Decoder-native resize public implementation

.github/workflows/lint.yaml

@@ -62,7 +62,7 @@ jobs:
         run: python -m pip install --upgrade pip
       - name: Install dependencies and FFmpeg
         run: |
-          python -m pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
+          python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cpu
           conda install "ffmpeg=7.0.1" pkg-config pybind11 -c conda-forge
           ffmpeg -version
       - name: Build and install torchcodec

docs/source/api_ref_transforms.rst

@@ -0,0 +1,17 @@
+.. _samplers:
+
+===================
+torchcodec.transforms
+===================
+
+.. currentmodule:: torchcodec.transforms
+
+For a tutorial, see: TODO_DECODER_TRANSFORMS_TUTORIAL.
+
+.. autosummary::
+    :toctree: generated/
+    :nosignatures:
+    :template: dataclass.rst
+
+    DecoderTransform
+    Resize

docs/source/index.rst

@@ -125,3 +125,4 @@ Encoding
    api_ref_decoders
    api_ref_encoders
    api_ref_samplers
+   api_ref_transforms

mypy.ini

@@ -4,3 +4,4 @@ files = src/torchcodec
 show_error_codes = True
 pretty = True
 allow_redefinition = True
+follow_untyped_imports = True

src/torchcodec/__init__.py

@@ -7,7 +7,7 @@
 # Note: usort wants to put Frame and FrameBatch after decoders and samplers,
 # but that results in circular import.
 from ._frame import AudioSamples, Frame, FrameBatch  # usort:skip # noqa
-from . import decoders, encoders, samplers  # noqa
+from . import decoders, encoders, samplers, transforms  # noqa
 
 try:
     # Note that version.py is generated during install.

src/torchcodec/decoders/_video_decoder.py

@@ -8,17 +8,18 @@
 import json
 import numbers
 from pathlib import Path
-from typing import Literal, Optional, Tuple, Union
+from typing import List, Literal, Optional, Sequence, Tuple, Union
 
 import torch
-from torch import device as torch_device, Tensor
+from torch import device as torch_device, nn, Tensor
 
 from torchcodec import _core as core, Frame, FrameBatch
 from torchcodec.decoders._decoder_utils import (
     _get_cuda_backend,
     create_decoder,
     ERROR_REPORTING_INSTRUCTIONS,
 )
+from torchcodec.transforms import DecoderTransform, Resize
 
 
 class VideoDecoder:
@@ -66,6 +67,13 @@ class VideoDecoder:
             probably is. Default: "exact".
             Read more about this parameter in:
             :ref:`sphx_glr_generated_examples_decoding_approximate_mode.py`
+        transforms (sequence of transform objects, optional): Sequence of transforms to be
+            applied to the decoded frames by the decoder itself, in order. Accepts both
+            :class:`~torchcodec.transforms.DecoderTransform` and
+            `torchvision.transforms.v2.Transform <https://docs.pytorch.org/vision/stable/transforms.html#v2-api-reference-recommended>`_
+            objects. All transforms are applied
+            in the ouput pixel format and colorspace. Read more about this parameter in:
+            TODO_DECODER_TRANSFORMS_TUTORIAL.
         custom_frame_mappings (str, bytes, or file-like object, optional):
             Mapping of frames to their metadata, typically generated via ffprobe.
             This enables accurate frame seeking without requiring a full video scan.
@@ -104,6 +112,7 @@ def __init__(
         num_ffmpeg_threads: int = 1,
         device: Optional[Union[str, torch_device]] = "cpu",
         seek_mode: Literal["exact", "approximate"] = "exact",
+        transforms: Optional[Sequence[Union[DecoderTransform, nn.Module]]] = None,
         custom_frame_mappings: Optional[
             Union[str, bytes, io.RawIOBase, io.BufferedReader]
         ] = None,
@@ -148,13 +157,16 @@ def __init__(
 
         device_variant = _get_cuda_backend()
 
+        transform_specs = _make_transform_specs(transforms)
+
         core.add_video_stream(
             self._decoder,
             stream_index=stream_index,
             dimension_order=dimension_order,
             num_threads=num_ffmpeg_threads,
             device=device,
             device_variant=device_variant,
+            transform_specs=transform_specs,
             custom_frame_mappings=custom_frame_mappings_data,
         )
 
@@ -432,6 +444,78 @@ def _get_and_validate_stream_metadata(
     )
 
