pixeltable 0.1.2__py3-none-any.whl → 0.2.0__py3-none-any.whl

pixeltable/type_system.py

@@ -1,15 +1,21 @@
+from __future__ import annotations
+
 import abc
-from typing import Any, Optional, Tuple, Dict, Callable, List, Union
-import enum
 import datetime
+import enum
 import json
+import typing
+import urllib.parse
+from copy import copy
+from pathlib import Path
+from typing import Any, Optional, Tuple, Dict, Callable, List, Union
 
-import os
-os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
-#import tensorflow as tf
 import PIL.Image
+import av
+import numpy as np
 import sqlalchemy as sql
 
+from pixeltable import exceptions as excs
 
 
 class ColumnType:
@@ -24,26 +30,16 @@ class ColumnType:
         ARRAY = 6
         IMAGE = 7
         VIDEO = 8
+        AUDIO = 9
+        DOCUMENT = 10
 
         # exprs that don't evaluate to a computable value in Pixeltable, such as an Image member function
-        INVALID = 9
-
-        def to_tf(self) -> 'tf.dtypes.DType':
-            import tensorflow as tf
-            if self == self.STRING:
-                return tf.string
-            if self == self.INT:
-                return tf.int64
-            if self == self.FLOAT:
-                return tf.float32
-            if self == self.BOOL:
-                return tf.bool
-            raise TypeError(f'Cannot convert {self} to TensorFlow')
+        INVALID = 255
 
         @classmethod
         def supertype(
                 cls, type1: 'Type', type2: 'Type',
-                # we need to pass this in because we can't easily add it as a class member
+                # we need to pass this in because we can't easily append it as a class member
                 common_supertypes: Dict[Tuple['Type', 'Type'], 'Type']
         ) -> Optional['Type']:
             if type1 == type2:
@@ -83,8 +79,9 @@ class ColumnType:
         (Type.INT, Type.FLOAT): Type.FLOAT,
     }
 
-    def __init__(self, t: Type):
+    def __init__(self, t: Type, nullable: bool = False):
         self._type = t
+        self.nullable = nullable
 
     @property
     def type_enum(self) -> Type:
@@ -94,7 +91,7 @@ class ColumnType:
         return json.dumps(self.as_dict())
 
     @classmethod
-    def serialize_list(cls, type_list: List['ColumnType']) -> str:
+    def serialize_list(cls, type_list: List[ColumnType]) -> str:
         return json.dumps([t.as_dict() for t in type_list])
 
     def as_dict(self) -> Dict:
@@ -104,33 +101,34 @@ class ColumnType:
         }
 
     def _as_dict(self) -> Dict:
-        return {}
+        return {'nullable': self.nullable}
 
     @classmethod
-    def deserialize(cls, type_str: str) -> 'ColumnType':
+    def deserialize(cls, type_str: str) -> ColumnType:
         type_dict = json.loads(type_str)
         return cls.from_dict(type_dict)
 
     @classmethod
-    def deserialize_list(cls, type_list_str: str) -> List['ColumnType']:
+    def deserialize_list(cls, type_list_str: str) -> List[ColumnType]:
         type_dict_list = json.loads(type_list_str)
         return [cls.from_dict(type_dict) for type_dict in type_dict_list]
 
     @classmethod
-    def from_dict(cls, type_dict: Dict) -> 'ColumnType':
+    def from_dict(cls, type_dict: Dict) -> ColumnType:
         assert '_classname' in type_dict
         type_class = globals()[type_dict['_classname']]
         return type_class._from_dict(type_dict)
 
     @classmethod
-    def _from_dict(cls, d: Dict) -> 'ColumnType':
+    def _from_dict(cls, d: Dict) -> ColumnType:
         """
-        Default implementation: simply invoke c'tor without arguments
+        Default implementation: simply invoke c'tor
         """
-        return cls()
+        assert 'nullable' in d
+        return cls(nullable=d['nullable'])
 
