tf_training

This repository uses scripts provided by tensorflow's object detection api to train Mask R-CNN with Mobilenet V1 as Backbone on the COCO Dataset.
Additionally: As many local machine have limited computing capacity (especially GPU Memory) I added a script that automatically restarts the training and evaluation process from the last saved checkpoint if training got killed by some out of memory / oom or any other error.

Note: You may need to install several dependencies (you'll face them while running the scripts)

Getting Started

• inside dataset/ run bash download_and_preprocess_mscoco.sh
• inside the wished configs/*.config modify all absolute paths to fit your directory structure
• inside train.sh modify the variables MODEL, TF_DIR and NUM_GPUS
• run train.sh
• watch the numbers get crunched!