Delete src/dima/dimax.py
Browse files- src/dima/dimax.py +0 -597
src/dima/dimax.py
DELETED
|
@@ -1,597 +0,0 @@
|
|
| 1 |
-
# src/dima/dima.py
|
| 2 |
-
from __future__ import annotations
|
| 3 |
-
|
| 4 |
-
import json
|
| 5 |
-
import time
|
| 6 |
-
from dataclasses import asdict, dataclass
|
| 7 |
-
from typing import Any, Dict, Optional, Union
|
| 8 |
-
|
| 9 |
-
import numpy as np
|
| 10 |
-
|
| 11 |
-
import jax
|
| 12 |
-
import jax.numpy as jnp
|
| 13 |
-
from jax import random
|
| 14 |
-
|
| 15 |
-
from flax import serialization as flax_ser
|
| 16 |
-
|
| 17 |
-
from .dmap import DMAP
|
| 18 |
-
from .gplm import GPLM
|
| 19 |
-
from .ddpm import DDPM
|
| 20 |
-
from .ann import ANNBackend, make_ann
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
# ----------------------------
|
| 24 |
-
# Optional: Hugging Face Hub
|
| 25 |
-
# ----------------------------
|
| 26 |
-
try:
|
| 27 |
-
from huggingface_hub import HfApi, HfFolder, upload_file, hf_hub_download
|
| 28 |
-
_HAS_HF = True
|
| 29 |
-
except Exception:
|
| 30 |
-
_HAS_HF = False
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
def _select_device(prefer: str = "auto"):
|
| 34 |
-
"""
|
| 35 |
-
Safe JAX device selection.
|
| 36 |
-
prefer: "auto" | "gpu" | "cpu"
|
| 37 |
-
"""
|
| 38 |
-
prefer = (prefer or "auto").lower()
|
| 39 |
-
devs = jax.devices()
|
| 40 |
-
gpu = [d for d in devs if d.platform == "gpu"]
|
| 41 |
-
cpu = [d for d in devs if d.platform == "cpu"]
|
| 42 |
-
|
| 43 |
-
if prefer in ("auto", "gpu"):
|
| 44 |
-
return gpu[0] if gpu else (cpu[0] if cpu else devs[0])
|
| 45 |
-
if prefer == "cpu":
|
| 46 |
-
return cpu[0] if cpu else devs[0]
|
| 47 |
-
# unknown -> auto
|
| 48 |
-
return gpu[0] if gpu else (cpu[0] if cpu else devs[0])
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
def _np_dtype_str(x) -> str:
|
| 52 |
-
try:
|
| 53 |
-
return str(np.dtype(x))
|
| 54 |
-
except Exception:
|
| 55 |
-
return "float32"
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
# ----------------------------
|
| 59 |
-
# Frozen inference-only models
|
| 60 |
-
# ----------------------------
|
| 61 |
-
class FrozenDMAP:
|
| 62 |
-
"""
|
| 63 |
-
Inference-only Nyström DMAP embedder built from saved DMAP state.
|
| 64 |
-
Uses kNN in ambient space against reference points.
|
| 65 |
-
"""
|
| 66 |
-
|
| 67 |
-
def __init__(
|
| 68 |
-
self,
|
| 69 |
-
state: Dict[str, Any],
|
| 70 |
-
*,
|
| 71 |
-
ann_backend: ANNBackend = "auto",
|
| 72 |
-
ann_params: Optional[Dict[str, Any]] = None,
|
| 73 |
-
n_jobs: int = -1,
|
| 74 |
-
):
|
| 75 |
-
self.k = int(state["k"])
|
| 76 |
-
self.beta = float(state["beta"])
|
| 77 |
-
self.alpha = float(state["alpha"])
|
| 78 |
-
self.eps = float(state["eps"])
|
| 79 |
-
self.dtype = np.dtype(state.get("dtype", "float32"))
|
| 80 |
-
|
| 81 |
-
self.R_iX = np.ascontiguousarray(np.asarray(state["R_iX"]).astype(self.dtype, copy=False))
|
| 82 |
-
self.qalpha_i = np.asarray(state["qalpha_i"], dtype=np.float64)
|
| 83 |
-
self.R_over_lam_ix = np.asarray(state["R_over_lam_ix"], dtype=np.float64) # (Nref,d)
|
| 84 |
-
|
| 85 |
-
self.ann, self.ann_backend = make_ann(ann_backend, ann_params=ann_params, n_jobs=n_jobs)
|
| 86 |
-
self.ann.build(self.R_iX)
|
| 87 |
-
|
| 88 |
-
def __call__(self, R_aX: Union[np.ndarray, list], *, batch_size: Optional[int] = None) -> np.ndarray:
|
| 89 |
-
R_aX = np.asarray(R_aX)
|
| 90 |
-