     @classmethod
-    def make_type(cls, t: Type) -> 'ColumnType':
+    def make_type(cls, t: Type) -> ColumnType:
         assert t != cls.Type.INVALID and t != cls.Type.ARRAY
         if t == cls.Type.STRING:
             return StringType()
@@ -148,21 +146,44 @@ class ColumnType:
             return ImageType()
         if t == cls.Type.VIDEO:
             return VideoType()
+        if t == cls.Type.AUDIO:
+            return AudioType()
+        if t == cls.Type.DOCUMENT:
+            return AudioType()
 
     def __str__(self) -> str:
         return self._type.name.lower()
 
     def __eq__(self, other: object) -> bool:
+        return self.matches(other) and self.nullable == other.nullable
+
+    def is_supertype_of(self, other: ColumnType) -> bool:
+        if type(self) != type(other):
+            return False
+        if self.matches(other):
+            return True
+        return self._is_supertype_of(other)
+
+    @abc.abstractmethod
+    def _is_supertype_of(self, other: ColumnType) -> bool:
+        return False
+
+    def matches(self, other: object) -> bool:
+        """Two types match if they're equal, aside from nullability"""
+        if not isinstance(other, ColumnType):
+            pass
         assert isinstance(other, ColumnType)
-        if False and type(self) != type(other):
+        if type(self) != type(other):
             return False
         for member_var in vars(self).keys():
+            if member_var == 'nullable':
+                continue
             if getattr(self, member_var) != getattr(other, member_var):
                 return False
         return True
 
     @classmethod
-    def supertype(cls, type1: 'ColumnType', type2: 'ColumnType') -> Optional['ColumnType']:
+    def supertype(cls, type1: ColumnType, type2: ColumnType) -> Optional[ColumnType]:
         if type1 == type2:
             return type1
 
@@ -184,16 +205,15 @@ class ColumnType:
 
     @classmethod
     @abc.abstractmethod
-    def _supertype(cls, type1: 'ColumnType', type2: 'ColumnType') -> Optional['ColumnType']:
+    def _supertype(cls, type1: ColumnType, type2: ColumnType) -> Optional[ColumnType]:
         """
         Class-specific implementation of determining the supertype. type1 and type2 are from the same subclass of
         ColumnType.
         """
         pass
 
-
     @classmethod
-    def get_value_type(cls, val: Any) -> 'ColumnType':
+    def infer_literal_type(cls, val: Any) -> Optional[ColumnType]:
         if isinstance(val, str):
             return StringType()
         if isinstance(val, int):
@@ -204,6 +224,85 @@ class ColumnType:
             return BoolType()
         if isinstance(val, datetime.datetime) or isinstance(val, datetime.date):
             return TimestampType()
+        if isinstance(val, np.ndarray):
+            col_type = ArrayType.from_literal(val)
+            if col_type is not None:
+                return col_type
+            # this could still be json-serializable
+        if isinstance(val, dict) or isinstance(val, np.ndarray):
+            try:
+                JsonType().validate_literal(val)
+                return JsonType()
+            except TypeError:
+                return None
+        return None
+
+
+    @classmethod
+    def from_python_type(cls, t: type) -> Optional[ColumnType]:
+        if t in _python_type_to_column_type:
+            return _python_type_to_column_type[t]
+        elif isinstance(t, typing._UnionGenericAlias) and t.__args__[1] is type(None):
+            # `t` is a type of the form Optional[T] (equivalently, Union[T, None]).
+            # We treat it as the underlying type but with nullable=True.
+            if t.__args__[0] in _python_type_to_column_type:
+                underlying = copy(_python_type_to_column_type[t.__args__[0]])
+                underlying.nullable = True
+                return underlying
+
+        return None
+
+
+    def validate_literal(self, val: Any) -> None:
+        """Raise TypeError if val is not a valid literal for this type"""
+        if val is None:
+            if not self.nullable:
+                raise TypeError('Expected non-None value')
+            else:
+                return
+        self._validate_literal(val)
+
+    def validate_media(self, val: Any) -> None:
+        """
+        Raise TypeError if val is not a path to a valid media file (or a valid in-memory byte sequence) for this type
+        """
+        if self.is_media_type():
+            raise NotImplementedError(f'validate_media() not implemented for {self.__class__.__name__}')
+
+    def _validate_file_path(self, val: Any) -> None:
+        """Raises TypeError if not a valid local file path or not a path/byte sequence"""
+        if isinstance(val, str):
+            parsed = urllib.parse.urlparse(val)
+            if parsed.scheme != '' and parsed.scheme != 'file':
+                return
+            path = Path(urllib.parse.unquote(parsed.path))
+            if not path.is_file():
+                raise TypeError(f'File not found: {str(path)}')
+        else:
+            if not isinstance(val, bytes):
+                raise TypeError(f'expected file path or bytes, got {type(val)}')
+
+    @abc.abstractmethod
+    def _validate_literal(self, val: Any) -> None:
+        """Raise TypeError if val is not a valid literal for this type"""
+        pass
+
+    @abc.abstractmethod
+    def _create_literal(self, val : Any) -> Any:
+        """Create a literal of this type from val, including any needed conversions.
+             val is guaranteed to be non-None"""
+        return val
+
+    def create_literal(self, val: Any) -> Any:
+        """Create a literal of this type from val or raise TypeError if not possible"""
+        if val is not None:
+            val = self._create_literal(val)
+
+        self.validate_literal(val)
+        return val
+
+    def print_value(self, val: Any) -> str:
+        return str(val)
 
