geolip-flow-predictions / test_cases.py
AbstractPhil's picture
Update test_cases.py
013d216 verified
"""
geolip.flows β€” Multi-flow ensemble for constellation geometry.
Each flow predicts the same geometric output using a different mathematical
formulation. The ensemble fuses predictions based on learned confidence.
Flows:
QuaternionFlow β€” Full MHA quaternion rotation (existing, heavyweight)
QuaternionLiteFlow β€” Staged quaternion with lighter spectral computation
VelocityFlow β€” Angular velocity dq/dt on the tangent bundle
MagnitudeFlow β€” Flow magnitude via Gram eigenvalue spectrum
OrbitalFlow β€” Omega-based orbital resonance using FL eigh
AlignmentFlow β€” SVD alignment via Procrustes rotation
Architecture:
Each flow: same input (anchors [B,k,d], queries [B,n,d]) β†’ output [B,n,d]
Ensemble: weighted fusion with learned per-flow confidence
Usage:
from geolip.flows import FlowEnsemble, OrbitalFlow, AlignmentFlow
ensemble = FlowEnsemble(
flows=[OrbitalFlow(d=256, k=128), AlignmentFlow(d=256, k=128)],
d_model=256,
)
output = ensemble(anchors, queries) # [B, n, d]
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from typing import List, Optional, Tuple
# Use geolip_core.linalg when available (FL eigh, Triton SVD, etc.)
# Falls back to torch.linalg transparently
try:
import geolip_core.linalg as LA
except ImportError:
import torch.linalg as LA
# ═══════════════════════════════════════════════════════════════════
# Base Flow
# ═══════════════════════════════════════════════════════════════════
class BaseFlow(nn.Module):
"""Base class for all geometric flows.
All flows share the same interface:
Input: anchors [B, k, d], queries [B, n, d]
Output: prediction [B, n, d], confidence [B, n, 1]
Subclasses implement _flow() with their specific math.
"""
def __init__(self, d_model: int, n_anchors: int, name: str = 'base'):
super().__init__()
self.d_model = d_model
self.n_anchors = n_anchors
self.name = name
# Confidence head: scalar per query position
self.confidence = nn.Sequential(
nn.Linear(d_model, d_model // 4),
nn.GELU(),
nn.Linear(d_model // 4, 1),
)
def forward(self, anchors: Tensor, queries: Tensor) -> Tuple[Tensor, Tensor]:
"""
Args:
anchors: [B, k, d] constellation anchor points
queries: [B, n, d] query embeddings
Returns:
prediction: [B, n, d] geometric prediction
confidence: [B, n, 1] per-query confidence score
"""
pred = self._flow(anchors, queries)
conf = torch.sigmoid(self.confidence(pred))
return pred, conf
def _flow(self, anchors: Tensor, queries: Tensor) -> Tensor:
raise NotImplementedError
# ═══════════════════════════════════════════════════════════════════
# QuaternionFlow β€” Full MHA quaternion rotation
# ═══════════════════════════════════════════════════════════════════
class QuaternionFlow(BaseFlow):
"""Full multi-head attention with quaternion geometric rotation.
Computes query-anchor attention, extracts rotation quaternion from
attention-weighted anchor geometry, applies rotation to queries.
Heavyweight β€” the full-fidelity path.
