PCDARTSOptions

Options & Configuration · AiDotNet.AutoML.NAS

Search settings of PCDARTS. The defaults are those of Xu et al. 2020, “PC-DARTS: Partial Channel Connections for Memory-Efficient Architecture Search” (Sec. 4.2) and its reference implementation.

Inheritance: DartsOptimizerOptions → PCDARTSOptions

For Beginners

PC-DARTS alternates two steps: one trains the network’s weights on the training data, the other trains the architecture choices on the validation data. During the first WarmupEpochs epochs only the weights are trained, so the architecture choices are not judged by an untrained network. The defaults reproduce the paper’s CIFAR-10 search; a smaller dataset or search space may need fewer epochs or a smaller batch.

How It Works

The optimizer settings reuse DartsOptimizerOptions, with PC-DARTS’ own defaults: the architecture parameters (alpha and the edge-normalization parameters beta) are trained with Adam at learning rate 6e-4, betas (0.5, 0.999) and weight decay 1e-3; the network weights with SGD, momentum 0.9, weight decay 3e-4, gradient-norm clipping at 5 and a learning rate cosine-annealed from 0.1 down to 0.

Constructors

Constructor Summary
PCDARTSOptions() Creates the options with PC-DARTS’ defaults.

Properties

Property Summary
bool UseEdgeNormalization { get; set; } Whether to weight each edge by softmax(beta), the edge normalization of PC-DARTS (Sec.
int BatchSize { get; set; } Mini-batch size of both the weight and the architecture steps (default: 256).
int Epochs { get; set; } Number of search epochs (default: 50).
int NumNodes { get; set; } The number of intermediate nodes of the searched cell (4 in Xu et al.
int PartialChannelK { get; set; } The partial-channel factor K: each edge’s operation mixture sees 1/K of the channels, the rest bypass it, and a channel shuffle with K groups follows (default: 4).
int WarmupEpochs { get; set; } Number of leading epochs that train only the network weights, leaving alpha and beta unchanged (default: 15).