     def is_scalar_type(self) -> bool:
         return self._type in self.scalar_types
@@ -241,6 +340,16 @@ class ColumnType:
     def is_video_type(self) -> bool:
         return self._type == self.Type.VIDEO
 
+    def is_audio_type(self) -> bool:
+        return self._type == self.Type.AUDIO
+
+    def is_document_type(self) -> bool:
+        return self._type == self.Type.DOCUMENT
+
+    def is_media_type(self) -> bool:
+        # types that refer to external media files
+        return self.is_image_type() or self.is_video_type() or self.is_audio_type() or self.is_document_type()
+
     @abc.abstractmethod
     def to_sql(self) -> str:
         """
@@ -274,6 +383,10 @@ class ColumnType:
             return sql.VARBINARY
         assert False
 
+    @abc.abstractmethod
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        assert False, f'Have not implemented {self.__class__.__name__} to Arrow'
+ 
     @staticmethod
     def no_conversion(v: Any) -> Any:
         """
@@ -282,21 +395,17 @@ class ColumnType:
         """
         assert False
 
-    def conversion_fn(self, target: 'ColumnType') -> Optional[Callable[[Any], Any]]:
+    def conversion_fn(self, target: ColumnType) -> Optional[Callable[[Any], Any]]:
         """
         Return Callable that converts a column value of type self to a value of type 'target'.
         Returns None if conversion isn't possible.
         """
         return None
 
-    @abc.abstractmethod
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        pass
-
 
 class InvalidType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.INVALID)
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.INVALID, nullable=nullable)
 
     def to_sql(self) -> str:
         assert False
@@ -304,13 +413,18 @@ class InvalidType(ColumnType):
     def to_sa_type(self) -> Any:
         assert False
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        raise TypeError(f'Invalid type cannot be converted to Tensorflow')
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        assert False
 
+    def print_value(self, val: Any) -> str:
+        assert False
+
+    def _validate_literal(self, val: Any) -> None:
+        assert False
 
 class StringType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.STRING)
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.STRING, nullable=nullable)
 
     def conversion_fn(self, target: ColumnType) -> Optional[Callable[[Any], Any]]:
         if not target.is_timestamp_type():
@@ -328,78 +442,111 @@ class StringType(ColumnType):
 
     def to_sa_type(self) -> str:
         return sql.String
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.string()
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        import tensorflow as tf
-        return tf.TensorSpec(shape=(), dtype=tf.string)
+    def print_value(self, val: Any) -> str:
+        return f"'{val}'"
+
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, str):
+            raise TypeError(f'Expected string, got {val.__class__.__name__}')
 
 
 class IntType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.INT)
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.INT, nullable=nullable)
 
     def to_sql(self) -> str:
-        return 'INTEGER'
+        return 'BIGINT'
 
     def to_sa_type(self) -> str:
-        return sql.Integer
+        return sql.BigInteger
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.int64() # to be consistent with bigint above
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        # TODO: how to specify the correct int subtype?
-        import tensorflow as tf
-        return tf.TensorSpec(shape=(), dtype=tf.int64)
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, int):
+            raise TypeError(f'Expected int, got {val.__class__.__name__}')
 
 
 class FloatType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.FLOAT)
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.FLOAT, nullable=nullable)
 
     def to_sql(self) -> str:
         return 'FLOAT'
 
     def to_sa_type(self) -> str:
         return sql.Float
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa
+        return pa.float32()
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        import tensorflow as tf
-        # TODO: how to specify the correct float subtype?
-        return tf.TensorSpec(shape=(), dtype=tf.float32)
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, float):
+            raise TypeError(f'Expected float, got {val.__class__.__name__}')
 
