Pytorch Classification

This project was created to train a single classification model.

Training

python helper/train.py +configs=<CONFIG>

Installation

$git clone https://github.com/pedrodiamel/pytorchvision.git
$cd pytorchvision
$python setup.py install

Training visualize

We now support Visdom for real-time loss visualization during training!

To use Visdom in the browser:

# First install Python server and client
pip install visdom
# Start the server (probably in a screen or tmux)
python -m visdom.server -env_path out/runs/visdom/ -port 6006
# http://localhost:6006/

For jupyter notebook

jupyter notebook --port 8080 --allow-root --ip 0.0.0.0 --no-browser

Docker

docker build -f "Dockerfile" -t torchcls:latest .
docker run -ti --privileged --ipc=host --name torchcls-dev -p 8080:8080 -p 6006:6006 -v $DATASETS:/.datasets torchcls:latest /bin/bash

Dockercompose

docker-compose up --build -d
docker-compose down
docker exec -it torchcls-dev /bin/bash

Benchmarking

Accuracy

Model	CIFAR10	CIFAR100	FERp	Affect
PreActResNet18	95.36%	77.02%	87.25	43.0
PreActResNet34	95.72%	78.83%

Accuracy FER problem

Model	Ferp(test)	AffectNet(val)	Ckp	Jaffe	BU3DFE	Models
PreActResNet18	82.372	26,100	55,307	36,318	39,828
FMPNet	79,535	29,200	65,363	46,766	41,379
CVGG	84,316	31,150	66,201	46,269	42,069
ResNet18	87,695	34,400	71,508	50,746	45,345
AlexNet	86,038	35,075	70,670	64,401	46,379
DeXpression	79,694	31,875	51,117	44,279	37,241

References

• https://github.com/kuangliu/pytorch-cifar
• https://github.com/Cadene/pretrained-models.pytorch
• http://rodrigob.github.io/are_we_there_yet/build/classification_datasets_results.html