"""
def __init__(self, d_model: int, n_anchors: int, n_heads: int = 4):
super().__init__(d_model, n_anchors, name='quaternion')
self.n_heads = n_heads
self.head_dim = d_model // n_heads
self.q_proj = nn.Linear(d_model, d_model)
self.k_proj = nn.Linear(d_model, d_model)
self.v_proj = nn.Linear(d_model, d_model)
self.out_proj = nn.Linear(d_model, d_model)
# Quaternion components: scalar + 3 imaginary from attention output
self.quat_proj = nn.Linear(d_model, 4)
def _flow(self, anchors, queries):
B, n, d = queries.shape
k = anchors.shape[1]
h = self.n_heads; hd = self.head_dim
Q = self.q_proj(queries).view(B, n, h, hd).transpose(1, 2)
K = self.k_proj(anchors).view(B, k, h, hd).transpose(1, 2)
V = self.v_proj(anchors).view(B, k, h, hd).transpose(1, 2)
attn = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(hd)
attn = F.softmax(attn, dim=-1)
ctx = torch.matmul(attn, V).transpose(1, 2).reshape(B, n, d)
# Extract quaternion and apply rotation
q = self.quat_proj(ctx) # [B, n, 4]
q = F.normalize(q, dim=-1)
rotated = self._quat_rotate(queries, q)
return self.out_proj(ctx + rotated)
def _quat_rotate(self, v, q):
"""Apply quaternion rotation to vectors. q: [B,n,4], v: [B,n,d]."""
# For d > 3: rotate first 3 dims, pass rest through
w, x, y, z = q[..., 0:1], q[..., 1:2], q[..., 2:3], q[..., 3:4]
v3 = v[..., :3]
# q * v * q^-1 via Rodriguez
t = 2.0 * torch.cross(torch.cat([x, y, z], dim=-1), v3, dim=-1)
v3_rot = v3 + w * t + torch.cross(torch.cat([x, y, z], dim=-1), t, dim=-1)
if v.shape[-1] > 3:
return torch.cat([v3_rot, v[..., 3:]], dim=-1)
return v3_rot
# ═══════════════════════════════════════════════════════════════════
# QuaternionLiteFlow β€” Staged lighter quaternion
# ═══════════════════════════════════════════════════════════════════
class QuaternionLiteFlow(BaseFlow):
"""Lightweight quaternion prediction without full MHA.
Uses anchor centroid + query projection to predict rotation directly.
Much lighter than full QuaternionFlow β€” trades attention resolution
for speed.
"""
def __init__(self, d_model: int, n_anchors: int):
super().__init__(d_model, n_anchors, name='quat_lite')
self.anchor_compress = nn.Linear(d_model, d_model)
self.query_proj = nn.Linear(d_model, d_model)
self.quat_head = nn.Sequential(
nn.Linear(d_model * 2, d_model),
nn.GELU(),
nn.Linear(d_model, 4),
)
self.out_proj = nn.Linear(d_model, d_model)
def _flow(self, anchors, queries):
B, n, d = queries.shape
# Compress anchors to single geometric summary
anchor_ctx = self.anchor_compress(anchors.mean(dim=1, keepdim=True)) # [B, 1, d]
anchor_ctx = anchor_ctx.expand(B, n, d)
q_proj = self.query_proj(queries)
combined = torch.cat([q_proj, anchor_ctx], dim=-1) # [B, n, 2d]
q = F.normalize(self.quat_head(combined), dim=-1)
rotated = self._quat_rotate_simple(queries, q)
return self.out_proj(rotated)
def _quat_rotate_simple(self, v, q):
w, xyz = q[..., 0:1], q[..., 1:4]
v3 = v[..., :3]
t = 2.0 * torch.cross(xyz, v3, dim=-1)
v3_rot = v3 + w * t + torch.cross(xyz, t, dim=-1)
if v.shape[-1] > 3:
return torch.cat([v3_rot, v[..., 3:]], dim=-1)
return v3_rot
# ═══════════════════════════════════════════════════════════════════
# VelocityFlow β€” Angular velocity on tangent bundle
# ═══════════════════════════════════════════════════════════════════
class VelocityFlow(BaseFlow):
"""Angular velocity flow on the tangent space of the constellation.
Models dq/dt: the rate of change of the query embedding induced by
the anchor geometry. Predicts velocity, integrates with Euler step.
The velocity is tangent to the hypersphere at each query point.
