High-Level Training framework for Pytorch

Pywick is a high-level Pytorch training framework that aims to get you up and running quickly with state of the art neural networks. Does the world need another Pytorch framework? Probably not. But we started this project when no good frameworks were available and it just kept growing. So here we are.

Pywick tries to stay on the bleeding edge of research into neural networks. If you just wish to run a vanilla CNN, this is probably going to be overkill. However, if you want to get lost in the world of neural networks, fine-tuning and hyperparameter optimization for months on end then this is probably the right place for you :)

Among other things Pywick includes:

  • State of the art normalization, activation, loss functions and optimizers not included in the standard Pytorch library (AdaBelief, Addsign, Apollo, Eve, Lookahead, Radam, Ralamb, RangerLARS etc).
  • A high-level module for training with callbacks, constraints, metrics, conditions and regularizers.
  • Hundreds of popular object classification and semantic segmentation models!
  • Comprehensive data loading, augmentation, transforms, and sampling capability.
  • Utility tensor functions.
  • Useful meters.
  • Basic GridSearch (exhaustive and random).


Hey, check this out, we now have docs! They’re still a work in progress though so apologies for anything that’s broken.

What’s New (highlights)

v0.6.5 - Docker all the things!

Another great improvement to the framework - docker! You can now run the 17flowers demo right out of the box!

  • Grab our docker image at docker hub: docker pull achaiah/pywick:latest. Pytorch 1.8 and cuda dependencies are pre-installed.

  • Run 17flowers demo with: docker run --rm -it --ipc=host -v your_local_out_dir:/jobs/17flowers --init -e demo=true achaiah/pywick:latest

  • Or run the container in standalone mode so you can use your own data (don’t forget to map your local dir to container):

      docker run --rm -it \
      --ipc=host \
      -v <your_local_data_dir>:<container_data_dir> \
      -v <your_local_out_dir>:<container_out_dir> \
      --init \

Older Notes

  • Oct. 11, 2021 - We thought ya might like YAML!
    • So you’re saying you like configuration files? You’re saying you like examples too? Well, we’ve got you covered! Huge release today with a configuration-based training example! All you have to do is:
      • Get your favorite dataset (or download 17 flowers to get started and pywick/examples/ to convert)
      • Adjust the configs/train_classifier.yaml file to fit your workspace
      • Then simply run: python3 configs/train_classifier.yaml and watch it train!
  • May 6, 2021
    • Many SoTA classification and segmentation models added: Swin-Transformer variants, NFNet variants (L0, L1), Halo nets, Lambda nets, ECA variants, Rexnet + others
    • Many new loss functions added: RecallLoss, SoftInvDiceLoss, OhemBCEDicePenalizeBorderLoss, RMIBCEDicePenalizeBorderLoss + others
    • Bug fixes
  • Jun. 15, 2020
    • 700+ models added from rwightman’s repo via torch.hub! See docs for all the variants!
    • Some minor bug fixes
  • Jan. 20, 2020
    • New release: 0.5.6 (minor fix from 0.5.5 for pypi)
    • Mish activation function (SoTA)
    • rwightman’s models of pretrained/ported variants for classification (44 total)
      • efficientnet Tensorflow port b0-b8, with and without AP, el/em/es, cc
      • mixnet L/M/S
      • mobilenetv3
      • mnasnet
      • spnasnet
    • Additional loss functions
  • Aug. 1, 2019
    • New segmentation NNs: BiSeNet, DANet, DenseASPP, DUNet, OCNet, PSANet
    • New Loss Functions: Focal Tversky Loss, OHEM CrossEntropy Loss, various combination losses
    • Major restructuring and standardization of NN models and loading functionality
    • General bug fixes and code improvements


Pywick requires pytorch >= 1.4

pip install pywick

or specific version from git:

pip install git+


The ModuleTrainer class provides a high-level training interface which abstracts away the training loop while providing callbacks, constraints, initializers, regularizers, and more.

See the example for a pretty complete configuration example. To get up and running with your own data quickly simply edit the configs/train_classifier.yaml file with your desired parameters and dataset location(s).

Note: Dataset needs to be organized for classification where each directory name is the name of a class and contains all images pertaining to that class

PyWick provides a wide range of callbacks, generally mimicking the interface found in Keras:

  • CSVLogger - Logs epoch-level metrics to a CSV file
  • CyclicLRScheduler - Cycles through min-max learning rate
  • EarlyStopping - Provides ability to stop training early based on supplied criteria
  • History - Keeps history of metrics etc. during the learning process
  • LambdaCallback - Allows you to implement your own callbacks on the fly
  • LRScheduler - Simple learning rate scheduler based on function or supplied schedule
  • ModelCheckpoint - Comprehensive model saver
  • ReduceLROnPlateau - Reduces learning rate (LR) when a plateau has been reached
  • SimpleModelCheckpoint - Simple model saver
  • Additionally, a TensorboardLogger is incredibly easy to implement via tensorboardX (now part of pytorch 1.1 release!)
from pywick.callbacks import EarlyStopping

callbacks = [EarlyStopping(monitor='val_loss', patience=5)]

PyWick also provides regularizers:

  • L1Regularizer
  • L2Regularizer
  • L1L2Regularizer

and constraints:

  • UnitNorm
  • MaxNorm
  • NonNeg

Both regularizers and constraints can be selectively applied on layers using regular expressions and the module_filter argument. Constraints can be explicit (hard) constraints applied at an arbitrary batch or epoch frequency, or they can be implicit (soft) constraints similar to regularizers where the constraint deviation is added as a penalty to the total model loss.

