File size: 6,890 Bytes
1e41dff | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 | # GPU experiments for the Geoderma rebuttal (GEODER-D-26-01032)
Deadline: **2026-05-29**.
## Context
These two experiments address the only remaining GPU-dependent reviewer
requests for this revision. Every CPU-based analysis is already complete
in `rebuttal/`; the headline numbers (Table 2 bootstrap CIs, β_year
sensitivities, NN-distance distribution, residual-SD breakdown, land-use
regression, spatial-vs-random split comparison, T3.1 multi-run spatial-CV
evidence) live in `rebuttal/rebuttal_numbers.md`.
| Reviewer concern | Experiment |
|------------------|------------|
| **R1.3** single spatial split is weak | Experiment 1 |
| **R3.6** no confidence intervals on Table 2 | Experiment 1 |
| **R3.8** no proper test set | Experiment 1 |
| **R3.9** no uncertainty quantification | Experiment 2 |
| **R4.4** uncertainty maps expected in DSM | Experiment 2 |
The two existing checkpoints (Model A, Model B) and which experiment
uses which are documented in `checkpoint_note.md`. **No retraining of the
existing checkpoints is needed.** Experiment 1 trains 5 fresh models
from scratch (one per fold). Experiment 2 uses the already-trained
Model B for stochastic inference only.
## Files
```
rebuttal/gpu_experiments/
├── README.md ← this file
├── checkpoint_note.md ← Model A vs Model B disambiguation
├── spatial_kfold/
│ └── run_kfold.py ← Experiment 1
└── uncertainty/
├── mc_dropout_inference.py ← Experiment 2, step 1: inference
└── plot_uncertainty.py ← Experiment 2, step 2: figure
```
After running, the outputs land in those same directories alongside the
scripts (see each script's docstring for the full output list).
## Experiment 1 — Spatial 5-fold CV (≈ 15 GPU hours)
- **Script:** `spatial_kfold/run_kfold.py`
- **Command:**
```bash
cd /home/valerian/SGTPublication
python rebuttal/gpu_experiments/spatial_kfold/run_kfold.py
```
- **Outputs (for the manuscript):**
- `spatial_kfold/kfold_results.md` — paste this Table into manuscript §2.5
- `spatial_kfold/figure_kfold.png` — new Figure (supplement or main)
- `spatial_kfold/kfold_predictions_all_folds.parquet` — per-row test
predictions (lon, lat, OC_actual, OC_predicted, fold_id, year, altitude)
- **Answers:** R1.3, R3.6, R3.8
## Experiment 2 — MC Dropout uncertainty (≈ 3 GPU hours)
- **Scripts:**
- `uncertainty/mc_dropout_inference.py` — run first; writes GeoTIFFs + parquet
- `uncertainty/plot_uncertainty.py` — run second; builds the 3-panel figure
- **Commands:**
```bash
cd /home/valerian/SGTPublication
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py
python rebuttal/gpu_experiments/uncertainty/plot_uncertainty.py
```
- **Outputs (for the manuscript):**
- `uncertainty/SGT_1mil_2023_mean_mc30.tif` — MC mean prediction (UTM 32N, 250 m)
- `uncertainty/SGT_1mil_2023_std_mc30.tif` — MC uncertainty (std), same grid
- `uncertainty/figure_uncertainty_3panel.png` — Figure for §4.6 (300 dpi)
- `uncertainty/figure_uncertainty_3panel.pdf` — vector for journal submission
- **Answers:** R3.9, R4.4
## Parallelisation — built in
Both scripts shard themselves across every visible CUDA device with no
extra launcher. