5

I was wondering if anyone here has ever tried to visualize a multidimensional tensor in numpy. If so, could you share with me how I might go about doing this? I was thinking of reducing it to a 2D visualization.

I've included some sample output. It's weirdly structured, there are ellipses "..." and it's got a 4D tensor layout [[[[ content here]]]]

Sample Data: 

[[[[ -9.37186633e-05  -9.89684777e-05  -8.97786958e-05 ...,
     -1.08984910e-04  -1.07056971e-04  -8.68257193e-05]
  [[ -9.61350961e-05  -8.75062251e-05  -9.39425736e-05 ...,
     -1.17737654e-04  -9.66376538e-05  -8.78447026e-05]
   [ -1.06558400e-04  -9.04031331e-05  -1.04479543e-04 ...,
     -1.02786013e-04  -1.07974607e-04  -1.07524407e-04]]
 [[[ -1.09648725e-04  -1.01073667e-04  -9.39013553e-05 ...,
     -8.94383265e-05  -9.06078858e-05  -9.83356076e-05]
   [ -9.76310257e-05  -1.04029998e-04  -1.01905476e-04 ...,
     -9.50643880e-05  -8.29156561e-05  -9.75912480e-05]]]
   [ -1.12038200e-04  -1.00154917e-04  -9.00980813e-05 ...,
     -1.10244124e-04  -1.16597665e-04  -1.10604939e-04]]]]
CodeinCafes
  • 107
  • 1
  • 1
  • 8

1 Answers1

5

  • For plotting high dimensional data there is a technique called as T-SNE

  • T-SNE is provided by tensorflow as a tesnorboard feature

  • You can just provide the tensor as an embedding and run tensorboard

  • You can visualize high dimensional data in either 3D or 2d

  • Here is a link for Data Visualization using Tensor-board: https://github.com/jayshah19949596/Tensorboard-Visualization-Freezing-Graph

  • Your code should be something like this :

      tensor_x = tf.Variable(mnist.test.images, name='images')
  config = projector.ProjectorConfig()
  # One can add multiple embeddings.
  embedding = config.embeddings.add()
  embedding.tensor_name = tensor_x.name
  # Link this tensor to its metadata file (e.g. labels).
  embedding.metadata_path = metadata
  # Saves a config file that TensorBoard will read during startup.
  projector.visualize_embeddings(tf.summary.FileWriter(logs_path), config)

  • Tensorboard visualization: enter image description here

  • You can use scikit learn's TSNE to plot high dimensional data

  • Below is sample coede to use scikit learn's TSNE

      # x is my data which is a nd-array
  # You have to convert your tensor to nd-array before using scikit-learn's tsne
  # Convert your tensor to x =====> x = tf.Session().run(tensor_x)
  standard = StandardScaler()
  x_std = standard.fit_transform(x)
  plt.figure()

  label_encoder = LabelEncoder()
  y = label_encoder.fit_transform(y)

  tsne = TSNE(n_components=2, random_state=0)  # n_components means you mean to plot your dimensional data to 2D
  x_test_2d = tsne.fit_transform(x_std)

  print()

  markers = ('s', 'd', 'o', '^', 'v', '8', 's', 'p', "_", '2')
  color_map = {0: 'red', 1: 'blue', 2: 'lightgreen', 3: 'purple', 4: 'cyan', 5: 'black', 6: 'yellow', 7: 'magenta',
           8: 'plum', 9: 'yellowgreen'}
  for idx, cl in enumerate(np.unique(y)):

      plt.scatter(x=x_test_2d[y == cl, 0], y=x_test_2d[y == cl, 1], c=color_map[idx], marker=markers[idx],
              label=cl)
  plt.xlabel('X in t-SNE')
  plt.ylabel('Y in t-SNE')
  plt.legend(loc='upper left')
  plt.title('t-SNE visualization of test data')
  plt.show()

  • ScikitLearn's TSNE Results: enter image description here

  • You can also use PCA for plotting high dimensional data to 2D

  • Here is implementation of PCA.

  • Scikit Learn PCA: https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html

Jai
  • 3,211
  • 2
  • 17
  • 26