Segformer

Here is the code of my project on segformer for the course of Deep Learning

I used the Hugging Face repository called Transformers and modified its code to test my variants. In order not to break the installation process (which is necessary for you to test the code and for me, since I had to reinstall the repository with pip so that the module recognizes my modifications), I did not remove or clean up unnecessary files. This document explains how to use the code and where to find the modifications that document my work.

Files to Examine

The important file is the notebook located in the hf_proj folder, named sergformer.ipynb. This is where I trained and evaluated my models.

To review the changes I made, navigate to the directory hf_proj/Modified_File, which contains copies of the final versions of the modified files. The added code sections are delimited by the comment: /!\ added code by student

I mainly modified the SegformerDecodeHead class in the file modeling_segformer.py (around line 675) as well as the SegformerConfig class in the file configuration_segformer.

Access to the Dataset

The dataset is freely accessible on Hugging Face, but you will need a Hugging Face account and a token (which is free) to access it. Since I do not know if you have an account, I have provided a token that will be deactivated at the end of the semester. If you need to run the notebook, you will need to copy and paste the token (instructions will be provided in the notebook).

token = see the report

Installation

Download my repository and follow the instructions below.

• Build a conda environment:

  • conda create --name seg
  • conda activate seg

• Install transformers using conda:

  • conda install conda-forge::transformers

• Go to the directory HF_SEGFORMER and then run:

  • pip install transformers[torch]
  • pip install pillow
  • pip install torchvision
  • pip install evaluate
  • pip install .

• If you prefer an alternative method, there is a file requir.txt in the hf_proj directory that lists all the dependencies.

Then download the weights with the google drive link from the report. Decompress the file and put each directory in the directoy HF_SEGFORMER/hf_proj.