+    def _create_literal(self, val: Any) -> Any:
+        if isinstance(val, int):
+            return float(val)
+        return val
 
 class BoolType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.BOOL)
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.BOOL, nullable=nullable)
 
     def to_sql(self) -> str:
         return 'BOOLEAN'
 
     def to_sa_type(self) -> str:
         return sql.Boolean
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.bool_()
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        import tensorflow as tf
-        # TODO: how to specify the correct int subtype?
-        return tf.TensorSpec(shape=(), dtype=tf.bool)
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, bool):
+            raise TypeError(f'Expected bool, got {val.__class__.__name__}')
 
+    def _create_literal(self, val: Any) -> Any:
+        if isinstance(val, int):
+            return bool(val)
+        return val
 
 class TimestampType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.TIMESTAMP)
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.TIMESTAMP, nullable=nullable)
 
     def to_sql(self) -> str:
         return 'INTEGER'
 
     def to_sa_type(self) -> str:
         return sql.TIMESTAMP
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.timestamp('us') # postgres timestamp is microseconds
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        raise TypeError(f'Timestamp type cannot be converted to Tensorflow')
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, datetime.datetime) and not isinstance(val, datetime.date):
+            raise TypeError(f'Expected datetime.datetime or datetime.date, got {val.__class__.__name__}')
 
+    def _create_literal(self, val: Any) -> Any:
+        if isinstance(val, str):
+            return datetime.datetime.fromisoformat(val)
+        return val
 
 class JsonType(ColumnType):
     # TODO: type_spec also needs to be able to express lists
-    def __init__(self, type_spec: Optional[Dict[str, ColumnType]] = None):
-        super().__init__(self.Type.JSON)
+    def __init__(self, type_spec: Optional[Dict[str, ColumnType]] = None, nullable: bool = False):
+        super().__init__(self.Type.JSON, nullable=nullable)
         self.type_spec = type_spec
 
     def _as_dict(self) -> Dict:
@@ -410,34 +557,52 @@ class JsonType(ColumnType):
         return result
 
     @classmethod
-    def _from_dict(cls, d: Dict) -> 'ColumnType':
+    def _from_dict(cls, d: Dict) -> ColumnType:
         type_spec = None
         if 'type_spec' in d:
             type_spec = {
                 field_name: cls.deserialize(field_type_dict) for field_name, field_type_dict in d['type_spec'].items()
             }
-        return cls(type_spec)
+        return cls(type_spec, nullable=d['nullable'])
 
     def to_sql(self) -> str:
         return 'JSONB'
 
     def to_sa_type(self) -> str:
         return sql.dialects.postgresql.JSONB
-
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        if self.type_spec is None:
-            raise TypeError(f'Cannot convert {self.__class__.__name__} with missing type spec to TensorFlow')
-        return {k: v.to_tf() for k, v in self.type_spec.items()}
-
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.string() # TODO: weight advantage of pa.struct type.
+
+    def print_value(self, val: Any) -> str:
+        val_type = self.infer_literal_type(val)
+        if val_type == self:
+            return str(val)
+        return val_type.print_value(val)
+
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, dict) and not isinstance(val, list):
+            raise TypeError(f'Expected dict or list, got {val.__class__.__name__}')
+        try:
+            _ = json.dumps(val)
+        except TypeError as e:
+            raise TypeError(f'Expected JSON-serializable object, got {val}')
+
+    def _create_literal(self, val: Any) -> Any:
+        if isinstance(val, tuple):
+            val = list(val)
+        return val
 
 class ArrayType(ColumnType):
     def __init__(
-            self, shape: Tuple[Union[int, None], ...], dtype: ColumnType.Type):
-        super().__init__(self.Type.ARRAY)
+            self, shape: Tuple[Union[int, None], ...], dtype: ColumnType, nullable: bool = False):
+        super().__init__(self.Type.ARRAY, nullable=nullable)
         self.shape = shape
-        self.dtype = dtype
+        assert dtype.is_int_type() or dtype.is_float_type() or dtype.is_bool_type() or dtype.is_string_type()
+        self.dtype = dtype._type
 
