summary method can save the training process and relevant parameters, which can be later visualized in tensorboard besides the loss and accuracy data in also be saved in np format for visualization in python later
summary class
tf.summary.scalar(tags, values, collections=None, name=None)
used to save loss and accuracy datas
tf.summary.scalar('accuracy',acc)
tf.summary.histogram(tags, values, collections=None, name=None)
# saved the variables distribution during training
tf.summary.histogram('histogram', var)
tf.summary.distribution
save weights distribution
tf.summary.image
save buffer with images
tf.summary.audio
save audio data in training
tf.summary.FileWritter(path,sess.graph)
define target file for FileWritter to write in, later this file can be loaded into tensorboard
if multi-plots are needed, then multi FileWritter need to be defined
train_writer = tf.summary.FileWriter(dir,sess.graph)
tf.summary.merge_all
save all datas during training
merge_summary = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(dir,sess.graph)
for step in xrange(training_step):
train_summary = sess.run(merge_summary,feed_dict = {...})
train_writer.add_summary(train_summary,step)
tf.summary.merge(inputs, collections=None, name=None)
selectively get summary information use tf.get_collection to get the data in collection
tf.summary.merge([tf.get_collection(tf.GraphKeys.SUMMARIES,'accuracy'),...])
np save method
train_loss=[]
for step in range(max_steps)
loss_in_each_step=...
train_loss.append(loss_in_each_step)
np.save(os.path.join(save_dir, 'accuracy_loss', 'train_loss'), train_loss)
