SENet (Keras implementation)

---

New information

• We provide a trained SEResNeXt model (training data: cifar10)
  Google drive
  You can try this model in evaluate-cifar10.ipynb.

---

Naive implementation of SENet models in Keras.

• Transplanting https://github.com/taki0112/SENet-Tensorflow to Keras.
• Only SE-ResNext at this stage.

Prerequisites

• nvidia-docker environment

Environment constuction

• Build a docker image (on the root directory of the repository)

  $ docker build -t [tag name] -f docker/Dockerfile .
  

• Create a container using the image

  $ nvidia-docker run -it -v $PWD:/work [tag name]
  

Train a model

• Train a model with cifar10 data.

  (in the container) $ pwd
  /work
  (in the container) $ python train-cifar10.py
  

Note that this script is written in an insufficient way; use data generator in consideration of expansion to general image data). The training speed is slow. On a p3.2xlarge instance, it takes about 1.5 days.

Evaluate the model

• Launch a jupyter notebook.

  (in the container) $ bash launch_notebook.sh
  

• Execute evaluate-cifar10.ipynb notebook.

Results

• Accuracy plot of train/val.
   

• Loss plot of train/val.
   

• Accuracy for the test data.
  92.38%