deeplab___'s Introduction

Steel Project

Kaggle Steel competition : https://www.kaggle.com/c/severstal-steel-defect-detection/overview Deeplab : https://github.com/tensorflow/models/tree/master/research/deeplab

Building dataset

First you need to download the data from the competition and organize your folder that way :

+ project
  + steel (code repo)
    + deeplab
      + ...
    + notebooks
      + ...
  + data
    + sample_submission.csv  
    + test_images/  
    + train.csv  
    + train_images/

Then, run from the steel repo

python3 deeplab/datasets/build_steel_data.py

This scripts will split the training set into train/val sets (default is 90%/10% split) and store the data into tfrecords.

Training

A local training job using xception_65 can be run with the following command:

# PROJECT_DIR should be the path to the project directory containing the code repo, the data, the models, and the experiments
VERSION_ID=0
MODEL=deeplabv3_pascal_train_aug

PATH_TO_INITIAL_CHECKPOINT=$PROJECT_DIR/models/$MODEL/model.ckpt
PATH_TO_TRAIN_DIR=$PROJECT_DIR/experiments/v$VERSION_ID
PATH_TO_DATASET=$PROJECT_DIR/data/tfrecord/

# From $PROJECT_DIR/steel/
python3 deeplab/train.py \
    --logtostderr \
    --training_number_of_steps=1000 \
    --train_split="train" \
    --model_variant="xception_65" \
    --atrous_rates=6 --atrous_rates=12 --atrous_rates=18 \
    --output_stride=16 --decoder_output_stride=4 \
    --train_crop_size="256,256" \
    --train_batch_size=1 \
    --dataset="steel" \
    --tf_initial_checkpoint=$PATH_TO_INITIAL_CHECKPOINT \
    --train_logdir=$PATH_TO_TRAIN_DIR \
    --dataset_dir=$PATH_TO_DATASET

Recommend Projects