Fix FPS truncated to integer in multiple locations

[AmitMY]

Summary

Fixes #201 — PoseBody.fps is typed as float (since v0.2), but several locations cast it to int, silently losing precision for non-integer framerates like 29.97 fps (NTSC standard).

Changes:

• utils/openpose.py: Remove int(fps) cast in load_openpose() — pass fps through as-is
• utils/generic.py: Fix fake_pose() signature from fps: int=25 to fps: float=25.0, remove int(fps) cast
• tensorflow/pose_body.py: Change TFRecord field from tf.int64 to tf.float32, serialize as FloatList instead of Int64List, remove unnecessary tf.cast on deserialization
• utils/holistic.py: Fix load_mediapipe_directory() type annotation from fps: int to fps: float
• numpy/pose_body.py: Fix interpolate() type annotation from new_fps: int to new_fps: float

Test plan

• Existing tests pass (72/72, excluding pre-existing TF import failures)
• Verified fake_pose(10, fps=29.97) now preserves 29.97 instead of truncating to 29

Note

Breaking change for TFRecord: Existing TFRecords serialized with int64 fps will not be compatible with the updated float32 schema. This is intentional — the old format was lossy.

🤖 Generated with Claude Code

[bricksdont]

looks good to me!

[bricksdont]

just one question:

Breaking change for TFRecord: Existing TFRecords serialized with int64 fps will not be compatible with the updated float32 schema. This is intentional — the old format was lossy.

does this have any impact on existing pose data from https://github.com/sign-language-processing/datasets?

[AmitMY]

does this have any impact on existing pose data from sign-language-processing/datasets?

No! this format was actually for when I worked for Google and needed to load the data with tensorflow.
datasets separately needs a fix https://github.com/sign-language-processing/datasets/blob/master/sign_language_datasets/utils/features/pose_feature.py#L127