robbytorch 0.1.3

Cool package for robust AI

Installation

Robbytorch requires Pytorch, however it's not specified in the dependencies - we recommend installing Pytorch manually via conda and only later installing Robbytorch by pip.

Use your conda env or create a new one:

conda create --name <ENV NAME> python=3.8 pip
conda activate <ENV NAME>

Install Pytorch. If you have older drivers for GPU you may want to require older version of CUDA, i.e.:

conda install pytorch torchvision torchaudio cudatoolkit=10.1 -c pytorch -c conda-forge

or even older Pytorch version:

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.1 -c pytorch

Then run:

pip install robbytorch

Usage

See jupyter notebooks in ipython/ for complete examples. For step-by-step introduction continue reading this file.

Prepare Dataset

Place your data into chosen root directory, i.e. "/dysk1/approx/robby".

You can subclass robbytorch.datasets.DictDateset and implement two methods - for more info please read the docstring for that class. Here's an example implementation:

import torch
from torchvision import transforms
from torch.utils.data import DataLoader

from robbytorch.datasets import DictDataset


class CreationsDataset(DictDataset):

    def load_data(self, idx):
        file_name = f"{self.metadata.iloc[idx]['creation_id']}.png"
        return self.load_image(file_name)
    
    def load_target_dict(self, idx):
        record = self.metadata.iloc[idx].to_dict()
        
        return {col: torch.tensor(record[col]).float() 
                for col in ['label', 'CR']
               }

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Resize((224, 224))
])
data_root = "/dysk1/approx/robby"
metadata = pd.DataFrame(data=[(3141, 0, 0.01), (2137, 1, 0.012)], columns=["creation_id", "label", "CR"])
dataset = CreationsDataset(data_root, metadata, transform=transform)
dataloader = DataLoader(dataset, batch_size=128, shuffle=False, num_workers=2)

Now whenever you iter through the dataloader you get a dict of batched tensors (with .shape[0] == 128):

{
    "data": batched_tensor_data,
    "label": batched_tensor_label,
    "CR": batched_CR_of_tensors
}

You can use this structure however you like during training/evaluation.

TODO - further eplanations:

• trenowanie: 3x forward
• configs z lib2
• Dodawanie auxiliary losses za pomocą magic hooks
• Writers objaśnić, livelossplot i mlflow
• wczytywanie robust networks
• opis utilities - notebook, visualization