DataScienceBowl2018-25th
This is the Solution for data science bowl 2018 challenge
the result for private LB is 0.557 at 39/3634(epoch 64) and 0.498 at 156/3634(epoch 65),
Because we have missed the merge deadline we have to make 2 different submit. the huge gap between two epoch may comes from inference bug
The code is based on mateerport's mask rcnn：https://github.com/matterport/Mask_RCNN
The main improvement comes from
- better roi align implementation which is modified from tensorpack faster rcnn
- large separable convolution from Light Head Rcnn Paper
- strong image augmentation especially random scale crop
- using clustering to select proper CV set
- divide large picture into small part during inference
- finetuning the trainging schedule and mask rcnn configuration
Dependencies
- Python 3; TensorFlow >= 1.5.0;keras>=2.0.5
- Pre-trained ResNet model from keras pretrain model .
- data science bowl data. It assumes the following directory structure:
DIR/
stage1_train/
images/
masks/
stage1_test/
images/
stage2_test/
images/
Usage
The Command line tool hasn't been test yet, if you encounter any bug, see main.ipynb as conference
Train a new model starting from ImageNet weights using train dataset (which is stage1_train minus validation set)
python3 nucleus.py train --dataset=/path/to/dataset --subset=train --weights=imagenet
Train a new model starting from specific weights file using the full stage1_train dataset
python3 nucleus.py train --dataset=/path/to/dataset --subset=stage1_train --weights=/path/to/weights.h5
Resume training a model that you had trained earlier
python3 nucleus.py train --dataset=/path/to/dataset --subset=train --weights=last
Generate submission file from stage1_test images
python3 nucleus.py detect --dataset=/path/to/dataset --subset=stage1_test --weights=<last or /path/to/weights.h5>