和單向LSTM搭建的步驟是一樣的：
- 設(shè)定超參數(shù)（隱藏向量的維度回挽，學(xué)習(xí)率猾普，輸入輸出維度等）
- 定義輸入輸出placeholder
- 定義網(wǎng)絡(luò)結(jié)構(gòu)皮获，寫好網(wǎng)絡(luò)計算語句（dynamic_run等乓梨，用于計算時生成LSTM單元的輸出狀態(tài)）
- 定義全連接層的權(quán)重和偏差牲剃，用于將LSTM單元狀態(tài)的輸出轉(zhuǎn)換成類別未規(guī)范化概率
- 計算輸出的未規(guī)范化概率
- 定義softmax層
- 定義損失
- 定義訓(xùn)練優(yōu)化器和優(yōu)化操作
- sess.run()
下面給出代碼：
learning_rate = 0.1
state_size = 128 # hidden layer num of features
n_classes = 19
n_features = 23
#輸入
x = tf.placeholder(tf.float32, [None, None, n_features], name='input_placeholder') #batch_size, time_step, feat_len
y = tf.placeholder(tf.float32, [None, None, n_classes], name='labels_placeholder') #batch_size, time_step, n_classes
batch_size = tf.placeholder(tf.int32, (), name='batch_size')
time_steps = tf.placeholder(tf.int32, (), name='times_step')
#雙向rnn
lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(state_size)
lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(state_size)
init_fw = lstm_fw_cell.zero_state(batch_size, dtype=tf.float32)
init_bw = lstm_bw_cell.zero_state(batch_size, dtype=tf.float32)
weights = tf.get_variable("weights", [2 * state_size, n_classes], dtype=tf.float32, #注意這里的維度
initializer = tf.random_normal_initializer(mean=0, stddev=1))
biases = tf.get_variable("biases", [n_classes], dtype=tf.float32,
initializer = tf.random_normal_initializer(mean=0, stddev=1))
outputs, final_states = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell,
lstm_bw_cell,
x,
initial_state_fw = init_fw,
initial_state_bw = init_bw)
outputs = tf.concat(outputs, 2) #將前向和后向的狀態(tài)連接起來
state_out = tf.matmul(tf.reshape(outputs, [-1, 2 * state_size]), weights) + biases #注意這里的維度
logits = tf.reshape(state_out, [batch_size, time_steps, n_classes])
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
logits=logits, labels=y)) #計算交叉熵
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate) #優(yōu)化方法
train_op = optimizer.minimize(loss_op)
#進(jìn)行softmax計算
probs = tf.nn.softmax(logits, -1) #-1也是默認(rèn)值遣疯，表示在最后一維進(jìn)行運算
predict = tf.argmax(probs, -1)#最大的概率在最后一維的哪一列，從0計數(shù)
#維度變?yōu)? batch_size * time_step
def train_network(num_epochs = 100):
with tf.Session() as sess:
sess.run(tf.global_variables_initializer()) #初始化variable
intervel = 5
for epoch in range(num_epochs):
#開始訓(xùn)練
for idx, (time_step, inputs, labels, idxes) in enumerate(get_dump_seq_data(1)):
_= sess.run([train_op],
feed_dict = {x: inputs,
y:labels,
batch_size:len(inputs),
time_steps: time_step})
print("epoch %d train done" % epoch)
#這一輪訓(xùn)練完畢凿傅，計算損失值和準(zhǔn)確率
if epoch % intervel == 0 and epoch > 1:
#訓(xùn)練集誤差
acc_record, total_df, total_acc, loss = compute_accuracy(sess, 1) #這里是我自定義的函數(shù)缠犀，與整個架構(gòu)關(guān)系不大
#驗證集誤差
acc_record_dev, total_df_dev, total_acc_dev, loss_dev = compute_accuracy(sess, 0)
print("train_acc: %.6f, train_loss: %.6f; dev_acc: %.6f, dev_loss: %.6f" % (total_acc, loss, total_acc_dev, loss_dev))
print("- " * 50)
if num_epochs - epoch <= intervel:
return acc_record, total_df, acc_record_dev, total_df_dev
