"""Predict yield for one plot from weekly captions + UAS images. Example: python scripts/inference_single_hf.py --hf_model_dir checkpoints \ --sample_json examples/sample_input.json --week 12 --output_json prediction.json The JSON holds planting_date, week, texts (one caption per week after planting) and images (path + date). Text and image embeddings are computed here. """ import sys from pathlib import Path PROJECT_ROOT = Path(__file__).resolve().parents[1] sys.path.insert(0, str(PROJECT_ROOT)) import argparse import json import torch from transformers import AutoModel from data.embeddings import ImageEmbedder, TextEmbedder, embed_samples from hf.auto import register_multimodal_autoclass def parse_args(): p = argparse.ArgumentParser() p.add_argument("--hf_model_dir", required=True, help="local checkpoint dir or Hub repo id") p.add_argument("--sample_json", required=True) p.add_argument("--week", type=int, default=None, help="overrides the sample's week (1..22)") p.add_argument("--output_json", default="inference_prediction.json") return p.parse_args() def main(): a = parse_args() register_multimodal_autoclass() device = torch.device("cuda" if torch.cuda.is_available() else "cpu") model = AutoModel.from_pretrained(a.hf_model_dir, trust_remote_code=True).to(device).eval() with open(a.sample_json) as f: sample = json.load(f) x = embed_samples( [sample], TextEmbedder(model.config.text_model, device), ImageEmbedder(model.config.image_model, device), week=a.week, ) week = int(x["week"][0]) with torch.no_grad(): out = model( text=torch.from_numpy(x["text"]).to(device), image=torch.from_numpy(x["image"]).to(device), img_mask=torch.from_numpy(x["img_mask"]).to(device), img_age=torch.from_numpy(x["img_age"]).to(device), cutoff=week, ) result = { "sample_id": sample.get("sample_id"), "week": week, "has_image_by_week": bool(x["img_mask"][0, week - 1] > 0), "image_age_weeks": int(x["img_age"][0, week - 1]), "yield_pred_bu_per_acre": float(out.predictions[0]), } print(json.dumps(result, indent=2)) with open(a.output_json, "w") as f: json.dump(result, f, indent=2) print(f"saved prediction to {a.output_json}") if __name__ == "__main__": main()