datachain 0.36.0__tar.gz → 0.36.1__tar.gz

{datachain-0.36.0 → datachain-0.36.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: datachain
-Version: 0.36.0
+Version: 0.36.1
 Summary: Wrangle unstructured AI data at scale
 Author-email: Dmitry Petrov <support@dvc.org>
 License-Expression: Apache-2.0
@@ -64,7 +64,6 @@ Requires-Dist: torch>=2.1.0; extra == "torch"
 Requires-Dist: torchvision; extra == "torch"
 Requires-Dist: transformers>=4.36.0; extra == "torch"
 Provides-Extra: audio
-Requires-Dist: torchaudio; extra == "audio"
 Requires-Dist: soundfile; extra == "audio"
 Provides-Extra: remote
 Requires-Dist: lz4; extra == "remote"
@@ -76,6 +75,7 @@ Requires-Dist: numba>=0.60.0; extra == "hf"
 Requires-Dist: datasets[vision]>=4.0.0; extra == "hf"
 Requires-Dist: datasets[audio]>=4.0.0; (sys_platform == "linux" or sys_platform == "darwin") and extra == "hf"
 Requires-Dist: fsspec>=2024.12.0; extra == "hf"
+Requires-Dist: torch<2.9.0; extra == "hf"
 Provides-Extra: video
 Requires-Dist: ffmpeg-python; extra == "video"
 Requires-Dist: imageio[ffmpeg,pyav]>=2.37.0; extra == "video"
@@ -117,6 +117,7 @@ Requires-Dist: huggingface_hub[hf_transfer]; extra == "examples"
 Requires-Dist: ultralytics; extra == "examples"
 Requires-Dist: open_clip_torch; extra == "examples"
 Requires-Dist: openai; extra == "examples"
+Requires-Dist: torchaudio<2.9.0; extra == "examples"
 Dynamic: license-file
 
 ================

{datachain-0.36.0 → datachain-0.36.1}/pyproject.toml

@@ -73,7 +73,6 @@ torch = [
   "transformers>=4.36.0"
 ]
 audio = [
-  "torchaudio",
   "soundfile"
 ]
 remote = [
@@ -88,7 +87,11 @@ hf = [
   "datasets[vision]>=4.0.0",
   # https://github.com/pytorch/torchcodec/issues/640
   "datasets[audio]>=4.0.0 ; (sys_platform == 'linux' or sys_platform == 'darwin')",
-  "fsspec>=2024.12.0"
+  "fsspec>=2024.12.0",
+  # Until datasets solve the issue, run test_hf_audio test to see if this can be removed
+  # https://github.com/meta-pytorch/torchcodec/issues/912
+  # https://github.com/huggingface/transformers/pull/41610
+  "torch<2.9.0"
 ]
 video = [
   "ffmpeg-python",
@@ -134,7 +137,9 @@ examples = [
   "huggingface_hub[hf_transfer]",
   "ultralytics",
   "open_clip_torch",
-  "openai"
+  "openai",
+  # Transformers still require it
+  "torchaudio<2.9.0"
 ]
 
 [project.urls]

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/data_storage/schema.py

@@ -11,7 +11,6 @@ from datachain.sql.types import (
     JSON,
     Boolean,
     DateTime,
-    Int,
     Int64,
     SQLType,
     String,
@@ -269,7 +268,7 @@ class DataTable:
     @classmethod
     def sys_columns(cls):
         return [
-            sa.Column("sys__id", Int, primary_key=True),
+            sa.Column("sys__id", UInt64, primary_key=True),
             sa.Column(
                 "sys__rand", UInt64, nullable=False, server_default=f.abs(f.random())
             ),

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/data_storage/sqlite.py

@@ -868,11 +868,8 @@ class SQLiteWarehouse(AbstractWarehouse):
                 if isinstance(c, BinaryExpression):
                     right_left_join = add_left_rows_filter(c)
 
-        # Use CTE instead of subquery to force SQLite to materialize the result
-        # This breaks deep nesting and prevents parser stack overflow.
         union_cte = sqlalchemy.union(left_right_join, right_left_join).cte()
-
-        return self._regenerate_system_columns(union_cte)
+        return sqlalchemy.select(*union_cte.c).select_from(union_cte)
 
     def _system_row_number_expr(self):
         return func.row_number().over()
@@ -884,11 +881,7 @@ class SQLiteWarehouse(AbstractWarehouse):
         """
         Create a temporary table from a query for use in a UDF.
         """
-        columns = [
-            sqlalchemy.Column(c.name, c.type)
-            for c in query.selected_columns
-            if c.name != "sys__id"
-        ]
+        columns = [sqlalchemy.Column(c.name, c.type) for c in query.selected_columns]
         table = self.create_udf_table(columns)
 
         with tqdm(desc="Preparing", unit=" rows", leave=False) as pbar:

