tensorflow應用:雙向LSTM神經網絡手寫數字識別
思路
將28X28的圖片看成28行像素,按行展開成28時間步,每時間步間對識別都有影響,故用雙向LSTM神經元,其實每列間對識別也有影響,用卷積神經網絡也許更合理,這裏只是學習LSTM的用法。應該也可以用兩個雙向LSTM神經網絡進行聯合預測,一個按行掃描,一個按列掃描。
Python程序1.建模訓練保存
# coding=utf-8
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"]='2' # 只顯示 warning 和 Error
import tensorflow as tf
from tensorflow.contrib import rnn
import numpy as np
###data (50000,784),(1000,784),(1000,784):
import pickle
import gzip
def load_data():
f = gzip.open('../data/mnist.pkl.gz', 'rb')
training_data, validation_data, test_data = pickle.load(f,encoding='bytes')
f.close()
return (training_data, validation_data, test_data)
def vectorized_result(j):
e = np.zeros(10)
e[j] = 1.0
return e
training_data, validation_data, test_data = load_data()
trainData_in=training_data[0][:50000]
trainData_out=[vectorized_result(j) for j in training_data[1][:50000]]
validData_in=validation_data[0]
validData_out=[vectorized_result(j) for j in validation_data[1]]
testData_in=test_data[0][:100]
testData_out=[vectorized_result(j) for j in test_data[1][:100]]
#define constants
#unrolled through 28 time steps 28行對應28個時間步:
TIME_STEPS=28
#hidden LSTM units
NUM_HIDDEN=128
#???rows of 28 pixels 每行28個像素:
NUM_INPUT=28
#learning rate for adam
LEARNING_RATE=0.001
#mnist is meant to be classified in 10 classes(0-9).
NUM_CLASSES=10
#size of batch
BATCH_SIZE=1024
TRAINING_EPOCHS=1#
#weights and biases of appropriate shape to accomplish above task
#out_weights=tf.Variable(tf.random_normal([NUM_HIDDEN,NUM_CLASSES]))
#雙向神經網絡的權重爲單向的2倍尺度:
out_weights=tf.Variable(tf.random_normal([2*NUM_HIDDEN,NUM_CLASSES]))
out_bias=tf.Variable(tf.random_normal([NUM_CLASSES]))
#defining placeholders
#input image placeholder:
x_input=tf.placeholder("float",[None,TIME_STEPS,NUM_INPUT],name='x_input')
#input label placeholder:
y_desired=tf.placeholder("float",[None,NUM_CLASSES])
#processing the input tensor from [BATCH_SIZE,NUM_STEPS,NUM_INPUT] to "TIME_STEPS" number of [BATCH-SIZE,NUM_INPUT] tensors!:
#對輸入的一個張量的第二維解包變成TIME_STEPS個張量!:
x_input_step=tf.unstack(x_input ,TIME_STEPS,1)
#defining the network:
#def BiRNN(x_input_step,out_weights,out_bias):
#lstm_layer=rnn.BasicLSTMCell(NUM_HIDDEN,forget_bias=1.0)
#正向神經元:
lstm_fw_cell=rnn.BasicLSTMCell(NUM_HIDDEN,forget_bias=1.0)
#反向神經元:
lstm_bw_cell=rnn.BasicLSTMCell(NUM_HIDDEN,forget_bias=1.0)
#outputs,_=rnn.static_rnn(lstm_layer,x_input_step,dtype="float32")
#構建雙向LSTM網絡:
outputs,_,_=rnn.static_bidirectional_rnn( lstm_fw_cell,lstm_bw_cell,x_input_step,dtype="float32")
#converting last output of dimension [batch_size,num_hidden] to [batch_size,num_classes] by out_weight multiplication
z_prediction= tf.add(tf.matmul(outputs[-1],out_weights),out_bias,name='z_prediction')
#z_prediction=BiRNN(x_input_step, out_weights, out_bias)
#注意!z_prediction經softmax歸一化後纔是最終的輸出,用於和標籤比較,下面的損失函數中用了softmax哈交叉熵,跳過了求y_output這一步:
y_output=tf.nn.softmax(z_prediction,name='y_output')
#loss_function:
loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=z_prediction,labels=y_desired),name='loss')
#optimization
opt=tf.train.AdamOptimizer(learning_rate=LEARNING_RATE).minimize(loss)
#model evaluation
correct_prediction=tf.equal(tf.argmax(z_prediction,1),tf.argmax(y_desired,1))
accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
