sawnergy 1.0.7__py3-none-any.whl → 1.0.8__py3-none-any.whl

sawnergy/embedding/embedder.py

@@ -1,15 +1,5 @@
 from __future__ import annotations
 
-"""
-Embedding orchestration for Skip-Gram (SG) and Skip-Gram with Negative Sampling (SGNS).
-
-This module consumes attractive/repulsive walk corpora produced by the walker
-pipeline and trains per-frame embeddings using either the PyTorch or PureML
-implementations of SG/SGNS. The resulting embeddings can be persisted back into
-an ``ArrayStorage`` archive along with rich metadata describing the training
-configuration.
-"""
-
 # third-pary
 import numpy as np
 
@@ -36,10 +26,8 @@ class Embedder:
 
     def __init__(self,
                  WALKS_path: str | Path,
-                 base: Literal["torch", "pureml"],
                  *,
                  seed: int | None = None,
-                 objective: Literal["sgns", "sg"] = "sgns"
                 ) -> None:
         """Initialize the embedder and load walk tensors.
 
@@ -51,24 +39,19 @@ class Embedder:
                 ``None`` if that collection is absent), and the metadata
                 ``num_RWs``, ``num_SAWs``, ``node_count``, ``time_stamp_count``,
                 ``walk_length``.
-            base: Which SGNS backend to use, either ``"torch"`` or ``"pureml"``.
             seed: Optional seed for the embedder's RNG. If ``None``, a random
                 32-bit seed is chosen.
-            objective: Training objective, either ``"sgns"`` (negative sampling)
-                or ``"sg"`` (plain full-softmax Skip-Gram).
 
         Raises:
             ValueError: If required metadata is missing or any loaded walk array
                 has an unexpected shape.
-            ImportError: If the requested backend is not installed.
-            NameError: If ``base`` is not one of ``{"torch","pureml"}``.
 
         Notes:
             - Walks in storage are 1-based (residue indexing). Internally, this
               class normalizes to 0-based indices for training utilities.
         """
         self._walks_path = Path(WALKS_path)
-        _logger.info("Initializing Embedder from %s (base=%s)", self._walks_path, base)
+        _logger.info("Initializing Embedder from %s", self._walks_path)
 
         # placeholders for optional walk collections
         self.attractive_RWs : np.ndarray | None = None
@@ -127,58 +110,76 @@ class Embedder:
         RWs_expected  = (time_stamp_count, node_count * num_RWs,  walk_length+1) if (num_RWs  > 0) else None
         SAWs_expected = (time_stamp_count, node_count * num_SAWs, walk_length+1) if (num_SAWs > 0) else None
 
-        self.vocab_size = int(node_count)
-        self.frame_count = int(time_stamp_count)
-        self.walk_length = int(walk_length)
+        self.vocab_size   = int(node_count)
+        self.frame_count  = int(time_stamp_count)
+        self.walk_length  = int(walk_length)
+        self.num_RWs      = int(num_RWs)
+        self.num_SAWs     = int(num_SAWs)
+        # Keep dataset names for metadata passthrough
+        self._attractive_RWs_name  = attractive_RWs_name
+        self._repulsive_RWs_name   = repulsive_RWs_name
+        self._attractive_SAWs_name = attractive_SAWs_name
+        self._repulsive_SAWs_name  = repulsive_SAWs_name
 
         # store walks if present
         if attractive_RWs is not None:
             if RWs_expected and attractive_RWs.shape != RWs_expected:
                 raise ValueError(f"ATTR RWs: expected {RWs_expected}, got {attractive_RWs.shape}")
             self.attractive_RWs = attractive_RWs
+            _logger.debug("ATTR RWs loaded: %s", self.attractive_RWs.shape)
 
         if repulsive_RWs is not None:
             if RWs_expected and repulsive_RWs.shape != RWs_expected:
                 raise ValueError(f"REP RWs: expected {RWs_expected}, got {repulsive_RWs.shape}")
             self.repulsive_RWs = repulsive_RWs
+            _logger.debug("REP  RWs loaded: %s", self.repulsive_RWs.shape)
 
         if attractive_SAWs is not None:
             if SAWs_expected and attractive_SAWs.shape != SAWs_expected:
                 raise ValueError(f"ATTR SAWs: expected {SAWs_expected}, got {attractive_SAWs.shape}")
             self.attractive_SAWs = attractive_SAWs
+            _logger.debug("ATTR SAWs loaded: %s", self.attractive_SAWs.shape)
 
         if repulsive_SAWs is not None:
             if SAWs_expected and repulsive_SAWs.shape != SAWs_expected:
                 raise ValueError(f"REP SAWs: expected {SAWs_expected}, got {repulsive_SAWs.shape}")
             self.repulsive_SAWs = repulsive_SAWs
+            _logger.debug("REP  SAWs loaded: %s", self.repulsive_SAWs.shape)
 