"""
def __init__(self, d_model: int, n_anchors: int):
super().__init__(d_model, n_anchors, name='velocity')
# Anchor-query interaction β†’ velocity field
self.anchor_proj = nn.Linear(d_model, d_model)
self.query_proj = nn.Linear(d_model, d_model)
self.vel_head = nn.Sequential(
nn.Linear(d_model, d_model),
nn.GELU(),
nn.Linear(d_model, d_model),
)
self.dt = nn.Parameter(torch.tensor(0.1)) # learnable step size
def _flow(self, anchors, queries):
B, n, d = queries.shape
# Compute direction from queries toward anchor centroid
a_proj = self.anchor_proj(anchors) # [B, k, d]
q_proj = self.query_proj(queries) # [B, n, d]
# Soft attention: query-anchor similarity β†’ weighted anchor direction
sim = torch.bmm(q_proj, a_proj.transpose(-2, -1)) # [B, n, k]
weights = F.softmax(sim / math.sqrt(d), dim=-1)
direction = torch.bmm(weights, a_proj) # [B, n, d]
# Velocity: project onto tangent space at query
velocity = self.vel_head(direction - q_proj)
# Tangent projection: remove component along query direction
q_norm = F.normalize(queries, dim=-1)
radial = (velocity * q_norm).sum(dim=-1, keepdim=True) * q_norm
tangent_vel = velocity - radial
# Euler integration
return queries + self.dt * tangent_vel
# ═══════════════════════════════════════════════════════════════════
# MagnitudeFlow β€” Gram eigenvalue spectrum
# ═══════════════════════════════════════════════════════════════════
class MagnitudeFlow(BaseFlow):
"""Flow based on the Gram matrix eigenvalue magnitude spectrum.
Computes the anchor Gram matrix, extracts eigenvalues via FL eigh,
uses the spectral profile to modulate query embeddings.
The eigenvalue magnitudes encode the constellation's energy distribution
across geometric modes.
"""
def __init__(self, d_model: int, n_anchors: int):
super().__init__(d_model, n_anchors, name='magnitude')
# Project anchors to small geometric space for Gram computation
self.geom_dim = min(n_anchors, 12) # FL eigh sweet spot
self.anchor_proj = nn.Linear(d_model, self.geom_dim)
# Spectral β†’ modulation
self.spec_proj = nn.Sequential(
nn.Linear(self.geom_dim, d_model),
nn.GELU(),
nn.Linear(d_model, d_model),
)
self.query_proj = nn.Linear(d_model, d_model)
self.gate = nn.Linear(d_model * 2, d_model)
def _flow(self, anchors, queries):
B, n, d = queries.shape
# Project anchors to geometric space
a_geom = self.anchor_proj(anchors) # [B, k, geom_dim]
# Gram matrix
G = torch.bmm(a_geom.transpose(-2, -1), a_geom) # [B, geom_dim, geom_dim]
# Eigendecomposition β€” differentiable through torch.linalg.eigh
eigenvalues, _ = LA.eigh(G, method='torch') # [B, geom_dim]
# Magnitude spectrum: how energy distributes across modes
magnitudes = eigenvalues.abs().sqrt() # [B, geom_dim] β€” the Ο‰ spectrum
spec_embed = self.spec_proj(magnitudes) # [B, d]
spec_embed = spec_embed.unsqueeze(1).expand(B, n, d)
# Gate: blend spectral modulation with query
q_proj = self.query_proj(queries)
gate_input = torch.cat([q_proj, spec_embed], dim=-1)
g = torch.sigmoid(self.gate(gate_input))
return queries + g * spec_embed
# ═══════════════════════════════════════════════════════════════════
# OrbitalFlow β€” Omega angular resonance via FL eigh
# ═══════════════════════════════════════════════════════════════════
class OrbitalFlow(BaseFlow):
"""Omega-based orbital resonance flow.
Computes the constellation's resonance frequencies (Ο‰α΅’ = √λᡒ from
Gram eigendecomposition), then uses the full eigendecomposition to
project queries into the resonance basis, apply frequency-dependent
modulation, and project back.
