TF-ProFM

PHMFD pretrained feature extractor exported from the local main-branch pretraining log. This is the PHMFD implementation of TF-ProFM, not an upstream official checkpoint.

• Training seed: 42; checkpoint epoch: 1.
• Feature dimension: 4096.
• Input length: 2048; resizing: resample_short_truncate_long.
• Per-channel normalization: sample_zscore.
• Pretraining-only heads and optimizer state are excluded. No fault-label classifier is included.
• Source checkpoint, SHA256 and source code commit: see provenance.json.
• All exported encoder tensors load strictly. Features and preprocessing were compared with the main-branch implementation (maximum feature error: 0).

Use the PHMFD inference-only branch:

python inference.py --model-id bojian1/TF-ProFM --revision seed-42 --input samples.npy --output features.npz --seed 42

main defaults to training seed 42; seed-42, seed-43, and seed-44 select training checkpoints. The CLI --seed controls inference randomness, independently of the training seed. TF-ProFM retains random patch sampling; reproducibility requires the same seed, device, batch size and input order. UniFault returns mean-pooled token features.

Artifacts contain only inference configuration and safetensors weights; load using from model import load_model, PHMFDProcessor from PHMFD, not Transformers AutoModel.