single = (R_aX.ndim == 1)
|
| 91 |
-
if single:
|
| 92 |
-
R_aX = R_aX[None, :]
|
| 93 |
-
|
| 94 |
-
R_aX = np.ascontiguousarray(R_aX.astype(self.dtype, copy=False))
|
| 95 |
-
|
| 96 |
-
if batch_size is None:
|
| 97 |
-
Z = self._embed(R_aX)
|
| 98 |
-
else:
|
| 99 |
-
out = []
|
| 100 |
-
bs = int(batch_size)
|
| 101 |
-
for s in range(0, R_aX.shape[0], bs):
|
| 102 |
-
out.append(self._embed(R_aX[s : s + bs]))
|
| 103 |
-
Z = np.vstack(out)
|
| 104 |
-
|
| 105 |
-
return Z[0] if single else Z
|
| 106 |
-
|
| 107 |
-
def _embed(self, R_aX: np.ndarray) -> np.ndarray:
|
| 108 |
-
j_aK, D2_aK = self.ann.search(R_aX, self.k) # (a,k)
|
| 109 |
-
K_ai = np.exp(-self.beta * (D2_aK.astype(np.float64) / self.eps)) # (a,k)
|
| 110 |
-
|
| 111 |
-
q_a = np.maximum(K_ai.sum(axis=1), 1e-30)
|
| 112 |
-
qalpha_a = np.maximum(np.power(q_a, self.alpha), 1e-30)
|
| 113 |
-
|
| 114 |
-
qalpha_i = np.maximum(self.qalpha_i[j_aK], 1e-30)
|
| 115 |
-
Kalpha_ai = K_ai / (qalpha_a[:, None] * qalpha_i)
|
| 116 |
-
|
| 117 |
-
d_a = np.maximum(Kalpha_ai.sum(axis=1), 1e-30)
|
| 118 |
-
P_ai = Kalpha_ai / d_a[:, None]
|
| 119 |
-
|
| 120 |
-
R_over = self.R_over_lam_ix[j_aK, :] # (a,k,d)
|
| 121 |
-
Z_ax = (P_ai[:, :, None] * R_over).sum(axis=1) # (a,d)
|
| 122 |
-
return Z_ax
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
class FrozenGPLM:
|
| 126 |
-
"""
|
| 127 |
-
Inference-only GPLM decoder built from saved GPLM state.
|
| 128 |
-
Supports optional pred_k via ANN over inducing points.
|
| 129 |
-
"""
|
| 130 |
-
|
| 131 |
-
def __init__(
|
| 132 |
-
self,
|
| 133 |
-
state: Dict[str, Any],
|
| 134 |
-
*,
|
| 135 |
-
ann_backend: ANNBackend = "auto",
|
| 136 |
-
ann_params: Optional[Dict[str, Any]] = None,
|
| 137 |
-
n_jobs: int = -1,
|
| 138 |
-
):
|
| 139 |
-
self.beta = float(state["beta"])
|
| 140 |
-
self.eps = float(state["eps"])
|
| 141 |
-
self.pred_k = None if state.get("pred_k", None) is None else int(state["pred_k"])
|
| 142 |
-
self.dtype = np.dtype(state.get("dtype", "float32"))
|
| 143 |
-
|
| 144 |
-
self.mean_X = np.asarray(state["mean_X"], dtype=np.float64)
|
| 145 |
-
self.M_mX = np.asarray(state["M_mX"], dtype=np.float64)
|
| 146 |
-
|
| 147 |
-
# latent whitening used at training time
|
| 148 |
-
self.lat_mean_x = np.asarray(state["lat_mean_x"], dtype=np.float64)
|
| 149 |
-
self.lat_std_x = np.asarray(state["lat_std_x"], dtype=np.float64)
|
| 150 |
-
|
| 151 |
-
# inducing points in whitened latent space
|
| 152 |
-
self.Z_mx_w = np.ascontiguousarray(np.asarray(state["Z_mx_w"]).astype(np.float64, copy=False))
|
| 153 |
-
self.m = int(self.Z_mx_w.shape[0])
|
| 154 |
-
|
| 155 |
-
self.ann_Z, self.ann_backend = make_ann(ann_backend, ann_params=ann_params, n_jobs=n_jobs)
|
| 156 |
-
self.ann_Z.build(self.Z_mx_w.astype(self.dtype, copy=False))
|
| 157 |
-
|
| 158 |
-
def __call__(self, R_ax: Union[np.ndarray, list], *, batch_size: Optional[int] = None) -> np.ndarray:
|
| 159 |
-
R_ax = np.asarray(R_ax)
|
| 160 |
-
single = (R_ax.ndim == 1)
|
| 161 |
-
if single:
|
| 162 |
-
R_ax = R_ax[None, :]
|
| 163 |
-
R_ax = np.ascontiguousarray(R_ax.astype(self.dtype, copy=False))
|
| 164 |
-
|
| 165 |
-
if batch_size is None:
|
| 166 |
-
Y = self._decode(R_ax)
|
| 167 |
-
else:
|
| 168 |
-
out = []
|
| 169 |
-
bs = int(batch_size)
|
| 170 |
-
for s in range(0, R_ax.shape[0], bs):
|
| 171 |
-
out.append(self._decode(R_ax[s : s + bs]))
|