from pywick.constraints import MaxNorm, NonNeg
from pywick.regularizers import L1Regularizer

# hard constraint applied every 5 batches
hard_constraint = MaxNorm(value=2., frequency=5, unit='batch', module_filter='*fc*')
# implicit constraint added as a penalty term to model loss
soft_constraint = NonNeg(lagrangian=True, scale=1e-3, module_filter='*fc*')
constraints = [hard_constraint, soft_constraint]

regularizers = [L1Regularizer(scale=1e-4, module_filter='*conv*')]

You can also fit directly on a and can have a validation set as well :

from pywick import TensorDataset
from import DataLoader

train_dataset = TensorDataset(x_train, y_train)
train_loader = DataLoader(train_dataset, batch_size=32)

val_dataset = TensorDataset(x_val, y_val)
val_loader = DataLoader(val_dataset, batch_size=32)

trainer.fit_loader(loader, val_loader=val_loader, num_epoch=100)

Image Segmentation Models

To load one of these models:

Read the docs for useful details! Then dive in:

# use the `get_model` utility
from pywick.models.model_utils import get_model, ModelType

model = get_model(model_type=ModelType.CLASSIFICATION, model_name='resnet18', num_classes=1000, pretrained=True)

For a complete list of models (including many experimental ones) you can call the get_supported_models method e.g. pywick.models.model_utils.get_supported_models(ModelType.SEGMENTATION)

Data Augmentation and Datasets

The PyWick package provides wide variety of good data augmentation and transformation tools which can be applied during data loading. The package also provides the flexible TensorDataset, FolderDataset and MultiFolderDataset classes to handle most dataset needs.

Torch Transforms

These transforms work directly on torch tensors

  • AddChannel
  • ChannelsFirst
  • ChannelsLast
  • Compose
  • ExpandAxis
  • Pad
  • PadNumpy
  • RandomChoiceCompose
  • RandomCrop
  • RandomFlip
  • RandomOrder
  • RangeNormalize
  • Slice2D
  • SpecialCrop
  • StdNormalize
  • ToFile
  • ToNumpyType
  • ToTensor
  • Transpose
  • TypeCast

Additionally, we provide image-specific manipulations directly on tensors:

  • Brightness
  • Contrast
  • Gamma
  • Grayscale
  • RandomBrightness
  • RandomChoiceBrightness
  • RandomChoiceContrast
  • RandomChoiceGamma
  • RandomChoiceSaturation
  • RandomContrast
  • RandomGamma
  • RandomGrayscale
  • RandomSaturation
  • Saturation

Affine Transforms (perform affine or affine-like transforms on torch tensors)

  • RandomAffine
  • RandomChoiceRotate
  • RandomChoiceShear
  • RandomChoiceTranslate
  • RandomChoiceZoom
  • RandomRotate
  • RandomShear
  • RandomSquareZoom
  • RandomTranslate
  • RandomZoom
  • Rotate
  • Shear
  • Translate
  • Zoom

We also provide a class for stringing multiple affine transformations together so that only one interpolation takes place:

  • Affine
  • AffineCompose

Blur and Scramble transforms (for tensors)

  • Blur
  • RandomChoiceBlur
  • RandomChoiceScramble
  • Scramble

Datasets and Sampling

We provide the following datasets which provide general structure and iterators for sampling from and using transforms on in-memory or out-of-memory data. In particular, the FolderDataset has been designed to fit most of your dataset needs. It has extensive options for data filtering and manipulation. It supports loading images for classification, segmentation and even arbitrary source/target mapping. Take a good look at its documentation for more info.

  • ClonedDataset
  • CSVDataset
  • FolderDataset
  • MultiFolderDataset
  • TensorDataset
  • tnt.BatchDataset
  • tnt.ConcatDataset
  • tnt.ListDataset
  • tnt.MultiPartitionDataset
  • tnt.ResampleDataset
  • tnt.ShuffleDataset
  • tnt.TensorDataset
  • tnt.TransformDataset

Imbalanced Datasets

In many scenarios it is important to ensure that your traing set is properly balanced, however, it may not be practical in real life to obtain such a perfect dataset. In these cases you can use the ImbalancedDatasetSampler as a drop-in replacement for the basic sampler provided by the DataLoader. More information can be found here

from pywick.samplers import ImbalancedDatasetSampler

train_loader =, 
    batch_size=args.batch_size, **kwargs)

Utility Functions

PyWick provides a few utility functions not commonly found:

Tensor Functions

  • th_iterproduct (mimics itertools.product)
  • th_gather_nd (N-dimensional version of torch.gather)
  • th_random_choice (mimics np.random.choice)
  • th_pearsonr (mimics scipy.stats.pearsonr)
  • th_corrcoef (mimics np.corrcoef)
  • th_affine2d and th_affine3d (affine transforms on torch.Tensors)

Acknowledgements and References

We stand on the shoulders of (github?) giants and couldn’t have done this without the rich github ecosystem and community. This framework is based in part on the excellent Torchsample framework originally published by @ncullen93. Additionally, many models have been gently borrowed/modified from @Cadene pretrained models repo as well as @Tramac segmentation repo.

Thank you to the following people and the projects they maintain:

  • @ncullen93
  • @cadene
  • @deallynomore
  • @recastrodiaz
  • @zijundeng
  • @Tramac
  • And many others! (attributions listed in the codebase as they occur)

Thank you to the following projects from which we gently borrowed code and models

| Thangs are broken matey! Arrr!!! | |———————–| | We’re working on this project as time permits so you might discover bugs here and there. Feel free to report them, or better yet, to submit a pull request! |