PyTorchNet is a Machine Learning framework that is built on top of PyTorch. And, it uses Tensorboard (or Visdom) for visualization.

PyTorchNet is easy to be customized by creating the necessary classes:

1. Data Loading: a dataset class is required to load the data.
2. Model Design: a nn.Module class that represents the network model.
3. Loss Method: an appropriate class for the loss, for example CrossEntropyLoss or MSELoss.
4. Evaluation Metric: a class to measure the accuracy of the results.

PyTorchNet consists of HAL library which has the following packages:

This is for loading and transforming datasets.

Network models are kept in this package. It already includes ResNet, PreActResNet, Stacked Hourglass and SphereFace.

There are number of different choices available for Classification or Regression. New loss methods can be put here.

There are number of different choices available for Classification or Regression. New accuracy metrics can be put here.

• main
• model

First, you need to download PyTorchNet by calling the following command:

git clone https://github.com/human-analysis/pytorchnet.git

PyTorchNet relies on several Python packages, such as Pytorch, Pytorch Lightning, tensorboard, Pillow Image, etc. you need to make sure that the requirements exist.

Notice

• If you do not have Pytorch or it does not meet the requirements, please follow the instruction on the Pytorch website.

Congratulations!!! You are now ready to use PyTorchNet!

PyTorchNet comes with a classification example in which a ResNet model is trained for the CIFAR10 dataset.

python main.py

PyTorchNet loads its parameters at the beginning via a config file and/or the command line.

When PyTorchNet is being run, it will automatically load all parameters from args.txt by default, if it exists. In order to load a custom config file, the following parameter can be used:

python main.py --config custom_args.txt

[Arguments] save_results = No

#project options
project_name=CIFAR10
save_dir=results/
logs_dir=results/

#dataset options
dataset=CIFAR10
dataroot=data/
cache_size=1000

#model options
precision=32
batch_size_test = 128
batch_size_train = 128
model_type = MobileNetV2
loss_type = Classification
evaluation_type = Accuracy

resolution_high = 32
resolution_wide = 32

manual_seed = 0
nepochs = 200

optim_method = SGD
learning_rate = 0.1
optim_options = {"momentum": 0.9, "weight_decay": 5e-4}

scheduler_method = CosineAnnealingLR
scheduler_options = {"T_max": 200}

#cpu/gpu settings
ngpu = 1
nthreads = 4

Parameters can also be set in the command line when invoking main.py. These parameters will precede the existing parameters in the configuration file.

python main.py --visualizer VisualizerTensorboard