Skip to content

Fallback to container duration in approximate mode #989

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 8 commits into from
Nov 19, 2025
Merged

Fallback to container duration in approximate mode #989

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
389c078
fallback to container duration in approximate mode when stream durati…
Oct 20, 2025
0539566
Merge branch 'main' into fallback-container-duration
mollyxu Oct 28, 2025
51b0a74
Merge branch 'meta-pytorch:main' into fallback-container-duration
mollyxu Nov 12, 2025
95eedb1
Merge branch 'meta-pytorch:main' into fallback-container-duration
mollyxu Nov 13, 2025
27718be
update pr to reflect refactor
Nov 13, 2025
2979900
Merge branch 'meta-pytorch:main' into fallback-container-duration
mollyxu Nov 17, 2025
ec12f87
Merge branch 'meta-pytorch:main' into fallback-container-duration
mollyxu Nov 19, 2025
68e48db
address feedback
Nov 19, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions src/torchcodec/_core/SingleStreamDecoder.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,6 +158,16 @@ void SingleStreamDecoder::initializeDecoder() {
ptsToSeconds(formatContext_->duration, defaultTimeBase);
}

// Use container duration as fallback for streams missing stream duration
if (containerMetadata_.durationSecondsFromHeader.has_value()) {
for (auto& streamMetadata : containerMetadata_.allStreamMetadata) {
if (!streamMetadata.durationSecondsFromHeader.has_value()) {
streamMetadata.durationSecondsFromHeader =
containerMetadata_.durationSecondsFromHeader;
Copy link
Contributor

@scotts scotts Oct 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Rather than assigning the container's duration to the stream's duration, I think we should do something like:

if (!streamMetadata.durationSectonsFromHeader.has_value()) {
  return containerMetadata_.durationSecondsFromHeader;
}

I think that keeps it clear where we're getting our information from. So far, we've kept that explicit in our existing metadata fields.

One wrinkle here is that the metadata we report to users in Python is VideoStreamMetadata. That is, it's explicitly the stream metadata. We don't have a hierarchically clean way to add the container metadata there. So while it's kinda odd, maybe we should add a field for duration_seconds_from_container on the Python side.

To be clear, my current thinking is that:

  1. In C++, we fall back to the container metadata where appropriate. This may be awkward because we don't pass around the container metadata, so we may want to add a durationSecondsFromContainer field to StreamMetadata.
  2. In Python, we should add a field for duration_seconds_from_container to StreamMetadata, and we have logic to fall back to it in duration_seconds.

@NicolasHug, @Dan-Flores, I'm curious to hear your thoughts as well.

Copy link
Contributor

@NicolasHug NicolasHug Oct 27, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@scotts can you help me understand where you think we should add this?

if (!streamMetadata.durationSectonsFromHeader.has_value()) {
  return containerMetadata_.durationSecondsFromHeader;
}

I agree with your general point that we should avoid setting the stream metadata to the container metadata.

We might be able to avoid defining a new *_from_container attribute on the StreamMetadata: the ContainerMetadata metaclass exists in Python. We just don't expose it currently.

The original error happens here:

torchcodec/src/torchcodec/decoders/_video_decoder.py

Lines 413 to 418 in 75a3325

if metadata.end_stream_seconds is None:
raise ValueError(
"The maximum pts value in seconds is unknown. "
+ ERROR_REPORTING_INSTRUCTIONS
)
end_stream_seconds = metadata.end_stream_seconds

and this is a place where we do know the ContainerMetadata. Maybe we can just fallback to it in that function only (I haven't checked that this would solve the original issue). And, whether we want to expose ContainerMetadata is kind of a separate issue.

As a side note, reasoning about all these fallbacks hurts my brain. We have fallbacks in the Python StreamMetadata class, in the C++ initializeDecoder(), and in the C++ helpers like getNumFrames. It might be time we address our long-time issue of centralizing all of the fallback logic in C++ (#125 (comment))

Copy link
Contributor Author

@mollyxu mollyxu Oct 27, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@scotts @NicolasHug Thank you for the feedback!

I tried implementing fallback at line 413-418 in _video_decoder.py but I think that num_frames would be none. This is due to metadata.num_frames not having the container duration when it computes num_frames using duration_seconds_from_header when num_frames is missing from the video.

