=(?, 5) dtype=float32>)
 
outputs_val:
 [[[0.         0.         0.         0.         0.        ]
  [0.         0.18740742 0.         0.2997518  0.        ]]
 
 [[0.         0.07222144 0.         0.11551574 0.        ]
  [0.         0.         0.         0.         0.        ]]
 
 [[0.         0.13463384 0.         0.21534224 0.        ]
  [0.03702604 0.18443246 0.         0.34539366 0.        ]]
 
 [[0.         0.54511094 0.         0.8718864  0.        ]
  [0.5382122  0.         0.04396425 0.4040263  0.        ]]] 
 
states_val:
 (array([[0.        , 0.83723307, 0.        , 0.        , 2.8518028 ],
       [0.        , 0.1996038 , 0.        , 0.        , 1.5456247 ],
       [0.        , 1.1372368 , 0.        , 0.        , 0.832613  ],
       [0.        , 0.7904129 , 2.4675028 , 0.        , 0.36980057]],
      dtype=float32), 
  array([[0.6524607 , 0.        , 0.        , 0.        , 0.        ],
       [0.25143963, 0.        , 0.        , 0.        , 0.        ],
       [0.5010576 , 0.        , 0.        , 0.        , 0.        ],
       [0.        , 0.3166597 , 0.4545995 , 0.        , 0.        ]],
      dtype=float32), 
  array([[0.        , 0.18740742, 0.        , 0.2997518 , 0.        ],
       [0.        , 0.07222144, 0.        , 0.11551574, 0.        ],
       [0.03702604, 0.18443246, 0.        , 0.34539366, 0.        ],
       [0.5382122 , 0.        , 0.04396425, 0.4040263 , 0.        ]],
      dtype=float32))


我們說過,outputs是最後一層的輸出,即 [batch_size,step,n_neurons] = [4,2,5] 

states是每一層的最後一個step的輸出,即三個結構為 [batch_size,n_neurons] = [4,5] 的tensor

繼續觀察資料,states中的最後一個array,正好是outputs的最後那個step的輸出

下面我們繼續講當由BasicLSTMCell構造單元工廠的時候,只講多層的情況,我們只需要將上面的BasicRNNCell替換成BasicLSTMCell就行了,列印資訊如下:

outputs_val.shape: 
Tensor("rnn/transpose_1:0", shape=(?, 2, 5), dtype=float32) 
 
states_val.shape:
(LSTMStateTuple(c=<tf.Tensor 'rnn/while/Exit_3:0' shape=(?, 5) dtype=float32>, 
                h=<tf.Tensor 'rnn/while/Exit_4:0' shape=(?, 5) dtype=float32>), 
LSTMStateTuple(c=<tf.Tensor 'rnn/while/Exit_5:0' shape=(?, 5) dtype=float32>, 
               h=<tf.Tensor 'rnn/while/Exit_6:0' shape=(?, 5) dtype=float32>), 
LSTMStateTuple(c=<tf.Tensor 'rnn/while/Exit_7:0' shape=(?, 5) dtype=float32>, 
               h=<tf.Tensor 'rnn/while/Exit_8:0' shape=(?, 5) dtype=float32>))
 
outputs_val: 
[[[1.2949290e-04 0.0000000e+00 2.7623639e-04 0.0000000e+00 0.0000000e+00]
  [9.4675866e-05 0.0000000e+00 2.0214770e-04 0.0000000e+00 0.0000000e+00]]
 
 [[4.3100454e-06 4.2123037e-07 1.4312843e-06 0.0000000e+00 0.0000000e+00]
  [0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]]
 
 [[0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]
  [0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]]
 
 [[0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]
  [0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]]] 
 
states_val: 
(LSTMStateTuple(
c=array([[0.        , 0.        , 0.04676079, 0.04284539, 0.        ],
       [0.        , 0.        , 0.0115245 , 0.        , 0.        ],
       [0.        , 0.        , 0.        , 0.        , 0.        ],
       [0.        , 0.        , 0.        , 0.        , 0.        ]],
      dtype=float32), 
h=array([[0.        , 0.        , 0.00035096, 0.04284406, 0.        ],
       [0.        , 0.        , 0.00142574, 0.        , 0.        ],
       [0.        , 0.        , 0.        , 0.        , 0.        ],
       [0.        , 0.        , 0.        , 0.        , 0.        ]],
      dtype=float32)), 
LSTMStateTuple(
c=array([[0.0000000e+00, 1.0477135e-02, 4.9871090e-03, 8.2785974e-04,
        0.0000000e+00],
       [0.0000000e+00, 2.3306280e-04, 0.0000000e+00, 9.9445322e-05,
        5.9535629e-05],
       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00],
       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00]], dtype=float32), 
h=array([[0.00000000e+00, 5.23016974e-03, 2.47756205e-03, 4.11730434e-04,
        0.00000000e+00],
       [0.00000000e+00, 1.16522635e-04, 0.00000000e+00, 4.97301044e-05,
        2.97713632e-05],
       [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
        0.00000000e+00],
       [0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 0.00000000e+00,
        0.00000000e+00]], dtype=float32)), 
LSTMStateTuple(
c=array([[1.8937115e-04, 0.0000000e+00, 4.0442235e-04, 0.0000000e+00,
        0.0000000e+00],
       [8.6200516e-06, 8.4243663e-07, 2.8625946e-06, 0.0000000e+00,
        0.0000000e+00],
       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00],
       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00]], dtype=float32), 
h=array([[9.4675866e-05, 0.0000000e+00, 2.0214770e-04, 0.0000000e+00,
        0.0000000e+00],
       [4.3100454e-06, 4.2123037e-07, 1.4312843e-06, 0.0000000e+00,
        0.0000000e+00],
       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00],
       [0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00]], dtype=float32)))


我們先看看LSTM單元的結構

如果您不檢視框內的內容,LSTM單元看起來與常規單元格完全相同,除了它的狀態分為兩個向量:h(t)和c(t)。你可以將h(t)視為短期狀態,將c(t)視為長期狀態。

因此我們的states包含三個LSTMStateTuple,每一個表示每一層的最後一個step的輸出,這個輸出有兩個資訊,一個是h表示短期記憶資訊,一個是c表示長期記憶資訊。維度都是[batch_size,n_neurons] = [4,5],states的最後一個LSTMStateTuple中的h就是outputs的最後一個step的輸出