|
|
""" |
|
|
Superposition Patch Classifier β Standalone Inference Module |
|
|
============================================================= |
|
|
Two-tier gated geometric transformer that extracts structural |
|
|
properties from (8, 16, 16) latent patches. |
|
|
|
|
|
No dependencies beyond PyTorch. All grid/gate constants inlined. |
|
|
|
|
|
Input: (B, 8, 16, 16) β adapted latent patches |
|
|
Output: gate_vectors (B, 64, 17), patch_features (B, 64, 256), logits |
|
|
|
|
|
Usage: |
|
|
from geometric_model import load_from_hub, extract_features |
|
|
|
|
|
model, config = load_from_hub() # reads config.json + model.pt from Hub |
|
|
out = model(patches) |
|
|
|
|
|
# Gate vectors: explicit geometric properties per patch |
|
|
local_gates = torch.cat([ |
|
|
F.softmax(out["local_dim_logits"], dim=-1), # 4d: dimensionality |
|
|
F.softmax(out["local_curv_logits"], dim=-1), # 3d: curvature class |
|
|
torch.sigmoid(out["local_bound_logits"]), # 1d: boundary flag |
|
|
torch.sigmoid(out["local_axis_logits"]), # 3d: active axes |
|
|
], dim=-1) # (B, 64, 11) |
|
|
|
|
|
structural_gates = torch.cat([ |
|
|
F.softmax(out["struct_topo_logits"], dim=-1), # 2d: topology |
|
|
torch.sigmoid(out["struct_neighbor_logits"]), # 1d: neighbor density |
|
|
F.softmax(out["struct_role_logits"], dim=-1), # 3d: surface role |
|
|
], dim=-1) # (B, 64, 6) |
|
|
|
|
|
gate_vectors = torch.cat([local_gates, structural_gates], dim=-1) # (B, 64, 17) |
|
|
patch_features = out["patch_features"] # (B, 64, embed_dim) |
|
|
""" |
|
|
|
|
|
import math |
|
|
import torch |
|
|
import torch.nn as nn |
|
|
import torch.nn.functional as F |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
GZ, GY, GX = 8, 16, 16 |
|
|
PATCH_Z, PATCH_Y, PATCH_X = 2, 4, 4 |
|
|
PATCH_VOL = PATCH_Z * PATCH_Y * PATCH_X |
|
|
MACRO_Z = GZ // PATCH_Z |
|
|
MACRO_Y = GY // PATCH_Y |
|
|
MACRO_X = GX // PATCH_X |
|
|
MACRO_N = MACRO_Z * MACRO_Y * MACRO_X |
|
|
|
|
|
|
|
|
NUM_LOCAL_DIMS = 4 |
|
|
NUM_LOCAL_CURVS = 3 |
|
|
NUM_LOCAL_BOUNDARY = 1 |
|
|
NUM_LOCAL_AXES = 3 |
|
|
LOCAL_GATE_DIM = NUM_LOCAL_DIMS + NUM_LOCAL_CURVS + NUM_LOCAL_BOUNDARY + NUM_LOCAL_AXES |
|
|
|
|
|
|
|
|
NUM_STRUCT_TOPO = 2 |
|
|
NUM_STRUCT_NEIGHBOR = 1 |
|
|
NUM_STRUCT_ROLE = 3 |
|
|
STRUCTURAL_GATE_DIM = NUM_STRUCT_TOPO + NUM_STRUCT_NEIGHBOR + NUM_STRUCT_ROLE |
|
|
|
|
|
TOTAL_GATE_DIM = LOCAL_GATE_DIM + STRUCTURAL_GATE_DIM |
|
|
|
|
|
|
|
|
CLASS_NAMES = [ |
|
|
"point", "line", "corner", "cross", "arc", "helix", "circle", |
|
|
"triangle", "quad", "plane", "disc", |
|
|
"tetrahedron", "cube", "pyramid", "prism", "octahedron", "pentachoron", "wedge", |
|
|
"sphere", "hemisphere", "torus", "bowl", "saddle", "capsule", "cylinder", "cone", "channel" |
|
|
] |
|
|
NUM_CLASSES = len(CLASS_NAMES) |
|
|
|
|
|
|
|
|
GATES = ["rigid", "curved", "combined", "open", "closed"] |
|
|
NUM_GATES = len(GATES) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class PatchEmbedding3D(nn.Module): |
|
|
def __init__(self, patch_dim=64): |
|
|
super().__init__() |
|
|
self.proj = nn.Linear(PATCH_VOL, patch_dim) |
|
|
pz = torch.arange(MACRO_Z).float() / MACRO_Z |
|
|
py = torch.arange(MACRO_Y).float() / MACRO_Y |
|
|
