Diveshj/sorghum-o1-foundation-classification

O1 -- Foundation Classification + Knowledge Distillation

97.87% Macro-F1 -- statistically ties its 27.87M-param teacher (p=0.52) at 1/7th the parameters. Multi-seed testing found teacher selection, not the loss-function tweak, drives most of the gain -- full story linked below.

What's in this folder

O1/
├── og_models/        <- original PyTorch .pt checkpoints, organized by task
├── onnx_models/       <- same checkpoints exported to ONNX (framework-independent)
├── results_csv/       <- training history, significance testing, and summary CSVs from every Kaggle source
├── sample_images/     <- a few real test images to try inference with
├── mapping.json       <- machine-readable index of every checkpoint
├── MAPPING.md         <- human-readable version of the same
├── infer_example.py   <- reference script: load an ONNX model, run a real prediction
├── test_inference.py  <- test harness: runs every ONNX model against the sample set,
│                        reports accuracy against ground truth
└── requirements.txt   <- minimal deps to run both scripts

102 checkpoints across 4 task categories: cnn_baselines, kd_students, vit_models (all from the original fixed-5-epoch-budget study), and convergence_checkpoints (new, 2026-09/10 -- the same recipes retrained to actual convergence with early stopping, to check whether 5 epochs had already converged; not yet exported to ONNX). See MAPPING.md's convergence_checkpoints section for the headline results: convergence training significantly beats the 5-epoch budget on every model tested (9 of 10 comparisons significant, p<0.05 to p<1e-24), resolves this project's previously-open SE-ResNet-50 anomaly (94.23% vs. 80.49% Macro-F1, McNemar p=1.8e-83, vs. GrainSet's own published 96.10%), and found the single largest convergence gain anywhere in the project on the plain EfficientNet-B0 baseline (+10.41pp, 88.14% -> 98.55%, p=7.2e-25) -- a caution flag for any KD-method comparison measured against that baseline.

Note: val_macro_f1 in MAPPING.md is only available for 57/82 of the original checkpoints -- not every training run saved it inside the checkpoint file itself. A missing value doesn't mean the checkpoint is untested, just that this one number isn't recorded in the file. The 10 convergence_checkpoints instead report test_macro_f1 in mapping.json (measured directly, not pulled from the checkpoint file).

Preprocessing (verified against the actual training notebooks, not assumed)

• Resize to 224x224, RGB
• Normalize with mean=[0.308562, 0.251994, 0.187898], std=[0.240441, 0.197289, 0.149387] (dataset-specific stats computed from GrainSet -- NOT generic ImageNet stats)
• Class order: ['NOR', 'F&S', 'SD', 'MY', 'AP', 'BN', 'HD', 'IM'] (index 0-7, argmax of the model output = this index)

Quick start

pip install -r requirements.txt
python infer_example.py      # single model, single image demo
python test_inference.py     # every model vs. the full sample set, with accuracy

See MAPPING.md for the full checkpoint list, architecture, and validation Macro-F1 per file.

Full results and discussion

../../O1 context folder/LAtest ccontext/05_o1_full_story_results_and_assessment.md -- the complete write-up this headline number comes from (plain path, not a clickable link -- the project's folder names have spaces in them). ../../SUMMARY.md has the whole project's overview across all 4 objectives.