pixeltable 0.2.24__py3-none-any.whl → 0.3.0__py3-none-any.whl

pixeltable/functions/huggingface.py

@@ -144,9 +144,9 @@ def cross_encoder_list(sentence1: str, sentences2: list, *, model_id: str) -> li
 
 
 @pxt.udf(batch_size=32)
-def clip_text(text: Batch[str], *, model_id: str) -> Batch[pxt.Array[(None,), pxt.Float]]:
+def clip(text: Batch[str], *, model_id: str) -> Batch[pxt.Array[(None,), pxt.Float]]:
     """
-    Computes a CLIP embedding for the specified text. `model_id` should be a reference to a pretrained
+    Computes a CLIP embedding for the specified text or image. `model_id` should be a reference to a pretrained
     [CLIP Model](https://huggingface.co/docs/transformers/model_doc/clip).
 
     __Requirements:__
@@ -164,7 +164,11 @@ def clip_text(text: Batch[str], *, model_id: str) -> Batch[pxt.Array[(None,), px
         Add a computed column that applies the model `openai/clip-vit-base-patch32` to an existing
         Pixeltable column `tbl.text` of the table `tbl`:
 
-        >>> tbl['result'] = clip_text(tbl.text, model_id='openai/clip-vit-base-patch32')
+        >>> tbl.add_computed_column(
+        ...     result=clip(tbl.text, model_id='openai/clip-vit-base-patch32')
+        ... )
+
+        The same would work with an image column `tbl.image` in place of `tbl.text`.
     """
     env.Env.get().require_package('transformers')
     device = resolve_torch_device('auto')
@@ -181,29 +185,8 @@ def clip_text(text: Batch[str], *, model_id: str) -> Batch[pxt.Array[(None,), px
     return [embeddings[i] for i in range(embeddings.shape[0])]
 
 
-@pxt.udf(batch_size=32)
-def clip_image(image: Batch[PIL.Image.Image], *, model_id: str) -> Batch[pxt.Array[(None,), pxt.Float]]:
-    """
-    Computes a CLIP embedding for the specified image. `model_id` should be a reference to a pretrained
-    [CLIP Model](https://huggingface.co/docs/transformers/model_doc/clip).
-
-    __Requirements:__
-
-    - `pip install torch transformers`
-
-    Args:
-        image: The image to embed.
-        model_id: The pretrained model to use for the embedding.
-
-    Returns:
-        An array containing the output of the embedding model.
-
-    Examples:
-        Add a computed column that applies the model `openai/clip-vit-base-patch32` to an existing
-        Pixeltable column `image` of the table `tbl`:
-
-        >>> tbl['result'] = clip_image(tbl.image, model_id='openai/clip-vit-base-patch32')
-    """
+@clip.overload
+def _(image: Batch[PIL.Image.Image], *, model_id: str) -> Batch[pxt.Array[(None,), pxt.Float]]:
     env.Env.get().require_package('transformers')
     device = resolve_torch_device('auto')
     import torch
@@ -219,8 +202,7 @@ def clip_image(image: Batch[PIL.Image.Image], *, model_id: str) -> Batch[pxt.Arr
     return [embeddings[i] for i in range(embeddings.shape[0])]
 
 
-@clip_text.conditional_return_type
-@clip_image.conditional_return_type
+@clip.conditional_return_type
 def _(model_id: str) -> pxt.ArrayType:
     try:
         from transformers import CLIPModel

pixeltable/functions/json.py

@@ -16,20 +16,15 @@ import pixeltable as pxt
 from pixeltable.utils.code import local_public_names
 
 
-@pxt.uda(
-    update_types=[pxt.JsonType(nullable=True)],
-    value_type=pxt.JsonType(),
-    requires_order_by=False,
-    allows_window=False,
-)
+@pxt.uda
 class make_list(pxt.Aggregator):
     """
     Collects arguments into a list.
     """
-    def __init__(self):
+    def __init__(self) -> None:
         self.output: list[Any] = []
 
-    def update(self, obj: Any) -> None:
+    def update(self, obj: pxt.Json) -> None:
         if obj is None:
             return
         self.output.append(obj)