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/data_storage/warehouse.py

@@ -5,7 +5,7 @@ import random
 import string
 from abc import ABC, abstractmethod
 from collections.abc import Callable, Generator, Iterable, Iterator, Sequence
-from typing import TYPE_CHECKING, Any, Union
+from typing import TYPE_CHECKING, Any, Union, cast
 from urllib.parse import urlparse
 
 import attrs
@@ -23,7 +23,7 @@ from datachain.node import DirType, DirTypeGroup, Node, NodeWithPath, get_path
 from datachain.query.batch import RowsOutput
 from datachain.query.schema import ColumnMeta
 from datachain.sql.functions import path as pathfunc
-from datachain.sql.types import Int, SQLType
+from datachain.sql.types import SQLType
 from datachain.utils import sql_escape_like
 
 if TYPE_CHECKING:
@@ -32,6 +32,7 @@ if TYPE_CHECKING:
         _FromClauseArgument,
         _OnClauseArgument,
     )
+    from sqlalchemy.sql.selectable import FromClause
     from sqlalchemy.types import TypeEngine
 
     from datachain.data_storage import schema
@@ -248,45 +249,56 @@ class AbstractWarehouse(ABC, Serializable):
 
     def _regenerate_system_columns(
         self,
-        selectable: sa.Select | sa.CTE,
+        selectable: sa.Select,
         keep_existing_columns: bool = False,
+        regenerate_columns: Iterable[str] | None = None,
     ) -> sa.Select:
         """
-        Return a SELECT that regenerates sys__id and sys__rand deterministically.
+        Return a SELECT that regenerates system columns deterministically.
 
-        If keep_existing_columns is True, existing sys__id and sys__rand columns
-        will be kept as-is if they exist in the input selectable.
-        """
-        base = selectable.subquery() if hasattr(selectable, "subquery") else selectable
-
-        result_columns: dict[str, sa.ColumnElement] = {}
-        for col in base.c:
-            if col.name in result_columns:
-                raise ValueError(f"Duplicate column name {col.name} in SELECT")
-            if col.name in ("sys__id", "sys__rand"):
-                if keep_existing_columns:
-                    result_columns[col.name] = col
-            else:
-                result_columns[col.name] = col
+        If keep_existing_columns is True, existing system columns will be kept as-is
+        even when they are listed in ``regenerate_columns``.
 
-        system_types: dict[str, sa.types.TypeEngine] = {
+        Args:
+            selectable: Base SELECT
+            keep_existing_columns: When True, reuse existing system columns even if
+                they are part of the regeneration set.
+            regenerate_columns: Names of system columns to regenerate. Defaults to
+                {"sys__id", "sys__rand"}. Columns not listed are left untouched.
+        """
+        system_columns = {
             sys_col.name: sys_col.type
             for sys_col in self.schema.dataset_row_cls.sys_columns()
         }
+        regenerate = set(regenerate_columns or system_columns)
+        generators = {
+            "sys__id": self._system_row_number_expr,
+            "sys__rand": self._system_random_expr,
+        }
+
+        base = cast("FromClause", selectable.subquery())
+
+        def build(name: str) -> sa.ColumnElement:
+            expr = generators[name]()
+            return sa.cast(expr, system_columns[name]).label(name)
+
+        columns: list[sa.ColumnElement] = []
+        present: set[str] = set()
+        changed = False
+
+        for col in base.c:
+            present.add(col.name)
+            regen = col.name in regenerate and not keep_existing_columns
+            columns.append(build(col.name) if regen else col)
+            changed |= regen
+
+        for name in regenerate - present:
+            columns.append(build(name))
+            changed = True
+
+        if not changed:
+            return selectable
 
-        # Add missing system columns if needed
-        if "sys__id" not in result_columns:
-            expr = self._system_row_number_expr()
-            expr = sa.cast(expr, system_types["sys__id"])
-            result_columns["sys__id"] = expr.label("sys__id")
-        if "sys__rand" not in result_columns:
-            expr = self._system_random_expr()
-            expr = sa.cast(expr, system_types["sys__rand"])
-            result_columns["sys__rand"] = expr.label("sys__rand")
-
-        # Wrap in subquery to materialize window functions, then wrap again in SELECT
-        # This ensures window functions are computed before INSERT...FROM SELECT
-        columns = list(result_columns.values())
         inner = sa.select(*columns).select_from(base).subquery()
         return sa.select(*inner.c).select_from(inner)
 
