Midterm

The repo of midterm project for DATA620004: Neural Network and Deep Learning

CNN on CUB dataset

Download dataset: https://data.caltech.edu/records/65de6-vp158, and then unzip into ./CUB_200_2011/.
Download weight: https://drive.google.com/drive/folders/1_8M-21TTVpdxizYT43qhybMJST82XOwE?usp=drive_link, put into './logs/0530_lr_0.0003_decay_0.0003_pretrained_True_dropout_True_cutoutFalse_modelResNet34'

Train

python ./Bird/main.py

Learning rate	Pretrain	Cutout	Dropout	Model	Test Accuracy $\uparrow$
$3e-4$	✔️	❌	✔️	ResNet34	83.12%
$3e-4$	✔️	✔️	✔️	ResNet34	82.27%
$3e-4$	✔️	✔️	❌	ResNet34	81.72%
$3e-5$	✔️	✔️	✔️	ResNet34	78.20%
$3e-4$	✔️	✔️	✔️	ResNet18	77.14%
$3e-4$	❌	✔️	✔️	ResNet34	43.17%

Table: The results on VOC0712 with different hyper-parameters combination. The method with pretrain, cutout augmentation, and Dropout achieved the best performance $83.12%$.

Inference

python ./Bird/infernce.py

Object detection on VOC

Ensure the enviroment is complete for mmdetection.
Download dataset: http://host.robots.ox.ac.uk/pascal/VOC/, and then unzip into ./data/. Notably, while training YOLOv3, the format should transfrom to COCO format.
Download weight: https://drive.google.com/drive/folders/1_8M-21TTVpdxizYT43qhybMJST82XOwE?usp=drive_link, put into './checkpoints/'