src/cross_features/prep.py

"""PReP — Perturbation-Robustness Probes (Phase 4, idx 10).

Run K=16 noisy forward passes per sample with Gaussian noise injected
on the input embeddings: e_k = e_0 + σ · N(0, I) where σ = SIGMA_REL ·
‖e_0‖ / √(numel(e_0)). Collect anchor-layer [CLS] hidden states
Z_l = {z_l^(k)}_{k=0}^{15} for l ∈ {0, 4, 8, 12} and summarise:

  per-anchor:
    effective rank of cov(Z_l)          (1)
    mean displacement from clean        (1)
  cross-anchor (4 layer pairs × 18 Ripser scalars):
    persistence on the 16×16 distance matrix between perturbations
    at each anchor pair                  (4 × 18 = 72)

Output dim = 4·2 + 4·18 = 80.

Memory: noise is shared across the batch, so the per-batch cost is
roughly 16 × (one forward). Disables ``torch.no_grad`` is unnecessary —
forwards are pure inference.
"""
from __future__ import annotations

import numpy as np
import torch

from .persistence_summary import ripser_summary_from_distance, column_names as pcol


ANCHORS = [0, 4, 8, 12]
K = 16
SIGMA_REL = 0.05


def _columns():
    cols = []
    for l in ANCHORS:
        cols.append(f"prep_L{l}_effrank")
        cols.append(f"prep_L{l}_displacement")
    for k in range(len(ANCHORS)):
        cols.extend(pcol(f"prep_anchor{ANCHORS[k]}"))
    return cols


COLUMNS = _columns()
DIM = len(COLUMNS)


def _input_embed_module(model):
    if hasattr(model, "roberta"):
        return model.roberta.embeddings
    if hasattr(model, "electra"):
        return model.electra.embeddings
    return model.embeddings


@torch.no_grad()
def _forward_with_embedding_noise(model, ids, att, noise, anchors_set):
    """Single forward with noise added on input embeddings; capture anchor outputs."""
    emb_mod = _input_embed_module(model)
    captured = {}
    handles = []

    def make_hook(l_idx, label):
        def fn(module, inputs, output):
            if isinstance(output, tuple): output = output[0]
            captured[label] = output.detach().clone()
        return fn

    # Intercept the embedding output and add noise
    def emb_post_hook(module, inputs, output):
        return output + noise

    handles.append(emb_mod.register_forward_hook(emb_post_hook))

    # Capture anchor layer outputs (using HF's hidden_states convention)
    # We use the encoder's per-layer hook (covers anchors > 0).
    if hasattr(model, "roberta"):
        encoder_layers = model.roberta.encoder.layer
    elif hasattr(model, "electra"):
        encoder_layers = model.electra.encoder.layer
    else:
        encoder_layers = model.encoder.layer
    for l in anchors_set:
        if l == 0:
            # anchor 0 == embedding output
            handles.append(emb_mod.register_forward_hook(make_hook(l, l)))
        else:
            handles.append(encoder_layers[l - 1].register_forward_hook(make_hook(l, l)))

    try:
        _ = model(input_ids=ids, attention_mask=att, return_dict=True)
    finally:
        for h in handles: h.remove()
    return captured


def extract_prep(model, input_ids, attention_mask, cache, pred_label=None):
    """Return (features (B, 80), columns)."""
    device = input_ids.device
    B = input_ids.shape[0]
    feats = np.zeros((B, DIM), dtype=np.float32)
    emb_mod = _input_embed_module(model)
    # Clean embedding to scale noise
    with torch.no_grad():
        e0 = emb_mod(input_ids)         # (B, T, D)
    sigma = SIGMA_REL * e0.norm() / float(np.sqrt(e0.numel()))

    # Anchors → list of (K+1) (B, T, D) tensors at each anchor
    anchors_set = set(ANCHORS)
    # Clean pass: anchor states
    clean = _forward_with_embedding_noise(model, input_ids, attention_mask,
                                           noise=torch.zeros_like(e0),
                                           anchors_set=anchors_set)

    # Noisy passes
    perturbed = {l: [] for l in ANCHORS}
    for k in range(K):
        noise = sigma * torch.randn_like(e0)
        cap = _forward_with_embedding_noise(model, input_ids, attention_mask,
                                              noise=noise, anchors_set=anchors_set)
        for l in ANCHORS:
            perturbed[l].append(cap[l][:, 0, :].cpu().numpy())   # (B, D)

    # Per-sample features
    for b in range(B):
        for ai, l in enumerate(ANCHORS):
            # Z_l: (K, D)  at sample b
            Z = np.stack([perturbed[l][k][b] for k in range(K)], axis=0)
            clean_b = clean[l][b, 0, :].cpu().numpy()
            # Effective rank of cov(Z)
            Zc = Z - Z.mean(axis=0, keepdims=True)
            cov = Zc.T @ Zc / max(K - 1, 1)
            try:
                sv = np.linalg.svd(cov, compute_uv=False)
                p = sv / (sv.sum() + 1e-12)
                eff_rank = float(np.exp(-(p * np.log(p + 1e-12)).sum()))
            except np.linalg.LinAlgError:
                eff_rank = 0.0
            displacement = float(np.linalg.norm(Z - clean_b[None, :], axis=1).mean())
            feats[b, ai * 2 + 0] = eff_rank
            feats[b, ai * 2 + 1] = displacement

            # Persistence on the K×K distance matrix of perturbations at anchor l
            Zn = Z / (np.linalg.norm(Z, axis=1, keepdims=True) + 1e-12)
            cos = Zn @ Zn.T
            D = 1.0 - cos
            np.fill_diagonal(D, 0.0)
            pers, _ = ripser_summary_from_distance(D, maxdim=1,
                                                      prefix=f"prep_anchor{l}")
            base = len(ANCHORS) * 2 + ai * 18
            feats[b, base : base + 18] = pers
    return feats, COLUMNS