Deep Learning for Computer Vision

Embedding visualization

The embedding layer, which is the pre-final layer, can be visualized in two or three dimensions using TensorBoard. The code snippets in this section are assumed to come after the convolution neural network model trained in the image classification chapter. First, we need a metadata file that is a tab separated file. Every line of the metadata file should have the labels of the images that are going to be visualized. A new variable is required for storing the embedding that is defined between session creation and initialization, as shown in the following code:

no_embedding_data = 1000embedding_variable = tf.Variable(tf.stack(    mnist.test.images[:no_embedding_data], axis=0), trainable=False)

We will take MNIST test data ...