Neural Piepline

Neural networks training pipeline based on PyTorch. Designed to standardize training process and to increase coding preformance.

• Core is about 2K lines, covered by tests, that you doesn't need to write again
• Flexible and customizable training process
• Checkpoints management and train process resuming (source and target device independent)
• Metrics processing and visualization by builtin (tensorboard, Matplotlib) or custom monitors
• Training best practices (e.g. learning rate decaying and hard negative mining)
• Metrics logging and comparison (DVC compatible)

Train MNIST example:

This code run MNIST image classification with Tensorboard monitoring. Code based on PyTorch example.

See full example there.

from neural_pipeline.builtin.monitors.tensorboard import TensorboardMonitor
from neural_pipeline import DataProducer, AbstractDataset, TrainConfig, TrainStage,\
    ValidationStage, Trainer, FileStructManager

import torch
from torch import nn
from torchvision import datasets, transforms

class Net(nn.Module):
    # Network implementation

class MNISTDataset(AbstractDataset):
    transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])

    def __init__(self, data_dir: str, is_train: bool):
        self.dataset = datasets.MNIST(data_dir, train=is_train, download=True)

    def __len__(self):
        return len(self.dataset)

    def __getitem__(self, item):
        data, target = self.dataset[item]
        return {'data': self.transforms(data), 'target': target}

fsm = FileStructManager(base_dir='data', is_continue=False)
model = Net()

train_dataset = DataProducer([MNISTDataset('data/dataset', True)], batch_size=4, num_workers=2)
validation_dataset = DataProducer([MNISTDataset('data/dataset', False)], batch_size=4, num_workers=2)

train_config = TrainConfig([TrainStage(train_dataset), ValidationStage(validation_dataset)], torch.nn.NLLLoss(),
                           torch.optim.SGD(model.parameters(), lr=1e-4, momentum=0.5))

trainer = Trainer(model, train_config, fsm, torch.device('cuda:0')).set_epoch_num(50)
trainer.monitor_hub.add_monitor(TensorboardMonitor(fsm, is_continue=False))
trainer.train()

Installation:

pip install neural-pipeline

For builtin module using install:

pip install tensorboardX matplotlib

Install latest version before it's published on PyPi

pip install -U git+https://github.com/toodef/neural-pipeline

Getting started:

Documentation

See the full documentation there

Data flow scheme:

See the examples

• MNIST classification - notebook, file
• Segmentation - notebook, file
• Resume training process - file