px = torch.arange(MACRO_X).float() / MACRO_X |
|
|
pos = torch.stack(torch.meshgrid(pz, py, px, indexing='ij'), dim=-1).reshape(MACRO_N, 3) |
|
|
self.register_buffer('pos_embed', pos) |
|
|
self.pos_proj = nn.Linear(3, patch_dim) |
|
|
|
|
|
def forward(self, x): |
|
|
B = x.shape[0] |
|
|
patches = x.view(B, MACRO_Z, PATCH_Z, MACRO_Y, PATCH_Y, MACRO_X, PATCH_X) |
|
|
patches = patches.permute(0, 1, 3, 5, 2, 4, 6).contiguous().view(B, MACRO_N, PATCH_VOL) |
|
|
return self.proj(patches) + self.pos_proj(self.pos_embed) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class TransformerBlock(nn.Module): |
|
|
def __init__(self, dim, n_heads, dropout=0.1): |
|
|
super().__init__() |
|
|
self.attn = nn.MultiheadAttention(dim, n_heads, dropout=dropout, batch_first=True) |
|
|
self.ff = nn.Sequential( |
|
|
nn.Linear(dim, dim * 4), nn.GELU(), nn.Dropout(dropout), |
|
|
nn.Linear(dim * 4, dim), nn.Dropout(dropout) |
|
|
) |
|
|
self.ln1, self.ln2 = nn.LayerNorm(dim), nn.LayerNorm(dim) |
|
|
|
|
|
def forward(self, x): |
|
|
x = x + self.attn(self.ln1(x), self.ln1(x), self.ln1(x))[0] |
|
|
return x + self.ff(self.ln2(x)) |
|
|
|
|
|
|
|
|
class GatedGeometricAttention(nn.Module): |
|
|
""" |
|
|
Multi-head attention with two-tier gate modulation. |
|
|
Q, K see both local and structural gates. |
|
|
V modulated by combined gate vector. |
|
|
Per-head compatibility bias from gate interactions. |
|
|
""" |
|
|
|
|
|
def __init__(self, embed_dim, gate_dim, n_heads, dropout=0.1): |
|
|
super().__init__() |
|
|
self.embed_dim = embed_dim |
|
|
self.n_heads = n_heads |
|
|
self.head_dim = embed_dim // n_heads |
|
|
|
|
|
self.q_proj = nn.Linear(embed_dim + gate_dim, embed_dim) |
|
|
self.k_proj = nn.Linear(embed_dim + gate_dim, embed_dim) |
|
|
self.v_proj = nn.Linear(embed_dim, embed_dim) |
|
|
|
|
|
self.gate_q = nn.Linear(gate_dim, n_heads) |
|
|
self.gate_k = nn.Linear(gate_dim, n_heads) |
|
|
self.v_gate = nn.Sequential(nn.Linear(gate_dim, embed_dim), nn.Sigmoid()) |
|
|
|
|
|
self.out_proj = nn.Linear(embed_dim, embed_dim) |
|
|
self.attn_drop = nn.Dropout(dropout) |
|
|
self.scale = math.sqrt(self.head_dim) |
|
|
|
|
|
def forward(self, h, gate_features): |
|
|
B, N, _ = h.shape |
|
|
hg = torch.cat([h, gate_features], dim=-1) |
|
|
Q = self.q_proj(hg).view(B, N, self.n_heads, self.head_dim).transpose(1, 2) |
|
|
K = self.k_proj(hg).view(B, N, self.n_heads, self.head_dim).transpose(1, 2) |
|
|
|
|
|
V = self.v_proj(h) |
|
|
V = (V * self.v_gate(gate_features)).view(B, N, self.n_heads, self.head_dim).transpose(1, 2) |
|
|
|
|
|
content_scores = (Q @ K.transpose(-2, -1)) / self.scale |
|
|
gq = self.gate_q(gate_features) |
|
|
gk = self.gate_k(gate_features) |
|
|
compat = torch.einsum('bih,bjh->bhij', gq, gk) |
|
|
|
|
|
attn = F.softmax(content_scores + compat, dim=-1) |
|
|
attn = self.attn_drop(attn) |
|
|
|
|
|
out = (attn @ V).transpose(1, 2).reshape(B, N, self.embed_dim) |
|
|
return self.out_proj(out) |
|
|
|
|
|
|
|
|
class GeometricTransformerBlock(nn.Module): |
|
|
def __init__(self, embed_dim, gate_dim, n_heads, dropout=0.1, ff_mult=4): |
|
|
super().__init__() |
|
|
self.ln1 = nn.LayerNorm(embed_dim) |
|
|