         # INTERNAL RNG
         self._seed = np.random.randint(0, 2**32 - 1) if seed is None else int(seed)
         self.rng = np.random.default_rng(self._seed)
         _logger.info("RNG initialized from seed=%d", self._seed)
 
-        # MODEL HANDLE
-        self.model_base: Literal["torch", "pureml"] = base
-        self.objective: Literal["sgns", "sg"] = objective
-        self.model_constructor = self._get_SGNS_constructor_from(base, objective)
-        _logger.info(
-            "SG backend resolved: %s (objective=%s)",
-            getattr(self.model_constructor, "__name__", repr(self.model_constructor)), self.objective
-        )
-
     # -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=- PRIVATE -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
 
     # HELPERS:
 
     @staticmethod
-    def _get_SGNS_constructor_from(base: Literal["torch", "pureml"],
-                                   objective: Literal["sgns", "sg"]):
-        """Resolve the SG/SGNS implementation class for the selected backend."""
+    def _get_NN_constructor_from(base: Literal["torch", "pureml"],
+                                 objective: Literal["sgns", "sg"]):
+        """Resolve the SG/SGNS implementation class for the selected backend.
+
+        Args:
+            base: Backend family to use, ``"torch"`` or ``"pureml"``.
+            objective: Training objective, ``"sgns"`` or ``"sg"``.
+
+        Returns:
+            A callable class (constructor) implementing the requested model.
+
+        Raises:
+            ImportError: If the requested backend cannot be imported.
+            NameError: If ``base`` is not one of the supported values.
+        """
+        _logger.debug("Resolving model constructor: base=%s objective=%s", base, objective)
         if base == "torch":
             try:
                 from .SGNS_torch import SGNS_Torch, SG_Torch
-                return SG_Torch if objective == "sg" else SGNS_Torch
+                ctor = SG_Torch if objective == "sg" else SGNS_Torch
+                _logger.debug("Resolved PyTorch class: %s", getattr(ctor, "__name__", str(ctor)))
+                return ctor
             except Exception:
+                _logger.exception("Failed to import PyTorch backend.")
                 raise ImportError(
                     "PyTorch is not installed, but base='torch' was requested. "
                     "Install PyTorch first, e.g.: `pip install torch` "
@@ -187,8 +188,11 @@ class Embedder:
         elif base == "pureml":
             try:
                 from .SGNS_pml import SGNS_PureML, SG_PureML
-                return SG_PureML if objective == "sg" else SGNS_PureML
+                ctor = SG_PureML if objective == "sg" else SGNS_PureML
+                _logger.debug("Resolved PureML class: %s", getattr(ctor, "__name__", str(ctor)))
+                return ctor
             except Exception:
+                _logger.exception("Failed to import PureML backend.")
                 raise ImportError(
                     "PureML is not installed, but base='pureml' was requested. "
                     "Install PureML first via `pip install ym-pure-ml` "
@@ -198,7 +202,18 @@ class Embedder:
 
     @staticmethod
     def _as_zerobase_intp(walks: np.ndarray, *, V: int) -> np.ndarray:
-        """Validate 1-based uint/int walks → 0-based intp; check bounds."""
+        """Validate and convert 1-based walks to 0-based ``intp``.
+
+        Args:
+            walks: 2D array of node ids with 1-based indexing.
+            V: Vocabulary size for bounds checking.
+
+        Returns:
+            2D array of dtype ``intp`` with 0-based indices.
+
+        Raises:
+            ValueError: If shape is not 2D or indices are out of bounds.
+        """
         arr = np.asarray(walks)
         if arr.ndim != 2:
             raise ValueError("walks must be 2D: (num_walks, walk_len)")
@@ -206,7 +221,9 @@ class Embedder:
             arr = arr.astype(np.int64, copy=False)
         # 1-based → 0-based
         arr = arr - 1
-        if arr.min() < 0 or arr.max() >= V:
+        mn, mx = int(arr.min()), int(arr.max())
+        _logger.debug("Zero-basing walks: min=%d max=%d V=%d", mn, mx, V)
+        if mn < 0 or mx >= V:
             raise ValueError("walk ids out of range after 1→0-based normalization")
         return arr.astype(np.intp, copy=False)
 
@@ -214,19 +231,29 @@ class Embedder:
     def _pairs_from_walks(walks0: np.ndarray, window_size: int) -> np.ndarray:
         """
         Skip-gram pairs including edge centers (one-sided when needed).
-        walks0: (W, L) int array (0-based ids).
-        Returns: (N_pairs, 2) int32 [center, context].
+
+        Args:
+            walks0: (W, L) int array (0-based ids).
+            window_size: Symmetric context window radius.
+
+        Returns:
+            Array of shape (N_pairs, 2) int32 with columns [center, context].
+
+        Raises:
+            ValueError: If shape is invalid or ``window_size`` <= 0.
         """
         if walks0.ndim != 2:
             raise ValueError("walks must be 2D: (num_walks, walk_len)")
 
