18. TF Visualization

18.1. 可视化的实现方式

Package summary, on github, provide a way to export condensed information about a model, which is then accessible in tools such as TensorBoard. 因为涉及到“数据导出”，即结构化数据的序列化和文件保存，所以不可避免的要用到 protobufs.

Package summary, api docs, 用于TensorFlow Event Processing, and is primarily being developed to support TensorBoard, but it can be used by anyone who wishes to analyze or visualize TensorFlow events files.

18.2. How to write event file

  1. 定义可视化参数

Create Summary Ops(on tf-guide) and add them to the collection of GraphKeys.SUMMARIES

_images/summary.png 
1
2
3
4
5
tf.summary.scalar(name="cls_loss", tensor=cls_loss_op)
tf.summary.scalar(name="bbox_loss", tensor=bbox_loss_op)
tf.summary.scalar(name="landmark_loss", tensor=landmark_loss_op)
tf.summary.scalar(name="cls_accuracy", tensor=accuracy_op)
summary_op = tf.summary.merge_all(key=tf.GraphKeys.SUMMARIES, scope=None)
  1. Creates a FileWriter

The FileWriter class provides a mechanism to create an event file in a given directory.

注意：(和tf.TFRecordWriter()的命名比较)并非tf.event.FileWriter()或者tf.EventWriter()。

1
2
3
4
# Launch the graph in a session.
sess = tf.Session()
# Create a summary writer, add the 'graph' to the event file.
writer = tf.summary.FileWriter(<some-directory>, sess.graph)
  1. Add data to summaries

在没有运行的时候这些操作是不会执行任何东西的，仅仅是定义了一下而已。

1
2
for step in range(MAX_STEP):
  _,_,summary = sess.run([summary_op])#for each iteration
  1. Write event protocol buffers to event files

class FileWriter updates the file contents asynchronously based on queues. This allows a training program to call methods, add_xxx(), to add data to the file directly from the training loop, without slowing down training.

注意:

  • add_xxx()仅仅是向FileWriter对象的缓存中存放event data, 而向disk上写数据是由FileWrite对象控制的。
  • add_xxx()并非在sess.run()中执行。
1
writer.add_summary(summary,global_step=step)
  1. flush
1
2
#Call this method to make sure that all pending events have been written to disk.
writer.flush()
  1. 涉及的protcol buffer
_images/summary-2.png