AMLS_II_assignment23_24

• Name: Zhaoyan LU
• Student ID: 23049710

Contents

• Overview
• Repo Structure
• Requirements
• Datasets
• Usage
• Models

Overview

This repo is for the final assignment of ELEC0135 Applied Machine Learning Systems II (23/24). This project aims at solving the Cassava Leaf Disease Classification task using fine-tuned pre-trained models or Mixtures of Experts (MoE).

Repo Structure

$ tree
.
├── A
│   ├── __init__.py
│   ├── config.json                             # training config
│   ├── confusion_matrix_of_base_model.png
│   ├── constants.py
│   ├── dataset.py                              # load datasets
│   ├── draw_confusion_matrix.py
│   ├── draw_learning_curves.py
│   ├── learning_curves.png
│   ├── logger.py
│   ├── model                                   # core module of Mixtures of Experts
│   │   ├── __init__.py
│   │   ├── configuration_moe.py                # MoE model configurations
│   │   └── modeling_moe.py                     # MoE model; See https://huggingface.co/docs/transformers/custom_models#sharing-custom-models
│   ├── moe.py                                  # entrypoint of moe action; train and test
│   ├── pretrained.py                           # entrypoint of pretrained action; train and test
│   ├── y.txt
│   └── y_pred.txt
├── Datasets
│   └── prepare.py                              # format dataset to the imagefolder format; See https://huggingface.co/docs/datasets/v2.0.0/en/image_process#imagefolder
├── Makefile
├── README.md
├── environment.yml
├── main.py                                     # entrypoint of the assignment
└── requirements.txt

Requirements

• Miniconda / Anaconda

• Nvidia GPU with cuda 11.8

• environment.yml

$ cat environment.yml
name: amls_ii-final-uceezl8
channels:
  - defaults
dependencies:
  - python=3.9
  - black                               # formatter
  - scikit-learn                        # confusion matrix
  - pandas                              # load dataframe
  - ipython                             # debugging
  - pillow                              # read images
  - torchvision                         # pre-processing
  - pytorch                             # build models
  - pytorch-cuda=11.8                   # pytorch for GPU
  - matplotlib                          # draw figures
  - pip:
    - -r requirements.txt

• requirements.txt

$ cat requirements.txt
# download the dataset from kaggle
kaggle
# preprocess
datasets
# metrics
evaluate
# fine-tuning and build custom models
transformers[torch]==4.39.3
# download and upload pretrained models
huggingface_hub
ipykernel

Create conda env:

$ make create-env
# or
$ conda env create -f environment.yml

IMPORTANT: If you want to load pre-trained models in Hugging Face, you need your own access tokens.

$ huggingface-cli login

Datasets

We are using the dataset provided by the competition. If you want to download the dataset, you may need to join this competition first.

# Step 0: enter conda env
$ conda activate amls_ii-final-uceezl8
# Step 1: download the dataset and extract it
$ make dataset
# or
$ cd Datasets
$ kaggle competitions download -c cassava-leaf-disease-classification
$ unzip -q cassava-leaf-disease-classification.zip
$ python prepare.py

Usage

• all options

$ python main.py --help
usage: main.py [-h] [-d] [-v] {pretrained,moe,info} ...

AMLS II Final Assignment

positional arguments:
  {pretrained,moe,info}
                        actions provided

options:
  -h, --help            show this help message and exit
  -d, --debug           Print lots of debugging statements
  -v, --verbose         Be verbose

info

$ python main.py info
-------------------------------------
|      AMLS II Assignment 23-24     |
|         Name: Zhaoyan Lu          |
|        Student No: 23049710       |
-------------------------------------

pretrained

$ python main.py pretrained --help
usage: main.py pretrained [-h] [--seed [S]] [--batch [B]] [--epoch [E]]
                          [--workers WORKERS] [--save] [--push_to_hub]
                          [--config CONFIG] [--dataset DATASET]
                          [--model_name MODEL_NAME]

options:
  -h, --help            show this help message and exit
  --seed [S]            random seed; default: 42
  --batch [B]           batch size per GPU; default: 70
  --epoch [E]           number of train epochs; default: 16
  --workers WORKERS     number of workers per GPU; default: 5
  --save                save model
  --push_to_hub         push model to huggingface
  --config CONFIG       path to config.json; default: A/config.json
  --dataset DATASET     path to dataset imagefolder; default:
                        Datasets/imagefolder
  --model_name MODEL_NAME
                        pretrained model name from huggingface; default:
                        facebook/convnextv2-base-1k-224

moe

$ python main.py moe --help
usage: main.py moe [-h] [--seed [S]] [--batch [B]] [--epoch [E]]
                   [--workers WORKERS] [--save] [--push_to_hub]
                   [--config CONFIG] [--dataset DATASET]
                   [--model_name MODEL_NAME]

options:
  -h, --help            show this help message and exit
  --seed [S]            random seed; default: 42
  --batch [B]           batch size per GPU; default: 70
  --epoch [E]           number of train epochs; default: 16
  --workers WORKERS     number of workers per GPU; default: 5
  --save                save model
  --push_to_hub         push model to huggingface
  --config CONFIG       path to config.json; default: A/config.json
  --dataset DATASET     path to dataset imagefolder; default:
                        Datasets/imagefolder
  --model_name MODEL_NAME
                        pretrained model name from huggingface; default:
                        facebook/convnextv2-base-1k-224

Models

Pretrained

• vit-base-patch16-224-finetuned
• resnet-101-finetuned
• resnet-152-finetuned

Mixtures of Experts

• MoE
• SwitchGate
• Expert 1
• Expert 2
• BaseModel