Image Classification Implementation

July 10, 2025 · View on GitHub

Setup

conda create -n tinynext python=3.9
conda activate tinynext
pip install -r requirements.txt

Image Classification for ImageNet-1K:

Model	Top-1 accuracy	#params	MACs	Latency	Logs
TinyNeXt-M	75.3%	2.3M	475M	19.4ms	M
TinyNeXt-S	72.7%	1.3M	304M	14.3ms	S
TinyNeXt-T	71.5%	1.0M	259M	12.7ms	T

Latency is measured on Nvidia Jetson Nano.

Training

Dataset Preparation

Download the ImageNet-1K dataset and structure the data as follows:

/path/to/imagenet-1k/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  validation/
    class1/
      img3.jpeg
    class2/
      img4.jpeg

Train TinyNeXt-M with 8 GPUs in one node:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 29501 --use_env main.py --model 'tinynext_m' --data-path '/data/imagenet' --reprob 0.0 --aa="" --mixup 0 --cutmix 0.0

Train TinyNeXt-S with 8 GPUs in one node:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 29501 --use_env main.py --model 'tinynext_s' --data-path '/data/imagenet' --reprob 0.0 --aa="" --mixup 0 --cutmix 0.0

Train TinyNeXt-T with 8 GPUs in one node:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 29501 --use_env main.py --model 'tinynext_t' --data-path '/data/imagenet' --reprob 0.0 --aa="" --mixup 0 --cutmix 0.0

Evaluation

TinyNeXt-M

Test with 8 GPUs in one node:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 29501 --use_env \
main.py --model 'tinynext_m' --data-path '/data/imagenet' --eval --resume "logs/tinynext_m/tinynext_m.pth"

This should give * eval loss: 1.0679 top1: 75.28 top5: 92.24

TinyNeXt-S

Test with 8 GPUs in one node:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 29501 --use_env \
main.py --model 'tinynext_s' --data-path '/data/imagenet' --eval --resume "logs/tinynext_s/tinynext_s.pth"

This should give * eval loss: 1.1817 top1: 72.70 top5: 90.93

TinyNeXt-T

Test with 8 GPUs in one node:

python -m torch.distributed.launch --nproc_per_node=8 --master_port 29501 --use_env \
main.py --model 'tinynext_t' --data-path '/data/imagenet' --eval --resume "logs/tinynext_t/tinynext_t.pth"

This should give * eval loss: 1.2419 top1: 71.54 top5: 90.24