| 172 |
-
Y = np.vstack(out)
|
| 173 |
-
|
| 174 |
-
return Y[0] if single else Y
|
| 175 |
-
|
| 176 |
-
@staticmethod
|
| 177 |
-
def _sqdist_ab(A: np.ndarray, B: np.ndarray) -> np.ndarray:
|
| 178 |
-
A = np.asarray(A, dtype=np.float64)
|
| 179 |
-
B = np.asarray(B, dtype=np.float64)
|
| 180 |
-
A2 = np.sum(A * A, axis=1, keepdims=True)
|
| 181 |
-
B2 = np.sum(B * B, axis=1, keepdims=True).T
|
| 182 |
-
G = A @ B.T
|
| 183 |
-
return np.maximum(A2 + B2 - 2.0 * G, 0.0)
|
| 184 |
-
|
| 185 |
-
def _decode(self, R_ax: np.ndarray) -> np.ndarray:
|
| 186 |
-
Za = R_ax.astype(np.float64)
|
| 187 |
-
Za_w = (Za - self.lat_mean_x[None, :]) / self.lat_std_x[None, :]
|
| 188 |
-
|
| 189 |
-
if self.pred_k is None or self.pred_k >= self.m:
|
| 190 |
-
D2_am = self._sqdist_ab(Za_w, self.Z_mx_w)
|
| 191 |
-
C_am = np.exp(-self.beta * (D2_am / self.eps))
|
| 192 |
-
Y = C_am @ self.M_mX
|
| 193 |
-
else:
|
| 194 |
-
j_aK, D2_aK = self.ann_Z.search(Za_w.astype(self.dtype, copy=False), self.pred_k)
|
| 195 |
-
W = np.exp(-self.beta * (D2_aK.astype(np.float64) / self.eps)) # (a,k)
|
| 196 |
-
M = self.M_mX[j_aK] # (a,k,D)
|
| 197 |
-
Y = np.sum(W[:, :, None] * M, axis=1) # (a,D)
|
| 198 |
-
|
| 199 |
-
return Y + self.mean_X[None, :]
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
# ----------------------------
|
| 203 |
-
# Config
|
| 204 |
-
# ----------------------------
|
| 205 |
-
@dataclass
|
| 206 |
-
class DIMAConfig:
|
| 207 |
-
d: int = 32
|
| 208 |
-
ddpm_device: str = "auto" # "auto" | "cpu" | "gpu"
|
| 209 |
-
# purely informational defaults (actual settings live in kwargs dicts)
|
| 210 |
-
version: str = "0.1.0"
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
# ----------------------------
|
| 214 |
-
# Main wrapper
|
| 215 |
-
# ----------------------------
|
| 216 |
-
class DIMA:
|
| 217 |
-
"""
|
| 218 |
-
DIMA: Diffusion–Intrinsic Manifold Autoencoder
|
| 219 |
-
|
| 220 |
-
Pipeline:
|
| 221 |
-
- DMAP encoder (NumPy/SciPy, CPU): ambient -> raw latent
|
| 222 |
-
- GPLM decoder (NumPy/SciPy, CPU): raw latent -> ambient
|
| 223 |
-
- DDPM latent model (JAX/Flax, CPU or GPU): trained on *normalized* latents
|
| 224 |
-
|
| 225 |
-
API:
|
| 226 |
-
model = DIMA(R_iX, dmap_kwargs=..., gplm_kwargs=..., ddpm_kwargs=..., ddpm_device="auto")
|
| 227 |
-
Z = model.encode(X) # normalized latents (jnp on ddpm device)
|
| 228 |
-
Xhat = model.decode(Z) # ambient (np on CPU), optionally with DDPM refinement
|
| 229 |
-
Xhat2 = model(X) # polymorphic: ambient -> latent
|
| 230 |
-
Xhat3 = model(Z) # polymorphic: latent -> ambient (refine=True default)
|
| 231 |
-
Xsamp = model.sample(1000) # unconditional samples (ambient np)
|
| 232 |
-
"""
|
| 233 |
-
|
| 234 |
-
def __init__(
|
| 235 |
-
self,
|
| 236 |
-
R_iX: np.ndarray,
|
| 237 |
-
*,
|
| 238 |
-
d: int = 32,
|
| 239 |
-
dmap_kwargs: Optional[Dict[str, Any]] = None,
|
| 240 |
-
gplm_kwargs: Optional[Dict[str, Any]] = None,
|
| 241 |
-
ddpm_kwargs: Optional[Dict[str, Any]] = None,
|
| 242 |
-
ddpm_device: str = "auto",
|
| 243 |
-
key: Optional[jax.Array] = None ):
|
| 244 |
-
self.config = DIMAConfig(d=int(d), ddpm_device=str(ddpm_device))
|
| 245 |
-
self.training_time = 0.0
|
| 246 |
-
|
| 247 |
-
t0 = time.time()
|
| 248 |
-
|
| 249 |
-
# data
|
| 250 |
-
self.R_iX = np.asarray(R_iX)
|
| 251 |
-
if self.R_iX.ndim != 2:
|
| 252 |
-
raise ValueError("R_iX must be 2D (N,D).")
