composer.models

Modules

composer.models.base	The ComposerModel base interface.
composer.models.bert	The BERT model family using Hugging Face Transformers.
composer.models.classify_mnist	A simple example convolutional neural network which can be used to classify MNIST data.
composer.models.deeplabv3	DeepLabV3 for image segmentation.
composer.models.efficientnetb0	The EfficientNet model family is a set of convolutional neural networks that can be used as the basis for a variety of vision tasks, but were initially designed for image classification.
composer.models.gpt2	The GPT-2 model family is set of transformer-based networks for autoregressive language modeling at various scales.
composer.models.initializers	composer.models.initializers
composer.models.model_hparams	General YAHP interface for Base ComposerModels.
composer.models.resnet	The ResNet model family is a set of convolutional neural networks described in Deep Residual Learning for Image Recognition (He et al, 2015).
composer.models.resnet_cifar	A ResNet model family adapted for CIFAR10 image sizes.
composer.models.ssd	SSD300 for object detection on MSCOCO.
composer.models.tasks	Model tasks are ComposerModels with forward passes and logging built-in for many common deep learning tasks.
composer.models.timm	A wrapper around timm.create_model() used to create ComposerClassifier.
composer.models.transformer_hparams	General YAHP interface for ComposerTransformers.
composer.models.transformer_shared	The ComposerModel base interface for Transformers.
composer.models.unet	The Unet architecture used in image segmentation.
composer.models.vit_small_patch16	ViT Small Patch 16 for image classification.

The models module contains the ComposerModel base class along with reference implementations of many common models. Additionally, it includes task-specific convenience ComposerModels that wrap existing Pytorch models with standard forward passes and logging to enable quick interaction with the Trainer.

See Composer Model for more details.

Classes

BERTModel	BERT model based on 🤗 Transformers.
ComposerClassifier	A convenience class that creates a ComposerModel for classification tasks from a vanilla PyTorch model.
ComposerDeepLabV3	DeepLabV3 model extending ComposerClassifier.
ComposerModel	The interface needed to make a PyTorch model compatible with composer.Trainer.
ComposerResNet	A ComposerClassifier wrapper around the torchvision implementations of the ResNet model family.
ComposerResNetCIFAR	ResNet models for CIFAR10 extending ComposerClassifier.
ComposerTransformer	The ComposerModel base interface for Transformers.
EfficientNetB0	A ComposerClassifier wrapper around the EfficientNet-b0 architecture.
GPT2Model	Implements ComposerTransformer to wrap Hugging Face GPT-2 transformers.
Initializer	Sets the initialization scheme for different layers of a PyTorch model.
MNIST_Classifier	A simple convolutional neural network extending ComposerClassifier.
SSD	Single Shot Object detection Model with pretrained ResNet34 backbone extending ComposerModel.
Timm	A wrapper around timm.create_model() used to create ComposerClassifier.
UNet	A U-Net model extending ComposerModel.
ViTSmallPatch16	Implements ViT-S/16 as a ComposerClassifier.

Hparams

These classes are used with yahp for YAML-based configuration.

BERTForClassificationHparams	YAHP classification interface for BERTModel.
BERTHparams	YAHP interface for BERTModel.
DeepLabV3Hparams	YAHP interface for ComposerDeepLabV3.
EfficientNetB0Hparams	YAHP interface for EfficientNetB0.
GPT2Hparams	YAHP interface for GPT2Model.
MnistClassifierHparams	YAHP interface for MNIST_Classifier.
ModelHparams	General YAHP interface for ComposerModels.
ResNetCIFARHparams	Hparams interface for ComposerResNetCIFAR.
ResNetHparams	YAHP interface for ComposerResNet.
SSDHparams	YAHP interface for SSD.
TimmHparams	YAHP interface for Timm.
TransformerHparams	Defines the necessary hyparameters for a Transformer base module.
UnetHparams	YAHP interface for UNet.
ViTSmallPatch16Hparams	YAHP interface for ViTSmallPatch16.