efficient-det 0.1.3

Efficient-Det Implementation in Keras

EfficientDet

EfficientDet Implementation in Keras focused on clean code and readability. Training will be logged with Tensorboard. To take a look at the training progress do: tensorboard --logdir logs This repo also includes the option of using wandb.ai for experiment tracking.

Installation

Via PIP (recommended)

Split into 3 ways to install. This is due to the way tensorflow needs to be installed to correctly work with CUDA. The first installation does not include tensorflow and is recommended to use.

pip install efficient-det

You can include a [cpu] or [gpu] tag to include the respective tensorflow version. Includes tensorflow dependency:

Via Docker

Runs with tensorflow:2.3.0-gpu. Depending on system CUDA version you might need to use another version. See this for more info.

Install docker and nvidia container toolkit on host system

sudo apt-get install -y docker.io nvidia-container-toolkit

Build Docker Image

sudo docker build . -t edet

From Source

1. Clone Repository

git clone [e-mail]:zeboz100/efficientdet.git

2. Build it

pip install -r requirements.txt

Usage

PIP

Training

python3 -m efficient_det.run_training --dataset_path /path/to/dataset

Run a Hyperparameter Search:

python3 -m efficient_det.run_hyper_parameter_search --dataset_path 
/path/to/dataset --num_tries 100 --gpus_per_trial 0.5

Source

Execute all commands in efficientdet/

Set PYTHONPATH :

export PYTHONPATH="$PWD/src"

Training

python3 src/efficient_det/train.py --dataset_path /path/to/dataset/

Hyperparameter Search:

python3 src/efficient_det/train.py --dataset_path /path/to/dataset/

Docker

Run Container

sudo docker run --gpus all -it edet bash

and then proceed with the PIP instructions.

To run all tests

python3 -m unittest

To test loaded model

You can test the loaded model via notebook or from a script.

• Notebook in examples/visualize_rsults.ipynb

• You need to set dataset path

• You need to set path to trained model

Execute Script from efficientdet/

python3 example/visualize_results.py