younader / vesuvius-first-letters

Vesuvius First Letters

This repository contains the second place solution to the Vesuvius First Letters challenge.

The approach uses an I3D architecture to detect ink from within the scrolls

How to use

Installation

I provide an image of my environment in the requirements.txt, I believe only the following packages need to be installed

pytorch-lightning   
typed-argument-parser   
segmentation_models_pytorch   
albumentations   
warmup_scheduler
wandb  

Training

Download the data and appropriate segments (instructions here).
Unzip the labels folder and place each {segmentid}_inklabel.png in its appropriate segment folder.

Adjust the CFG class with your compute in 64x64_256stride_i3d.py, these are the typical configs used in the kaggle competition.

python 64x64_256stride_i3d.py

Inference

the inference script runs a trained model , for more info about the arguments check the InferenceArgumentParser class in the inference script. a pretrained checkpoint is available here

e.g: python inference.py --segment_id 123 --model_path 'model.ckpt'

vesuvius-first-letters's Issues

Inferencing

Hi Youssef,

Thank you for providing your code. I tried to regenerate the mask of the segment with ID "20230826135043" using the provided inference script and model checkpoint. I end up with this:

This looks different than what was posted here. Does the default parameter of the inference script match the model checkpoint?

Training

Hi Youssef,

It seems like you are training the I3d model without pretrained weights based on the 64x64_256stride_i3d.py script and not using the provided fragments from the Kaggle challenge, can you please clarify why?