Decode timestamp batches without frame index conversion

[mavenlin]

get_frames_played_at currently converts timestamps to indices before decoding. That assumes a single global timestamp-to-index mapping, which is not valid for VFR video. The singular get_frame_played_at already seeks and decodes by requested timestamp; batch lookup should match it.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/meta-pytorch/torchcodec/1629

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[mavenlin]

Sharing some tests, I compare the approximate mode with the exact mode, and the new code after this PR is better in the alignment with the exact mode.

Tested on a sampled set of videos from diverse sources. The 4/5 columns reports the number of frames that differs from the exact mode.

dataset	videos	frames	old code	new code
dataset_01	200	17,202	273	0
dataset_02	10	749	214	0
dataset_03	200	13,983	256	0
dataset_04	200	25,199	206	0
dataset_05	200	16,901	448	0
dataset_06	200	14,984	201	0
dataset_07	102	5,598	682	0
dataset_08	200	17,556	256	0
dataset_09	200	15,979	233	0
dataset_10	200	17,902	351	0
dataset_11	200	21,600	427	0
dataset_12	200	20,043	247	0
dataset_13	200	13,551	314	0
dataset_14	200	12,805	255	0
dataset_15	200	25,555	6,398	0
dataset_16	200	13,662	851	0
dataset_17	200	18,109	1,461	0
dataset_18	200	16,961	404	0
dataset_19	200	20,901	724	0
dataset_20	200	22,690	0	0
dataset_21	200	24,441	934	0
dataset_22	200	14,807	1,209	1
dataset_23	200	7,454	56	0
dataset_24	200	17,940	544	0
TOTAL	4,512	396,572	16,944	1