MaskRCNN_Tensorflow_Docker

Integration of FastMaskRCNN + Tensorflow + Nvidia-docker. (Tested in Ubuntu 16.04)

• Creates docker image "user/tensorflow_gpu_mrcnn" using Tensorflow_gpu + FastMaskRCNN dependencies
• Creates docker container "user-mrcnn_tf1.1" from "user/tensorflow_gpu_mrcnn"
• Clones FastMaskRCNN code inside MaskRCNN_Tensorflow_Docker/MRCNN/ (host)
• Mounts MRCNN/ (host) at /home/user/ (docker) inside "user-mrcnn_tf1.1" container. (Host and Docker are now sharing the same code i.e. FastMaskRCNN. So code changes can be made at host side and code can be run inside the docker container)
• Downloads/Copies datasets required for FastMaskRCNN inside FastMaskRCNN (visible at both host and docker due to mounting)
• Runs FastMaskRCNN code to:
  • Generate annotations
  • Train the network

Prerequisite

• Install Docker

Installation

$ cd MaskRCNN_Tensorflow_Docker
$ ./install.py

Training with CPU:
$ sudo nvidia-docker exec -it user-mrcnn_tf1.1 bash -c "cd MRCNN/FastMaskRCNN; export CUDA_VISIBLE_DEVICES= ; python train/train.py"

Training with GPU:
$ sudo nvidia-docker exec -it user-mrcnn_tf1.1 bash -c "cd MRCNN/FastMaskRCNN; python train/train.py"

Notes

• Using Tensorflow 1.1 (due to issue #88 with TF 1.2)
• Tested on
  • Ubuntu 16.04.1 x86_64
  • Kernel 4.8.0-56-generic
  • Cuda 8.0.61
  • CuDNN 5.1.10

Acknowledgment

• FastMaskRCNN
• Tensorflow
• Nvidia-docker