gsvvcompressor 1.2.0__cp310-cp310-win_amd64.whl

gsvvcompressor/draco/twopass.py

@@ -0,0 +1,359 @@
+"""
+Two-pass Draco-based serialization and deserialization.
+
+This module accumulates multiple frames during serialization, then compresses
+all data as a single block during flush. This enables better compression by
+treating all frames as one continuous point cloud.
+
+WARNING: cloudpickle uses pickle under the hood. Do NOT use with untrusted
+data sources. Only use with trusted data (local files, same-process, etc.).
+"""
+
+import struct
+from dataclasses import dataclass
+from typing import Iterator, List, Optional, Self
+
+import cloudpickle as pickle
+import numpy as np
+import zstandard as zstd
+
+from ..deserializer import AbstractDeserializer
+from ..payload import Payload
+from ..serializer import AbstractSerializer
+from .interface import DracoPayload
+
+from . import dracoreduced3dgs
+
+# 4-byte big-endian length prefix (max 4GB per object)
+_LEN = struct.Struct(">I")
+
+
+@dataclass
+class TwoPassDracoPayload(DracoPayload):
+    """
+    Payload containing data in Draco-compatible format with point count.
+
+    This extends DracoPayload with a num_points field to track the number
+    of points in each frame, enabling reconstruction after batched compression.
+
+    Attributes:
+        positions: Point positions, shape (N, 3), dtype float32/float64
+        scales: Scale indices or values, shape (N, 1), dtype int32 for VQ indices
+        rotations: Rotation indices or values, shape (N, 2), dtype int32 for VQ indices
+        opacities: Opacity indices or values, shape (N, 1), dtype int32 for VQ indices
+        features_dc: DC feature indices or values, shape (N, 1), dtype int32 for VQ indices
+        features_rest: Rest feature indices or values, shape (N, 9), dtype int32 for VQ indices
+        extra: Optional additional payload for codec-specific data
+        num_points: Number of points in this frame
+    """
+
+    num_points: int = 0
+
+    def to(self, device) -> Self:
+        """
+        Move the Payload to the specified device.
+
+        Since TwoPassDracoPayload uses numpy arrays (CPU-only), only the extra
+        payload (if present) is moved to the target device.
+
+        Args:
+            device: The target device (e.g., 'cpu', 'cuda', torch.device).
+
+        Returns:
+            A new TwoPassDracoPayload with the extra payload moved to the target device.
+        """
+        return TwoPassDracoPayload(
+            positions=self.positions,
+            scales=self.scales,
+            rotations=self.rotations,
+            opacities=self.opacities,
+            features_dc=self.features_dc,
+            features_rest=self.features_rest,
+            extra=self.extra.to(device) if self.extra is not None else None,
+            num_points=self.num_points,
+        )
+
+
+class TwoPassDracoSerializer(AbstractSerializer):
+    """
+    Two-pass serializer that accumulates frames and compresses as a single block.
+
+    This serializer collects all frame data during serialize_frame calls,
+    concatenating positions, scales, rotations, etc. The actual compression
+    happens only during flush(), where all data is encoded as one Draco
+    point cloud and compressed with zstd.
+
+    Frame format after flush:
+    [num_frames][num_points_per_frame...][draco_len][draco_bytes][extras_len][extras_bytes]
+    """
+
+    def __init__(
+        self,
+        zstd_level: int = 7,
+        draco_level: int = 0,
+        qp: int = 30,
+        qscale: int = 30,
+        qrotation: int = 30,
+        qopacity: int = 30,
+        qfeaturedc: int = 30,
+        qfeaturerest: int = 30,
+    ):
+        """
+        Initialize the serializer.
+
+        Args:
+            zstd_level: Zstd compression level (1-22). Default is 7.
+            draco_level: Draco compression level (0-10). Default is 0.
+            qp: Quantization bits for positions (0 to disable).
+            qscale: Quantization bits for scales (0 to disable).
+            qrotation: Quantization bits for rotations (0 to disable).
+            qopacity: Quantization bits for opacities (0 to disable).
+            qfeaturedc: Quantization bits for features_dc (0 to disable).
+            qfeaturerest: Quantization bits for features_rest (0 to disable).
