- Internet connection for downloading CIFAR10 dataset and TensorFlow docker image.
- Docker and Nvidia runtime installed as described on TensorFlow website
- 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.
- Clone or download/unzip this repository to a folder.
- Navigate into the folder created in Step 1.
- Open a command shell/terminal and start the TensorFlow docker by running start_tf2.sh:
source ./start_tf2.sh
- When the docker starts, execute run_train.sh like this:
source ./run_train.sh
- Huang et al. "Densely Connected Convolutional Networks" (v5) Jan 28 2018.
- Krizhevsky, Alex. "Learning Multiple Layers of Features from Tiny Images".