         _, L = walks0.shape
         k = int(window_size)
+        _logger.debug("Building SG pairs: L=%d, window=%d", L, k)
 
         if k <= 0:
             raise ValueError("window_size must be positive")
         
         if L == 0:
+            _logger.debug("Empty walks length; returning 0 pairs.")
             return np.empty((0, 2), dtype=np.int32)
 
         out_chunks = []
@@ -244,18 +271,42 @@ class Embedder:
             out_chunks.append(np.stack((centers_l, ctx_l), axis=2).reshape(-1, 2))
 
         if not out_chunks:
+            _logger.debug("No offsets produced pairs; returning empty.")
             return np.empty((0, 2), dtype=np.int32)
 
-        return np.concatenate(out_chunks, axis=0).astype(np.int32, copy=False)
+        pairs = np.concatenate(out_chunks, axis=0).astype(np.int32, copy=False)
+        _logger.debug("Built %d training pairs", pairs.shape[0])
+        return pairs
 
     @staticmethod
     def _freq_from_walks(walks0: np.ndarray, *, V: int) -> np.ndarray:
-        """Node frequencies from walks (0-based)."""
-        return np.bincount(walks0.ravel(), minlength=V).astype(np.int64, copy=False)
+        """Node frequencies from walks (0-based).
+
+        Args:
+            walks0: 2D array of 0-based node ids.
+            V: Vocabulary size (minlength for bincount).
+
+        Returns:
+            1D array of int64 frequencies with length ``V``.
+        """
+        freq = np.bincount(walks0.ravel(), minlength=V).astype(np.int64, copy=False)
+        _logger.debug("Frequency mass: total=%d nonzero=%d", int(freq.sum()), int(np.count_nonzero(freq)))
+        return freq
 
     @staticmethod
     def _soft_unigram(freq: np.ndarray, *, power: float = 0.75) -> np.ndarray:
-        """Return normalized Pn(w) ∝ f(w)^power as float64 probs."""
+        """Return normalized Pn(w) ∝ f(w)^power as float64 probs.
+
+        Args:
+            freq: 1D array of token frequencies.
+            power: Exponent used for smoothing (default 0.75 à la word2vec).
+
+        Returns:
+            1D array of probabilities summing to 1.0.
+
+        Raises:
+            ValueError: If mass is invalid (all zeros or non-finite).
+        """
         p = np.asarray(freq, dtype=np.float64)
         if p.sum() == 0:
             raise ValueError("all frequencies are zero")
@@ -263,13 +314,31 @@ class Embedder:
         s = p.sum()
         if not np.isfinite(s) or s <= 0:
             raise ValueError("invalid unigram mass")
-        return p / s
+        probs = p / s
+        _logger.debug("Noise distribution ready (power=%.3f)", power)
+        return probs
 
     def _materialize_walks(self, frame_id: int, rin: Literal["attr", "repuls"],
                            using: Literal["RW", "SAW", "merged"]) -> np.ndarray:
+        """Materialize the requested set of walks for a frame.
+
+        Args:
+            frame_id: 1-based frame index.
+            rin: Which RIN to pull from: ``"attr"`` or ``"repuls"``.
+            using: Which walk sets to include: ``"RW"``, ``"SAW"``, or ``"merged"``.
+                If ``"merged"``, concatenate available RW and SAW along axis 0.
+
+        Returns:
+            A 2D array of walks with shape (num_walks, walk_length+1).
+
+        Raises:
+            IndexError: If ``frame_id`` is out of range.
+            ValueError: If no matching walks are available.
+        """
         if not 1 <= frame_id <= int(self.frame_count):
             raise IndexError(f"frame_id must be in [1, {self.frame_count}]; got {frame_id}")
 
+        _logger.debug("Materializing %s walks at frame=%d using=%s", rin, frame_id, using)
         frame_id -= 1
 
         if rin == "attr":
@@ -296,8 +365,12 @@ class Embedder:
         if not parts:
             raise ValueError(f"No walks available for {rin=} with {using=}")
         if len(parts) == 1:
-            return parts[0]
-        return np.concatenate(parts, axis=0)
+            out = parts[0]
+        else:
+            out = np.concatenate(parts, axis=0)
+
+        _logger.debug("Materialized walks shape: %s", getattr(out, "shape", None))
+        return out
 