This flow directly uses the Ο‰ spectrum to shape the geometric response.
Modes in the CV band [0.447, 0.480] (corresponding to λ ∈ [0.20, 0.23])
are amplified. Modes outside are attenuated.
"""
def __init__(self, d_model: int, n_anchors: int, cv_lo: float = 0.20, cv_hi: float = 0.23):
super().__init__(d_model, n_anchors, name='orbital')
self.geom_dim = min(n_anchors, 12)
self.anchor_proj = nn.Linear(d_model, self.geom_dim)
self.cv_lo = cv_lo
self.cv_hi = cv_hi
# Per-mode learnable response curve
self.mode_response = nn.Parameter(torch.ones(self.geom_dim))
# Projection back to d_model
self.query_to_geom = nn.Linear(d_model, self.geom_dim)
self.geom_to_query = nn.Linear(self.geom_dim, d_model)
self.out_proj = nn.Linear(d_model, d_model)
def _flow(self, anchors, queries):
B, n, d = queries.shape
a_geom = self.anchor_proj(anchors) # [B, k, geom_dim]
G = torch.bmm(a_geom.transpose(-2, -1), a_geom) # [B, gd, gd]
# Eigendecomposition β€” the Ο‰ spectrum (differentiable via torch.linalg.eigh)
eigenvalues, eigenvectors = LA.eigh(G, method='torch') # [B, gd], [B, gd, gd]
# Ο‰ = √|Ξ»|
omega = eigenvalues.abs().sqrt() # [B, gd]
# CV band resonance: modes near the attractor basin get amplified
in_band = ((eigenvalues >= self.cv_lo) & (eigenvalues <= self.cv_hi)).float()
near_binding = torch.exp(-10.0 * (eigenvalues - 0.29154).pow(2))
# Mode weighting: learned response Γ— geometric structure
mode_weight = self.mode_response.unsqueeze(0) * (1.0 + in_band + near_binding)
# Project queries into resonance basis
q_geom = self.query_to_geom(queries) # [B, n, gd]
# Rotate into eigenbasis: q_eigen = q_geom @ V
q_eigen = torch.bmm(q_geom, eigenvectors) # [B, n, gd]
# Apply frequency-dependent modulation
q_modulated = q_eigen * mode_weight.unsqueeze(1) # [B, n, gd]
# Rotate back: q_out = q_modulated @ V^T
q_out = torch.bmm(q_modulated, eigenvectors.transpose(-2, -1))
# Project back to d_model
return self.out_proj(self.geom_to_query(q_out) + queries)
# ═══════════════════════════════════════════════════════════════════
# AlignmentFlow β€” SVD-based Procrustes alignment
# ═══════════════════════════════════════════════════════════════════
class AlignmentFlow(BaseFlow):
"""SVD alignment flow via soft Procrustes rotation in projected space.
Projects to geom_dim, computes optimal rotation via SVD of the
cross-covariance in the small space, applies rotation, projects back.
"""
def __init__(self, d_model: int, n_anchors: int):
super().__init__(d_model, n_anchors, name='alignment')
self.geom_dim = min(n_anchors, 12) # FL eigh sweet spot
self.anchor_proj = nn.Linear(d_model, self.geom_dim)
self.query_proj = nn.Linear(d_model, self.geom_dim)
self.geom_to_query = nn.Linear(self.geom_dim, d_model)
self.strength = nn.Parameter(torch.tensor(0.1))
def _flow(self, anchors, queries):
B, n, d = queries.shape
# Project to small geometric space
a_proj = self.anchor_proj(anchors) # [B, k, geom_dim]
q_proj = self.query_proj(queries) # [B, n, geom_dim]
# Attention-weighted anchors β†’ per-query targets [B, n, geom_dim]
sim = torch.bmm(q_proj, a_proj.transpose(-2, -1)) / math.sqrt(self.geom_dim)
weights = F.softmax(sim, dim=-1) # [B, n, k]
targets = torch.bmm(weights, a_proj) # [B, n, geom_dim]
# Cross-covariance in small space: [B, geom_dim, geom_dim]
C = torch.bmm(q_proj.transpose(-2, -1), targets)
# SVD β†’ optimal rotation via gram_eigh (differentiable, no in-place ops)
U, _, Vh = LA.svd(C, method='gram_eigh')
R = torch.bmm(U, Vh) # [B, geom_dim, geom_dim]
# Rotate queries in small space, project back to d_model
q_rotated = torch.bmm(q_proj, R) # [B, n, geom_dim]
delta = self.geom_to_query(q_rotated - q_proj) # [B, n, d]
return queries + self.strength * delta
# ═══════════════════════════════════════════════════════════════════
# Flow Ensemble
# ═══════════════════════════════════════════════════════════════════
class FlowEnsemble(nn.Module):
"""Ensemble fusion of multiple geometric flows.
