Mask R-CNN for table detection training on google colab GPU

Checklist!!

• In custom.py updated class name to table(If not already edited).

• Also replace images and annotation json in customeimages/ train and val folders!

• Add the following into the code cell and execute.

!rm mask-rcnn-training/
!git clone https://github.com/Gerald2077/mask-rcnn-training
!pip install mrcnn
%cd mask-rcnn-training/

The above steps do the following to prepare jupyter notebook on colab-

• Clear the previous repo folder !rm mask-rcnn-training/

• Clone the repo, !git clone https://github.com/Gerald2077/mask-rcnn-training

• Install mask rcnn using pip !pip install mrcnn

• Navigate to the main directory %cd mask-rcnn-training/

Don't need this now,

import sys 
sys.path.insert(0, '/mask-rcnn-training')

Steps to view dataset using notebook

• Open inspect_custom_data.ipynb in google colab by providing github repo link.

• Select GPU from Edit -> Notebook Settings

• Do copy to drive action.

Steps to train model

• Run this in a separate code block after making sure GPU is enabled in EDIT->Notebook Settings & you have the initial steps like deleting folder, cloning repo & making sure a couple of images from the dataset are viewable with mask, before proceeding with training(taking 50 mins at the moment on GPU)

!python3 custom.py train --dataset=customImages/ --weights=coco

• After training we will save the generated h5 file to our local machine & use it for inference using CPU. We can also verify the same with another notebook, where we need to upload the h5 file & update the filename in that notebook as well.

• To download the h5 weights file, do the following-

Keep updating the same code block with below lines & repeat individually,
%cd ..
%cd logs/
%cd {into latest table, check the training log for h5 file name & folder it is enclosed in.}
#then run the below,
from google.colab import files
files.download('mask_rcnn_table_0010.h5')

Saving weight file on google drive and download

• Mounting google drive

from google.colab import drive 
drive.mount('/content/gdrive')

• Copy the file to google drive %cp mask_rcnn_table_0010.h5 /content/gdrive/'My Drive'/