     # INTERFACES: (private)
 
@@ -306,6 +379,17 @@ class Embedder:
                                     using: Literal["RW", "SAW", "merged"],
                                     window_size: int,
                                     alpha: float = 0.75) -> tuple[np.ndarray, np.ndarray]:
+        """Construct (center, context) pairs and noise distribution for ATTR.
+
+        Args:
+            frame_id: 1-based frame index.
+            using: Walk subset to include.
+            window_size: Skip-gram window radius.
+            alpha: Unigram smoothing exponent.
+
+        Returns:
+            Tuple of (pairs, noise_probs).
+        """
         walks = self._materialize_walks(frame_id, "attr", using)
         walks0 = self._as_zerobase_intp(walks, V=self.vocab_size)
         attractive_corpus = self._pairs_from_walks(walks0, window_size)
@@ -319,6 +403,17 @@ class Embedder:
                                    using: Literal["RW", "SAW", "merged"],
                                    window_size: int,
                                    alpha: float = 0.75) -> tuple[np.ndarray, np.ndarray]:
+        """Construct (center, context) pairs and noise distribution for REP.
+
+        Args:
+            frame_id: 1-based frame index.
+            using: Walk subset to include.
+            window_size: Skip-gram window radius.
+            alpha: Unigram smoothing exponent.
+
+        Returns:
+            Tuple of (pairs, noise_probs).
+        """
         walks = self._materialize_walks(frame_id, "repuls", using)
         walks0 = self._as_zerobase_intp(walks, V=self.vocab_size)
         repulsive_corpus = self._pairs_from_walks(walks0, window_size)
@@ -333,61 +428,53 @@ class Embedder:
             frame_id: int,
             RIN_type: Literal["attr", "repuls"],
             using: Literal["RW", "SAW", "merged"],
-            window_size: int,
-            num_negative_samples: int,
             num_epochs: int,
-            batch_size: int,
+            negative_sampling: bool = False,
+            window_size: int = 5,
+            num_negative_samples: int = 10,
+            batch_size: int = 1024,
             *,
             lr_step_per_batch: bool = False,
             shuffle_data: bool = True,
             dimensionality: int = 128,
             alpha: float = 0.75,
             device: str | None = None,
-            sgns_kwargs: dict[str, object] | None = None,
+            model_base: Literal["torch", "pureml"] = "pureml",
+            model_kwargs: dict[str, object] | None = None,
             kind: Literal["in", "out", "avg"] = "in",
-            _seed: int | None = None,
-            objective: Literal["sgns", "sg"] | None = None
+            _seed: int | None = None
             ) -> np.ndarray:
-        """Train embeddings for a single frame and return the input embedding matrix.
+        """Train embeddings for a single frame and return the matrix.
 
         Args:
-            frame_id: 1-based frame index to train on.
-            RIN_type: Interaction channel to use: ``"attr"`` (attractive) or
-                ``"repuls"`` (repulsive).
-            using: Which walk collections to include: ``"RW"``, ``"SAW"``, or
-                ``"merged"`` (concatenates both if available).
-            window_size: Symmetric skip-gram window size ``k``.
-            num_negative_samples: Number of negative samples per positive pair.
-                Ignored when ``objective="sg"``.
-            num_epochs: Number of passes over the pair dataset.
-            batch_size: Mini-batch size for training.
-            shuffle_data: Whether to shuffle pairs each epoch.
-            dimensionality: Embedding dimensionality ``D``.
-            alpha: Noise distribution exponent (``Pn ∝ f^alpha``).
-            device: Optional device string for the Torch backend (e.g., ``"cuda"``).
-            sgns_kwargs: Extra keyword arguments forwarded to the backend SGNS
-                constructor. For PureML, required keys are:
-                ``{"optim", "optim_kwargs"}``; ``lr_sched`` is optional, but if
-                provided then ``lr_sched_kwargs`` must also be provided.
-            kind: Which embedding matrix to return: ``"in"``, ``"out"``, or ``"avg"``.
-            _seed: Optional child seed for this frame's model initialization.
-            objective: Training objective override for this call (``"sgns"`` or
-                ``"sg"``). If ``None``, uses the value set at construction.
+            frame_id: 1-based frame index to embed.
+            RIN_type: ``"attr"`` or ``"repuls"`` - which corpus to use.
+            using: Which walks to use (``"RW"``, ``"SAW"``, or ``"merged"``).
+            num_epochs: Number of passes over the pairs.
+            negative_sampling: If ``True``, use SGNS objective; else plain SG.
+            window_size: Skip-gram symmetric window radius.
+            num_negative_samples: Negatives per positive pair (SGNS only).
+            batch_size: Minibatch size for training.
+            lr_step_per_batch: If ``True``, step LR every batch (else per epoch).
+            shuffle_data: Shuffle pairs each epoch.
+            dimensionality: Embedding dimension ``D``.
+            alpha: Unigram smoothing power for noise distribution.
+            device: Optional backend device hint (e.g., ``"cuda"``).
+            model_base: Backend family (``"torch"`` or ``"pureml"``).
+            model_kwargs: Passed through to backend model constructor.
+            kind: Which embedding to return: ``"in"``, ``"out"``, or ``"avg"``.
+            _seed: Optional override seed for this frame.
 
