Pywick

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 (Addsign, Eve, Lookahead, Radam, Ralamb, RangerLARS etc).
• A high-level module for training with callbacks, constraints, metrics, conditions and regularizers.
• Dozens 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).

Docs

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)

• Jun. 15, 2020
  • 200+ 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

Install

Pywick requires pytorch >= 1.0

pip install pywick

or specific version from git:

pip install git+https://github.com/achaiah/[email protected]

ModuleTrainer

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

Example:

from pywick.modules import ModuleTrainer
from pywick.initializers import XavierUniform
from pywick.metrics import CategoricalAccuracySingleInput
import torch.nn as nn
import torch.functional as F

# Define your model EXACTLY as normal
class Network(nn.Module):
    def __init__(self):
        super(Network, self).__init__()
        self.conv1 = nn.Conv2d(1, 32, kernel_size=3)
        self.conv2 = nn.Conv2d(32, 64, kernel_size=3)
        self.fc1 = nn.Linear(1600, 128)
        self.fc2 = nn.Linear(128, 10)

    def forward(self, x):
        x = F.relu(F.max_pool2d(self.conv1(x), 2))
        x = F.relu(F.max_pool2d(self.conv2(x), 2))
        x = x.view(-1, 1600)
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training=self.training)
        x = self.fc2(x)
        return F.log_softmax(x)

model = Network()
trainer = ModuleTrainer(model)   # optionally supply cuda_devices as a parameter

initializers = [XavierUniform(bias=False, module_filter='fc*')]

# initialize metrics with top1 and top5 
metrics = [CategoricalAccuracySingleInput(top_k=1), CategoricalAccuracySingleInput(top_k=5)]

trainer.compile(loss='cross_entropy',
                # callbacks=callbacks,          # define your callbacks here (e.g. model saver, LR scheduler)
                # regularizers=regularizers,    # define regularizers
                # constraints=constraints,      # define constraints
                optimizer='sgd',
                initializers=initializers,
                metrics=metrics)

trainer.fit_loader(train_dataset_loader, 
            val_loader=val_dataset_loader,
            num_epoch=20,
            verbose=1)

You also have access to the standard evaluation and prediction functions:

loss = trainer.evaluate(x_train, y_train)
y_pred = trainer.predict(x_train)

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 the TensorboardX (now part of pytorch 1.1 release!)

from pywick.callbacks import EarlyStopping

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

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 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]
trainer.set_constraints(constraints)

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

You can also fit directly on a torch.utils.data.DataLoader and can have a validation set as well :

from pywick import TensorDataset
from torch.utils.data 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)

Extensive Library of Image Classification Models (most are pretrained!)

• All standard models from Pytorch:
  • Densenet
  • Inception v3
  • MobileNet v2
  • ResNet
  • ShuffleNet v2
  • SqueezeNet
  • VGG
• BatchNorm Inception
• Deep High-Resolution Representation Learning for Human Pose Estimation
• Deep Layer Aggregation
• Dual Path Networks
• EfficientNet variants (b0-b8, el, em, es, lite1-lite4, pruned, AP/NS)
• ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks
• FBResnet
• FBNet-C
• Inception v4
• InceptionResnet v2
• Mixnet variants (l, m, s, xl, xxl)
• MnasNet
• MobileNet V3
• NasNet variants (mnas, pnas, mobile)
• PNASNet
• Polynet
• Pyramid Resnet
• RegNet - Designing Network Design Spaces
• Resnet variants (gluon, res2net, se, ssl, tv, wide)
• ResNeSt: Split-Attention Networks
• ResNext variants (ig, se, ssl, swsl, tv)
• SE Net variants (gluon, resnet, resnext, inception)
• SelecSLS Convolutional Net
• Selective Kernel Networks
• Semi-Supervised and Semi-Weakly Supervised ImageNet Models
• Single-Pass NAS Net
• TResNet: High Performance GPU-Dedicated Architecture
• Wide Resnet
• XCeption

Image Segmentation Models

• BiSeNet (Bilateral Segmentation Network for Real-time Semantic Segmentation)
• DANet (Dual Attention Network for Scene Segmentation)
• Deeplab v2 (DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs)
• Deeplab v3 (Rethinking Atrous Convolution for Semantic Image Segmentation)
• DenseASPP (DenseASPP for Semantic Segmentation in Street Scenes)
• DRNNet (Dilated Residual Networks)
• DUC, HDC (understanding convolution for semantic segmentation)
• DUNet (Decoders Matter for Semantic Segmentation)
• ENet (ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation)
• Vanilla FCN: FCN32, FCN16, FCN8, in the versions of VGG, ResNet and OptDenseNet respectively (Fully convolutional networks for semantic segmentation)
• FRRN (Full Resolution Residual Networks for Semantic Segmentation in Street Scenes)
• FusionNet (FusionNet in Tensorflow by Hyungjoo Andrew Cho)
• GCN (Large Kernel Matters)
• LinkNet (Link-Net)
• OCNet (Object Context Network for Scene Parsing)
• PSPNet (Pyramid scene parsing network)
• RefineNet (RefineNet)
• SegNet (Segnet: A deep convolutional encoder-decoder architecture for image segmentation)
• Tiramisu (The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation)
• U-Net (U-net: Convolutional networks for biomedical image segmentation)
• Additional variations of many of the above

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 = torch.utils.data.DataLoader(train_dataset, 
    sampler=ImbalancedDatasetSampler(train_dataset),
    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

• PyTorchNet
• pretrained-models.pytorch
• DeepLab_pytorch
• Pytorch for Semantic Segmentation
• Binseg Pytorch
• awesome-semantic-segmentation-pytorch
• And many others! (attributions listed in the codebase as they occur)