|
| 253 |
-
self.N, self.D = self.R_iX.shape
|
| 254 |
-
self.d = int(d)
|
| 255 |
-
|
| 256 |
-
dmap_kwargs = {} if dmap_kwargs is None else dict(dmap_kwargs)
|
| 257 |
-
gplm_kwargs = {} if gplm_kwargs is None else dict(gplm_kwargs)
|
| 258 |
-
ddpm_kwargs = {} if ddpm_kwargs is None else dict(ddpm_kwargs)
|
| 259 |
-
|
| 260 |
-
# choose device for DDPM
|
| 261 |
-
self.ddpm_device = _select_device(self.config.ddpm_device)
|
| 262 |
-
self.cpu_device = _select_device("cpu")
|
| 263 |
-
|
| 264 |
-
if key is None:
|
| 265 |
-
key = random.PRNGKey(0)
|
| 266 |
-
self.rng = key
|
| 267 |
-
|
| 268 |
-
# -------------------------
|
| 269 |
-
# 1) Train DMAP (CPU)
|
| 270 |
-
# -------------------------
|
| 271 |
-
self.enc = DMAP(self.R_iX, d=self.d, **dmap_kwargs)
|
| 272 |
-
R_ix = np.asarray(self.enc(self.R_iX)) # raw latents (N,d) float64
|
| 273 |
-
|
| 274 |
-
# -------------------------
|
| 275 |
-
# 2) Fit latent normalization (ONCE)
|
| 276 |
-
# -------------------------
|
| 277 |
-
self.lat_mean_np = R_ix.mean(axis=0, keepdims=True).astype(np.float64)
|
| 278 |
-
self.lat_std_np = (R_ix.std(axis=0, keepdims=True) + 1e-6).astype(np.float64)
|
| 279 |
-
|
| 280 |
-
Z_ix = (R_ix - self.lat_mean_np) / self.lat_std_np # normalized latents (np)
|
| 281 |
-
|
| 282 |
-
# keep norms on ddpm device too
|
| 283 |
-
self.lat_mean_j = jax.device_put(jnp.asarray(self.lat_mean_np, dtype=jnp.float32), self.ddpm_device)
|
| 284 |
-
self.lat_std_j = jax.device_put(jnp.asarray(self.lat_std_np, dtype=jnp.float32), self.ddpm_device)
|
| 285 |
-
|
| 286 |
-
# -------------------------
|
| 287 |
-
# 3) Train GPLM (CPU)
|
| 288 |
-
# -------------------------
|
| 289 |
-
self.dec = GPLM(R_ix, self.R_iX, **gplm_kwargs)
|
| 290 |
-
|
| 291 |
-
# -------------------------
|
| 292 |
-
# 4) Train DDPM on normalized latents (DDPM device)
|
| 293 |
-
# -------------------------
|
| 294 |
-
Z_ix_j = jax.device_put(jnp.asarray(Z_ix, dtype=jnp.float32), self.ddpm_device)
|
| 295 |
-
|
| 296 |
-
# defaults (can be overridden by ddpm_kwargs)
|
| 297 |
-
ddpm_init = dict(
|
| 298 |
-
T=200,
|
| 299 |
-
hidden_dim=128,
|
| 300 |
-
t_embed_dim=64,
|
| 301 |
-
learning_rate=3e-4,
|
| 302 |
-
n_iter=200_000,
|
| 303 |
-
ema_decay=0.999,
|
| 304 |
-
beta_max=0.02,
|
| 305 |
-
batch_size=256,
|
| 306 |
-
key=self.rng,
|
| 307 |
-
verbose_every=0,
|
| 308 |
-
eps=1e-5,
|
| 309 |
-
)
|
| 310 |
-
ddpm_init.update(ddpm_kwargs)
|
| 311 |
-
|
| 312 |
-
with jax.default_device(self.ddpm_device):
|
| 313 |
-
self.dm = DDPM(Z_ix_j, **ddpm_init)
|
| 314 |
-
|
| 315 |
-
self.training_time = time.time() - t0
|
| 316 |
-
|
| 317 |
-
# -------------------------
|
| 318 |
-
# Encode / Decode
|
| 319 |
-
# -------------------------
|
| 320 |
-
def encode(self, R_aX: Union[np.ndarray, jnp.ndarray], *, to_device: bool = True) -> jnp.ndarray:
|
| 321 |
-
"""
|
| 322 |
-
ambient -> normalized latent (DDPM space)
|
| 323 |
-
Returns jnp on ddpm_device by default.
