嵌入可视化 #TensorFlow #Python

embedding.py

# -*- coding: utf-8 -*-
# 嵌入可视化
import numpy as np
import tensorflow as tf
from tensorboard.plugins import projector
from tensorflow.examples.tutorials.mnist import input_data
import os

PATH_TO_MNIST_DATA = "MNIST_data"
LOG_DIR = "projector/data"
IMAGE_NUM = 300

# 读取测试数据
mnist = input_data.read_data_sets(PATH_TO_MNIST_DATA, one_hot=False)

# 获取数据
plot_array = mnist.test.images[:IMAGE_NUM]  # shape: (n_observations, n_features)

# 产生 meta data
np.savetxt(os.path.join(LOG_DIR, 'metadata.tsv'), mnist.test.labels[:IMAGE_NUM], fmt='%d')

# Download sprite image
# https://www.tensorflow.org/images/mnist_10k_sprite.png, 100x100 thumbnails
PATH_TO_SPRITE_IMAGE = os.path.join(LOG_DIR, 'mnist_10k_sprite.png')

# 可视化嵌入
# 1) 创建2D张量变量
session = tf.InteractiveSession()
embedding_var = tf.Variable(plot_array, name='embedding')
tf.global_variables_initializer().run()

# 2) 定期保存嵌入数据到LOG_DIR
# 这里我们只保存一次 设置global_step为固定数值
saver = tf.train.Saver()
saver.save(session, os.path.join(LOG_DIR, "model.ckpt"), global_step=0)

# 3) 用嵌入链接meta数据和预览图片
# 使用保存检查点文件的同样的LOG_DIR
summary_writer = tf.summary.FileWriter(LOG_DIR)

config = projector.ProjectorConfig()
# 可以添加多个embeddings. 这里我们只添加一个
embedding = config.embeddings.add()
embedding.tensor_name = embedding_var.name
# 链接张量到metadata文件 (e.g. labels).
embedding.metadata_path = os.path.join(LOG_DIR, 'metadata.tsv')
# 链接张量到预览图.
embedding.sprite.image_path = PATH_TO_SPRITE_IMAGE
embedding.sprite.single_image_dim.extend([28, 28])
# 保存配置文件以让tensorboard读取 
projector.visualize_embeddings(summary_writer, config)

"""
在.tsv的目录下输入 tensorboard --logdir=。
或者 tensorboard --logdir=绝对路径
来启动tb
"""