TimeSeAD Extensions

This repository contains extension models and utilities for the TimeSeAD library.

NeutralAD encoder configuration

NeutralAD supports multiple encoder backends via the encoder_type argument. The encoder is configured with encoder_cfg, which is passed directly to the selected encoder class.

encoder_type="patchtst"

Uses PatchTSTEncoder, which tokenizes each channel into patches before running a transformer stack.

Example:

from timesead_ext.models.other import NeutralAD

model = NeutralAD(
    ts_channels=3,
    seq_len=96,
    encoder_type="patchtst",
    encoder_cfg={
        "patch_len": 16,
        "patch_stride": 8,
        "d_model": 128,
        "num_heads": 4,
        "num_layers": 2,
        "d_ff": 256,
        "dropout": 0.1,
    },
    pooling="mean",
    proj_head="mlp",
    proj_cfg={"hidden_dim": 256, "dropout": 0.1},
)

encoder_type="itransformer"

Uses ITransformerEncoder, which projects each variate (channel) into the model dimension and applies transformer blocks across variates.

Example:

from timesead_ext.models.other import NeutralAD

model = NeutralAD(
    ts_channels=5,
    seq_len=64,
    encoder_type="itransformer",
    encoder_cfg={
        "d_model": 128,
        "num_heads": 4,
        "num_layers": 2,
        "d_ff": 256,
        "dropout": 0.1,
    },
    pooling="meanmax",
    proj_head="linear",
)

Pooling and projection heads

NeutralAD combines encoder outputs with a pooling strategy and an optional projection head:

• pooling
  • "base": use the encoder's internal pooling (configure via encoder_cfg["pooling"]).
  • "mean": mean pooling over tokens.
  • "meanmax": concatenated mean and max pooling.
  • "attn": attention pooling.
• proj_head
  • "base" or "linear": linear projection into the latent dimension.
  • "mlp": MLP projection head, configure with proj_cfg (hidden_dim, dropout).
  • "identity"/"none": skip projection.