|
| 324 |
-
"""
|
| 325 |
-
X = np.asarray(R_aX)
|
| 326 |
-
Z_raw = np.asarray(self.enc(X)) # CPU
|
| 327 |
-
Z_norm = (Z_raw - self.lat_mean_np) / self.lat_std_np
|
| 328 |
-
Zj = jnp.asarray(Z_norm, dtype=jnp.float32)
|
| 329 |
-
return jax.device_put(Zj, self.ddpm_device) if to_device else Zj
|
| 330 |
-
|
| 331 |
-
def decode(
|
| 332 |
-
self,
|
| 333 |
-
Z_ax: Union[np.ndarray, jnp.ndarray],
|
| 334 |
-
*,
|
| 335 |
-
refine: bool = True,
|
| 336 |
-
t_start: int = 10,
|
| 337 |
-
add_noise: bool = True,
|
| 338 |
-
key: Optional[jax.Array] = None,
|
| 339 |
-
batch_size: Optional[int] = None,
|
| 340 |
-
) -> np.ndarray:
|
| 341 |
-
"""
|
| 342 |
-
normalized latent -> (optional DDPM refine) -> raw latent -> ambient
|
| 343 |
-
Returns numpy array (CPU).
|
| 344 |
-
"""
|
| 345 |
-
Z = jnp.asarray(Z_ax, dtype=jnp.float32)
|
| 346 |
-
if Z.ndim == 1:
|
| 347 |
-
Z = Z[None, :]
|
| 348 |
-
Z = jax.device_put(Z, self.ddpm_device)
|
| 349 |
-
|
| 350 |
-
if refine:
|
| 351 |
-
Z = self.dm.refine_latents(Z, t_start=int(t_start), key=key, add_noise=bool(add_noise))
|
| 352 |
-
|
| 353 |
-
# unnormalize to raw latent (still on device)
|
| 354 |
-
R_ax = (Z * self.lat_std_j) + self.lat_mean_j
|
| 355 |
-
R_ax_np = np.asarray(jax.device_get(R_ax)) # to CPU
|
| 356 |
-
|
| 357 |
-
X_hat = self.dec(R_ax_np, batch_size=batch_size) # CPU decode
|
| 358 |
-
return np.asarray(X_hat)
|
| 359 |
-
|
| 360 |
-
def reconstruct(
|
| 361 |
-
self,
|
| 362 |
-
R_aX: Union[np.ndarray, jnp.ndarray],
|
| 363 |
-
*,
|
| 364 |
-
refine: bool = False,
|
| 365 |
-
t_start: int = 10,
|
| 366 |
-
add_noise: bool = True,
|
| 367 |
-
key: Optional[jax.Array] = None,
|
| 368 |
-
batch_size: Optional[int] = None,
|
| 369 |
-
) -> np.ndarray:
|
| 370 |
-
"""Convenience: decode(encode(X))."""
|
| 371 |
-
Z = self.encode(R_aX, to_device=True)
|
| 372 |
-
return self.decode(Z, refine=refine, t_start=t_start, add_noise=add_noise, key=key, batch_size=batch_size)
|
| 373 |
-
|
| 374 |
-
def sample(self, n: int, *, decode: bool = True, batch_size: Optional[int] = None) -> Union[jnp.ndarray, np.ndarray]:
|
| 375 |
-
"""
|
| 376 |
-
Unconditional samples from latent DDPM.
|
| 377 |
-
If decode=True, returns ambient samples as np.ndarray.
|
| 378 |
-
Otherwise returns normalized latents as jnp.ndarray.
