Fine-grained Image Classification via Pytorch

Simple code based on Pytorch pre-trained Resnet50.

You can alse use any other Resnet, Densenet, VGG models by changing only a few lines of code.

Resnet50 accuracy

Dataset	w/o amp	w/ apex.amp	w/ torch.cuda.amp	SOTA
CUB-200-2011	86.74	86.68	86.59	91.7
FGVC Aircraft	93.25	93.58	92.86	94.7
Stanford Cars	94.09	94.30	94.32	96.32

Data preparation

data
├ ─ dataset_dir0
│	├ ─ ─ train
│	│	├ ─ class0
│	│	│	├ ─ img0
│	│	│	└ ─ img1
│	│	└ ─ class1
│	└ ─ ─ test
│		├ ─ class0
│		│	├ ─ img0
│		│	└ ─ img1
│		└ ─ class1
└ ─ dataset_dir1

For collated dataset, see:

CUB-200-2011: https://github.com/cyizhuo/CUB-200-2011-dataset

FGVC Aircraft: https://github.com/cyizhuo/FGVC-Aircraft-dataset

Stanford Cars: https://github.com/cyizhuo/Stanford-Cars-dataset

Python env requirements

numpy

tqdm

pytorch

torchvision

P.S. torch.cuda.amp requires pytorch ≥ 1.6

my env:

python == 3.8.10

pytorch == 1.8.1

numpy == 1.20.2

Usage

Simple usage:

python train.py -d dataset_dir

Full parameters:

python train.py -d dataset_dir -b batch_size -g gpu_id -w num_workers -s seed -a amp -n note

Tricks

Label Smoothing

Cosine Learning Rate Decay

A good paper about tricks:

Bag of Tricks for Image Classification with Convolutional Neural Networks