Atmospheric Profile Transformer Codebase
Overview
This codebase trains a FiLM-conditioned transformer model for atmospheric profile regression
predicting net thermal and reflected fluxes from pressure-temperature profiles and global variables 
(e.g., orbital distance). It uses preprocessed data from PICASO radiative transfer simulations, 
focusing on efficiency for GPU training (e.g., A100) via normalized .npy shards.

Goals:

Emulate expensive radiative transfer computations with a fast neural surrogate.
Support reproducible training with configurable normalization, model architecture, and hyperparameters.
Enable inference via exported TorchScript models.
Pipeline
Data Generation (Optional): Use create_data.py to process input HDF5 profiles with PICASO, saving raw results to data/raw/ (e.g., picaso_results_1.h5).
Preprocessing: Load raw HDF5, compute normalization stats from train split, normalize all data, and shard into .npy files in data/processed/. Skips if hash matches.
Training: Load shards, train transformer with AMP/grad accumulation/early stopping, save best model and JIT export to models/<folder>/.
Evaluation: Test on held-out shards, log metrics.
Run via python src/main.py --config config/config.jsonc.

Directory Structure
src/: Core code.
config/: Configuration files (e.g., config.jsonc).
data/raw/: Raw HDF5 inputs.
data/processed/: Normalized .npy shards and metadata.
models/: Trained models, logs, and exports.
testing/: (Future use for tests).
Key Files
src/utils.py: Config loading, logging, seeding, splits, hashing.
src/normalizer.py: Computes/applies normalization (e.g., log-min-max, symlog).
src/preprocess.py: Converts HDF5 to normalized shards.
src/dataset.py: Loads shards, pads sequences.
src/model.py: FiLM-transformer definition and JIT export.
src/train.py: Training loop with optimizers/schedulers.
src/hardware.py: Device detection.
src/main.py: Pipeline orchestrator.
create_data.py: Generates raw HDF5 from profiles (standalone).
Setup
Install dependencies: PyTorch, h5py, numpy, tqdm, etc.
Ensure HDF5 files exist in data/raw/.
Customize config/config.jsonc for your setup.