camgoz-nslt

Neural Sign Language Translation reproduction

This is a faithful run of the authors' published neccam/nslt Luong recipe for Camgoz et al., Neural Sign Language Translation (CVPR 2018). It targets the Luong row of Table 5, not the separate Bahdanau system in Table 6.

Pipeline status: complete. Numerical agreement: fully reproduced.

The authors' full 150,000-step run completed. Its terminal best-model loop selected checkpoint 118,000 on development BLEU-4 and evaluated that fixed checkpoint on both splits.

Target

The pinned repository's unsmoothed, case-sensitive, whitespace-tokenized corpus BLEU implementation is used. It prints BLEU-1 through BLEU-4 and returns BLEU-4.

These are the terminal upstream evaluations of checkpoint 118,000, not the final-step checkpoint. Exact values, raw prediction hashes, run IDs, and provenance are in reproduction.json.

Pinned sources

Data and execution

The run uses RWTH-PHOENIX-Weather 2014T v3 on the shared Modal datasets v2 Volume: 7,096 train, 519 dev, and 642 test videos. It uses the shared huggingface-cache v2 Volume at /cache/huggingface.

The upstream manifests expect MATLAB-resized frame directories. With approval for this reproduction, the retained one-concern patch reads the supplied MP4s using simple-video-utils, resizes frames to 227x227 with OpenCV cubic interpolation, and preserves the upstream BGR mean-subtraction convention. This is a documented preprocessing deviation: direct MP4 decode/OpenCV resize is not pixel-identical to the authors' MATLAB resize and intermediate image files.

Run from the repository root with authenticated repro-sign Modal access:

.agents/skills/reproduce-paper/scripts/modal_repro_sign.sh run --detach \
  papers/camgoz-2018-nslt/scripts/modal_app.py::launch_train

The launcher invokes the authors' module directly with its README recipe: four residual 1,000-unit GRU layers, Luong attention, reversed source, batch size 1, Adam 1e-5, seed 285, and 150,000 training steps. The results volume is camgoz-nslt-results; a 24-hour Modal segment resumes from the verified upstream checkpoint rather than restarting.

Retained patches

• 0001-compatibility.patch: TensorFlow/Python compatibility required to execute the pinned source.
• 0002-alexnet-initialization.patch: preserves the paper-required ImageNet initialization, which upstream otherwise overwrites with global variable initialization on a fresh run.
• 0003-simple-video-utils.patch: approved direct-video input path described above.
• 0004-checkpoint-resume.patch: saves paired resume state with every checkpoint and retains the upstream evaluation cadence over controlled Modal segments.

No datasets, raw videos, outputs, or checkpoints are committed to this repository. The permitted selected checkpoint bundle, hparams, and exact predictions are preserved at repro-sign/camgoz-nslt revision 7851726; the repository's latest revision carries this model card and evidence record. The 503,848,900-byte model data file has SHA-256 bbace325a51197db547c18493870c5c5a87b289939b7505bd7ce73ed79831a05.

Execution evidence

The first single-A100 segment reached a verified checkpoint at step 91,000 before Modal's 24-hour function limit. The second segment resumed it without restarting, skipped the 6,132 already-seen examples, trained through step 150,000, and completed the upstream terminal best-model evaluation. The selected checkpoint is step 118,000.

The source checkpoint, prediction files, and configuration are recorded with immutable Hugging Face revision URLs and SHA-256 hashes in reproduction.json.