@@ -950,10 +962,15 @@ class AbstractWarehouse(ABC, Serializable):
         SQLite TEMPORARY tables cannot be directly used as they are process-specific,
         and UDFs are run in other processes when run in parallel.
         """
+        columns = [
+            c
+            for c in columns
+            if c.name not in [col.name for col in self.dataset_row_cls.sys_columns()]
+        ]
         tbl = sa.Table(
             name or self.udf_table_name(),
             sa.MetaData(),
-            sa.Column("sys__id", Int, primary_key=True),
+            *self.dataset_row_cls.sys_columns(),
             *columns,
         )
         self.db.create_table(tbl, if_not_exists=True)

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/diff/__init__.py

@@ -24,7 +24,7 @@ class CompareStatus(str, Enum):
     SAME = "S"
 
 
-def _compare(  # noqa: C901, PLR0912
+def _compare(  # noqa: C901
     left: "DataChain",
     right: "DataChain",
     on: str | Sequence[str],
@@ -151,11 +151,7 @@ def _compare(  # noqa: C901, PLR0912
     if status_col:
         cols_select.append(diff_col)
 
-    if not dc_diff._sys:
-        # TODO workaround when sys signal is not available in diff
-        dc_diff = dc_diff.settings(sys=True).select(*cols_select).settings(sys=False)
-    else:
-        dc_diff = dc_diff.select(*cols_select)
+    dc_diff = dc_diff.select(*cols_select)
 
     # final schema is schema from the left chain with status column added if needed
     dc_diff.signals_schema = (

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/lib/audio.py

@@ -1,4 +1,5 @@
 import posixpath
+import re
 from typing import TYPE_CHECKING
 
 from datachain.lib.file import FileError
@@ -9,7 +10,7 @@ if TYPE_CHECKING:
     from datachain.lib.file import Audio, AudioFile, File
 
 try:
-    import torchaudio
+    import soundfile as sf
 except ImportError as exc:
     raise ImportError(
         "Missing dependencies for processing audio.\n"
@@ -26,18 +27,25 @@ def audio_info(file: "File | AudioFile") -> "Audio":
 
     try:
         with file.open() as f:
-            info = torchaudio.info(f)
+            info = sf.info(f)
+
+            sample_rate = int(info.samplerate)
+            channels = int(info.channels)
+            frames = int(info.frames)
+            duration = float(info.duration)
 
-            sample_rate = int(info.sample_rate)
-            channels = int(info.num_channels)
-            frames = int(info.num_frames)
-            duration = float(frames / sample_rate) if sample_rate > 0 else 0.0
+            # soundfile provides format and subtype
+            if info.format:
+                format_name = info.format.lower()
+            else:
+                format_name = file.get_file_ext().lower()
 
-            codec_name = getattr(info, "encoding", "")
-            file_ext = file.get_file_ext().lower()
-            format_name = _encoding_to_format(codec_name, file_ext)
+            if not format_name:
+                format_name = "unknown"
+            codec_name = info.subtype if info.subtype else ""
 
-            bits_per_sample = getattr(info, "bits_per_sample", 0)
+            # Calculate bit rate from subtype
+            bits_per_sample = _get_bits_per_sample(info.subtype)
             bit_rate = (
                 bits_per_sample * sample_rate * channels if bits_per_sample > 0 else -1
             )
@@ -58,44 +66,39 @@ def audio_info(file: "File | AudioFile") -> "Audio":
     )
 
 
-def _encoding_to_format(encoding: str, file_ext: str) -> str:
+def _get_bits_per_sample(subtype: str) -> int:
     """
-    Map torchaudio encoding to a format name.
+    Map soundfile subtype to bits per sample.
 
     Args:
-        encoding: The encoding string from torchaudio.info()
-        file_ext: The file extension as a fallback
+        subtype: The subtype string from soundfile
 
     Returns:
-        Format name as a string
+        Bits per sample, or 0 if unknown
     """
-    # Direct mapping for formats that match exactly
-    encoding_map = {
-        "FLAC": "flac",
-        "MP3": "mp3",
-        "VORBIS": "ogg",
-        "AMR_WB": "amr",
-        "AMR_NB": "amr",
-        "OPUS": "opus",
-        "GSM": "gsm",
+    if not subtype:
+        return 0
+
+    # Common PCM and floating-point subtypes
+    pcm_bits = {
+        "PCM_16": 16,
+        "PCM_24": 24,
+        "PCM_32": 32,
+        "PCM_S8": 8,
+        "PCM_U8": 8,
+        "FLOAT": 32,
+        "DOUBLE": 64,
     }
 