pixeltable/functions/math.py

@@ -0,0 +1,67 @@
+import builtins
+import math
+from typing import Optional
+
+import sqlalchemy as sql
+
+import pixeltable as pxt
+from pixeltable.utils.code import local_public_names
+
+
+@pxt.udf(is_method=True)
+def abs(self: float) -> float:
+    return builtins.abs(self)
+
+
+@abs.to_sql
+def _(self: sql.ColumnElement) -> sql.ColumnElement:
+    return sql.func.abs(self)
+
+
+@pxt.udf(is_method=True)
+def ceil(self: float) -> float:
+    # This ensures the same behavior as SQL
+    if math.isfinite(self):
+        return float(math.ceil(self))
+    else:
+        return self
+
+
+@ceil.to_sql
+def _(self: sql.ColumnElement) -> sql.ColumnElement:
+    return sql.func.ceiling(self)
+
+
+@pxt.udf(is_method=True)
+def floor(self: float) -> float:
+    # This ensures the same behavior as SQL
+    if math.isfinite(self):
+        return float(math.floor(self))
+    else:
+        return self
+
+
+@floor.to_sql
+def _(self: sql.ColumnElement) -> sql.ColumnElement:
+    return sql.func.floor(self)
+
+
+@pxt.udf(is_method=True)
+def round(self: float, digits: Optional[int] = None) -> float:
+    # Set digits explicitly to 0 to guarantee a return type of float; this ensures the same behavior as SQL
+    return builtins.round(self, digits or 0)
+
+
+@round.to_sql
+def _(self: sql.ColumnElement, digits: Optional[sql.ColumnElement] = None) -> sql.ColumnElement:
+    if digits is None:
+        return sql.func.round(self)
+    else:
+        return sql.func.round(sql.cast(self, sql.Numeric), sql.cast(digits, sql.Integer))
+
+
+__all__ = local_public_names(__name__)
+
+
+def __dir__():
+    return __all__

pixeltable/functions/mistralai.py

@@ -36,7 +36,6 @@ def chat_completions(
     temperature: Optional[float] = 0.7,
     top_p: Optional[float] = 1.0,
     max_tokens: Optional[int] = None,
-    min_tokens: Optional[int] = None,
     stop: Optional[list[str]] = None,
     random_seed: Optional[int] = None,
     response_format: Optional[dict] = None,
@@ -75,7 +74,6 @@ def chat_completions(
         temperature=temperature,
         top_p=top_p,
         max_tokens=_opt(max_tokens),
-        min_tokens=_opt(min_tokens),
         stop=stop,
         random_seed=_opt(random_seed),
         response_format=response_format,  # type: ignore[arg-type]

pixeltable/functions/ollama.py

@@ -34,7 +34,7 @@ def generate(
     template: str = '',
     context: Optional[list[int]] = None,
     raw: bool = False,
-    format: str = '',
+    format: Optional[str] = None,
     options: Optional[dict] = None,
 ) -> dict:
     """
@@ -44,7 +44,7 @@ def generate(
         prompt: The prompt to generate a response for.
         model: The model name.
         suffix: The text after the model response.
-        format: The format of the response; must be one of `'json'` or `''` (the empty string).
+        format: The format of the response; must be one of `'json'` or `None`.
         system: System message.
         template: Prompt template to use.
         context: The context parameter returned from a previous call to `generate()`.
@@ -68,7 +68,7 @@ def generate(
         raw=raw,
         format=format,
         options=options,
-    )  # type: ignore[call-overload]
+    ).dict()  # type: ignore[call-overload]
 
 
 @pxt.udf
@@ -77,7 +77,7 @@ def chat(
     *,
     model: str,
     tools: Optional[list[dict]] = None,
-    format: str = '',
+    format: Optional[str] = None,
     options: Optional[dict] = None,
 ) -> dict:
     """
@@ -87,7 +87,7 @@ def chat(
         messages: The messages of the chat.
         model: The model name.
         tools: Tools for the model to use.
-        format: The format of the response; must be one of `'json'` or `''` (the empty string).
+        format: The format of the response; must be one of `'json'` or `None`.
         options: Additional options to pass to the `chat` call, such as `max_tokens`, `temperature`, `top_p`, and `top_k`.
             For details, see the
             [Valid Parameters and Values](https://github.com/ollama/ollama/blob/main/docs/modelfile.md#valid-parameters-and-values)
@@ -103,7 +103,7 @@ def chat(
         tools=tools,
         format=format,
         options=options,
-    )  # type: ignore[call-overload]
+    ).dict()  # type: ignore[call-overload]
 
 
 @pxt.udf(batch_size=16)
@@ -135,8 +135,8 @@ def embed(
         model=model,
         input=input,
         truncate=truncate,
-        options=options,  # type: ignore[arg-type]
-    )
+        options=options,
+    ).dict()
     return [np.array(data, dtype=np.float64) for data in results['embeddings']]
 