Copy link
Contributor

@scotts scotts Oct 28, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@NicolasHug,

@scotts can you help me understand where you think we should add this?

if (!streamMetadata.durationSectonsFromHeader.has_value()) {
  return containerMetadata_.durationSecondsFromHeader;
}

Right now, there's only one place we really use the duration's value in C++: getMaxSeconds(). That would become:

std::optional<double> SingleStreamDecoder::getMaxSeconds(
    const StreamMetadata& streamMetadata) {
  switch (seekMode_) {
    case SeekMode::custom_frame_mappings:
    case SeekMode::exact:
      return streamMetadata.endStreamPtsSecondsFromContent.value();
    case SeekMode::approximate: {
      //
      // Fallback to container duration if stream does not have one
      //
      if (!streamMetadata.durationSecondsFromHeader.has_value()) {
        return containerMetadata_.durationSecondsFromHeader;
      }
      return std::nullopt;
    }
    default:
      TORCH_CHECK(false, "Unknown SeekMode");
  }
}

Agreed that now may be the time to consolidate all of this fall back logic in C++. The fallback logic in Python is not straight-forward to understand because it's close to being a state machine. We're trying to be as tolerant of missing values as possible, which means that value a may try to fallback to value b in some instances, but value b may also try to fallback to value a in other instances. Implementing this logic in one place, for all uses, will at least make sure that we're consistent across Python and C++.

Part of why this logic is convoluted in Python is that the Python metadata class has no knowledge of seek modes. Instead, it has fallback logic exclusively based on what values are present.

}
}
}

if (formatContext_->bit_rate > 0) {
containerMetadata_.bitRate = formatContext_->bit_rate;
}
Expand Down
11 changes: 0 additions & 11 deletions test/test_decoders.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,6 @@
SINE_MONO_S32,
SINE_MONO_S32_44100,
SINE_MONO_S32_8000,
supports_approximate_mode,
TEST_SRC_2_720P,
TEST_SRC_2_720P_H265,
TEST_SRC_2_720P_MPEG4,
Expand Down Expand Up @@ -1492,8 +1491,6 @@ def test_get_frames_at_tensor_indices(self):
def test_beta_cuda_interface_get_frame_at(
self, asset, contiguous_indices, seek_mode
):
if seek_mode == "approximate" and not supports_approximate_mode(asset):
pytest.skip("asset doesn't work with approximate mode")

if in_fbcode() and asset is AV1_VIDEO:
pytest.skip("AV1 CUDA not supported internally")
Expand Down Expand Up @@ -1540,8 +1537,6 @@ def test_beta_cuda_interface_get_frame_at(
def test_beta_cuda_interface_get_frames_at(
self, asset, contiguous_indices, seek_mode
):
if seek_mode == "approximate" and not supports_approximate_mode(asset):
pytest.skip("asset doesn't work with approximate mode")
if in_fbcode() and asset is AV1_VIDEO:
pytest.skip("AV1 CUDA not supported internally")

Expand Down Expand Up @@ -1585,8 +1580,6 @@ def test_beta_cuda_interface_get_frames_at(
)
@pytest.mark.parametrize("seek_mode", ("exact", "approximate"))
def test_beta_cuda_interface_get_frame_played_at(self, asset, seek_mode):
if seek_mode == "approximate" and not supports_approximate_mode(asset):
pytest.skip("asset doesn't work with approximate mode")
if in_fbcode() and asset is AV1_VIDEO:
pytest.skip("AV1 CUDA not supported internally")

Expand Down Expand Up @@ -1627,8 +1620,6 @@ def test_beta_cuda_interface_get_frame_played_at(self, asset, seek_mode):
)
@pytest.mark.parametrize("seek_mode", ("exact", "approximate"))
def test_beta_cuda_interface_get_frames_played_at(self, asset, seek_mode):
if seek_mode == "approximate" and not supports_approximate_mode(asset):
pytest.skip("asset doesn't work with approximate mode")
if in_fbcode() and asset is AV1_VIDEO:
pytest.skip("AV1 CUDA not supported internally")

Expand Down Expand Up @@ -1670,8 +1661,6 @@ def test_beta_cuda_interface_get_frames_played_at(self, asset, seek_mode):
)
@pytest.mark.parametrize("seek_mode", ("exact", "approximate"))
def test_beta_cuda_interface_backwards(self, asset, seek_mode):
if seek_mode == "approximate" and not supports_approximate_mode(asset):
pytest.skip("asset doesn't work with approximate mode")
if in_fbcode() and asset is AV1_VIDEO:
pytest.skip("AV1 CUDA not supported internally")

Expand Down
7 changes: 0 additions & 7 deletions test/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -798,10 +798,3 @@ def sample_format(self) -> str:
},
frames={0: {}}, # Not needed for now
)


def supports_approximate_mode(asset: TestVideo) -> bool:
# Those are missing the `duration` field so they fail in approximate mode (on all devices).
# TODO: we should address this, see
# https://github.com/meta-pytorch/torchcodec/issues/945
return asset not in (AV1_VIDEO, TEST_SRC_2_720P_VP9, TEST_SRC_2_720P_VP8)
Loading