| """Numeric and log state encoders. |
| |
| NumericStateEncoder: a fixed-length feature vector (metrics, sensor readings, a |
| tabular row) -> one token per feature (value * feature embedding + feature id), |
| so the core can attend to individual features. Missing values are NaN -> masked. |
| |
| LogStateEncoder: a text state encoder for logs. Logs are text, so it reuses the |
| text encoder; it only normalises volatile tokens (hex ids, numbers, timestamps) |
| and keeps the last `max_lines` lines before tokenisation (see minijev.text_task_utils). |
| """ |
| from __future__ import annotations |
|
|
| import torch |
| from torch import nn |
|
|
|
|
| class NumericStateEncoder(nn.Module): |
| uses_backbone = False |
|
|
| def __init__(self, d: int, n_features: int, **_): |
| super().__init__() |
| self.feat = nn.Parameter(torch.zeros(n_features, d)) |
| self.scale = nn.Parameter(torch.zeros(n_features, d)) |
| nn.init.normal_(self.feat, std=0.02) |
| nn.init.normal_(self.scale, std=0.02) |
| self.mlp = nn.Sequential(nn.LayerNorm(d), nn.Linear(d, d), nn.GELU(), nn.Linear(d, d)) |
|
|
| def forward(self, x): |
| valid = ~torch.isnan(x) |
| v = torch.nan_to_num(x).unsqueeze(-1) |
| tok = self.feat + v * self.scale |
| tok = tok + self.mlp(tok) |
| return tok, valid |
|
|