+        """
+        self._zstd_level = zstd_level
+        self._draco_level = draco_level
+        self._qp = qp
+        self._qscale = qscale
+        self._qrotation = qrotation
+        self._qopacity = qopacity
+        self._qfeaturedc = qfeaturedc
+        self._qfeaturerest = qfeaturerest
+
+        # Accumulators for frame data
+        self._positions_list: List[np.ndarray] = []
+        self._scales_list: List[np.ndarray] = []
+        self._rotations_list: List[np.ndarray] = []
+        self._opacities_list: List[np.ndarray] = []
+        self._features_dc_list: List[np.ndarray] = []
+        self._features_rest_list: List[np.ndarray] = []
+        self._num_points_list: List[int] = []
+        self._extras_list: List[Optional[Payload]] = []
+
+    def serialize_frame(self, payload: DracoPayload) -> Iterator[bytes]:
+        """
+        Accumulate a DracoPayload for later compression.
+
+        This method does not yield any bytes - all compression happens in flush().
+
+        Args:
+            payload: A DracoPayload instance to accumulate.
+
+        Yields:
+            Nothing (compression is deferred to flush).
+        """
+        num_points = payload.positions.shape[0]
+
+        self._positions_list.append(payload.positions)
+        self._scales_list.append(payload.scales)
+        self._rotations_list.append(payload.rotations)
+        self._opacities_list.append(payload.opacities)
+        self._features_dc_list.append(payload.features_dc)
+        self._features_rest_list.append(payload.features_rest)
+        self._num_points_list.append(num_points)
+        self._extras_list.append(payload.extra)
+
+        # Yield nothing - compression happens in flush
+        return
+        yield  # Make this a generator
+
+    def flush(self) -> Iterator[bytes]:
+        """
+        Compress all accumulated frames as a single block.
+
+        Concatenates all frame data, encodes with Draco, pickles extras,
+        and compresses everything with zstd.
+
+        Yields:
+            Compressed byte chunks containing all frames.
+        """
+        if not self._positions_list:
+            return
+
+        # Concatenate all arrays
+        all_positions = np.concatenate(self._positions_list, axis=0)
+        all_scales = np.concatenate(self._scales_list, axis=0)
+        all_rotations = np.concatenate(self._rotations_list, axis=0)
+        all_opacities = np.concatenate(self._opacities_list, axis=0)
+        all_features_dc = np.concatenate(self._features_dc_list, axis=0)
+        all_features_rest = np.concatenate(self._features_rest_list, axis=0)
+
+        # Encode all point cloud data using Draco
+        draco_encoded = dracoreduced3dgs.encode(
+            all_positions,
+            all_scales,
+            all_rotations,
+            all_opacities,
+            all_features_dc,
+            all_features_rest,
+            self._draco_level,
+            self._qp,
+            self._qscale,
+            self._qrotation,
+            self._qopacity,
+            self._qfeaturedc,
+            self._qfeaturerest,
+        )
+
+        # Pickle extras list
+        extras_pickled = pickle.dumps(self._extras_list, protocol=pickle.DEFAULT_PROTOCOL)
+
+        # Build frame structure:
+        # [num_frames][num_points_per_frame...][draco_len][draco_bytes][extras_len][extras_bytes]
+        num_frames = len(self._num_points_list)
+
+        frame_header = _LEN.pack(num_frames)
+        for n in self._num_points_list:
+            frame_header += _LEN.pack(n)
+
+        framed = (
+            frame_header
+            + _LEN.pack(len(draco_encoded))
+            + draco_encoded
+            + _LEN.pack(len(extras_pickled))
+            + extras_pickled
+        )
+
+        # Compress with zstd
+        compressor = zstd.ZstdCompressor(level=self._zstd_level)
+        compressed = compressor.compress(framed)
+
+        yield compressed
+
+        # Clear accumulators
+        self._positions_list.clear()
+        self._scales_list.clear()
+        self._rotations_list.clear()
+        self._opacities_list.clear()
+        self._features_dc_list.clear()
+        self._features_rest_list.clear()
+        self._num_points_list.clear()
+        self._extras_list.clear()
+
+
+class TwoPassDracoDeserializer(AbstractDeserializer):
+    """
+    Deserializer for two-pass Draco-compressed frame blocks.
+
+    This deserializer takes a single compressed block containing multiple frames
+    and returns them as a list of TwoPassDracoPayload objects.
