Demo for VRDL HW2

Timing Benchmark

Colab Demo Link: https://colab.research.google.com/drive/1ox_6ci-s-CTuOB7bGSD-VMFPJ_y45LEQ?usp=sharing
Timing: 0.083s per image

References

yolov5 Github Repo

Prerequists

python >= 3.6.0
torch >= 1.7

Use Instructions

Example using Anaconda:

Create and activate new environmeent:

conda create -n hw2_demo python>=3.6.0 
conda activate hw2_demo

Install torch and torchvision

conda install pytorch>=1.7 torchvision cudatoolkit={your cuda version} -c pytorch

Install additional dependencies

make install

Download Digit datasets and Generate COCO-Style annotation

make getdataset

This script download training data from TA's google drive, and arrange it into COCO format. I also host a preprocessed annotation file in my own google drive, so the .mat file is ignored.

Reproduce answer.json

make reproduce

This will download model weights from my google drive and reproduce answer.json

Train the model (fine-tune)

make train

If the training is interrupted, you can resume the training by:

cd yolov5/ && python train.py --resume

The training result will be stored in the ./yolov5/runs/train/exp directory.

(If you want to restart the trainning, you need to rename or delete the directory.)

To visualize the training result, you can use the following command:

tensorboard --logdir=./yolov5/runs/train/exp

Test the model

Test the model by running the following command:

make inference

This will use the best model in the ./yolov5/runs/train/exp directory by default.

You can also specify the model by:

cd yolov5/ && python inference.py --model_path={path to the model}

Nemo1999/NYCU_VRDL_HW2