DenseNet implementation and training with TensorFlow 2.x

Requirements

• Internet connection for downloading CIFAR10 dataset and TensorFlow docker image.
• Docker and Nvidia runtime installed as described on TensorFlow website

Scripts in this repository

• densenet.py: Description of the DenseNet architecture
• train.py: Training and evaluation of the densenet model using the well-known CIFAR-10 dataset.
• run_train.py: Linux shell script that sets the arguments for train.py and executes it.
  • The python command can be executed directly if preferred, for example:

  python train.py --opt rms --epochs 140 --learnrate 0.001 --batchsize 125 --tboard ./tb_log --keras_hdf5 ./densenet.h5

• start_tf2.sh: Linux shell script that pulls latest TensorFlow nightly build and starts docker.

Instructions

1. Clone or download/unzip this repository to a folder.
2. Navigate into the folder created in Step 1.
3. Open a command shell/terminal and start the TensorFlow docker by running start_tf2.sh:

source ./start_tf2.sh

4. When the docker starts, execute run_train.sh like this:

source ./run_train.sh

References

1. Huang et al. "Densely Connected Convolutional Networks" (v5) Jan 28 2018.
2. Krizhevsky, Alex. "Learning Multiple Layers of Features from Tiny Images".