ChronosOptions<T>
Options & Configuration · AiDotNet.TimeSeries
Options for ChronosFoundationModel. The defaults are Chronos-T5 (Small) and the paper’s training
recipe (Ansari et al. 2024, §5.2); the released sizes are Tiny (d_model 256, d_ff 1024, 4 heads, 4 layers), Mini
(384, 1536, 8, 4), Small (512, 2048, 8, 6), Base (768, 3072, 12, 12) and Large (1024, 4096, 16, 24), all with
d_kv 64.
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → ChronosOptions<T>
Known derived types: ChronosFinanceOptions<T>
Constructors
| Constructor | Summary |
|---|---|
ChronosOptions() |
|
ChronosOptions(ChronosOptions<T> other) |
Properties
| Property | Summary |
|---|---|
double DropoutRate { get; set; } |
T5’s dropout rate (the released configurations’ 0.1). |
double HighLimit { get; set; } |
The highest bin center, in units of the mean scale (the paper’s 15). |
double LearningRate { get; set; } |
AdamW’s learning rate, decayed linearly to zero over TrainingSteps (the paper’s 0.001). |
double LowLimit { get; set; } |
The lowest bin center, in units of the mean scale (the paper’s −15). |
double MaxGradNorm { get; set; } |
The global gradient norm clip (the transformers default 1; 0 disables it). |
double MissingValueProbability { get; set; } |
The largest per-window probability of dropping a value to missing in training (the official 0.2). |
double Temperature { get; set; } |
The sampling temperature (the official 1). |
double TopP { get; set; } |
The nucleus cut of sampling (the official 1, none). |
double WeightDecay { get; set; } |
AdamW’s weight decay (the paper’s 0.01). |
int BatchSize { get; set; } |
The windows per step (the paper’s effective batch of 256). |
int ContextLength { get; set; } |
The context the model reads: the last ContextLength values (the paper’s 512). |
int FeedForwardDim { get; set; } |
T5’s d_ff. |
int KeyValueDim { get; set; } |
T5’s d_kv, the width of each attention head. |
int MinPast { get; set; } |
The values a training window’s split point is at least past the start (the official 60). |
int ModelDim { get; set; } |
T5’s d_model. |
int NumHeads { get; set; } |
T5’s attention heads. |
int NumLayers { get; set; } |
The encoder’s and the decoder’s blocks. |
int NumSamples { get; set; } |
The sample paths a forecast draws (the paper’s 20). |
int NumTokens { get; set; } |
The vocabulary: the bins plus PAD and EOS (the paper’s 4096). |
int PredictionLength { get; set; } |
The steps one sampling pass predicts, and the label length in training (the paper’s 64). |
int RelativeAttentionBuckets { get; set; } |
T5’s relative attention buckets. |
int RelativeAttentionMaxDistance { get; set; } |
T5’s relative attention maximum distance. |
int TopK { get; set; } |
The top-k cut of sampling (the official 50). |
int TrainingSteps { get; set; } |
The optimizer steps (the paper’s 200K). |
Methods
| Method | Summary |
|---|---|
ChronosOptions<T> FromPretrainedConfig(string configJsonPath) |
The options of a released checkpoint, read from its config.json; the training settings keep their defaults. |