Image Classification

---

[Content]

1. Description
2. Usage
   2-1. Model Training
   2-2. Detection Evaluation
3. Contact

---

[Description]

This is a repository for PyTorch training implementation of general purposed classifier. With the training code, various feature extractor backbones such as Resnet, Darknet, CSP-Darknet53, Mobilenet series can be created. It can provide 6x faster dataloader rather than torchvision.datasets.ImageFolder.

• 224 x 224 Model Performance Table

Model	Dataset	Train	Valid	Size (pixel)	Accuracy (Top-1)	Params (M)	FLOPs (G)
Darknet19	ImageNet	train2012	val2012	224	73.50	20.84	5.62
Darknet53	ImageNet	train2012	val2012	224	77.67	41.61	14.29
Darknet53-tiny	ImageNet	train2012	val2012	224	52.75	2.09	0.64
CSP-Darknet53	ImageNet	train2012	val2012	224	77.83	21.74	6.72
Mobilenet-v1	ImageNet	train2012	val2012	224	71.49	4.23	1.18
Mobilenet-v2	ImageNet	train2012	val2012	224	66.61	3.50	0.66
Mobilenet-v3-small	ImageNet	train2012	val2012	224	-	-	-
Mobilenet-v3-large	ImageNet	train2012	val2012	224	69.93	5.48	0.47

Pretrained Model Weights Download

• Darknet19
• Darknet53
• Darknet53-tiny
• CSP-Darknet53
• Mobilenet-v1
• Mobilenet-v2
• Mobilenet-v3-large
• Mobilenet-v3-small(NOT YET)

• 448 x 448 Model Performance Table

Model	Dataset	Train	Valid	Size (pixel)	Accuracy (Top-1)	Params (M)	FLOPs (G)
Darknet19	ImageNet	train2012	val2012	448	76.22	20.84	22.47
Darknet53	ImageNet	train2012	val2012	448	79.77	41.61	57.17
Darknet53-tiny	ImageNet	train2012	val2012	448	-	-	-
CSP-Darknet53	ImageNet	train2012	val2012	448	79.33	21.74	26.86
Mobilenet-v1	ImageNet	train2012	val2012	448	-	-	-
Mobilenet-v2	ImageNet	train2012	val2012	448	-	-	-
Mobilenet-v3-small	ImageNet	train2012	val2012	448	-	-	-
Mobilenet-v3-large	ImageNet	train2012	val2012	448	-	-	-

Pretrained Model Weights Download

• Darknet19-448
• Darknet53-448
• Darknet53-tiny-448(NOT YET)
• CSP-Darknet53-448
• Mobilenet-v1-448(NOT YET)
• Mobilenet-v2-448(NOT YET)
• Mobilenet-v3-large-448(NOT YET)
• Mobilenet-v3-small-448(NOT YET)

[Usage]

Model Training

• You can train various classifier architectures (ResNet18, ResNet34, ResNet50, ResNet101, DarkNet19, DarkNet53, and CSP-DarkNet53, Mobilenetv1-v3). In addition, you can finetune classifier adding "--pretrained" in training command after putting the weight files into "./weights" directory with {model name}.
• In case of fine-tuning with 448px image, we finetune pretrained 224px model for 15~20 epochs without warming up learning rate. That is the comment like shown below. python3 train.py --exp darknet19-448 --model darknet19 --batch-size 128 --base-lr 0.001 --warmup 0 --img-size 448 --num-epochs 15 --pretrained

python train.py --exp my_test --data imagenet.yaml --model resnet18
                                                   --model resnet34
                                                   --model resnet50
                                                   --model resnet101
                                                   --model darknet19
                                                   --model darknet53
                                                   --model darknet53-tiny
                                                   --model csp-darknet53 --width_multiple 1.0 --depth_multiple 1.0
                                                   --model mobilenet-v1 --width_multiple 1.0
                                                   --model mobilenet-v2 --width_multiple 1.0
                                                   --model mobilenet-v3 --mode {large, small} --width_multiple 1.0

Classification Evaluation

• It computes Top-1 accuracy. Top-1 accuracy is the conventional accuracy, the model answer (the one with highest probability) must be exactly the expected answer.

python val.py --exp my_test --data voc.yaml --ckpt-name best.pt

---

[Contact]

• Author: Jiho Park
• Email: [email protected]