tensorboard的使用

1、keras 与 tensorboard

from tensorflow.keras.callbacks import TensorBoard

filepath="snapshot/v2_weights-improvement-{epoch:02d}-{val_acc:.2f}.hdf5"
tensorboard = TensorBoard(log_dir='logs/{}'.format(filepath))
callbacks_list = [checkpoint, tensorboard]

model.fit(train_imgs, train_labels, validation_data=(test_imgs, test_labels) ,
          batch_size=256, callbacks=callbacks_list, epochs=15, shuffle=True, verbose=0)

　终端  tensorboard --logdir="./log"

   http://localhost:6006（如果访问的是服务器，别忘记修改ip）

   https://stackoverflow.com/questions/42112260/how-do-i-use-the-tensorboard-callback-of-keras

    更多参考：http://nooverfit.com/wp/tensorboard%E4%B8%8A%E6%89%8B%EF%BC%8Ctensorboard%E5%8F%AF%E8%A7%86%E5%8C%96%E6%99%AE%E5%8F%8A%E8%B4%B4%EF%BC%88%E4%BB%A3%E7%A0%81%E5%9F%BA%E4%BA%8Etensorflow1-2%E4%BB%A5%E4%B8%8A%EF%BC%89/

                          https://zhuanlan.zhihu.com/p/33178205

                       https://zhuanlan.zhihu.com/p/36946874

                      https://zhuanlan.zhihu.com/p/37022051

     

     监督的信息主要包括lr, acc，loss等

      

# 静态流图
summary_writer = tf.summary.FileWriter(train_dir, sess.graph)
# 动态scalars
tf.summary.scalar("loss", loss)
tf.summary.scalar("accuracy", acc)

with tf.name_scope("input_shape"):
    image_shape_input = tf.reshape(x, [-1, 100, 100, 3])
    tf.summary.image('input', image_shape_input, 2)

summary_op = tf.summary.merge_all()

 # 记录耗时信息
    run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
    run_metadata = tf.RunMetadata()

    _, summary, train_loss, train_acc = sess.run([train_op, summary_op, loss, acc],
                                                 feed_dict={x: x_train_a, y_: y_train_a},
                                                 options=run_options,
                                                 run_metadata=run_metadata)
    end = time.time()
    summary_writer.add_run_metadata(run_metadata, 'step %04d' % steps)
    summary_writer.add_summary(summary, steps)

2、模型可视化

     pb模型可视化 

v1 = tf.Variable(tf.constant(1, shape=[1]), name='v1')
v2 = tf.Variable(tf.constant(1, shape=[1]), name='v2')
result = v1 + v2


with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    graph = tf.get_default_graph()
    graph_def = graph.as_graph_def()
    output_grapg_def = tf.graph_util.convert_variables_to_constants(sess, graph_def, ["add"])

    with tf.gfile.GFile("Model/freeze_model.pb", 'wb') as f:
        f.write(output_grapg_def.SerializeToString())

    graph_def.ParseFromString(tf.gfile.GFile("Model/freeze_model.pb", 'rb').read())
    tf.import_graph_def(graph_def, name='graph')
    summary_writer = tf.summary.FileWriter('log/', graph)

    meta可视化

    

v1 = tf.Variable(tf.constant(1, shape=[1]), name='v1')
v2 = tf.Variable(tf.constant(1, shape=[1]), name='v2')
result = v1 + v2

saver = tf.train.Saver()
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    print(sess.run(result))
    saver.save(sess, "Model/model.ckpt")

    graph = tf.get_default_graph()
    graph_def = graph.as_graph_def()
    _ = tf.train.import_meta_graph("Model/model.ckpt.meta")
    summary_writer = tf.summary.FileWriter("log/", graph)