Tensorflow Implementation of Convolutional Autoencoder

Project details

Autoencoder is used as powerful tool to extract important features from the input data and then use the encoded (latent) represntation as a new input for classification model. The nice thing about autoencoder is that we don't need labels to train it and it belong to unsupervised learning technically. In simple words, it does the dimensionality reduction while preserving the vital features that are transformed into a latent space. For an image classification/detection task, a particular type of autoencoder called convolutional autoencoder (CAE) is used.

This project demonstrates how CAE can be implemented in tensorflow framework as a feature extractor for an image classification task. It also contains a sample pretrained model which is trained on Fashion-MNIST Dataset.

Getting started

1. Make sure python 3.x is installed
2. To install other dependencies:

pip install -r requirements.txt

3. To test a pretrained autoencoder model (like examples in out/autoencoder folder) by feeding a random image from validation set, run

python autoencoder.py test -m ./out/autoencoder/sample/autoencoder_latest.meta

4. To train a new autoencoder model, run

python autoencoder.py train

5. To continue training a pretrained autoencoder model, run

python autoencoder.py train -m ./out/autoencoder/sample/autoencoder_latest.meta

6. To get these details in terminal, simply execute

python autoencoder.py -h

7. To test a pretrained classifer model by (like examples in out/classifier folder) by feeding a random image from validation set, run

python classifier.py test -m ./out/classifier/sample_pretrained_trainable/classifier_best.meta

8. To train a new classifier model with pretrained autoencoder weights, run

python classifier.py train start -a ./out/autoencoder/sample/autoencoder_best.meta

To make autoencoder layers trainable,

python classifier.py train start -a ./out/autoencoder/sample/autoencoder_best.meta -t

with random weights replacing pretrianed autoencoder (only using autoencoder model grpah)

python classifier.py train start -a ./out/autoencoder/sample/autoencoder_best.meta -r

9. To continue training a pretrained classifier model, run

python classifier.py train continue -m ./out/classifier/sample_pretrained_trainable/classifier_latest.meta

10.To get these details in terminal, simply execute

python classifier.py -h

Results

1. When the pretrained autoencoder model is tested, you should see the output something like this
2. The training and validation loss convergence of the trained autoencoder model is shown below:
3. When the pretrained classifier model is tested, you should see the output something like this
4. The Training convergence of classifier models with pretrained autoencoder weights and random weights:

Instructions

1. data folder contains the fashion-MNIST dataset which can be replaced with custom dataset as well.
2. out folder gets automatically created when the model is getting trained. It contains the latest and the best model checkpoints. It also stores the log necessary for tensorboard visualization.
3. autoencoder.py contians the main code for training and testing the autoencoder model.
4. autoencoder_model_db.py file contains some example model architectures of CAE. Please modify or add new definitions here. Please esure that autoencoder.py file is modified accordingly to call the new function. Please name the layers with proper prefix as shown in example model as it is required to import the encoder during classifier model creation.
5. fashion_mnist_utils.py contains the functions related to pre-processing and loading of fashion-MNIST dataset.
6. classifier.py contians the main code for training and testing the classifier model.
7. classifier_model_db.py file contains some example model architectures of classifier by importing the encoder graph from autoencoder mdoel. Please modify or add new definitions here. Please esure that classifier.py file is modified accordingly to call the new function.
8. To see the training progress/summary and also compare between different models, use tensorboard for visualization by running

tensorboard --logdir='./out/'