kexiii/pytorch-hymenoptera

Pytorch CV project template for deep learning researchers

pytorch-hymenoptera

Pytorch CV project template for deep learning researchers, take hymenoptera dataset as an example.

Requirement

• Pytorch 0.3+
• Torch vision 0.1.9+
• Python 2.7

Dataset

This template takes the Pytorch official tutorial's hymenoptera dataset as an example, to run the example:

• Create a directory to store the dataset

cd pytorch-hymenoptera/

mkdir dataset

• Update the config.json file in /pytorch-hymenoptera/data_loader

{
    "data_dir": "../dataset/hymenoptera_data",
    "batch_size": 4,
    "num_workers": 4
}

• Download the dataset form here, put it into dataset directory and unzip it

 cd pytorch-hymenoptera/
 cd dataset/
 unzip hymenoptera_data.zip

• Train and test

Example

• Train the model, if you use the default model settings(including seed in config.json), you will get 94.118% val accuracy

python train.py

• Test the model, in this example, we still use the validation set to test our model(don't do this in practice), so you will get the test accuracy near 94.118% too. Note that you should choose the checkpoint file manually

python test.py --checkpoint logging/checkpoint.pth

• Plot the curve

python plot.py

This will produce two image files: accuracy.jpg, loss.jpg

• Resume from previous work

python train.py --resume logging/checkpoint_to_resume.pth