-    if encoding in encoding_map:
-        return encoding_map[encoding]
+    if subtype in pcm_bits:
+        return pcm_bits[subtype]
 
-    # For PCM variants, use file extension to determine format
-    if encoding.startswith("PCM_"):
-        # Common PCM formats by extension
-        pcm_formats = {
-            "wav": "wav",
-            "aiff": "aiff",
-            "au": "au",
-            "raw": "raw",
-        }
-        return pcm_formats.get(file_ext, "wav")  # Default to wav for PCM
+    # Handle variants such as PCM_S16LE, PCM_F32LE, etc.
+    match = re.search(r"PCM_(?:[A-Z]*?)(\d+)", subtype)
+    if match:
+        return int(match.group(1))
 
-    # Fallback to file extension if encoding is unknown
-    return file_ext if file_ext else "unknown"
+    return 0
 
 
 def audio_to_np(
@@ -114,27 +117,27 @@ def audio_to_np(
 
     try:
         with audio.open() as f:
-            info = torchaudio.info(f)
-            sample_rate = info.sample_rate
+            info = sf.info(f)
+            sample_rate = info.samplerate
 
             frame_offset = int(start * sample_rate)
             num_frames = int(duration * sample_rate) if duration is not None else -1
 
             # Reset file pointer to the beginning
-            # This is important to ensure we read from the correct position later
             f.seek(0)
 
-            waveform, sr = torchaudio.load(
-                f, frame_offset=frame_offset, num_frames=num_frames
+            # Read audio data with offset and frame count
+            audio_np, sr = sf.read(
+                f,
+                start=frame_offset,
+                frames=num_frames,
+                always_2d=False,
+                dtype="float32",
             )
 
-            audio_np = waveform.numpy()
-
-            if audio_np.shape[0] > 1:
-                audio_np = audio_np.T
-            else:
-                audio_np = audio_np.squeeze()
-
+            # soundfile returns shape (frames,) for mono or
+            # (frames, channels) for multi-channel
+            # We keep this format as it matches expected output
             return audio_np, int(sr)
     except Exception as exc:
         raise FileError(
@@ -152,11 +155,9 @@ def audio_to_bytes(
 
     If duration is None, converts from start to end of file.
     If start is 0 and duration is None, converts entire file."""
-    y, sr = audio_to_np(audio, start, duration)
-
     import io
 
-    import soundfile as sf
+    y, sr = audio_to_np(audio, start, duration)
 
     buffer = io.BytesIO()
     sf.write(buffer, y, sr, format=format)

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/lib/dc/datachain.py

@@ -856,7 +856,9 @@ class DataChain:
                 udf_obj.to_udf_wrapper(self._settings.batch_size),
                 **self._settings.to_dict(),
             ),
-            signal_schema=self.signals_schema | udf_obj.output,
+            signal_schema=SignalSchema({"sys": Sys})
+            | self.signals_schema
+            | udf_obj.output,
         )
 
     def gen(
@@ -894,7 +896,7 @@ class DataChain:
                 udf_obj.to_udf_wrapper(self._settings.batch_size),
                 **self._settings.to_dict(),
             ),
-            signal_schema=udf_obj.output,
+            signal_schema=SignalSchema({"sys": Sys}) | udf_obj.output,
         )
 
     @delta_disabled
@@ -1031,7 +1033,7 @@ class DataChain:
                 partition_by=processed_partition_by,
                 **self._settings.to_dict(),
             ),
-            signal_schema=udf_obj.output,
+            signal_schema=SignalSchema({"sys": Sys}) | udf_obj.output,
         )
 
     def batch_map(
@@ -1097,11 +1099,7 @@ class DataChain:
         sign = UdfSignature.parse(name, signal_map, func, params, output, is_generator)
         DataModel.register(list(sign.output_schema.values.values()))
 
-        signals_schema = self.signals_schema
-        if self._sys:
-            signals_schema = SignalSchema({"sys": Sys}) | signals_schema
-
-        params_schema = signals_schema.slice(
+        params_schema = self.signals_schema.slice(
             sign.params, self._setup, is_batch=is_batch
         )
 
@@ -1156,11 +1154,9 @@ class DataChain:
             )
         )
 
