peal.architectures.interfaces¶
Pydantic config templates for PEAL architectures and training tasks.
ArchitectureConfig describes a network as a list of layer entries
(fc, vgg, resnet, transformer), each backed by one of the
per-layer config classes here; peal.architectures.predictors turns such a
config into a module. TaskConfig describes what the network is trained
and evaluated on: the weighted mix of criterions from
peal.training.criterions, the output type and channel count, and optional
restrictions of the input/output variables or classes. Every class carries a
config_name field so that a yaml config can be mapped back to its class.
Classes
The config template for a neural architecture. |
|
The config template for a Fully Connected Layer. |
|
The config template for a ResNet layer. |
|
A dict of critirion names (that have to be implemented in peal.training.criterions) mapped to the weight. |
|
The config template for a transformer layer. |
|
The config template for a VGG Layer. |