Skip to main content

The flexible training toolbox

Project description

Trainable: The Flexible PyTorch Training Toolbox

If you're sick of dealing with all of the boilerplate code involved in training, evaluation, visualization, and preserving your models, then you're in luck. Trainable offers a simple, yet extensible framework to make understanding the latest papers the only headache of Neural Network training.


pip install trainable


The typical workflow for trainable involves defining a callable Algorithm to describe how to train your network on a batch, and how you'd like to label your losses:

class MSEAlgorithm(Algorithm):
    def __init__(self, eval=False, **args):
        self.mse = nn.MSELoss()

    def __call__(self, model, batch, device):
        x, target = batch
        x, target =,
        y = model(x)

        loss = self.mse(y, target)

        metrics = { self.key("MSE Loss"):loss.item() }
        return metrics

Then you simply instantiate your model, dataset, and optimizer...

device = torch.device('cuda')

model = MyModel().to(device)
optim = FancyPantsOptimizer(model.parameters(), lr=1e-4)

train_data = DataLoader(SomeTrainingDataset('path/to/your/data'), batch_size=32)
test_data = DataLoader(SomeTestingDataset('path/to/your/data'), batch_size=32)

...and let trainable take care of the rest!

trainer = Trainer(
  visualizer=MyVisualizer(),  # Typically Plotter() or Saver()

save_path = "desired/save/path/for/your/session.sesh"
trainer.start_session(model, optim, path)


A beautiful detailed description of what the heck 
you were trying to accomplish with this training.

metrics = trainer.train(train_data, test_data, epochs=200)

Plotting your data is simple as well:

import matplotlib.pyplot as plt

for key in metrics:

Tunable Options

The Trainer interface gives you a nice handful of options to configure your training experience. They include:

  • Display Frequency: How often (in batches) information such as your training loss is updated in your progress bar.
  • Visualization Frequency: How often (in batches) the training produces a visualization of your model's outputs.
  • Validation Frequency: How often (in epochs) the trainer performs validation with your test data.
  • Autosave Frequency: How often your session is saved out to disk.
  • Device: On which hardware your training should occur.


Do you want a little more granularity in how you visualize your data? Or perhaps running an epoch with your model is a little more involved than just training on each batch of data? Wondering why the heck pytorch doesn't have a built-in dataset for unsupervised images? Maybe your training algorithm involves VGG? Got you covered. Check out the source for the various submodules:


Find any other headaches in neural net training that you think you can simplify with Trainable? Feel free to make a pull request from my github repo.


Email me anytime at

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for trainable, version 0.1.3.post9
Filename, size File type Python version Upload date Hashes
Filename, size trainable-0.1.3.post9-py3-none-any.whl (21.0 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size trainable-0.1.3.post9.tar.gz (17.5 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page