|
| 379 |
-
"""
|
| 380 |
-
with jax.default_device(self.ddpm_device):
|
| 381 |
-
Z = self.dm.sample(int(n))
|
| 382 |
-
if not decode:
|
| 383 |
-
return Z
|
| 384 |
-
return self.decode(Z, refine=False, batch_size=batch_size)
|
| 385 |
-
|
| 386 |
-
def __call__(self, R_any: Union[np.ndarray, jnp.ndarray]) -> Union[jnp.ndarray, np.ndarray]:
|
| 387 |
-
"""
|
| 388 |
-
Polymorphic call:
|
| 389 |
-
- if trailing dim == D: ambient -> normalized latent
|
| 390 |
-
- if trailing dim == d: normalized latent -> ambient (refine=True)
|
| 391 |
-
"""
|
| 392 |
-
X = np.asarray(R_any)
|
| 393 |
-
last = X.shape[-1]
|
| 394 |
-
if last == self.D:
|
| 395 |
-
return self.encode(X)
|
| 396 |
-
if last == self.d:
|
| 397 |
-
return self.decode(X, refine=True)
|
| 398 |
-
raise ValueError(f"Trailing dim must be D={self.D} (ambient) or d={self.d} (latent), got {last}.")
|
| 399 |
-
|
| 400 |
-
# -------------------------
|
| 401 |
-
# Save / Load
|
| 402 |
-
# -------------------------
|
| 403 |
-
def _pack_encoder(self) -> Dict[str, Any]:
|
| 404 |
-
# minimal state for Nyström OOS
|
| 405 |
-
return dict(
|
| 406 |
-
R_iX=np.asarray(self.enc.R_iX),
|
| 407 |
-
qalpha_i=np.asarray(getattr(self.enc, "qalpha_i", getattr(self.enc, "qα_i"))),
|
| 408 |
-
R_over_lam_ix=np.asarray(getattr(self.enc, "R_over_λ_ix", getattr(self.enc, "R_over_lam_ix"))),
|
| 409 |
-
k=int(self.enc.k),
|
| 410 |
-
beta=float(getattr(self.enc, "beta", getattr(self.enc, "β"))),
|
| 411 |
-
alpha=float(getattr(self.enc, "alpha", getattr(self.enc, "α"))),
|
| 412 |
-
eps=float(getattr(self.enc, "eps", getattr(self.enc, "ε"))),
|
| 413 |
-
dtype=_np_dtype_str(getattr(self.enc, "dtype", np.float32)),
|
| 414 |
-
)
|
| 415 |
-
|
| 416 |
-
def _pack_decoder(self) -> Dict[str, Any]:
|
| 417 |
-
return dict(
|
| 418 |
-
Z_mx_w=np.asarray(getattr(self.dec, "Z_mx_w", getattr(self.dec, "Z_mx_w", None))),
|
| 419 |
-
M_mX=np.asarray(self.dec.M_mX),
|
| 420 |
-
mean_X=np.asarray(getattr(self.dec, "mean_X", np.zeros((self.D,), dtype=np.float64))),
|
| 421 |
-
lat_mean_x=np.asarray(getattr(self.dec, "lat_mean_x", np.zeros((self.d,), dtype=np.float64))),
|
| 422 |
-
lat_std_x=np.asarray(getattr(self.dec, "lat_std_x", np.ones((self.d,), dtype=np.float64))),
|
| 423 |
-
beta=float(getattr(self.dec, "beta", getattr(self.dec, "β"))),
|
| 424 |
-
eps=float(getattr(self.dec, "eps", getattr(self.dec, "ε"))),
|
| 425 |
-
pred_k=getattr(self.dec, "pred_k", getattr(self.dec, "pred_κ", None)),
|
| 426 |
-
dtype=_np_dtype_str(getattr(self.dec, "dtype", np.float32)),
|
| 427 |
-
)
|
| 428 |
-
|
| 429 |
-
def state_dict(self) -> Dict[str, Any]:
|
| 430 |
-
"""
|
| 431 |
-
Serializable state (msgpack via flax.serialization).
