camaro/solafune-precipitation-nowcasting

Solafune Precipitation Nowcasting — final submission weights

The 11 model checkpoints behind the final submission (equal-weight blend, Public LB 0.6360204).

Each file is the EMA state_dict at epoch 10 of a model trained on all 20 training locations. No optimizer, scheduler or RNG state is included.

file	temporal fusion	stem stride	frames	extras	seed	TTA
exp175/model_best.pt	stack2d	1	3	—	42	8
exp176/model_best.pt	conv3d	2	3	spectral adapter	42	8
exp177/model_best.pt	stack2d	2	3	—	1234	8
exp178/model_best.pt	stack2d	1	3	multiscale skip	42	8
exp179/model_best.pt	stack2d	2	3	satellite FiLM	42	8
exp180/model_best.pt	conv3d	4	3	spectral adapter, rot360 aug	42	3
exp182/model_best.pt	stack2d	1	3	—	1234	8
exp183/model_best.pt	conv3d	2	3	spectral adapter	1234	8
exp184/model_best.pt	stack2d	2	3	satellite FiLM	1234	8
exp185/model_best.pt	stack2d	1	3	multiscale skip	1234	8
exp186/model_best.pt	conv3d	2	6	multiscale skip, target alignment, IR temporal stats	42	8

Usage

The checkpoints only load against the matching model definition. Place them under src/output/expNNN/model_best.pt in the solution repository, then run stages 2 and 3 of src/run.sh.

from configs import load_config
from data import data_input_channels, data_input_frames
from model import NowcastModel
import torch

cfg = load_config("configs/exp175.py")
model = NowcastModel(cfg.model, cfg.data.target_size,
                     data_input_channels(cfg.data), data_input_frames(cfg.data))
model.load_state_dict(torch.load("output/exp175/model_best.pt", map_location="cpu", weights_only=True))