stable_pretraining.backbone

The backbone module provides neural network architectures and utilities for self-supervised learning.

Architectures

MLP(in_channels, hidden_channels, norm_layer)	This block implements the multi-layer perceptron (MLP) module.
Resnet9(num_classes, num_channels, *args, ...)	A Residual network.
ConvMixer([in_channels, num_classes, dim, ...])	ConvMixer model from [Trockman and Kolter, 2022].

Utility Functions

from_timm(model_name[, low_resolution])
from_torchvision(model_name[, low_resolution])	Load a backbone model.
set_embedding_dim(module, dim[, bias, ...])

Specialized Modules

TeacherStudentWrapper(student[, warm_init, ...])	Backbone wrapper that implements teacher-student distillation via EMA.
EvalOnly(backbone)	Wrapper that forces a module to remain in evaluation mode.

Modules

mae