|
| 432 |
-
"""
|
| 433 |
-
dd = dict(
|
| 434 |
-
T=int(self.dm.T),
|
| 435 |
-
D=int(self.dm.D),
|
| 436 |
-
hidden_dim=int(getattr(self.dm.model, "hidden", 128)),
|
| 437 |
-
t_embed_dim=int(getattr(self.dm.model, "t_dim", 64)),
|
| 438 |
-
ema_decay=float(getattr(self.dm, "ema_decay", 0.999)),
|
| 439 |
-
beta_max=float(getattr(self.dm, "beta_max", 0.02)),
|
| 440 |
-
eps=float(getattr(self.dm, "eps", 1e-5)),
|
| 441 |
-
params=self.dm.state.params,
|
| 442 |
-
ema_params=self.dm.state.ema_params,
|
| 443 |
-
)
|
| 444 |
-
|
| 445 |
-
state = dict(
|
| 446 |
-
meta=dict(
|
| 447 |
-
N=int(self.N),
|
| 448 |
-
D=int(self.D),
|
| 449 |
-
d=int(self.d),
|
| 450 |
-
training_time=float(self.training_time),
|
| 451 |
-
),
|
| 452 |
-
config=asdict(self.config),
|
| 453 |
-
latent_norm=dict(
|
| 454 |
-
mean=np.asarray(self.lat_mean_np, dtype=np.float64),
|
| 455 |
-
std=np.asarray(self.lat_std_np, dtype=np.float64),
|
| 456 |
-
),
|
| 457 |
-
encoder=self._pack_encoder(),
|
| 458 |
-
decoder=self._pack_decoder(),
|
| 459 |
-
ddpm=dd,
|
| 460 |
-
)
|
| 461 |
-
return state
|
| 462 |
-
|
| 463 |
-
def save_local(self, weights_file: str = "dima.msgpack", config_file: str = "config.json") -> None:
|
| 464 |
-
state = self.state_dict()
|
| 465 |
-
blob = flax_ser.msgpack_serialize(state)
|
| 466 |
-
with open(weights_file, "wb") as f:
|
| 467 |
-
f.write(blob)
|
| 468 |
-
with open(config_file, "w") as f:
|
| 469 |
-
json.dump(state["config"], f, indent=2)
|
| 470 |
-
return None
|
| 471 |
-
|
| 472 |
-
@classmethod
|
| 473 |
-
def load_local(
|
| 474 |
-
cls,
|
| 475 |
-
weights_file: str = "dima.msgpack",
|
| 476 |
-
*,
|
| 477 |
-
ddpm_device: str = "auto",
|
| 478 |
-
ann_backend: ANNBackend = "auto",
|
| 479 |
-
ann_params: Optional[Dict[str, Any]] = None,
|
| 480 |
-
n_jobs: int = -1,
|
| 481 |
-
key: Optional[jax.Array] = None,
|
| 482 |
-
) -> "DIMA":
|
| 483 |
-
with open(weights_file, "rb") as f:
|
| 484 |
-
state = flax_ser.msgpack_restore(f.read())
|
| 485 |
-
|
| 486 |
-
obj = cls.__new__(cls) # bypass __init__
|
| 487 |
-
|
| 488 |
-
obj.config = DIMAConfig(**state["config"])
|
| 489 |
-
obj.ddpm_device = _select_device(ddpm_device)
|
| 490 |
-
obj.cpu_device = _select_device("cpu")
|
| 491 |
-
obj.training_time = float(state["meta"].get("training_time", 0.0))
|
| 492 |
-
|
| 493 |
-
obj.N = int(state["meta"]["N"])
|
| 494 |
-
obj.D = int(state["meta"]["D"])
|
| 495 |
-
obj.d = int(state["meta"]["d"])
|
| 496 |
-
|
| 497 |
-
# RNG
|
| 498 |
-
if key is None:
|
| 499 |
-
key = random.PRNGKey(0)
|
| 500 |
-
obj.rng = key
|
| 501 |
-
|
| 502 |
-
# latent norm
|
| 503 |
-
obj.lat_mean_np = np.asarray(state["latent_norm"]["mean"], dtype=np.float64)
|
| 504 |
-
obj.lat_std_np = np.asarray(state["latent_norm"]["std"], dtype=np.float64)
|
| 505 |
-
|
| 506 |
-
obj.lat_mean_j = jax.device_put(jnp.asarray(obj.lat_mean_np, dtype=jnp.float32), obj.ddpm_device)
|
| 507 |
-
obj.lat_std_j = jax.device_put(jnp.asarray(obj.lat_std_np, dtype=jnp.float32), obj.ddpm_device)
|
| 508 |
-
|
| 509 |
-
# encoder/decoder frozen
|
| 510 |
-
obj.enc = FrozenDMAP(state["encoder"], ann_backend=ann_backend, ann_params=ann_params, n_jobs=n_jobs)
|
| 511 |
-
obj.dec = FrozenGPLM(state["decoder"], ann_backend=ann_backend, ann_params=ann_params, n_jobs=n_jobs)
|
| 512 |
