from keras.callbacks import TensorBoard 
from keras.models import Sequential 
…… 
model = Sequential() 
…… 
tbCallBack = keras.callbacks.TensorBoard(log_dir='./Graph', 
histogram_freq=1, 
write_graph=True, 
write_images=True)

model_history = model.fit(X_train_train, 
y_train_train, 
batch_size=batch_size, 
epochs=epochs, 
verbose=1, 
validation_data=(X_train_val, y_train_val), 
callbacks = [EarlyStopping(patience=patience,mode='min',verbose=1), 
history, 
tbCallBack]) 

%预先写好writer，定义好model 
writer = tf.summary.FileWriter(…) 
model = … 
%训练时 
loss = model.train_on_batch(…) 
summary = tf.Summary(value=[ 
tf.Summary.Value(tag=”d_loss”, simple_value=d_loss), 
tf.Summary.Value(tag=”g_loss”, simple_value=g_loss), 
]) 
writer.add_summary(summary) 

import tensorflow as tf 
import datetime 
%在训练开始之前，预先定义好可视化的东西，用的是原生的tensorflow方法，这里我们以一个GAN模型为例，让它显示整张模型图，两个标量损失函数，以及5个生成图像。方法是预先用placeholder声明所要显示的那些东西，然后在训练过程中将训练结果来填充它们。你可以自行加入histgram，都和tensorflow一样 
#start tensorboard 
sess=tf.Session() 
logdir = “tensorboard/” + datetime.datetime.now().strftime(“%Y%m%d-%H%M%S”) + “/” 
writer = tf.summary.FileWriter(logdir, sess.graph) 
D_LOSS = tf.placeholder(tf.float32, []) 
G_LOSS = tf.placeholder(tf.float32, []) 
IMAGES = tf.placeholder(tf.float32,shape=[None,28,28,1]) 
tf.summary.scalar(“D_LOSS”, D_LOSS) 
tf.summary.scalar(“G_LOSS”, G_LOSS) 
tf.summary.image(“IMAGES”, IMAGES, 5) 
merged=tf.summary.merge_all() 
#end tensorboard 

for epoch in range(100): 
% 用keras的train_on_batch方法进行训练 
d_loss = d.train_on_batch(。。。。。。) 
g_loss = d_on_g.train_on_batch(。。。。。。) 
generated_images = g.predict(。。。。。。) 
if index%10==0: #tensorboard 
% 将训练结果填充可视化数据 
summary=sess.run(merged,feed_dict={D_LOSS:d_loss, G_LOSS:g_loss, IMAGES:generated_images}) 
writer.add_summary(summary,index)