Each flow produces a prediction and a confidence score.
The ensemble fuses predictions weighted by confidence.
The fusion can be:
'weighted': confidence-weighted average
'gated': learned gate over concatenated predictions
'residual': sum of confidence-weighted residuals from input
"""
def __init__(self, flows: List[BaseFlow], d_model: int, fusion: str = 'weighted'):
super().__init__()
self.flows = nn.ModuleList(flows)
self.d_model = d_model
self.fusion = fusion
self.n_flows = len(flows)
if fusion == 'gated':
self.gate = nn.Sequential(
nn.Linear(d_model * self.n_flows, d_model),
nn.GELU(),
nn.Linear(d_model, d_model),
)
# Per-flow learnable temperature
self.temperature = nn.Parameter(torch.ones(self.n_flows))
def forward(self, anchors: Tensor, queries: Tensor) -> Tensor:
"""
Args:
anchors: [B, k, d] constellation anchors
queries: [B, n, d] query embeddings
Returns:
fused: [B, n, d] ensemble prediction
"""
predictions = []
confidences = []
for i, flow in enumerate(self.flows):
pred, conf = flow(anchors, queries)
predictions.append(pred)
confidences.append(conf * self.temperature[i])
if self.fusion == 'weighted':
return self._weighted_fusion(predictions, confidences)
elif self.fusion == 'gated':
return self._gated_fusion(predictions, confidences)
elif self.fusion == 'residual':
return self._residual_fusion(predictions, confidences, queries)
else:
raise ValueError(f"Unknown fusion: {self.fusion}")
def _weighted_fusion(self, preds, confs):
# Stack confidences and normalize
conf_stack = torch.cat(confs, dim=-1) # [B, n, n_flows]
weights = F.softmax(conf_stack, dim=-1) # [B, n, n_flows]
pred_stack = torch.stack(preds, dim=-1) # [B, n, d, n_flows]
return (pred_stack * weights.unsqueeze(-2)).sum(dim=-1)
def _gated_fusion(self, preds, confs):
cat = torch.cat(preds, dim=-1) # [B, n, d * n_flows]
return self.gate(cat)
def _residual_fusion(self, preds, confs, queries):
conf_stack = torch.cat(confs, dim=-1)
weights = F.softmax(conf_stack, dim=-1)
residuals = torch.stack([p - queries for p in preds], dim=-1)
fused_residual = (residuals * weights.unsqueeze(-2)).sum(dim=-1)
return queries + fused_residual
def flow_diagnostics(self, anchors: Tensor, queries: Tensor) -> dict:
"""Run all flows and return per-flow diagnostics."""
diag = {}
for i, flow in enumerate(self.flows):
pred, conf = flow(anchors, queries)
diag[flow.name] = {
'pred_norm': pred.norm(dim=-1).mean().item(),
'confidence_mean': conf.mean().item(),
'confidence_std': conf.std().item(),
'residual_norm': (pred - queries).norm(dim=-1).mean().item(),
'temperature': self.temperature[i].item(),
}
return diag