Deep Learning for Clustering

Code for project "Deep Learning for Clustering" under lab course "Deep Learning for Computer Vision and Biomedicine" - TUM. Depends on numpy, theano, lasagne, scikit-learn, matplotlib.

Contributors

• Mohd Yawar Nihal Siddiqui
• Elie Aljalbout
• Vladimir Golkov (Supervisor)

Related Papers:

This repository is an implementation of the paper : Elie Aljalbout, Vladimir Golkov, Yawar Siddiqui, Daniel Cremers "Clustering with Deep Learning: Taxonomy and new methods"

• arxiv: https://arxiv.org/abs/1801.07648

Usage

Use the main script for training, visualizing clusters and/or reporting clustering metrics

python main.py <options>

Option
-d DATASET_NAME, --dataset DATASET_NAME	(Required) Dataset on which autoencoder is to be trained trained, or metrics/visualizations are to be performed [MNIST,COIL20]
-a ARCH_IDX, --architecture ARCH_IDX	(Required) Index of architecture of autoencoder in the json file (archs/)
--pretrain EPOCHS	Pretrain the autoencoder for specified #epochs specified by architecture on specified dataset
--cluster EPOCHS	Refine the autoencoder for specified #epochs with clustering loss, assumes that pretraining results are available
--metrics	Report k-means clustering metrics on the clustered latent space, assumes pretrain and cluster based training have been performed
--visualize	Visualize the image space and latent space, assumes pre-training and cluster based training have been performed

Project Structure

Folder / File	Description
archs	Contains json files specifying architectures for autoencoder networks used. File mnist.json contains architectures for MNIST dataset. We use the second architecture for the reported results (command line argument -a 1)
coil, mnist	Contains the datasets COIL20 and MNIST respectively
logs	Output folder for logs generated by the scripts. Named by date and time of script execution
plots	Scatter plots showing the raw, pre-trained latent space, and the final latent space clusters
saved_params	Contains saved network parameters and saved representation of inputs in latent space
custom_layers.py	Custom lasagne layers, Unpool2D - which performs inverse max pooling by replicating input pixels as dictated by the filter size, and the ClusteringLayer - a layer that outputs soft cluster assignments based on k-means cluster distance
main.py	The main python script for training and evaluating the network
misc.py	Contains dataset handlers and other utility methods
network.py	Contains classes for parsing and building the network from json files and also for training the network

Autoencoder Builder

We've implemented a NetworkBuilder class that can be used to quickly describe the architecture of an autoencoder through a json file. The json specification of the architecture is a dictionary with the following fields

Field	Description
name	Name identifier given to the architecture, used for file naming while saving parameters
batch_size	Batch size to be used while training the network
use_batch_norm	Whether to use batch normalization for convolutional/deconvolutional layers
network_type	Type of network - convolutional or fully connected
layers	A list describing the encoder part of the autoencoder

Further, each item in the layers list is a dictionary with the following fields

Field	Description
type	Can be Input, Conv2D, MaxPool2D, MaxPool2D*, Dense, Reshape, Deconv2D
num_filters	For Conv2D/MaxPool2D/MaxPool2D*/Deconv2D layers this field specifies number of filters
filter_size	Dimensions of kernel for the above layers
num_units	For Dense layer number of hidden units
non_linearity	Non-Linearity function used at output of the layer
conv_mode	Can be used to specify the convolution mode like same, valid etc. for convolutional layers
output_non_linearity	If you want a different non linearity function at the output than the one which would be obtained by mirroring

Only the encoder part of the autoencoder needs to be specified, the decoder will be automatically generated by the class.

Example of a network description

{
    "name": "c-5-6_p_c-5-16_p_c-4-120",
    "use_batch_norm": 1,
    "batch_size": 100,
    "layers": [
      {
        "type": "Input",
        "output_shape":[1, 28, 28]
      },
      {
        "type": "Conv2D",
        "num_filters": 50,
        "filter_size": [5, 5],
        "non_linearity": "rectify"
      },
      {
        "type": "MaxPool2D*",
        "filter_size": [2, 2]
      },
      {
        "type": "Conv2D",
        "num_filters": 50,
        "filter_size": [5, 5],
        "non_linearity": "rectify"
      },
      {
        "type": "MaxPool2D*",
        "filter_size": [2, 2]
      },
      {
        "type": "Conv2D",
        "num_filters": 120,
        "filter_size": [4, 4],
        "non_linearity": "linear"
      }
    ]
  }

This would generate the network 50[5x5] 50[5x5]_bn max[2x2] 50[5x5] 50[5x5]_bn max[2x2] **120[4x4] 120[4x4]_bn **50[4x4] 50[4x4]_bn ups*[2x2] 50[5x5] 50[5x5]_bn ups*[2x2] 1[5x5]

Experiments and Results

We trained and tested the network on two datasets - MNIST and COIL20

Dataset	Image size	Number of samples	Number of clusters
MNIST	28x28x1	60000	10
COIL20	128x128x1	1440	20

Clustering was performed with two different loss functions -

• Loss = KL-Divergence(soft assignment distribution, target distribution) + Autoencoder Reconstruction loss , where the target distribution is a distribution that improves cluster purity and puts more emphasis on data points assigned with a high confidence. For more details check out the DEC paper [1].
• Loss = k-Means loss + Autoencoder Reconstruction loss

MNIST

Our network

Clustering space	Clustering Accuracy	Normalized Mutual Information
Image pixels	0.542	0.480
Autoencoder	0.760	0.667
Autoencoder + k-Means Loss	0.781	0.796
Autoencoder + KLDiv Loss	0.859	0.825

Other networks

Method	Clustering Accuracy	Normalized Mutual Information
DEC	0.843	0.800
DCN	0.830	0.810
CNN-RC	-	0.915
CNN-FD	-	0.876
DBC	0.964	0.917

Note: The commit b34743114f68624b5371cd0d4c059b141422902f gives upto 0.96 accuracy and 0.92 NMI on the MNIST dataset. We will include it to the main branch once we can get better results with the COIL architecture

Latent space visualizations

Pixel space

Autoencoder

Autoencoder Latent Space Evolution (video)

Autoencoder + KLDivergence

Autoencoder + KLDivergence Latent Space Evolution (video)

Autoencoder + k-Means

COIL20

Our network

Clustering space	Clustering Accuracy	Normalized Mutual Information
Image pixels	0.689	0.793
Autoencoder	0.739	0.828
Autoencoder + k-Means Loss	0.745	0.846
Autoencoder + KLDiv Loss	0.762	0.848

Other networks

Method	Clustering Accuracy	Normalized Mutual Information
DEN	0.725	0.870
CNN-RC	-	1.000
DBC	0.793	0.895

Latent space visualizations

Pixel space

Autoencoder

Autoencoder + k-Means

Autoencoder + KLDivergence