self.attn = GatedGeometricAttention(embed_dim, gate_dim, n_heads, dropout) |
|
|
self.ln2 = nn.LayerNorm(embed_dim) |
|
|
self.ff = nn.Sequential( |
|
|
nn.Linear(embed_dim, embed_dim * ff_mult), nn.GELU(), nn.Dropout(dropout), |
|
|
nn.Linear(embed_dim * ff_mult, embed_dim), nn.Dropout(dropout) |
|
|
) |
|
|
|
|
|
def forward(self, h, gate_features): |
|
|
h = h + self.attn(self.ln1(h), gate_features) |
|
|
h = h + self.ff(self.ln2(h)) |
|
|
return h |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
class SuperpositionPatchClassifier(nn.Module): |
|
|
""" |
|
|
Two-tier gated geometric transformer. |
|
|
|
|
|
Stage 0: Local gates from raw patch embeddings (what IS in this patch) |
|
|
Stage 1: Bootstrap attention with local gate context |
|
|
Stage 1.5: Structural gates from post-attention features (what ROLE this patch plays) |
|
|
Stage 2: Geometric gated attention with both gate tiers |
|
|
Stage 3: Classification heads |
|
|
|
|
|
For feature extraction (no classification), use outputs: |
|
|
- gate vectors: cat(local_gates, structural_gates) β (B, 64, 17) |
|
|
- patch_features: out["patch_features"] β (B, 64, embed_dim) |
|
|
- global_features: out["global_features"] β (B, embed_dim) |
|
|
""" |
|
|
|
|
|
def __init__(self, embed_dim=128, patch_dim=64, n_bootstrap=2, n_geometric=2, |
|
|
n_heads=4, dropout=0.1): |
|
|
super().__init__() |
|
|
self.embed_dim = embed_dim |
|
|
|
|
|
|
|
|
self.patch_embed = PatchEmbedding3D(patch_dim) |
|
|
|
|
|
|
|
|
local_hidden = patch_dim * 2 |
|
|
self.local_encoder = nn.Sequential( |
|
|
nn.Linear(patch_dim, local_hidden), nn.GELU(), nn.Dropout(dropout), |
|
|
nn.Linear(local_hidden, local_hidden), nn.GELU(), nn.Dropout(dropout), |
|
|
) |
|
|
self.local_dim_head = nn.Linear(local_hidden, NUM_LOCAL_DIMS) |
|
|
self.local_curv_head = nn.Linear(local_hidden, NUM_LOCAL_CURVS) |
|
|
self.local_bound_head = nn.Linear(local_hidden, NUM_LOCAL_BOUNDARY) |
|
|
self.local_axis_head = nn.Linear(local_hidden, NUM_LOCAL_AXES) |
|
|
|
|
|
|
|
|
self.proj = nn.Linear(patch_dim + LOCAL_GATE_DIM, embed_dim) |
|
|
|
|
|
|
|
|
self.bootstrap_blocks = nn.ModuleList([ |
|
|
TransformerBlock(embed_dim, n_heads, dropout) |
|
|
for _ in range(n_bootstrap) |
|
|
]) |
|
|
|
|
|
|
|
|
self.struct_topo_head = nn.Linear(embed_dim, NUM_STRUCT_TOPO) |
|
|
self.struct_neighbor_head = nn.Linear(embed_dim, NUM_STRUCT_NEIGHBOR) |
|
|
self.struct_role_head = nn.Linear(embed_dim, NUM_STRUCT_ROLE) |
|
|
|
|
|
|
|
|
self.geometric_blocks = nn.ModuleList([ |
|
|
GeometricTransformerBlock(embed_dim, TOTAL_GATE_DIM, n_heads, dropout) |
|
|
for _ in range(n_geometric) |
|
|
]) |
|
|
|
|
|
|
|
|
gated_dim = embed_dim + TOTAL_GATE_DIM |
|
|
|
|
|
self.patch_shape_head = nn.Sequential( |
|
|
nn.Linear(gated_dim, embed_dim), nn.GELU(), nn.Dropout(dropout), |
|
|
nn.Linear(embed_dim, NUM_CLASSES) |
|
|
) |
|
|
|
|
|
self.global_pool = nn.Sequential( |
|
|
nn.Linear(gated_dim, embed_dim), nn.GELU(), |
|
|
nn.Linear(embed_dim, embed_dim) |
|
|
) |
|
|
self.global_gate_head = nn.Linear(embed_dim, NUM_GATES) |
|
|
self.global_shape_head = nn.Linear(embed_dim, NUM_CLASSES) |
|
|
|
|
|
def forward(self, x): |
|
|
|
|
|
e = self.patch_embed(x) |
|
|