         Returns:
-            np.ndarray: Learned embedding matrix (selected by ``kind``) of shape ``(V, D)``.
-
-        Raises:
-            ValueError: If requested walks are missing, if no training pairs are
-                generated, or if required ``sgns_kwargs`` for PureML are absent.
-            AttributeError: If the SGNS model does not expose embeddings via
-                ``.embeddings`` or ``.parameters[0]``.
+            ``(V, D)`` float32 embedding matrix.
         """
         _logger.info(
-            "Preparing frame %d (rin=%s using=%s window=%d neg=%d epochs=%d batch=%d)",
-            frame_id, RIN_type, using, window_size, num_negative_samples, num_epochs, batch_size
+            "embed_frame: frame=%d RIN=%s using=%s base=%s D=%d epochs=%d batch=%d sgns=%s k=%d alpha=%.3f",
+            frame_id, RIN_type, using, model_base, dimensionality, num_epochs, batch_size,
+            str(negative_sampling), window_size, alpha
         )
 
+        # ------------------ resolve training data -----------------
         if RIN_type == "attr":
             if self.attractive_RWs is None and self.attractive_SAWs is None:
                 raise ValueError("Attractive random walks are missing")
@@ -397,162 +484,124 @@ class Embedder:
                 raise ValueError("Repulsive random walks are missing")
             pairs, noise_probs = self._repulsive_corpus_and_prob(frame_id=frame_id, using=using, window_size=window_size, alpha=alpha)
         else:
-            raise ValueError(f"Unknown RIN_type: {RIN_type!r}")
-
+            raise NameError(f"Unknown RIN_type: {RIN_type!r}")
         if pairs.size == 0:
             raise ValueError("No training pairs generated for the requested configuration")
+        # ----------------------------------------------------------
 
+        # ---------------- construct training corpus ---------------
         centers  = pairs[:, 0].astype(np.int64, copy=False)
         contexts = pairs[:, 1].astype(np.int64, copy=False)
+        _logger.debug("Pairs split: centers=%s contexts=%s", centers.shape, contexts.shape)
+        # ----------------------------------------------------------
 
-        model_kwargs: dict[str, object] = dict(sgns_kwargs or {})
-        if self.model_base == "pureml":
-            required = {"optim", "optim_kwargs"}
-            missing = required.difference(model_kwargs)
-            if missing:
-                raise ValueError(f"PureML backend requires {sorted(missing)} in sgns_kwargs.")
-            has_sched = ("lr_sched" in model_kwargs and model_kwargs["lr_sched"] is not None)
-            has_sched_kwargs = ("lr_sched_kwargs" in model_kwargs and model_kwargs["lr_sched_kwargs"] is not None)
-            if has_sched and not has_sched_kwargs:
-                raise ValueError("When providing lr_sched for PureML, you must also provide lr_sched_kwargs.")
-
-        child_seed = int(self._seed if _seed is None else _seed)
-        model_kwargs.update({
+        # ------------ resolve model_constructor kwargs ------------
+        if model_kwargs is not None:
+            if (("lr_sched" in model_kwargs and model_kwargs.get("lr_sched", None) is not None)
+                and ("lr_sched_kwargs" in model_kwargs and model_kwargs.get("lr_sched_kwargs", None) is None)):
+                raise ValueError("When `lr_sched`, you must also provide `lr_sched_kwargs`.")
+
+        constructor_kwargs: dict[str, object] = dict(model_kwargs or {})
+        constructor_kwargs.update({
             "V": self.vocab_size,
             "D": dimensionality,
-            "seed": child_seed
+            "seed": int(self._seed if _seed is None else _seed),
+            "device": device
         })
-
-        if self.model_base == "torch" and device is not None:
-            model_kwargs["device"] = device
-
-        # Resolve objective (call-level override beats constructor default)
-        obj = self.objective if objective is None else objective
-        self.model_constructor = self._get_SGNS_constructor_from(self.model_base, obj)
-        self.model = self.model_constructor(**model_kwargs)
-
-        _logger.info(
-            "Training SG base=%s constructor=%s objective=%s frame=%d pairs=%d dim=%d epochs=%d batch=%d neg=%d shuffle=%s",
-            self.model_base,
-            getattr(self.model_constructor, "__name__", repr(self.model_constructor)),
-            obj,
-            frame_id,
-            pairs.shape[0],
-            dimensionality,
-            num_epochs,
-            batch_size,
-            num_negative_samples,
-            shuffle_data
-        )
-
-        if obj == "sgns":
-            self.model.fit(
-                centers,
-                contexts,
-                num_epochs,
-                batch_size,
-                num_negative_samples,
-                noise_probs,
-                shuffle_data,
-                lr_step_per_batch
-            )
-        else:
-            self.model.fit(
-                centers,
-                contexts,
-                num_epochs,
-                batch_size,
-                shuffle_data,
-                lr_step_per_batch
+        _logger.debug("Model constructor kwargs: %s", {k: constructor_kwargs[k] for k in ("V","D","seed","device")})
+        # ----------------------------------------------------------
+
+        # --------------- resolve model constructor ----------------
+        model_constructor = self._get_NN_constructor_from(
+            model_base, objective=("sgns" if negative_sampling else "sg"))
+        # ----------------------------------------------------------
+
+        # ------------------ initialize the model ------------------
+        model = model_constructor(**constructor_kwargs)
+        _logger.debug("Model initialized: %s", model_constructor.__name__ if hasattr(model_constructor,"__name__") else str(model_constructor))
+        # ----------------------------------------------------------
+
+        # -------------------- fitting the data --------------------
+        _logger.info("Fitting model on %d pairs ...", pairs.shape[0])
+        model.fit(centers=centers,
+                  contexts=contexts, 
+                  num_epochs=num_epochs, 
+                  batch_size=batch_size,
+                  # -- optional; for SGNS; safely ignored by SG via **_ignore -- 
+                  num_negative_samples=num_negative_samples,
+                  noise_dist=noise_probs,
+                  # -----------------------------------------
+                  shuffle_data=shuffle_data,
+                  lr_step_per_batch=lr_step_per_batch
             )
-
-        # Select embedding matrix by kind
-        if kind == "in":
-            embeddings = getattr(self.model, "in_embeddings", None)
-            if embeddings is None:
-                embeddings = getattr(self.model, "embeddings", None)
-                if embeddings is None:
-                    params = getattr(self.model, "parameters", None)
-                    if isinstance(params, tuple) and params:
-                        embeddings = params[0]
-        elif kind == "out":
-            embeddings = getattr(self.model, "out_embeddings", None)
-        else:  # "avg"
-            embeddings = getattr(self.model, "avg_embeddings", None)
+        _logger.info("Training complete for frame %d", frame_id)
+        # ----------------------------------------------------------
+
+        # OUTPUT:
+        embeddings = (model.in_embeddings  if kind == "in" else  
+                      model.out_embeddings if kind == "out" else
+                      model.avg_embeddings if kind == "avg" else
+                      None
+                )
 