 

pixeltable/functions/openai.py

@@ -7,17 +7,18 @@ the [Working with OpenAI](https://pixeltable.readme.io/docs/working-with-openai)
 
 import base64
 import io
+import json
 import pathlib
 import uuid
-from typing import TYPE_CHECKING, Callable, Optional, TypeVar, Union
+from typing import TYPE_CHECKING, Any, Callable, Optional, TypeVar, Union
 
 import numpy as np
 import PIL.Image
 import tenacity
 
 import pixeltable as pxt
-from pixeltable import env
-from pixeltable.func import Batch
+from pixeltable import env, exprs
+from pixeltable.func import Batch, Tools
 from pixeltable.utils.code import local_public_names
 
 if TYPE_CHECKING:
@@ -225,6 +226,33 @@ def chat_completions(
             ]
             tbl['response'] = chat_completions(messages, model='gpt-4o-mini')
     """
+
+    if tools is not None:
+        tools = [
+            {
+                'type': 'function',
+                'function': tool
+            }
+            for tool in tools
+        ]
+
+    tool_choice_: Union[str, dict, None] = None
+    if tool_choice is not None:
+        if tool_choice['auto']:
+            tool_choice_ = 'auto'
+        elif tool_choice['required']:
+            tool_choice_ = 'required'
+        else:
+            assert tool_choice['tool'] is not None
+            tool_choice_ = {
+                'type': 'function',
+                'function': {'name': tool_choice['tool']}
+            }
+
+    extra_body: Optional[dict[str, Any]] = None
+    if tool_choice is not None and not tool_choice['parallel_tool_calls']:
+        extra_body = {'parallel_tool_calls': False}
+
     result = _retry(_openai_client().chat.completions.create)(
         messages=messages,
         model=model,
@@ -241,8 +269,9 @@ def chat_completions(
         temperature=_opt(temperature),
         top_p=_opt(top_p),
         tools=_opt(tools),
-        tool_choice=_opt(tool_choice),
+        tool_choice=_opt(tool_choice_),
         user=_opt(user),
+        extra_body=extra_body,
     )
     return result.dict()
 
@@ -453,6 +482,24 @@ def moderations(input: str, *, model: Optional[str] = None) -> dict:
     return result.dict()
 
 
+def invoke_tools(tools: Tools, response: exprs.Expr) -> exprs.InlineDict:
+    """Converts an OpenAI response dict to Pixeltable tool invocation format and calls `tools._invoke()`."""
+    return tools._invoke(_openai_response_to_pxt_tool_calls(response))
+
+
+@pxt.udf
+def _openai_response_to_pxt_tool_calls(response: dict) -> Optional[dict]:
+    openai_tool_calls = response['choices'][0]['message']['tool_calls']
+    if openai_tool_calls is not None:
+        return {
+            tool_call['function']['name']: {
+                'args': json.loads(tool_call['function']['arguments'])
+            }
+            for tool_call in openai_tool_calls
+        }
+    return None
+
+
 _T = TypeVar('_T')
 
 

pixeltable/functions/timestamp.py

@@ -232,7 +232,7 @@ def _(
         sql.cast(day, sql.Integer),
         sql.cast(hour, sql.Integer),
         sql.cast(minute, sql.Integer),
-        sql.cast(second + microsecond / 1000000.0, sql.Double))
+        sql.cast(second + microsecond / 1000000.0, sql.Float))
 
 # @pxt.udf
 # def date(self: datetime) -> datetime:

pixeltable/functions/video.py

@@ -47,13 +47,7 @@ _format_defaults = {  # format -> (codec, ext)
 #         output_container.mux(packet)
 
 
-@pxt.uda(
-    init_types=[pxt.IntType()],
-    update_types=[pxt.ImageType()],
-    value_type=pxt.VideoType(),
-    requires_order_by=True,
-    allows_window=False,
-)
+@pxt.uda(requires_order_by=True)
 class make_video(pxt.Aggregator):
     """
     Aggregator that creates a video from a sequence of images.
@@ -80,7 +74,7 @@ class make_video(pxt.Aggregator):
         for packet in self.stream.encode(av_frame):
             self.container.mux(packet)
 
-    def value(self) -> str:
+    def value(self) -> pxt.Video:
         for packet in self.stream.encode():
             self.container.mux(packet)
         self.container.close()
@@ -132,7 +126,7 @@ def _get_metadata(path: str) -> dict:
         assert isinstance(container, av.container.InputContainer)
         streams_info = [__get_stream_metadata(stream) for stream in container.streams]
         result = {
-            'bit_exact': container.bit_exact,
+            'bit_exact': getattr(container, 'bit_exact', False),
             'bit_rate': container.bit_rate,
             'size': container.size,
             'metadata': container.metadata,

pixeltable/functions/vision.py

@@ -220,7 +220,7 @@ def eval_detections(
     return result
 
 
-@pxt.uda(update_types=[pxt.JsonType()], value_type=pxt.JsonType(), allows_std_agg=True, allows_window=False)
+@pxt.uda
 class mean_ap(pxt.Aggregator):
     """
     Calculates the mean average precision (mAP) over