The overall pipeline consists of two major steps:
- Feature learning
- Clustering and analysis
All setups are configured through .task JSON files located in tasks directory. Results are stored in results folder,
TensorBoard logs and checkpoints stored in logs folder.
Three types of feature learning are provided: convolutional autoencoder, pre-trained CNN, vector quantization based on K-means.
Default set of layers includes convolutional, batch normalization and max-pooling layers. Optionally dense layers can be added in the middle of the model. Available options:
- Random input rotation during training
- Max pooling with or without argmax (preservation of positions during decode). Argmax works on GPU only.
- Same weights for encoder and decoder
- Any number of dense layers
For instance, MNIST dataset is split into two: 1-5 digits and 6-0. The CNN classifier is trained on the first one (1-5), than feature-learning part of the model is evaluated on the second one (6-0). Feature-learning part - everything but the last Softmax layer.
Feature learning described in this paper. Based on k-means' centroids learning.
Output includes t-SNE embedding for three types of clustering (K-means, DBSCAN, HDBSCAN) and ground truth coloring. Here goes the example of ground truth scatter plot for MNIST dataset (output from Quick start example):
3D PCA reduction using Tensorboard Embedding:
Python 3 and the following libraries are required. Installation using pip:
pip install tensorflow, matplotlib, scikit_image, hdbscan, imageio, numpy, scipy, Pillow, skimage, scikit_learn
To run convolutional autoencoder on MNIST dataset:
python cnn_feature_cluster.py --task=examples/cnnAE
After training is finished, test results will be saved in results folder. The feature learning quality is evaluated using clustering
and V-measure entropy score.
To start tensorboard:
tensorboard --logdir=logs/train
Tensorboard supports input/output samples, computational graph, loss/accuracy plots, feature vector embeddings (projector)