         if embeddings is None:
-            raise AttributeError(
-                "SG/SGNS model does not expose the requested embeddings: "
-                f"kind={kind!r}"
-            )
+            if kind not in ("in", "out", "avg"):
+                raise NameError(f"Unknown {kind} embeddings kind. Expected: one of ['in', 'out', 'avg']")
 
-        embeddings = np.asarray(embeddings)
-        _logger.info("Frame %d embeddings ready: shape=%s dtype=%s", frame_id, embeddings.shape, embeddings.dtype)
-        return embeddings
+        E = np.asarray(embeddings, dtype=np.float32)
+        _logger.debug("Returned embeddings shape: %s dtype=%s", E.shape, E.dtype)
+        return E
 
     def embed_all(
         self,
         RIN_type: Literal["attr", "repuls"],
         using: Literal["RW", "SAW", "merged"],
-        window_size: int,
-        num_negative_samples: int,
         num_epochs: int,
-        batch_size: int,
+        negative_sampling: bool = False,
+        window_size: int = 2,
+        num_negative_samples: int = 10,
+        batch_size: int = 1024,
         *,
+        lr_step_per_batch: bool = False,
         shuffle_data: bool = True,
         dimensionality: int = 128,
         alpha: float = 0.75,
         device: str | None = None,
-        sgns_kwargs: dict[str, object] | None = None,
+        model_base: Literal["torch", "pureml"] = "pureml",
+        model_kwargs: dict[str, object] | None = None,
+        kind: Literal["in", "out", "avg"] = "in",
         output_path: str | Path | None = None,
         num_matrices_in_compressed_blocks: int = 20,
         compression_level: int = 3,
-        objective: Literal["sgns", "sg"] | None = None,
-        kind: Literal["in", "out", "avg"] = "in"):
-        """Train embeddings for all frames and persist them to compressed storage.
+        ) -> str:
+        """Embed all frames and persist a self-contained archive.
 
-        Iterates through all frames (``1..frame_count``), trains an SGNS model
-        per frame using the configured backend, collects the resulting input
-        embeddings, and writes them into a new compressed ``ArrayStorage`` archive.
+        The resulting file stores a block named ``FRAME_EMBEDDINGS`` with a
+        compressed sequence of per-frame matrices (each ``(V, D)``), alongside
+        rich metadata mirroring the style of other SAWNERGY modules.
 
