songyaochen / Enhanced-DETR-for-Yellow-Leaf-Disease-Detection

SSY340 final project from Chalmers Technology of University

Enhanced-DETR-for-Yellow-Leaf-Disease-Detection

SSY340 final project from Chalmers Technology of University

Explore the Poster · View the report

Requirements

• Python version: 3.8 (tested)
• CUDA version: 11.7 (tested)
• CUDA version: 11.5 (this version maybe conflict with build MSDeformAttn)

Set up environment

create conda environment

conda create -n env_name python=3.8

activate conda environment

conda activate env_name

install dependency

pip install -r requirements.txt

install MSDeformAttn package (required)

pip install models/ops

Download coco dataset

• http://images.cocodataset.org/zips/train2017.zip
• http://images.cocodataset.org/zips/val2017.zip
• http://images.cocodataset.org/annotations/annotations_trainval2017.zip

After download unzip all in path/to/coco
Folder structure should look like this

path/to/coco/
├── train2017/
├── val2017/
└── annotations/
    ├── instances_train2017.json
    └── instances_val2017.json

Backbone

ResNet-50

MobileNet

EfficiencyNet

Training

bash train_efficientnet.sh

content of train_efficientnet.sh

#!/usr/bin/env bash

set -x

python -u main.py \
    --output_dir "exps/effi-v2S_ddetr" \
    --backbone "efficientnet" \
    --batch_size 2

Note
Use arg --coco_path "path/to/coco" to update coco path
path/to/coco is coco dataset directory in previous step

Download model weight

download weight folder and move it inside repo folder (same level as main.py)

https://drive.google.com/drive/folders/1F4WzkFU0agqYoJIAwQ3FMZNb7-PQbuI1?usp=drive_link

In each model there are 3 files

File Name	Description
checkpoint0049.pth	model weight
log.txt	evaluation at each epoch on COCO train dataset
train.log	log during training

Inference

• Inference a image

python inference.py --backbone name_backbone  --inf_path path/to/image.png

• Inference a folder with multiple images

python inference.py --backbone name_backbone  --folder --inf_path path/to/folder

Note
replace name_backbone with one of the following:

name_backbone	Model using
resnet50	res-50_ddetr
efficientnet	effi-v2S_ddetr
mobilenet	mb-v3L_ddetr
swin	swin-T_ddetr

Inference with GUI

• Run backend using flask (This will load 4 model and 4 debug model. Config file backend.py if needed)

python backend

• Run fronend using streamlit

streamlit run frontend.py  

Video demo: https://www.youtube.com/watch?v=8VNTvP90f5Y

Run bench mark on google colab

https://colab.research.google.com/drive/1rxxCjrq1dEC8e7-bGUoSTMd1CkZ4CGwR

Explain file

Python file

File name	Description
main.py	train and evaluate model
inference.py	running inference on images
frontend.py	web UI to inference image using streamlit
backend.py	host a backend using Flask
draw_coco.py	draw lablel of COCO train dataset
draw_comparision.py	combine COCO label and result of detection as an image
benchmark.py	benchmark to know the FPS of a model
debug.py	Show attention weight of encoder and decoder
extract_train_time.py	extract training time of file weight/train.log
inspect_eval.py	get AP of each class at an evaluation
visualize.py	plot loss and AP of the training process

Bash file

File name	Description
benchmark_resnet.sh	example of running benchmark.py
train_dist*.sh	train model distributively on 2 GPU
train_*.sh	train model
view.sh	view the last line of weight/train.log