-    def _supertype(cls, type1: 'ArrayType', type2: 'ArrayType') -> Optional['ArrayType']:
+    def _supertype(cls, type1: ArrayType, type2: ArrayType) -> Optional[ArrayType]:
         if len(type1.shape) != len(type2.shape):
             return None
         base_type = ColumnType.supertype(type1.dtype, type2.dtype)
@@ -452,54 +617,94 @@ class ArrayType(ColumnType):
         return result
 
     def __str__(self) -> str:
-        return f'{self.__class__.__name__}({self.shape}, dtype={self.dtype.name})'
+        return f'{self._type.name.lower()}({self.shape}, dtype={self.dtype.name})'
 
     @classmethod
-    def _from_dict(cls, d: Dict) -> 'ColumnType':
+    def _from_dict(cls, d: Dict) -> ColumnType:
         assert 'shape' in d
         assert 'dtype' in d
         shape = tuple(d['shape'])
-        dtype = cls.Type(d['dtype'])
-        return cls(shape, dtype)
-
-    def to_sql(self) -> str:
-        return 'BYTEA'
+        dtype = cls.make_type(cls.Type(d['dtype']))
+        return cls(shape, dtype, nullable=d['nullable'])
 
-    def to_sa_type(self) -> str:
-        return sql.VARBINARY
+    @classmethod
+    def from_literal(cls, val: np.ndarray) -> Optional[ArrayType]:
+        # determine our dtype
+        assert isinstance(val, np.ndarray)
+        if np.issubdtype(val.dtype, np.integer):
+            dtype = IntType()
+        elif np.issubdtype(val.dtype, np.floating):
+            dtype = FloatType()
+        elif val.dtype == np.bool_:
+            dtype = BoolType()
+        elif val.dtype == np.str_:
+            dtype = StringType()
+        else:
+            return None
+        return cls(val.shape, dtype=dtype, nullable=True)
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        import tensorflow as tf
-        return tf.TensorSpec(shape=self.shape, dtype=self.dtype.to_tf())
+    def is_valid_literal(self, val: np.ndarray) -> bool:
+        if not isinstance(val, np.ndarray):
+            return False
+        if len(val.shape) != len(self.shape):
+            return False
+        # check that the shapes are compatible
+        for n1, n2 in zip(val.shape, self.shape):
+            if n1 is None:
+                return False
+            if n2 is None:
+                # wildcard
+                continue
+            if n1 != n2:
+                return False
+        return val.dtype == self.numpy_dtype()
 
+    def _validate_literal(self, val: Any) -> None:
+        if not isinstance(val, np.ndarray):
+            raise TypeError(f'Expected numpy.ndarray, got {val.__class__.__name__}')
+        if not self.is_valid_literal(val):
+            raise TypeError((
+                f'Expected ndarray({self.shape}, dtype={self.numpy_dtype()}), '
+                f'got ndarray({val.shape}, dtype={val.dtype})'))
 
-class ImageType(ColumnType):
-    @enum.unique
-    class Mode(enum.Enum):
-        L = 0,
-        RGB = 1
+    def _create_literal(self, val: Any) -> Any:
+        if isinstance(val, (list,tuple)):
+            return np.array(val)
+        return val
 
-        @classmethod
-        def from_pil(cls, pil_mode: str) -> 'Mode':
-            if pil_mode == 'L':
-                return cls.L
-            if pil_mode == 'RGB':
-                return cls.RGB
+    def to_sql(self) -> str:
+        return 'BYTEA'
 
-        def to_pil(self) -> str:
-            return self.name
+    def to_sa_type(self) -> str:
+        return sql.LargeBinary
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        if any([n is None for n in self.shape]):
+            raise TypeError(f'Cannot convert array with unknown shape to Arrow')        
+        return pa.fixed_shape_tensor(pa.from_numpy_dtype(self.numpy_dtype()), self.shape)
+
+    def numpy_dtype(self) -> np.dtype:
+        if self.dtype == self.Type.INT:
+            return np.dtype(np.int64)
+        if self.dtype == self.Type.FLOAT:
+            return np.dtype(np.float32)
+        if self.dtype == self.Type.BOOL:
+            return np.dtype(np.bool_)
+        if self.dtype == self.Type.STRING:
+            return np.dtype(np.str_)
+        assert False
 
