DenseNet implementation using Tensorflow 2

We implemented Densenet using squeeze and excitation layers in tensorflow 2 for our experiments.

For more information about densenet please refer to the original paper.

Model Usage

from densenet import densenet_model
model = densenet_model(classes=n_clases)

you can disable the se layers by setting the argument with_se_layers to false. Changing nb_layers lenght will change the number of dense layers.

References

Inspired by flyyufelix keras implementation (https://github.com/flyyufelix/DenseNet-Keras).

For more information about densenet please refer to the original paper (https://arxiv.org/abs/1608.06993).

Dependencies and Installation

The code has been tested on Ubuntu 18.04 with Python 3.6.8 and TensorFlow 2
The two main dependencies are TensorFlow and Pillow package (Pillow is included in dependencies)
To install densenet lib run python setup.py install

Repository Structure

The repository organized as follows.

data directory contains scripts for dataset downloading and used as a default directory for datasets.
densenet is the library containing the model itself.
src/datasets logic for datasets loading and processing.
src/scripts directory contains scripts for launching the training. train/run_train.py and eval/run_eval.py launch training and evaluation respectively. tests folder contains basic training procedure on small-valued parameters to check general correctness. results folder contains .md file with current configuration and details of conducted experiments.

Training

Training and evaluation configurations are specified through config files, each config describes single train+eval evnironment.

Run the following command to run training on <config> with default parameters.

$ ./bin/densenet --mode train --config <config>

<config> = cifar10 | cifar100

Evaluating

To run evaluation on a specific dataset

$ ./bin/densenet --mode eval --config <config>

<config> = cifar10 | cifar100

Results

In the results/<ds> directory you can find the following results of training processes on a specific dataset <ds>:

In the /results directory you can find the results of a training processes using a <model> on a specific <dataset>:

.
├─ . . .
├─ results
│  ├─ <dataset>                            # results for an specific dataset.
│  │  ├─ <model>                           # results training a <model> on a <dataset>.
│  │  │  ├─ models                         # ".h5" files for trained models.
│  │  │  ├─ results                        # ".csv" files with the different metrics for each training period.
│  │  │  ├─ summaries                      # tensorboard summaries.
│  │  │  ├─ config                         # optional configuration files.
│  │  └─ └─ <dataset>_<model>_results.csv  # ".csv" file in which the relationships between configurations, models, results and summaries are listed by date.
│  └─ summary.csv                          # contains the summary of all the training
└─ . . .

where

<dataset> = cifar10 | cifar100
<model> = densenet

To run TensorBoard, use the following command

$ tensorboard --logdir=./results/<ds>/summaries/

Environment

Quickstart

$ ./bin/start [-t <tag-name>] [--sudo] [--build]

<tag-name> = cpu | devel-cpu | gpu | nightly-gpu-py3

or setup and use docker on your own

Build the docker image,

$ docker build --rm -f dockerfiles/tf-py3-jupyter.Dockerfile -t <name>:latest .

and now run the image

$ docker run --rm -u $(id -u):$(id -g) -p 6006:6006 -p 8888:8888 <name>:latest

Visit that link, hey look your jupyter notebooks are ready to be created.

If you want, you can attach a shell to the running container

$ docker exec -it <container-id> /bin/sh -c "[ -e /bin/bash ] && /bin/bash || /bin/sh"

And then you can find the entire source code in /develop.

$ cd /develop

To run TensorBoard, use the following command (alternatively python -m tensorboard.main)

$ tensorboard --logdir=/path/to/summaries

okason97/DenseNet-Tensorflow2