|
|
|
|
|
|
e_local = self.local_encoder(e) |
|
|
local_dim_logits = self.local_dim_head(e_local) |
|
|
local_curv_logits = self.local_curv_head(e_local) |
|
|
local_bound_logits = self.local_bound_head(e_local) |
|
|
local_axis_logits = self.local_axis_head(e_local) |
|
|
|
|
|
local_gates = torch.cat([ |
|
|
F.softmax(local_dim_logits, dim=-1), |
|
|
F.softmax(local_curv_logits, dim=-1), |
|
|
torch.sigmoid(local_bound_logits), |
|
|
torch.sigmoid(local_axis_logits), |
|
|
], dim=-1) |
|
|
|
|
|
|
|
|
h = self.proj(torch.cat([e, local_gates], dim=-1)) |
|
|
for blk in self.bootstrap_blocks: |
|
|
h = blk(h) |
|
|
|
|
|
|
|
|
struct_topo_logits = self.struct_topo_head(h) |
|
|
struct_neighbor_logits = self.struct_neighbor_head(h) |
|
|
struct_role_logits = self.struct_role_head(h) |
|
|
|
|
|
structural_gates = torch.cat([ |
|
|
F.softmax(struct_topo_logits, dim=-1), |
|
|
torch.sigmoid(struct_neighbor_logits), |
|
|
F.softmax(struct_role_logits, dim=-1), |
|
|
], dim=-1) |
|
|
|
|
|
all_gates = torch.cat([local_gates, structural_gates], dim=-1) |
|
|
|
|
|
|
|
|
for blk in self.geometric_blocks: |
|
|
h = blk(h, all_gates) |
|
|
|
|
|
|
|
|
h_gated = torch.cat([h, all_gates], dim=-1) |
|
|
shape_logits = self.patch_shape_head(h_gated) |
|
|
g = self.global_pool(h_gated.mean(dim=1)) |
|
|
|
|
|
return { |
|
|
"local_dim_logits": local_dim_logits, |
|
|
"local_curv_logits": local_curv_logits, |
|
|
"local_bound_logits": local_bound_logits, |
|
|
"local_axis_logits": local_axis_logits, |
|
|
"struct_topo_logits": struct_topo_logits, |
|
|
"struct_neighbor_logits": struct_neighbor_logits, |
|
|
"struct_role_logits": struct_role_logits, |
|
|
"patch_shape_logits": shape_logits, |
|
|
"patch_features": h, |
|
|
"global_features": g, |
|
|
"global_gates": self.global_gate_head(g), |
|
|
"global_shapes": self.global_shape_head(g), |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def load_config(repo_id="AbstractPhil/geovocab-patch-maker", config_file="config.json"): |
|
|
"""Load model config from HuggingFace Hub.""" |
|
|
import json |
|
|
from huggingface_hub import hf_hub_download |
|
|
|
|
|
path = hf_hub_download(repo_id=repo_id, filename=config_file) |
|
|
with open(path, "r") as f: |
|
|
return json.load(f) |
|
|
|
|
|
|
|
|
def from_config(config, device="cpu"): |
|
|
"""Instantiate model from config dict (no weights).""" |
|
|
return SuperpositionPatchClassifier( |
|
|
embed_dim=config["embed_dim"], |
|
|
patch_dim=config["patch_dim"], |
|
|
n_bootstrap=config["n_bootstrap"], |
|
|
n_geometric=config["n_geometric"], |
|
|
n_heads=config["n_heads"], |
|
|
dropout=config.get("dropout", 0.0), |
|
|
).to(device) |
|
|
|
|
|
|
|
|
def load_from_hub( |
|
|
repo_id="AbstractPhil/geovocab-patch-maker", |
|
|
weights_file="model.pt", |
|
|
config_file="config.json", |
|
|
device="cuda" if torch.cuda.is_available() else "cpu", |
|
|
): |
|
|
""" |
|
|
Load pretrained model from HuggingFace Hub. |
|
|
|
|
|
Reads config.json for architecture, model.pt for weights. |
|
|
Falls back to config embedded in checkpoint if config.json missing. |
|
|
""" |
|
|
from huggingface_hub import hf_hub_download |
|
|
|
|
|
|
|
|
try: |
|
|