#以下彙總一些參數用於TensorBoard:
for value in [loss]:
tf.summary.scalar(value.op.name,value) #彙總的標籤及值
summary_op=tf.summary.merge_all() #彙總合併
#initialize variables:
init=tf.global_variables_initializer()
with tf.Session() as sess:
# 生成一個寫日誌的writer,並將當前的tensorflow計算圖寫入日誌。
# tensorflow提供了多種寫日誌文件的API
summary_writer=tf.summary.FileWriter(r'C:\temp\log_simple_stats',sess.graph)
sess.run(init)
num_batches=int(len(trainData_in)/BATCH_SIZE)
for epoch in range(TRAINING_EPOCHS):
for i in range(num_batches):
batch_x=trainData_in[i*BATCH_SIZE:(i+1)*BATCH_SIZE]
batch_x=batch_x.reshape((BATCH_SIZE,TIME_STEPS,NUM_INPUT))#
batch_y=trainData_out[i*BATCH_SIZE:(i+1)*BATCH_SIZE]
#優化及日誌結果!!!!!!:::::
_,summary=sess.run([opt,summary_op], feed_dict={x_input: batch_x, y_desired: batch_y})
#寫日誌,將結果添加到彙總:
summary_writer.add_summary(summary,global_step=epoch*num_batches+i)
if i %10==0:
acc=sess.run(accuracy,feed_dict={x_input:batch_x,y_desired:batch_y})
los=sess.run(loss,feed_dict={x_input:batch_x,y_desired:batch_y})
print('epoch:%4d,'%epoch,'%4d'%i)
print("Accuracy ",acc)
print("Loss ",los)
print("__________________")
print("Finished!")
print("Test Accuracy ",sess.run(accuracy,\
feed_dict={x_input:testData_in.reshape((-1,TIME_STEPS,NUM_INPUT)),\
y_desired:testData_out}))
saver=tf.train.Saver()
save_path=saver.save(sess,'../data')
print('Model saved to %s' % save_path)
summary_writer.close()
Tensorboard檢查計算圖及訓練結果
在終端運行:
Tensorboard --logdir= C:\temp\log_simple_stats
C:\Users\li\AppData\Local\Programs\Python\Python36\Scripts>tensorboard --logdir=C:\temp\log_simple_stats
TensorBoard 1.10.0 at http://li-PC:6006 (Press CTRL+C to quit)
用谷歌瀏覽器打開http://li-pc:6006/
打開訓練好的模型進行預測
# coding=utf-8
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"]='2' # 只顯示 warning 和 Error
###data (50000,784),(1000,784),(1000,784):
import pickle
import gzip
import numpy as np
def load_data():
f = gzip.open('../data/mnist.pkl.gz', 'rb')
training_data, validation_data, test_data = pickle.load(f,encoding='bytes')
f.close()
return (training_data, validation_data, test_data)
def vectorized_result(j):
e = np.zeros(10)
e[j] = 1.0
return e
import tensorflow as tf
import matplotlib.pyplot as plt
#unrolled through 28 time steps 28行對應28個時間步:
TIME_STEPS=28
#???rows of 28 pixels 每行28個像素:
NUM_INPUT=28
training_data, validation_data, test_data = load_data()
testData_in=test_data[0]
testData_out=[vectorized_result(j) for j in test_data[1]]
sess=tf.InteractiveSession()
new_saver=tf.train.import_meta_graph('../data.meta')
new_saver.restore(sess, '../data')
tf.get_default_graph().as_graph_def()
x_input=sess.graph.get_tensor_by_name('x_input:0')
y_output=sess.graph.get_tensor_by_name('y_output:0')
try_input=testData_in[6]
try_desired=testData_out[6]
print(try_desired)
print(y_output.eval(feed_dict={x_input:\
np.array([try_input]).reshape((-1,TIME_STEPS,NUM_INPUT))}))
try_input.resize(28,28)
plt.imshow(try_input,cmap='Greys_r')
plt.show()
[0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
[[1.13495190e-07 7.10399399e-06 5.81623426e-05 1.16373285e-05
7.91627526e-01 5.11910184e-04 1.04986066e-04 1.08990945e-01
1.73597573e-03 9.69514474e-02]]
模型只進行了一輪訓練,40次更新,40X1024個樣本,就準確識別了手寫數字4,判斷爲數字4的概率是 7.91627526e-01