Engineering / research
Engineer learning over video streams
Analyze decoding, temporal sampling, clip construction, model state, and storage so video experiments preserve the information their tasks require.
Your route through engineering
Follow the path or begin with a question. Every lesson is open.
Your place, at your pace. Progress is optional and saved only in this browser when available.
Read / explore / reason
Inside this module
Module 01 research
Engineer learning over video streams
Analyze decoding, temporal sampling, clip construction, model state, and storage so video experiments preserve the information their tasks require.
- 01
Keep integer timestamps attached to their rational timebases
Distinguish PTS, DTS, stream timebases, epoch choices, and variable frame intervals before using decoded video as a temporal dataset.
30 min - 02
Decode the dependencies before selecting the clip
Separate demuxing, packet/frame flow, random-access entry points, reordering, decoder draining, and presentation-time clip selection.
30 min - 03
Sample presentation time without promising all motion survives
Distinguish timestamp-based selection from frame strides and demonstrate temporal aliasing under a precisely stated uniform-sampling model.
30 min - 04
Align clip support, event labels, and the prediction cutoff
Define half-open source-time intervals, overlap policies, temporal IoU, and feature availability without treating offline annotations as already observable inputs.
30 min - 05
Budget the actual patch and tubelet sequence
Count spatial-temporal tokens, dense score tensors, and masked-encoder scope without turning a simplified attention allocation into a complete video-model memory claim.
30 min - 06
Carry temporal state only within its declared stream
Separate causal feature buffers, chunk boundaries, lookahead, and reset policy so streaming predictions match the model's actual information boundary.
30 min - 07
Transform a clip and its labels under one explicit rule
Preserve coherent spatial transforms across a clip while checking which label meanings survive cropping, reflection, temporal reversal, and speed changes.
30 min - 08
Count source identities and sampler multiplicities across workers
Track clip provenance, rank/worker assignment, sampler padding, and repeated views so distributed training and evaluation consume the intended video cohort.
30 min - 09
Version video embeddings together with their temporal meaning
Count vector payload, retain source intervals, define metric normalization, and validate retrieval against the same eligible cohort and encoder/index versions.
30 min - 10
Test the temporal evidence boundary before interpreting quality
Separate source-group leakage, feature-support overlap, timestamp integrity, class imbalance, and temporal localization so evaluation matches its declared generalization task.
30 min