create_scheme: Create Model Scheme

In neuroimaginador/dl4ni: Deep Learning Flows for Neuroimaging

Description

This function creates an object of class DLscheme that contains the basic configuration of a DLmodel

Usage

create_scheme(...)

Arguments

...	attributes to add to the scheme, as key, value pairs.

Details

All schemes should provide at least a optimizer argument. The rest of parameters, if not provided, are set to their default values, from get_dl4ni_config.

Another interesting attribute is memory_limit. The default, "512M" is rather conservative. Please feel free to increase this quantity (it's a character string) to a reasonable amounto of memory inside the limits of your system (RAM or GPU memory, wherever your system is configured to train). The automatic estimation of batch size in the training phase uses this memory_limit to adjust the batch size that fits in memory, thus increasing it will allow larger batch sizes and faster training phases. In case you overestimate the available memory, the system may become inestable, so use with caution. Reasonable values could be "2G" in older systems (like mine).

We don't need to tell, in the scheme, the number of input volumes in the model. This will be made in the "instantiation" of the model using instantiate_model, which will use additional information about the problem to solve in order to build the actual model.

This strategy is most useful when creating a DLflow with create_flow. In this case, we can define a complete flow using just functions and schemes (although pretrained models are supported) and the actual models will be instantiated (built) in the training phase, according to the inputs and outputs provided.

Value

An object of the class DLscheme.

neuroimaginador/dl4ni documentation built on May 3, 2019, 5:47 p.m.