dcher95/phenofield-e98

PhenoField e98 (e98_large_specdrop_dyn4l)

Frozen phenology field encoder (CrossModalVAE, AdaLN-DiT dynamic branch, species dropout). Image-free interface: forward_field(species_id, prism_window, alphaearth_emb) → (z_static [576], z_dynamic [192]). The dynamic encoder consumes only the 365×7 trailing PRISM window and the 64-d AlphaEarth embedding (no doy / lat-lon / climplicit weights exist in this checkpoint) — season enters through the PRISM window alone.

Files

file	what
last.ckpt	Lightning checkpoint (weights + full training config in hyper_parameters)
hparams.yaml	same config, human-readable
models/*.py	the exact model code (CrossModalVAE, AdaLNDynamicEncoder) — self-contained, relative imports only
species_embeddings_v2.pt	BioCLIP text matrix [6825, 768]; the constructor loads it via config['species_embedding_path']
species_vocab.json	species name → id (indexes rows of the embedding matrix)
inputs/prism_weekly.npz	climatological per-(cell, week) 365×7 PRISM windows for the CONUS 0.5° grid
inputs/alphaearth_2017.parquet	per-point AlphaEarth 64-d embeddings (nearest-cell lookup)
inputs/grid_centroids_0.5deg.csv, inputs/plant_flowering_events.parquet	grid definition + plant flowering observations

Loading

import sys, torch
from huggingface_hub import snapshot_download
from omegaconf import OmegaConf

d = snapshot_download('dcher95/phenofield-e98')
sys.path.insert(0, d)
from models.cross_modal_vae import CrossModalVAE

ckpt = torch.load(f'{d}/last.ckpt', map_location='cpu', weights_only=False)
cfg = OmegaConf.to_container(OmegaConf.create(ckpt['hyper_parameters']).model, resolve=True)
cfg['species_embedding_path'] = f'{d}/species_embeddings_v2.pt'
model = CrossModalVAE(cfg)
state = {k[len('model.'):]: v for k, v in ckpt['state_dict'].items() if k.startswith('model.')}
missing, unexpected = model.load_state_dict(state, strict=False)
assert not missing and not unexpected   # this models/ copy loads clean
model.eval()

Do not load this checkpoint with the models/ in dcher95/PhenoField's top level — that class silently drops dynamic_encoder.null_species_emb (the parameter species dropout trains). Use the models/ shipped here (a copy of vendor/bwei_ppe).

Downstream pipeline: pipelines/ in the ANTHEIA repo (Summer-2026-CfE project).