Sign inSign up
train-rnn-asd logo

bmaged23/train-rnn-asd

By bmaged23

•Updated 23 days ago

Checkpoint-required visual ASD API/CLI. CPU, pinned runtime, tested container delivery.

Image
0

55

bmaged23/train-rnn-asd repository overview

⁠Active Speaker Detection — checkpoint-required inference API

GitHub source, historical model results and documentation⁠ · API guide⁠ · Docker and CI/CD⁠

CPU service for a supplied windowed BiLSTM checkpoint. Predicts visual speaking status from completed, preprocessed landmark feature windows. No raw-video upload, training, pretrained checkpoint or automatic model/dataset download is included. Linux x86-64 Docker is the tested target. No GPU is required by this image.

⁠Pull and run

Supply models/model.pt and a matching models/manifest.json using the checkpoint/input schema⁠. Create a private .env with ASD_API_TOKEN=your_private_token if desired.

docker pull bmaged23/train-rnn-asd:cpu
docker run --rm --read-only --tmpfs /tmp --env-file .env \
  -p 127.0.0.1:8000:8000 -v "$PWD/models:/models:ro" \
  -e ASD_CHECKPOINT=/models/model.pt -e ASD_MANIFEST=/models/manifest.json \
  bmaged23/train-rnn-asd:cpu

Open http://127.0.0.1:8000/docs⁠. Use POST /v1/predict for JSON feature windows; GET /v1/model describes the exact feature order. GET /metrics exposes process-local prediction counters/timing. No Git clone or extra dependency install is needed. The image runs as UID 10001; model files must be readable.

Readiness: http://127.0.0.1:8000/health/ready⁠. Without a valid checkpoint and manifest, the process reports 503 and the container is unhealthy. It never falls back to random weights. Stop with Ctrl+C.

⁠Verification and delivery

34 offline tests cover software/model contracts and API behavior. The candidate image runs actual CLI and HTTP inference with tiny temporary synthetic checkpoints, checks parity against native PyTorch and validates readiness, errors and auth. The published immutable image is pulled anonymously and tested again before the cpu tag is promoted. Runtime pushes automate these gates and publication through GitHub Actions. Original trained-model accuracy is not re-evaluated by these tests.

The image excludes weights, research datasets, personal .env files and caches. This is a toolkit for supplied checkpoints, not a ready-trained ASD detector.

Tag summary

Content type

Image

Digest

sha256:ec2d7420b…

Size

450.5 MB

Last updated

23 days ago

docker pull bmaged23/train-rnn-asd:cpu