Object-Classification-Detection

Environment Setup

• Pytorch >= 1.10.2 (CUDA 11.3 / CPU)
• torchvision
• tensorboard
• numpy
• matlablib
• tqdm (Optional)

Download dataset into data/ folder.

Model is saved into logs/ by default, you can customized it via --save-folder argument.

Training and Testing

Simply run python train.py --method baseline Other method could be mixup, cutmix and cutout

• Batch size: 128
• Optimizer: Adam
• Scheduler: StepLR
• Loss function: CrossEntrophyLoss
• Metrics: Top5 Accuracy

model	batch size	lr	epoch	method	train loss	valid loss	acc@5
ResNet18	128	1e-3	30	baseline	1.14729	1.967063	0.788469
ResNet18	128	1e-3	80	baseline	0.10185	3.024865	0.774229
ResNet101	128	1e-3	30	baseline	2.29751	14.033395	0.620847
ResNet18	128	1e-3	30	cutmix	3.57314	2.194764	0.752769
ResNet18	128	1e-3	30	cutout	1.85152	2.012332	0.775514
ResNet18	128	1e-3	30	mixup	3.15058	2.177603	0.748517

Visualization results

• Tensorboard results: Given the file in /logs, open tensorboard via tensorboard --logdir ./tensorboard

For example:

Model architecture: ResNet18

Method: Baseline

(Smoothing == 0.6)