 
+def _convert_to_decoder_transforms(
+    transforms: Sequence[Union[DecoderTransform, nn.Module]],
+) -> List[DecoderTransform]:
+    """Convert a sequence of transforms that may contain TorchVision transform
+    objects into a list of only TorchCodec transform objects.
+
+    Args:
+        transforms: Squence of transform objects. The objects can be one of two
+        types:
+                1. torchcodec.transforms.DecoderTransform
+                2. torchvision.transforms.v2.Transform, but our type annotation
+                   only mentions its base, nn.Module. We don't want to take a
+                   hard dependency on TorchVision.
+
+    Returns:
+        List of DecoderTransform objects.
+    """
+    try:
+        from torchvision.transforms import v2
+
+        tv_available = True
+    except ImportError:
+        tv_available = False
+
+    converted_transforms: list[DecoderTransform] = []
+    for transform in transforms:
+        if not isinstance(transform, DecoderTransform):
+            if not tv_available:
+                raise ValueError(
+                    f"The supplied transform, {transform}, is not a TorchCodec "
+                    " DecoderTransform. TorchCodec also accept TorchVision "
+                    "v2 transforms, but TorchVision is not installed."
+                )
+            if isinstance(transform, v2.Resize):
+                converted_transforms.append(Resize._from_torchvision(transform))
+            else:
+                raise ValueError(
+                    f"Unsupported transform: {transform}. Transforms must be "
+                    "either a TorchCodec DecoderTransform or a TorchVision "
+                    "v2 transform."
+                )
+        else:
+            converted_transforms.append(transform)
+
+    return converted_transforms
+
+
+def _make_transform_specs(
+    transforms: Optional[Sequence[Union[DecoderTransform, nn.Module]]],
+) -> str:
+    """Given a sequence of transforms, turn those into the specification string
+       the core API expects.
+
+    Args:
+        transforms: Optional sequence of transform objects. The objects can be
+            one of two types:
+                1. torchcodec.transforms.DecoderTransform
+                2. torchvision.transforms.v2.Transform, but our type annotation
+                   only mentions its base, nn.Module. We don't want to take a
+                   hard dependency on TorchVision.
+
+    Returns:
+        String of transforms in the format the core API expects: transform
+        specifications separate by semicolons.
+    """
+    if transforms is None:
+        return ""
+
+    transforms = _convert_to_decoder_transforms(transforms)
+    return ";".join([t._make_params() for t in transforms])
+
+
 def _read_custom_frame_mappings(
     custom_frame_mappings: Union[str, bytes, io.RawIOBase, io.BufferedReader]
 ) -> tuple[Tensor, Tensor, Tensor]:

src/torchcodec/transforms/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from ._decoder_transforms import DecoderTransform, Resize  # noqa

src/torchcodec/transforms/_decoder_transforms.py

@@ -0,0 +1,85 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Sequence
+
+from torch import nn
+
+
+@dataclass
+class DecoderTransform(ABC):
+    """Base class for all decoder transforms.
+
+    A *decoder transform* is a transform that is applied by the decoder before
+    returning the decoded frame.  Applying decoder transforms to frames
+    should be both faster and more memory efficient than receiving normally
+    decoded frames and applying the same kind of transform.
+
+    Most `DecoderTransform` objects have a complementary transform in TorchVision,
+    specificially in
+    `torchvision.transforms.v2 <https://docs.pytorch.org/vision/stable/transforms.html#v2-api-reference-recommended>`_.
+    For such transforms, we ensure that:
+
+      1. The names are the same.
+      2. Default behaviors are the same.
+      3. The parameters for the `DecoderTransform` object are a subset of the
+         TorchVision transform object.
+      4. Parameters with the same name control the same behavior and accept a
+         subset of the same types.
+      5. The difference between the frames returned by a decoder transform and
+         the complementary TorchVision transform are small.
+
+    All decoder transforms are applied in the output pixel format and colorspace.
+    """
+
+    @abstractmethod
+    def _make_params(self) -> str:
+        pass
+
+
+@dataclass
+class Resize(DecoderTransform):
+    """Resize the decoded frame to a given size.
+
+    Complementary TorchVision transform:
+    `torchvision.transforms.v2.Resize <https://docs.pytorch.org/vision/stable/generated/torchvision.transforms.v2.Resize.html#torchvision.transforms.v2.Resize>`_.
+    Interpolation is always bilinear. Anti-aliasing is always on.
+
+    Args:
+        size: (sequence of int): Desired output size. Must be a sequence of
+            the form (height, width).
+    """
+
+    size: Sequence[int]
+
+    def _make_params(self) -> str:
+        assert len(self.size) == 2
+        return f"resize, {self.size[0]}, {self.size[1]}"
+
+    @classmethod
+    def _from_torchvision(cls, resize_tv: nn.Module):
+        from torchvision.transforms import v2
+
+        assert isinstance(resize_tv, v2.Resize)
+
+        if resize_tv.interpolation is not v2.InterpolationMode.BILINEAR:
+            raise ValueError(
+                "TorchVision Resize transform must use bilinear interpolation."
+            )
+        if resize_tv.antialias is False:
+            raise ValueError(
+                "TorchVision Resize transform must have antialias enabled."
+            )
+        if resize_tv.size is None:
+            raise ValueError("TorchVision Resize transform must have a size specified.")
+        if len(resize_tv.size) != 2:
+            raise ValueError(
+                "TorchVision Resize transform must have a (height, width) "
+                f"pair for the size, got {resize_tv.size}."
+            )
+        return cls(size=resize_tv.size)