LSTM 权重和偏置在tensorflow、keras的保存结构

https://blog.csdn.net/koibiki/article/details/83116596

 

@tf_export("nn.rnn_cell.BasicLSTMCell")
class BasicLSTMCell(LayerRNNCell):
 
input_depth = inputs_shape[1].value
h_depth = self._num_units
self._kernel = self.add_variable(
    _WEIGHTS_VARIABLE_NAME,
    shape=[input_depth + h_depth, 4 * self._num_units])
self._bias = self.add_variable(
    _BIAS_VARIABLE_NAME,
    shape=[4 * self._num_units],
    initializer=init_ops.zeros_initializer(dtype=self.dtype))
以上是tensorflow LSTMCell中的一段代码，里面可以看到lstm的kernel的shape为 [input_depth + h_depth, 4 * self._num_units]

即[输入深度 + 输出深度， 4 × cell个数]，在第二维上的排列顺序为 i、c、f、o。

一个minist的例子验证
为了方便验证，设置 timesteps = 1，这样就可以不用考虑 和的对结果的影响

为了获得中间层的结果，将所有层输出均通过一个字典保存

from __future__ import print_function
import tensorflow as tf
from tensorflow.contrib import rnn
# Import MNIST data
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
from keras.utils import np_utils
import math
 
mnist = input_data.read_data_sets("./data/", one_hot=True)
 
# Training Parameters
learning_rate = 0.001
training_steps = 10000
batch_size = 128
display_step = 200
 
# Network Parameters
num_input = 784 # MNIST data input (img shape: 28*28)
timesteps = 1 # timesteps
num_hidden = 128 # hidden layer num of features
num_classes = 10 # MNIST total classes (0-9 digits)
 
# tf Graph input
X = tf.placeholder("float", [None, timesteps, num_input])
Y = tf.placeholder("float", [None, num_classes])
 
# Define weights
weights = {
    'w': tf.Variable(tf.random_normal([num_hidden, num_classes]))
}
biases = {
    'b': tf.Variable(tf.random_normal([num_classes]))
}
 
def RNN(x, weights, biases):
    x = tf.unstack(x, timesteps, 1)
    lstm_cell = rnn.BasicLSTMCell(num_hidden, forget_bias=1.0)
    outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
    logits = tf.matmul(outputs[-1], weights['w']) + biases['b']
    out = {'lstm':lstm_cell, 'lstm_out':outputs, 'states':states, 'logits': logits}
    return out
 
out = RNN(X, weights, biases)
logits = out['logits']
prediction = tf.nn.softmax(logits)
 
# Define loss and optimizer
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
    logits=logits, labels=Y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op)
 
# Evaluate model (with test logits, for dropout to be disabled)
correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
 
# Initialize the variables (i.e. assign their default value)
init = tf.global_variables_initializer()
 
# Start training
sess =  tf.Session()
# Run the initializer
sess.run(init)
 
for step in range(1, training_steps+1):
    batch_x, batch_y = mnist.train.next_batch(batch_size)
    # Reshape data to get 28 seq of 28 elements
    batch_x = batch_x.reshape((batch_size, timesteps, num_input))
    # Run optimization op (backprop)
#     print(batch_x.shape)
    sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})
    if step % display_step == 0 or step == 1:
        # Calculate batch loss and accuracy
        loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x,
                                                                 Y: batch_y})
        print("Step " + str(step) + ", Minibatch Loss= " + \
                  "{:.4f}".format(loss) + ", Training Accuracy= " + \
                  "{:.3f}".format(acc))
print("Optimization Finished!")
 
# Calculate accuracy for 1 mnist test images
test_data = mnist.test.images[:1].reshape((-1, timesteps, num_input))
test_label = mnist.test.labels[:1]
训练完成，运行lstm cell的tensor，传入test数据，获得中间层lstm cell输出结果，其中test数据只有1个。

lstm_out = sess.run(out['lstm_out'], feed_dict={X:test_data})
得到lstm cell结果

[array([[ 0.0598105 , -0.14341736,  0.02396348, -0.08234564, -0.04004124,
          ....]], dtype=float32)]
运行kernel和bias tersor获得权重

lstm_kernel = sess.run(out['lstm'].weights[0])
lstm_bias = sess.run(out['lstm'].weights[1])
分拆权重参数

lstm_k_i = lstm_kernel[:784, :128]
lstm_k_h_i = lstm_kernel[784:, :128]
lstm_k_c = lstm_kernel[:784, 128:128*2]
lstm_k_h_c = lstm_kernel[784:, 128:128*2]
lstm_k_f = lstm_kernel[:784, 128*2:128*3]
lstm_k_h_f = lstm_kernel[784:, 128*2:128*3]
lstm_k_o = lstm_kernel[:784, 128*3:]
lstm_k_h_o = lstm_kernel[784:, 128*3:]
lstm_b_i = lstm_bias[:128]
lstm_b_c = lstm_bias[128:128*2]
lstm_b_f = lstm_bias[128*2:128*3]
lstm_b_o = lstm_bias[128*3:]
定义sigmoid函数

def sigmoid(v):
    return 1 /(1 + math.exp(-v))
获得第四个lstm cell的输出，注意，因为timesteps = 1，所以此处没有考虑的影响。

i = sigmoid(np.dot(lstm_k_i[:,3], test_data[0,0]) + lstm_b_i[3])
f = sigmoid(np.dot(lstm_k_f[:,3], test_data[0,0]) + lstm_b_f[3])
ct_ = math.tanh(np.dot(lstm_k_c[:,3], test_data[0,0]) + lstm_b_c[3])
ct = i * ct_ + f*0
o = sigmoid(np.dot(lstm_k_o[:,3], test_data[0,0]) + lstm_b_o[3])
ht = math.tanh(ct) * o
-0.08234565254093663
与模型测试输出一致。