-
|
| 513 |
-
# rebuild DDPM skeleton with dummy data, then load params
|
| 514 |
-
dd = state["ddpm"]
|
| 515 |
-
T = int(dd["T"])
|
| 516 |
-
D = int(dd["D"])
|
| 517 |
-
hidden_dim = int(dd["hidden_dim"])
|
| 518 |
-
t_embed_dim = int(dd["t_embed_dim"])
|
| 519 |
-
ema_decay = float(dd.get("ema_decay", 0.999))
|
| 520 |
-
beta_max = float(dd.get("beta_max", 0.02))
|
| 521 |
-
eps = float(dd.get("eps", 1e-5))
|
| 522 |
-
|
| 523 |
-
dummy = jnp.zeros((1, D), dtype=jnp.float32)
|
| 524 |
-
with jax.default_device(obj.ddpm_device):
|
| 525 |
-
obj.dm = DDPM(
|
| 526 |
-
dummy,
|
| 527 |
-
T=T,
|
| 528 |
-
hidden_dim=hidden_dim,
|
| 529 |
-
t_embed_dim=t_embed_dim,
|
| 530 |
-
learning_rate=1e-3,
|
| 531 |
-
n_iter=0, # skip training on load
|
| 532 |
-
ema_decay=ema_decay,
|
| 533 |
-
beta_max=beta_max,
|
| 534 |
-
batch_size=1,
|
| 535 |
-
key=obj.rng,
|
| 536 |
-
verbose_every=0,
|
| 537 |
-
eps=eps,
|
| 538 |
-
)
|
| 539 |
-
obj.dm.state = obj.dm.state.replace(params=dd["params"], ema_params=dd["ema_params"])
|
| 540 |
-
|
| 541 |
-
obj.R_iX = None # training data not stored by default
|
| 542 |
-
return obj
|
| 543 |
-
|
| 544 |
-
# -------------------------
|
| 545 |
-
# Hugging Face helpers
|
| 546 |
-
# -------------------------
|
| 547 |
-
def upload_to_huggingface(
|
| 548 |
-
self,
|
| 549 |
-
repo_id: str,
|
| 550 |
-
*,
|
| 551 |
-
hf_token: Optional[str] = None,
|
| 552 |
-
weights_file: str = "dima.msgpack",
|
| 553 |
-
config_file: str = "config.json",
|
| 554 |
-
) -> None:
|
| 555 |
-
if not _HAS_HF:
|
| 556 |
-
raise ImportError("huggingface_hub not installed. pip install dima[hf] (or huggingface_hub).")
|
| 557 |
-
if hf_token is not None:
|
| 558 |
-
HfFolder.save_token(hf_token)
|
| 559 |
-
api = HfApi()
|
| 560 |
-
_ = api.whoami(token=hf_token) # sanity check token (no-op if public)
|
| 561 |
-
|
| 562 |
-
self.save_local(weights_file=weights_file, config_file=config_file)
|
| 563 |
-
for fname in (weights_file, config_file):
|
| 564 |
-
upload_file(path_or_fileobj=fname, path_in_repo=fname, repo_id=repo_id, token=hf_token)
|
| 565 |
-
return None
|
| 566 |
-
|
| 567 |
-
@classmethod
|
| 568 |
-
def load_from_huggingface(
|
| 569 |
-
cls,
|
| 570 |
-
repo_id: str,
|
| 571 |
-
*,
|
| 572 |
-
hf_token: Optional[str] = None,
|
| 573 |
-
weights_file: str = "dima.msgpack",
|
| 574 |
-
config_file: str = "config.json",
|
| 575 |
-
ddpm_device: str = "auto",
|
| 576 |
-
ann_backend: ANNBackend = "auto",
|
| 577 |
-
ann_params: Optional[Dict[str, Any]] = None,
|
| 578 |
-
n_jobs: int = -1,
|
| 579 |
-
key: Optional[jax.Array] = None,
|
| 580 |
-
) -> "DIMA":
|
| 581 |
-
if not _HAS_HF:
|
| 582 |
-
raise ImportError("huggingface_hub not installed. pip install dima[hf] (or huggingface_hub).")
|
| 583 |
-
|
| 584 |
-
wpath = hf_hub_download(repo_id, weights_file, token=hf_token)
|
| 585 |
-
_ = hf_hub_download(repo_id, config_file, token=hf_token) # for humans / future-proofing
|
| 586 |
-
|
| 587 |
-
return cls.load_local(
|
| 588 |
-
wpath,
|
| 589 |
-
ddpm_device=ddpm_device,
|
| 590 |
-
ann_backend=ann_backend,
|
| 591 |
-
ann_params=ann_params,
|
| 592 |
-
n_jobs=n_jobs,
|
| 593 |
-
key=key,
|
| 594 |
-
)
|
| 595 |
-
|
| 596 |
-
|
| 597 |
-
__all__ = ["DIMA", "DIMAConfig"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|