Image classification with Tensorflow

This practical exercise aims to make me as a learner understand several topics when using Tensorflow to do image classification. Since there are so many available codes doing the same thing, including the Tensorflow tutorials for MNIST Tensorflow tutorials for CIFAR10 and others, it seems to me they are either pure Tensorflow implementation with very high-level maturity and all those non-trivial advanced coding frameworks, or just using Tensorflow to glue high-level APIs like Keras. Therefore I started this exercise to get familiar with all those basic tools Tensorflow provide to make me have a solid idea how things work, which might be necessary while reading more and more Tensorflow implementations. Topics include:

• To learn how to package data into tfrecord files
• Using tf.data.Dataset to implement to data pipeline for training, validation, and testing sets with k-fold cross validation
• To model neural network with tf.contrib.slim, which is an easy-to-use and clean package of common neural layers
• To build a graph for model inference, loss & predection computation, and the optimization process (or trainer).
• To save a trained model as checkpoint files, with inputs as tf.placeholder and necessary operators stored by tf.add_to_collection, and later loaded back to make inference.

Versions Tested

• Python 3.5 or 3.6
• CUDA 9.0
• Tensorflow-GPU 1.12 (with cuDNN 7.1.4) or 1.9 (with cuDNN 7.0.5)
• NumPy 1.15

Usage

Currently two image datasets (MNIST and CIFAR10) and their corresponding models are implemented.

It is suggested to use the virtual environment wrapper for python packages. So the following commands are under such virtual environment, which runs the Python 3 with simply python commands

First download the dataset, note the argument dataset can be either cifar10 or mnist.

python download_data.py --dataset=cifar10

After this step, train.tfrecord and test.tfrecord will appear in data folder.

Then train the dataset, note the argument dataset should be chosen the same as previous step. Arguments also include batch size (batch_size), initial learning rate (init_lr), number of epochs (num_epochs), which are free to choose. By default the checkpoints will be stored in train folder, which is also adaptable with argument log_dir.

python train.py --dataset=cifar10 

During the training, you can use Tensorboard on the other terminal to inspect the training details.

tensorboard --logdir=train

By default it will show up at http://localhost:6006.

After training, the model graph and checkpoints are stored in log_dir. Finally test the model with test dataset. Use argument cpkt_path to specify the model we would like to test.

python test.py --dataset=cifar10 --ckpt_path=train/cifar10_bs_50_lr_0.1_l2s_1e-08/model_epoch20.ckpt

For sanity check, CIFAR10 can get around 70% accuracy with default model and hyperparameters.

Note: Running with CUDA-based GPU requires no extra changes in codes or running procedures. Refer to this for checking if CUDA driver and cuDNN is properly installed and configured.

Future work

• To learn how to freeze the model into .pb file
• To visualize training versus validation for each epoch

Useful references

• Finetune AlexNet with Tensorflow, an implementation with nicely structured training graph as an implementation example
• TensorFlow: saving/restoring and mixing multiple models, a clear explanation on how model checkpoints are saved and stored
• Stanford CS 20: Tensorflow for Deep Learning Research, the course which aims to explore more details about the practical aspects of Tensorflow