+    """
+
+    def __init__(self):
+        """Initialize the deserializer."""
+        self._buffer = bytearray()
+
+    def deserialize_frame(self, data: bytes) -> Iterator[TwoPassDracoPayload]:
+        """
+        Decompress and deserialize a block of frames.
+
+        Args:
+            data: Compressed bytes containing all frames.
+
+        Yields:
+            TwoPassDracoPayload objects for each frame in the block.
+        """
+        # Accumulate data
+        self._buffer.extend(data)
+
+        # Try to decompress and parse
+        if not self._buffer:
+            return
+
+        try:
+            # Decompress with zstd
+            decompressor = zstd.ZstdDecompressor()
+            decompressed = decompressor.decompress(bytes(self._buffer))
+        except zstd.ZstdError:
+            # Not enough data yet, wait for more
+            return
+
+        # Clear buffer after successful decompression
+        self._buffer.clear()
+
+        # Parse frame structure
+        offset = 0
+
+        # Read num_frames
+        if len(decompressed) < offset + _LEN.size:
+            raise ValueError("Incomplete data: missing num_frames")
+        (num_frames,) = _LEN.unpack(decompressed[offset: offset + _LEN.size])
+        offset += _LEN.size
+
+        # Read num_points for each frame
+        num_points_list = []
+        for _ in range(num_frames):
+            if len(decompressed) < offset + _LEN.size:
+                raise ValueError("Incomplete data: missing num_points")
+            (n,) = _LEN.unpack(decompressed[offset: offset + _LEN.size])
+            num_points_list.append(n)
+            offset += _LEN.size
+
+        # Read draco data
+        if len(decompressed) < offset + _LEN.size:
+            raise ValueError("Incomplete data: missing draco_len")
+        (draco_len,) = _LEN.unpack(decompressed[offset: offset + _LEN.size])
+        offset += _LEN.size
+
+        if len(decompressed) < offset + draco_len:
+            raise ValueError("Incomplete data: missing draco_bytes")
+        draco_bytes = decompressed[offset: offset + draco_len]
+        offset += draco_len
+
+        # Read extras
+        if len(decompressed) < offset + _LEN.size:
+            raise ValueError("Incomplete data: missing extras_len")
+        (extras_len,) = _LEN.unpack(decompressed[offset: offset + _LEN.size])
+        offset += _LEN.size
+
+        if len(decompressed) < offset + extras_len:
+            raise ValueError("Incomplete data: missing extras_bytes")
+        extras_bytes = decompressed[offset: offset + extras_len]
+        offset += extras_len
+
+        # Decode draco data
+        pc = dracoreduced3dgs.decode(draco_bytes)
+
+        # Unpickle extras list
+        extras_list: List[Optional[Payload]] = pickle.loads(extras_bytes)
+
+        # Split arrays by num_points
+        all_positions = pc.positions
+        all_scales = pc.scales
+        all_rotations = pc.rotations
+        all_opacities = pc.opacities
+        all_features_dc = pc.features_dc
+        all_features_rest = pc.features_rest
+
+        start_idx = 0
+        for i, num_points in enumerate(num_points_list):
+            end_idx = start_idx + num_points
+
+            yield TwoPassDracoPayload(
+                positions=all_positions[start_idx:end_idx],
+                scales=all_scales[start_idx:end_idx],
+                rotations=all_rotations[start_idx:end_idx],
+                opacities=all_opacities[start_idx:end_idx],
+                features_dc=all_features_dc[start_idx:end_idx],
+                features_rest=all_features_rest[start_idx:end_idx],
+                extra=extras_list[i] if i < len(extras_list) else None,
+                num_points=num_points,
+            )
+
+            start_idx = end_idx
+
+    def flush(self) -> Iterator[TwoPassDracoPayload]:
+        """
+        Flush any remaining buffered data.
+
+        Yields:
+            Any remaining TwoPassDracoPayload objects.
+        """
+        # If there's leftover data, it's incomplete/corrupted
+        if self._buffer:
+            raise ValueError(
+                f"Incomplete data in buffer: {len(self._buffer)} bytes remaining"
+            )
+
+        return
+        yield  # Make this a generator

gsvvcompressor/encoder.py

@@ -0,0 +1,122 @@
+from abc import ABC, abstractmethod
+from typing import Iterator
+
+from gaussian_splatting import GaussianModel
+
+from .payload import Payload
+from .serializer import AbstractSerializer
+
+
+class AbstractEncoder(ABC):
+    """
+    Abstract base class for compression algorithms.