         Args:
-            RIN_type: Interaction channel to use: ``"attr"`` or ``"repuls"``.
-            using: Walk collections: ``"RW"``, ``"SAW"``, or ``"merged"``.
-            window_size: Symmetric skip-gram window size ``k``.
-            num_negative_samples: Number of negative samples per positive pair.
-                Ignored when ``objective="sg"``.
-            num_epochs: Number of epochs for each frame.
-            batch_size: Mini-batch size used during training.
-            shuffle_data: Whether to shuffle pairs each epoch.
-            dimensionality: Embedding dimensionality ``D``.
-            alpha: Noise distribution exponent (``Pn ∝ f^alpha``).
-            device: Optional device string for Torch backend.
-            sgns_kwargs: Extra constructor kwargs for the SGNS backend (see
-                :meth:`embed_frame` for PureML requirements).
-            output_path: Destination path. If ``None``, a new file named
-                ``EMBEDDINGS_<timestamp>.zip`` is created next to the source
-                WALKS archive. If the provided path lacks a suffix, ``.zip`` is
-                appended.
-            num_matrices_in_compressed_blocks: Number of per-frame matrices to
-                store per compressed chunk in the output archive.
-            compression_level: Blosc Zstd compression level (0-9).
-            objective: Training objective for all frames (``"sgns"`` or ``"sg"``).
-                If ``None``, uses the value set at construction.
-            kind: Which embedding matrix to persist for each frame: ``"in"``,
-                ``"out"``, or ``"avg"``.
+            RIN_type: ``"attr"`` or ``"repuls"`` - which corpus to use.
+            using: Which walks to use (``"RW"``, ``"SAW"``, or ``"merged"``).
+            num_epochs: Number of epochs to train per frame.
+            negative_sampling: If ``True``, use SGNS; otherwise plain SG.
+            window_size: Skip-gram window radius.
+            num_negative_samples: Negatives per positive pair (SGNS).
+            batch_size: Minibatch size for training.
+            lr_step_per_batch: If ``True``, step LR per batch (else per epoch).
+            shuffle_data: Shuffle pairs each epoch.
+            dimensionality: Embedding dimension.
+            alpha: Unigram smoothing power for noise distribution.
+            device: Backend device hint (e.g., ``"cuda"``).
+            model_base: Backend family (``"torch"`` or ``"pureml"``).
+            model_kwargs: Passed through to backend model constructor.
+            kind: Which embedding to store: ``"in"``, ``"out"``, or ``"avg"``.
+            output_path: Optional path for the output archive (``.zip`` inferred).
+            num_matrices_in_compressed_blocks: How many frames per compressed chunk.
+            compression_level: Integer compression level for the archive.
 
         Returns:
-            str: Filesystem path to the written embeddings archive (``.zip``).
-
-        Raises:
-            ValueError: If configuration produces no pairs for a frame or if
-                PureML kwargs are incomplete.
-            RuntimeError: Propagated from storage operations on failure.
-
-        Notes:
-            - A deterministic child seed is spawned per frame from the master
-            seed using ``np.random.SeedSequence`` to ensure reproducibility
-            across runs.
+            Path to the created embeddings archive, as ``str``.
         """
         current_time = sawnergy_util.current_time()
         if output_path is None:
@@ -563,71 +612,102 @@ class Embedder:
                 output_path = output_path.with_suffix(".zip")
 
         _logger.info(
-            "Embedding all frames -> %s | frames=%d dim=%d base=%s",
-            output_path, self.frame_count, dimensionality, self.model_base
+            "embed_all: frames=%d D=%d base=%s RIN=%s using=%s out=%s",
+            self.frame_count, dimensionality, model_base, RIN_type, using, output_path
         )
 
+        # Per-frame deterministic seeds
         master_ss = np.random.SeedSequence(self._seed)
         child_seeds = master_ss.spawn(self.frame_count)
 
-        embeddings = []
-        for frame_idx, seed_seq in enumerate(child_seeds, start=1):
+        embeddings: list[np.ndarray] = []
+        for frame_id, seed_seq in enumerate(child_seeds, start=1):
             child_seed = int(seed_seq.generate_state(1, dtype=np.uint32)[0])
-            _logger.info("Processing frame %d/%d (child_seed=%d entropy=%d)", frame_idx, self.frame_count, child_seed, seed_seq.entropy)
-            embeddings.append(
-                self.embed_frame(
-                    frame_idx,
-                    RIN_type,
-                    using,
-                    window_size,
-                    num_negative_samples,
-                    num_epochs,
-                    batch_size,
+            _logger.info("Embedding frame %d/%d with seed=%d", frame_id, self.frame_count, child_seed)
+            E = self.embed_frame(
+                    frame_id=frame_id,
+                    RIN_type=RIN_type,
+                    using=using,
+                    num_epochs=num_epochs,
+                    negative_sampling=negative_sampling,
+                    window_size=window_size,
+                    num_negative_samples=num_negative_samples,
+                    batch_size=batch_size,
+                    lr_step_per_batch=lr_step_per_batch,
                     shuffle_data=shuffle_data,
                     dimensionality=dimensionality,
                     alpha=alpha,
                     device=device,
-                    sgns_kwargs=sgns_kwargs,
-                    _seed=child_seed,
-                    objective=objective,
-                    kind=kind
+                    model_base=model_base,
+                    model_kwargs=model_kwargs,
+                    kind=kind,
+                    _seed=child_seed
                 )
-            )
+            embeddings.append(np.asarray(E, dtype=np.float32, copy=False))
+            _logger.debug("Frame %d embedded: E.shape=%s", frame_id, E.shape)
 
