keras-object-recognition

Minimalist Keras implementation for deep learning object recognition.

Installation

(optional) Create a new anaconda environment:

conda create --name keras-object-recognition python=3
source activate keras-object-recognition

Install the requirements:

pip install -r requirements.txt

Make sure keras uses tensorflow backend. Edit ~/.keras/keras.json like this:

{
    "floatx": "float32",
    "epsilon": 1e-07,
    "backend": "tensorflow",
    "image_dim_ordering": "tf"
}

Training

Train a model with:

python train.py

Default options (see train.py for the available options):

1. --savepath results
2. --dataset cifar10
3. --net_type resnet
4. --depth 16
5. --widen 1
6. --weight_decay 5e-4
7. --randomcrop 4
8. --randomcrop_type reflect
9. --hflip (pass to remove hflip)
10. --epoch_max 200
11. --epoch_init 0
12. --bs 128
13. --nthreads 2
14. --lr 0.1
15. --lr_decay 0.2
16. --lr_schedule 60 120 160
17. --momentum 0.9
18. --nesterov (pass to remove nesterov)

Tensorboard Visualization

In a new terminal, call tensorboard and use the value of --savepath as logdir:

tensorboard --logdir=results

Open your internet browser at localhost with the provided port number (like 6006), as follows: http://localhost:6006/.