-        def num_channels(self) -> int:
-            return len(self.name)
 
+class ImageType(ColumnType):
     def __init__(
             self, width: Optional[int] = None, height: Optional[int] = None, size: Optional[Tuple[int, int]] = None,
-            mode: Optional[Mode] = None
+            mode: Optional[str] = None, nullable: bool = False
     ):
         """
         TODO: does it make sense to specify only width or height?
         """
-        super().__init__(self.Type.IMAGE)
+        super().__init__(self.Type.IMAGE, nullable=nullable)
         assert not(width is not None and size is not None)
         assert not(height is not None and size is not None)
         if size is not None:
@@ -510,22 +715,53 @@ class ImageType(ColumnType):
             self.height = height
         self.mode = mode
 
+    def __str__(self) -> str:
+        if self.width is not None or self.height is not None or self.mode is not None:
+            params_str = ''
+            if self.width is not None:
+                params_str = f'width={self.width}'
+            if self.height is not None:
+                if len(params_str) > 0:
+                    params_str += ', '
+                params_str += f'height={self.height}'
+            if self.mode is not None:
+                if len(params_str) > 0:
+                    params_str += ', '
+                params_str += f'mode={self.mode}'
+            params_str = f'({params_str})'
+        else:
+            params_str = ''
+        return f'{self._type.name.lower()}{params_str}'
+
+    def _is_supertype_of(self, other: ImageType) -> bool:
+        if self.mode != other.mode:
+            return False
+        if self.width is None and self.height is None:
+            return True
+        if self.width != other.width and self.height != other.height:
+            return False
+
+    @property
+    def size(self) -> Optional[Tuple[int, int]]:
+        if self.width is None or self.height is None:
+            return None
+        return (self.width, self.height)
+
     @property
     def num_channels(self) -> Optional[int]:
         return None if self.mode is None else self.mode.num_channels()
 
     def _as_dict(self) -> Dict:
         result = super()._as_dict()
-        result.update(width=self.width, height=self.height, mode=self.mode.value if self.mode is not None else None)
+        result.update(width=self.width, height=self.height, mode=self.mode)
         return result
 
     @classmethod
-    def _from_dict(cls, d: Dict) -> 'ColumnType':
+    def _from_dict(cls, d: Dict) -> ColumnType:
         assert 'width' in d
         assert 'height' in d
         assert 'mode' in d
-        mode_val = d['mode']
-        return cls(width=d['width'], height=d['height'], mode=cls.Mode(mode_val) if mode_val is not None else None)
+        return cls(width=d['width'], height=d['height'], mode=d['mode'], nullable=d['nullable'])
 
     def conversion_fn(self, target: ColumnType) -> Optional[Callable[[Any], Any]]:
         if not target.is_image_type():
@@ -552,23 +788,111 @@ class ImageType(ColumnType):
 
     def to_sa_type(self) -> str:
         return sql.String
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.binary()
+
+    def _validate_literal(self, val: Any) -> None:
+        if isinstance(val, PIL.Image.Image):
+            return
+        self._validate_file_path(val)
+
+    def validate_media(self, val: Any) -> None:
+        assert isinstance(val, str)
+        try:
+            _ = PIL.Image.open(val)
+        except PIL.UnidentifiedImageError:
+            raise excs.Error(f'Not a valid image: {val}') from None
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        import tensorflow as tf
-        return tf.TensorSpec(shape=(self.height, self.width, self.num_channels), dtype=tf.uint8)
+class VideoType(ColumnType):
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.VIDEO, nullable=nullable)
 
+    def to_sql(self) -> str:
+        # stored as a file path
+        return 'VARCHAR'
 