-        embeddings = [np.asarray(e) for e in embeddings]
         block_name = "FRAME_EMBEDDINGS"
         with sawnergy_util.ArrayStorage.compress_and_cleanup(output_path, compression_level=compression_level) as storage:
+            _logger.info("Writing %d frame matrices to block '%s' ...", len(embeddings), block_name)
             storage.write(
                 these_arrays=embeddings,
                 to_block_named=block_name,
                 arrays_per_chunk=num_matrices_in_compressed_blocks
             )
-            storage.add_attr("time_created", current_time)
-            storage.add_attr("seed", int(self._seed))
-            storage.add_attr("rng_scheme", "SeedSequence.spawn_per_frame_v1")
-            storage.add_attr("source_walks_path", str(self._walks_path))
-            storage.add_attr("model_base", self.model_base)
-            storage.add_attr("rin_type", RIN_type)
-            storage.add_attr("using_mode", using)
+
+            # Core dataset discovery (for consumers like the Embeddings Visualizer)
+            storage.add_attr("frame_embeddings_name", block_name)
+            storage.add_attr("time_stamp_count", int(self.frame_count))
+            storage.add_attr("node_count", int(self.vocab_size))
+            storage.add_attr("embedding_dim", int(dimensionality))
+
+            # Provenance of input WALKS
+            storage.add_attr("source_WALKS_path", str(self._walks_path))
+            storage.add_attr("walk_length", int(self.walk_length))
+            storage.add_attr("num_RWs", int(self.num_RWs))
+            storage.add_attr("num_SAWs", int(self.num_SAWs))
+            storage.add_attr("attractive_RWs_name", self._attractive_RWs_name)
+            storage.add_attr("repulsive_RWs_name",  self._repulsive_RWs_name)
+            storage.add_attr("attractive_SAWs_name", self._attractive_SAWs_name)
+            storage.add_attr("repulsive_SAWs_name",  self._repulsive_SAWs_name)
+
+            # Training configuration (sufficient to reproduce)
+            storage.add_attr("objective", "sgns" if negative_sampling else "sg")
+            storage.add_attr("model_base", model_base)
+            storage.add_attr("embedding_kind", kind)  # 'in' | 'out' | 'avg'
+            storage.add_attr("num_epochs", int(num_epochs))
+            storage.add_attr("batch_size", int(batch_size))
             storage.add_attr("window_size", int(window_size))
             storage.add_attr("alpha", float(alpha))
-            storage.add_attr("dimensionality", int(dimensionality))
+            storage.add_attr("negative_sampling", bool(negative_sampling))
             storage.add_attr("num_negative_samples", int(num_negative_samples))
-            storage.add_attr("num_epochs", int(num_epochs))
-            storage.add_attr("batch_size", int(batch_size))
+            storage.add_attr("lr_step_per_batch", bool(lr_step_per_batch))
             storage.add_attr("shuffle_data", bool(shuffle_data))
-            storage.add_attr("frames_written", int(len(embeddings)))
-            storage.add_attr("vocab_size", int(self.vocab_size))
-            storage.add_attr("frame_count", int(self.frame_count))
-            storage.add_attr("embedding_dtype", str(embeddings[0].dtype))
-            storage.add_attr("frame_embeddings_name", block_name)
+            storage.add_attr("device_hint", device if device is not None else "")
+            storage.add_attr("model_kwargs_repr", repr(model_kwargs) if model_kwargs is not None else "{}")
+
+            # Which walks were used to train
+            storage.add_attr("RIN_type", RIN_type)   # 'attr' or 'repuls'
+            storage.add_attr("using", using)         # 'RW' | 'SAW' | 'merged'
+
+            # Reproducibility
+            storage.add_attr("master_seed", int(self._seed))
+            # Note: this records seeds derived from child SeedSequences at metadata time.
+            storage.add_attr("per_frame_seeds", [int(s.generate_state(1, dtype=np.uint32)[0]) for s in child_seeds])
+
+            # Archive/IO details
             storage.add_attr("arrays_per_chunk", int(num_matrices_in_compressed_blocks))
             storage.add_attr("compression_level", int(compression_level))
-            storage.add_attr("objective", self.objective if objective is None else objective)
+            storage.add_attr("created_at", current_time)
+
+            _logger.info(
+                "Stored embeddings archive: %s | shape=(T,N,D)=(%d,%d,%d)",
+                output_path, self.frame_count, self.vocab_size, dimensionality
+            )
 
-        _logger.info("Embedding archive written to %s", output_path)
         return str(output_path)
 
+
 __all__ = ["Embedder"]
 
 if __name__ == "__main__":