File size: 1,515 Bytes
b48ebe5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 | """
example_usage.py — Minimal inference with BeMAE-Hα.
This script downloads the checkpoint from the HuggingFace Hub and computes
the z_halpha embedding of a dummy batch with the expected input format
(128 bins over Hα ± 50 Å, flux normalised to the pseudo-continuum).
See `stage1_encoder/dataset.py` in the code repository for the
preprocessing of real BESS-Bench spectra.
"""
import json
import sys
import torch
from huggingface_hub import snapshot_download
MODEL_ID = "matthieulel/bemae-halpha-v1"
# 1. Download the snapshot
ckpt_dir = snapshot_download(MODEL_ID)
sys.path.insert(0, ckpt_dir)
from model import SpectralEncoderHalpha, ModelConfig # noqa: E402
# 2. Load config and weights (seed 42 at the root; seed123/ and seed456/ also available)
seed_dir = ckpt_dir
with open(f"{seed_dir}/config.json") as f:
meta = json.load(f)
cfg = ModelConfig(**meta["model_config"])
encoder = SpectralEncoderHalpha(cfg)
state = torch.load(f"{seed_dir}/pytorch_model.bin", map_location="cpu")
encoder.load_state_dict(state)
encoder.eval()
# 3. Prepare a dummy spectrum (Hα-centred, 128 bins, pseudo-continuum at 1.0)
B = 4
flux = torch.ones(B, 128) + 0.3 * torch.randn(B, 128) * 0.01
wavelengths = torch.linspace(6512.8, 6612.8, 128).unsqueeze(0).expand(B, -1)
validity = torch.ones(B, 128)
# 4. Inference
with torch.no_grad():
z_halpha, *_ = encoder(flux, wavelengths, validity, mask=None)
print(f"z_halpha shape : {tuple(z_halpha.shape)}")
print(f"z_halpha[0, :8] : {z_halpha[0, :8].numpy()}")
|