5 Best Ways to Use TensorFlow and Pre-Trained Models for Data Visualization in Python

Rate this post

πŸ’‘ Problem Formulation: How do we employ the robustness of TensorFlow and the efficiency of pre-trained models for visualizing datasets in Python? For developers and analysts, the input is their data in any form, such as images or text. The desired output is a visual representation that reveals underlying patterns or features to aid in interpretation and decision making.

Method 1: Using Feature Vectors from Pre-Trained Models for Dimensionality Reduction

Data visualization can often be enhanced by reducing the dimensionality of complex datasets. TensorFlow, with its library of pre-trained models, can be used to extract feature vectors from data, which can then be visualized using techniques like PCA or t-SNE.

Here’s an example:

import matplotlib.pyplot as plt
import tensorflow as tf
from sklearn.manifold import TSNE

# Load pre-trained model
model = tf.keras.applications.VGG16(include_top=False, input_shape=(224, 224, 3))

# Assume 'data' is a batch of images with shape (batch_size, 224, 224, 3)
features = model.predict(data)

# Use t-SNE for dimensionality reduction
tsne = TSNE(n_components=2, random_state=0)
reduced_features = tsne.fit_transform(features.reshape(len(features), -1))

# Plotting the results
plt.scatter(reduced_features[:, 0], reduced_features[:, 1])

The code snippet starts with importing necessary modules and loading a pre-trained VGG16 model. It then uses the model to predict features from a batch of images and applies t-SNE to reduce these features to two dimensions. Finally, it plots these reduced features using matplotlib.

Method 2: Visualizing CNN Filters

Convoluted Neural Networks (CNNs) filters can be visualized to understand what features your network has learned. TensorFlow provides functions to access the filters and layers within pre-trained models which can then be visualized using Python’s plotting libraries.

Here’s an example:

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

# Load a pre-trained model
model = tf.keras.applications.VGG16(weights='imagenet', include_top=False)

# Get the weights of the first convolutional layer
filters, biases = model.layers[1].get_weights()

# Normalize filter values between 0 and 1 for visualization
f_min, f_max = filters.min(), filters.max()
filters = (filters - f_min) / (f_max - f_min)

# Plot the first few filters
n_filters = 6
for i in range(n_filters):
    # Get the filter
    f = filters[:, :, :, i]
    # Plotting each channel separately
    for j in range(3):
        plt.subplot(n_filters, 3, i * 3 + j + 1)
        plt.imshow(f[:, :, j], cmap='gray')

The example loads the VGG16 pre-trained model, obtains the weights of the first convolutional layer, normalizes these weights, and plots the first six filters of three channels in grayscale to visualize the kinds of features the CNN is detecting.

Method 3: Activations Visualization

Visualizing activations can provide insights into which features activate certain layers of a neural network. TensorFlow makes it possible to compute and visualize the activations of specific layers in response to certain input data, thus enabling an investigative look into layer responses.

Here’s an example:

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

# Load a pre-trained model
model = tf.keras.applications.VGG16(weights='imagenet')

# Define a new model to output features from the first block
layer_outputs = [layer.output for layer in model.layers[1:3]]
activation_model = tf.keras.models.Model(inputs=model.input, outputs=layer_outputs)

# Suppose 'img' is a preprocessed input image ready for the model
activations = activation_model.predict(img)

# Visualize the first layer activation
first_layer_activation = activations[0]

# Plot the activation of the first channel
plt.matshow(first_layer_activation[0, :, :, 0], cmap='viridian')

This demonstrates how to create a new model that outputs the activations from the first two layers of VGG16 when given an input image. It predicts the activations and visualizes the activation map of the first channel of the first layer.

Method 4: Visualizing Embeddings with TensorBoard

TensorBoard provides powerful tools for visualization, including the embedding projector. You can visualize high-dimensional embeddings, which is particularly useful for understanding deep learning models and the relationships between data points.

Here’s an example:

import tensorflow as tf
import numpy as np
from tensorboard.plugins import projector

# Assume 'embeddings' is your embeddings matrix, and 'metadata.tsv' contains the labels
embeddings = tf.Variable(np.random.randn(100, 256), name='word_embeddings')
labels = [f'Word {i}' for i in range(100)]

# Set up the config
config = projector.ProjectorConfig()
embedding_config = config.embeddings.add()
embedding_config.tensor_name = embeddings.name
embedding_config.metadata_path = 'metadata.tsv'

# Save the metadata file needed for TensorBoard
with open(embedding_config.metadata_path, 'w') as meta:
    for label in labels:

# Use the TensorFlow session to save the model
sess = tf.compat.v1.InteractiveSession()
saver = tf.compat.v1.train.Saver([embeddings])
saver.save(sess, 'embeddings.ckpt')
projector.visualize_embeddings(tf.summary.FileWriter('.'), config)

This code snippet initializes an embeddings matrix as a tf.Variable, creates a list of labels, and sets up the projector config with these embeddings and labels. It then writes the inputs to a metadata file, initializes a TensorFlow session, and uses a Saver to save the model checkpoint. Finally, TensorBoard’s projector is used to visualize the embeddings.

Bonus One-Liner Method 5: Quick Plot with Seaborn

For a straightforward and effective visualization, TensorFlow’s output can be plotted as statistical graphics using Seaborn, a Python data visualization library that’s built on top of matplotlib.

Here’s an example:

import seaborn as sns
import tensorflow as tf

# Assume 'data' is your pre-processed dataset ready for the model
model = tf.keras.applications.VGG16(include_top=False)
features = model.predict(data)

sns.pairplot(pd.DataFrame(features.reshape(len(features), -1)), diag_kind='kde')

This snippet flattens the features extracted by the VGG16 model and visualizes pairwise relationships in a dataset using Seaborn’s pairplot function, plotting a kernel density estimate (KDE) on the diagonal.


  • Method 1: Dimensionality Reduction with t-SNE. Strengths: Reveals clustering and patterns effectively. Weaknesses: Can be computationally expensive for large datasets.
  • Method 2: CNN Filters Visualization. Strengths: Provides intuitive insights into the learned features. Weaknesses: Limited to certain types of layers and may not convey complete information about the network’s operation.
  • Method 3: Activations Visualization. Strengths: Shows how different inputs activate network layers. Weaknesses: Can be overwhelming with deep networks and requires selection of relevant layers.
  • Method 4: Embeddings with TensorBoard. Strengths: Offers interactive visualization and exploration of embeddings. Weaknesses: Initial setup may require some effort and is overkill for simple visualizations.
  • Method 5: Quick Plot with Seaborn. Strengths: Fast and easy to implement. Weaknesses: May not be suitable for highly complex visualizations needed for deep learning models.