config = load_config(repo_id, config_file) |
|
|
print(f"β Config loaded from {config_file}") |
|
|
except Exception: |
|
|
config = None |
|
|
|
|
|
|
|
|
weights_path = hf_hub_download(repo_id=repo_id, filename=weights_file) |
|
|
ckpt = torch.load(weights_path, map_location=device, weights_only=False) |
|
|
|
|
|
|
|
|
if config is None: |
|
|
config = ckpt["config"] |
|
|
print(f" Config from checkpoint (no {config_file} found)") |
|
|
|
|
|
model = from_config(config, device=device) |
|
|
model.load_state_dict(ckpt["model_state_dict"]) |
|
|
model.eval() |
|
|
|
|
|
epoch = ckpt.get("epoch", "?") |
|
|
n_params = sum(p.numel() for p in model.parameters()) |
|
|
print(f"β Loaded {repo_id} (epoch {epoch}, {n_params:,} params)") |
|
|
return model, config |
|
|
|
|
|
|
|
|
@torch.no_grad() |
|
|
def extract_features(model, patches, batch_size=256): |
|
|
""" |
|
|
Convenience: patches β (gate_vectors, patch_features) |
|
|
|
|
|
Args: |
|
|
model: SuperpositionPatchClassifier (eval mode) |
|
|
patches: (N, 8, 16, 16) tensor |
|
|
batch_size: inference batch size |
|
|
|
|
|
Returns: |
|
|
gate_vectors: (N, 64, 17) β explicit geometric properties |
|
|
patch_features: (N, 64, embed_dim) β learned representations |
|
|
""" |
|
|
device = next(model.parameters()).device |
|
|
all_gates, all_patch = [], [] |
|
|
|
|
|
for s in range(0, patches.shape[0], batch_size): |
|
|
batch = patches[s:s + batch_size].to(device) |
|
|
out = model(batch) |
|
|
|
|
|
local = torch.cat([ |
|
|
F.softmax(out["local_dim_logits"], dim=-1), |
|
|
F.softmax(out["local_curv_logits"], dim=-1), |
|
|
torch.sigmoid(out["local_bound_logits"]), |
|
|
torch.sigmoid(out["local_axis_logits"]), |
|
|
], dim=-1) |
|
|
|
|
|
struct = torch.cat([ |
|
|
F.softmax(out["struct_topo_logits"], dim=-1), |
|
|
torch.sigmoid(out["struct_neighbor_logits"]), |
|
|
F.softmax(out["struct_role_logits"], dim=-1), |
|
|
], dim=-1) |
|
|
|
|
|
all_gates.append(torch.cat([local, struct], dim=-1).cpu()) |
|
|
all_patch.append(out["patch_features"].cpu()) |
|
|
|
|
|
return torch.cat(all_gates), torch.cat(all_patch) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
|
import json |
|
|
|
|
|
|
|
|
model = SuperpositionPatchClassifier() |
|
|
n_params = sum(p.numel() for p in model.parameters()) |
|
|
print(f"SuperpositionPatchClassifier: {n_params:,} parameters") |
|
|
|
|
|
x = torch.randn(2, 8, 16, 16) |
|
|
out = model(x) |
|
|
print(f" Input: {x.shape}") |
|
|
print(f" patch_features: {out['patch_features'].shape}") |
|
|
print(f" local_dim: {out['local_dim_logits'].shape}") |
|
|
print(f" struct_topo: {out['struct_topo_logits'].shape}") |
|
|
print(f" patch_shapes: {out['patch_shape_logits'].shape}") |
|
|
print(f" global_features: {out['global_features'].shape}") |
|
|
|
|
|
|
|
|
import os |
|
|
cfg_path = os.path.join(os.path.dirname(__file__), "config.json") |
|
|
if os.path.exists(cfg_path): |
|
|
with open(cfg_path) as f: |
|
|
config = json.load(f) |
|
|
model2 = from_config(config) |
|
|
print(f"\n from_config: {sum(p.numel() for p in model2.parameters()):,} params") |
|
|
print(f" config: {config['model_type']} embed={config['embed_dim']} patches={config['num_patches']}") |