composer.models

Models provided to Trainer must use the basic interface specified by BaseMosaicModel.

Additionally, for convience we provide a number of extensions of BaseMosaicModel as detailed below.

Base Models

`BaseMosaicModel`	The minimal interface needed to use a model with `composer.trainer.Trainer`.
`MosaicClassifier`	Implements the base logic that all classifiers can build on top of.
`MosaicTransformer`	Implements the base logic that all Transformers can build on top of.

Image Models

`CIFAR10_ResNet56`	A ResNet-56 model extending `MosaicClassifier`.
`MNIST_Classifier`	A simple convolutional neural network.
`EfficientNetB0`	An EfficientNet-b0 model extending `MosaicClassifier`.
`ResNet18`	A ResNet-18 model extending `MosaicClassifier`.
`ResNet50`	A ResNet-50 model extending `MosaicClassifier`.
`ResNet101`	A ResNet-101 model extending `MosaicClassifier`.
`UNet`	A U-Net model extending `MosaicClassifier`.

Language Models

GPT2Model

Implements a GPT-2 wrapper around a MosaicTransformer.

Metrics and Loss Functions

Evaluation metrics for common tasks are in torchmetrics and are directly compatible with BaseMosaicModel. Additionally, we provide implementations of the following metrics and loss functions.

`Dice`	The Dice Coefficient for evaluating image segmentation.
`CrossEntropyLoss`	Torchmetric cross entropy loss implementation.
`soft_cross_entropy`	Drop-in replacement for `torch.CrossEntropy` that can handle dense labels.
`LanguageCrossEntropyLoss`	Hugging Face compatible cross entropy loss.
`Perplexity`	Subclasses `LanguageCrossEntropyLoss` to implement perplexity.