-class VideoType(ColumnType):
-    def __init__(self):
-        super().__init__(self.Type.VIDEO)
+    def to_sa_type(self) -> str:
+        return sql.String
+    
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa # pylint: disable=import-outside-toplevel
+        return pa.string()
+
+    def _validate_literal(self, val: Any) -> None:
+        self._validate_file_path(val)
+
+    def validate_media(self, val: Any) -> None:
+        assert isinstance(val, str)
+        try:
+            with av.open(val, 'r') as fh:
+                if len(fh.streams.video) == 0:
+                    raise excs.Error(f'Not a valid video: {val}')
+                # decode a few frames to make sure it's playable
+                # TODO: decode all frames? but that's very slow
+                num_decoded = 0
+                for frame in fh.decode(video=0):
+                    _ = frame.to_image()
+                    num_decoded += 1
+                    if num_decoded == 10:
+                        break
+                if num_decoded < 2:
+                    # this is most likely an image file
+                    raise excs.Error(f'Not a valid video: {val}')
+        except av.AVError:
+            raise excs.Error(f'Not a valid video: {val}') from None
+
+class AudioType(ColumnType):
+    def __init__(self, nullable: bool = False):
+        super().__init__(self.Type.AUDIO, nullable=nullable)
 
-    def _as_dict(self) -> Dict:
-        result = super()._as_dict()
-        return result
+    def to_sql(self) -> str:
+        # stored as a file path
+        return 'VARCHAR'
 
-    @classmethod
-    def _from_dict(cls, d: Dict) -> 'ColumnType':
-        return cls()
+    def to_sa_type(self) -> str:
+        return sql.String
+
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa  # pylint: disable=import-outside-toplevel
+        return pa.string()
+
+    def _validate_literal(self, val: Any) -> None:
+        self._validate_file_path(val)
+
+    def validate_media(self, val: Any) -> None:
+        try:
+            with av.open(val) as container:
+                if len(container.streams.audio) == 0:
+                    raise excs.Error(f'No audio stream in file: {val}')
+                audio_stream = container.streams.audio[0]
+
+                # decode everything to make sure it's playable
+                # TODO: is there some way to verify it's a playable audio file other than decoding all of it?
+                for packet in container.demux(audio_stream):
+                    for _ in packet.decode():
+                        pass
+        except av.AVError as e:
+            raise excs.Error(f'Not a valid audio file: {val}\n{e}') from None
+
+class DocumentType(ColumnType):
+    @enum.unique
+    class DocumentFormat(enum.Enum):
+        HTML = 0
+        MD = 1
+        PDF = 2
+
+    def __init__(self, nullable: bool = False, doc_formats: Optional[str] = None):
+        super().__init__(self.Type.DOCUMENT, nullable=nullable)
+        if doc_formats is not None:
+            type_strs = doc_formats.split(',')
+            for type_str in type_strs:
+                if not hasattr(self.DocumentFormat, type_str):
+                    raise ValueError(f'Invalid document type: {type_str}')
+            self._doc_formats = [self.DocumentFormat[type_str.upper()] for type_str in type_strs]
+        else:
+            self._doc_formats = [t for t in self.DocumentFormat]
 
     def to_sql(self) -> str:
         # stored as a file path
@@ -577,5 +901,38 @@ class VideoType(ColumnType):
     def to_sa_type(self) -> str:
         return sql.String
 
-    def to_tf(self) -> Union['tf.TypeSpec', Dict[str, 'tf.TypeSpec']]:
-        assert False
+    def to_arrow_type(self) -> 'pyarrow.DataType':
+        import pyarrow as pa  # pylint: disable=import-outside-toplevel
+        return pa.string()
+
+    def _validate_literal(self, val: Any) -> None:
+        self._validate_file_path(val)
+
+    def validate_media(self, val: Any) -> None:
+        assert isinstance(val, str)
+        from pixeltable.utils.documents import get_document_handle
+        with open(val, 'r') as fh:
+            try:
+                s = fh.read()
+                dh = get_document_handle(s)
+                if dh is None:
+                    raise excs.Error(f'Not a recognized document format: {val}')
+            except Exception as e:
+                raise excs.Error(f'Not a recognized document format: {val}') from None
+
+
+# A dictionary mapping various Python types to their respective ColumnTypes.
+# This can be used to infer Pixeltable ColumnTypes from type hints on Python
+# functions. (Since Python functions do not necessarily have type hints, this
+# should always be an optional/convenience inference.)
+_python_type_to_column_type: dict[type, ColumnType] = {
+    str: StringType(),
+    int: IntType(),
+    float: FloatType(),
+    bool: BoolType(),
+    datetime.datetime: TimestampType(),
+    datetime.date: TimestampType(),
+    list: JsonType(),
+    dict: JsonType(),
+    PIL.Image.Image: ImageType()
+}