-    def select(self, *args: str, _sys: bool = True) -> "Self":
+    def select(self, *args: str) -> "Self":
         """Select only a specified set of signals."""
         new_schema = self.signals_schema.resolve(*args)
-        if self._sys and _sys:
-            new_schema = SignalSchema({"sys": Sys}) | new_schema
         columns = new_schema.db_signals()
         return self._evolve(
             query=self._query.select(*columns), signal_schema=new_schema
@@ -1710,9 +1706,11 @@ class DataChain:
 
         signals_schema = self.signals_schema.clone_without_sys_signals()
         right_signals_schema = right_ds.signals_schema.clone_without_sys_signals()
-        ds.signals_schema = SignalSchema({"sys": Sys}) | signals_schema.merge(
-            right_signals_schema, rname
-        )
+
+        ds.signals_schema = signals_schema.merge(right_signals_schema, rname)
+
+        if not full:
+            ds.signals_schema = SignalSchema({"sys": Sys}) | ds.signals_schema
 
         return ds
 
@@ -1723,6 +1721,7 @@ class DataChain:
         Parameters:
             other: chain whose rows will be added to `self`.
         """
+        self.signals_schema = self.signals_schema.clone_without_sys_signals()
         return self._evolve(query=self._query.union(other._query))
 
     def subtract(  # type: ignore[override]

{datachain-0.36.0 → datachain-0.36.1}/src/datachain/query/dataset.py

@@ -438,9 +438,6 @@ class UDFStep(Step, ABC):
         """
 
     def populate_udf_table(self, udf_table: "Table", query: Select) -> None:
-        if "sys__id" not in query.selected_columns:
-            raise RuntimeError("Query must have sys__id column to run UDF")
-
         if (rows_total := self.catalog.warehouse.query_count(query)) == 0:
             return
 
@@ -634,12 +631,11 @@ class UDFStep(Step, ABC):
 
         # Apply partitioning if needed.
         if self.partition_by is not None:
-            if "sys__id" not in query.selected_columns:
-                _query = query = self.catalog.warehouse._regenerate_system_columns(
-                    query,
-                    keep_existing_columns=True,
-                )
-
+            _query = query = self.catalog.warehouse._regenerate_system_columns(
+                query_generator.select(),
+                keep_existing_columns=True,
+                regenerate_columns=["sys__id"],
+            )
             partition_tbl = self.create_partitions_table(query)
             temp_tables.append(partition_tbl.name)
             query = query.outerjoin(
@@ -960,28 +956,23 @@ class SQLUnion(Step):
         q2 = self.query2.apply_steps().select().subquery()
         temp_tables.extend(self.query2.temp_table_names)
 
-        columns1, columns2 = _order_columns(q1.columns, q2.columns)
-
-        union_select = sqlalchemy.select(*columns1).union_all(
-            sqlalchemy.select(*columns2)
-        )
-        union_cte = union_select.cte()
-        regenerated = self.query1.catalog.warehouse._regenerate_system_columns(
-            union_cte
-        )
-        result_columns = tuple(regenerated.selected_columns)
+        columns1 = _drop_system_columns(q1.columns)
+        columns2 = _drop_system_columns(q2.columns)
+        columns1, columns2 = _order_columns(columns1, columns2)
 
         def q(*columns):
-            if not columns:
-                return regenerated
+            selected_names = [c.name for c in columns]
+            col1 = [c for c in columns1 if c.name in selected_names]
+            col2 = [c for c in columns2 if c.name in selected_names]
+            union_query = sqlalchemy.select(*col1).union_all(sqlalchemy.select(*col2))
 
-            names = {c.name for c in columns}
-            selected = [c for c in result_columns if c.name in names]
-            return regenerated.with_only_columns(*selected)
+            union_cte = union_query.cte()
+            select_cols = [union_cte.c[name] for name in selected_names]
+            return sqlalchemy.select(*select_cols)
 
         return step_result(
             q,
-            result_columns,
+            columns1,
             dependencies=self.query1.dependencies | self.query2.dependencies,
         )
 
@@ -1070,7 +1061,7 @@ class SQLJoin(Step):
         q1 = self.get_query(self.query1, temp_tables)
         q2 = self.get_query(self.query2, temp_tables)
 
-        q1_columns = list(q1.c)
+        q1_columns = _drop_system_columns(q1.c) if self.full else list(q1.c)
         q1_column_names = {c.name for c in q1_columns}
 
         q2_columns = []
@@ -1211,6 +1202,10 @@ def _order_columns(
     return [[d[n] for n in column_order] for d in column_dicts]
 
 
+def _drop_system_columns(columns: Iterable[ColumnElement]) -> list[ColumnElement]:
+    return [c for c in columns if not c.name.startswith("sys__")]
+
+
 @attrs.define
 class ResultIter:
     _row_iter: Iterable[Any]