Setup for local source processing
This directory contains the shared download and image/video conversion helpers
used by the six recipe-only directories under sample/. Their source data and
generated records are not distributed by this repository.
Use Python 3.10 or later. From the directory containing MMTA-v1.0/, install
the dependencies:
python -m pip install -r MMTA-v1.0/processing/requirements.txt
Install ffmpeg and ffprobe for video processing. Credentialed PhysioNet
downloads also require wget and an interactive terminal.
| Domain | Download and generation instructions |
|---|---|
| MIMIC mortality | mimic_death |
| MIMIC discharge | mimic_disch |
| PixelRec | pixelrec |
| SP500 / FinMultiTime | sp500 |
| CoVLA | covla |
| CUHK-X HAR | cuhk_x_har |
Complete each upstream source's access requirements with your own account before
downloading. download.py --dry-run prints a plan without contacting a source or
writing data; actual downloads require --accept-terms. Use --help on each
domain's download.py and generate.py to inspect all available options.
Downloaded archives must be unpacked before building.
Keep source data and outputs outside MMTA-v1.0/, for example in sibling
local_sources/ and local_work/ directories. Generation writes a new domain
directory containing task JSONL, sensor CSVs and intermediate prompts. Video
recipes also write cropped videos; image recipes reference the local source
images rather than copying them. Generated paths are relative to that local
domain directory, and can refer to source images outside it.
common/local_generation.py orchestrates the per-domain builders and the common
converters. Optional --input-jsonl skips source prompt construction for all
recipes except SP500, which rebuilds inputs using source-dated news strictly
before the forecast cutoff. MIMIC discharge also uses the table, bucketing and
splitting helpers in sample/image-timeseries/mimic_death/; retain both MIMIC
directories when copying scripts.
The recipes support local reconstruction, whose exact records may depend on the
available upstream versions. They are not additional load_dataset
configurations and do not change the five released test splits.
See CODE_PROVENANCE.md and provenance.json for script origins and retained notices.