+
+    This encoder uses a two-stage process:
+    1. Pack frames into Payload objects (via `pack`)
+    2. Serialize Payloads to bytes (via the serializer)
+
+    Subclasses must implement `pack` and `flush_pack`, and provide a serializer.
+    This design separates frame packing logic from serialization format.
+    """
+
+    def __init__(self, serializer: AbstractSerializer, payload_device=None):
+        """
+        Initialize the encoder.
+
+        Args:
+            serializer: The serializer to use for converting Payload to bytes.
+            payload_device: The target device for encoded Payloads before
+                serialization (e.g., 'cpu', 'cuda'). If None, no device
+                transfer is performed.
+        """
+        self._serializer = serializer
+        self._payload_device = payload_device
+
+    @abstractmethod
+    def pack(self, frame: GaussianModel) -> Iterator[Payload]:
+        """
+        Pack a single frame into Payload objects.
+
+        This method transforms the input frame into Payload objects that
+        can be serialized.
+
+        Args:
+            frame: A GaussianModel instance to pack.
+
+        Yields:
+            Packed Payload instances. May yield zero, one, or multiple
+            payloads. When the iterator is exhausted, all payloads for this
+            frame have been yielded.
+        """
+        pass
+
+    @abstractmethod
+    def flush_pack(self) -> Iterator[Payload]:
+        """
+        Flush any remaining buffered payloads from the packing stage.
+
+        This method should be called after all frames have been packed
+        to ensure any remaining buffered payloads are output.
+
+        Yields:
+            Remaining buffered Payload instances. May yield zero, one, or
+            multiple payloads until all buffered data has been flushed.
+        """
+        pass
+
+    def encode_frame(self, frame: GaussianModel) -> Iterator[bytes]:
+        """
+        Encode a single frame of GaussianModel.
+
+        This method packs the frame into Payloads and then serializes them.
+
+        Args:
+            frame: A GaussianModel instance to encode.
+
+        Yields:
+            Encoded byte chunks. May yield zero, one, or multiple chunks.
+            When the iterator is exhausted, all data for this frame has been
+            encoded.
+        """
+        for payload in self.pack(frame):
+            if self._payload_device is not None:
+                payload = payload.to(self._payload_device)
+            yield from self._serializer.serialize_frame(payload)
+
+    def flush(self) -> Iterator[bytes]:
+        """
+        Flush any remaining buffered data from both packing and serialization.
+
+        This method should be called after all frames have been encoded
+        to ensure any remaining buffered data is output.
+
+        Yields:
+            Remaining buffered byte chunks. May yield zero, one, or multiple
+            chunks until all buffered data has been flushed.
+        """
+        # Flush packing stage and serialize any remaining payloads
+        for payload in self.flush_pack():
+            if self._payload_device is not None:
+                payload = payload.to(self._payload_device)
+            yield from self._serializer.serialize_frame(payload)
+
+        # Flush serialization stage
+        yield from self._serializer.flush()
+
+    def encode_stream(self, stream: Iterator[GaussianModel]) -> Iterator[bytes]:
+        """
+        Encode a stream of GaussianModel frames.
+
+        This method packs each frame and serializes the payloads.
+        It handles the flush logic for both packing and serialization stages.
+
+        Args:
+            stream: An iterator that yields GaussianModel instances to encode.
+
+        Yields:
+            Encoded bytes for each packed frame or flush operation.
+        """
+        for frame in stream:
+            yield from self.encode_frame(frame)
+
+        yield from self.flush()

gsvvcompressor/interframe/__init__.py

@@ -0,0 +1,11 @@
+from .interface import InterframeEncoderInitConfig, InterframeCodecContext, InterframeCodecInterface
+from .encoder import InterframeEncoder
+from .decoder import InterframeDecoder
+
+__all__ = [
+    'InterframeEncoderInitConfig',
+    'InterframeCodecContext',
+    'InterframeCodecInterface',
+    'InterframeEncoder',
+    'InterframeDecoder',
+]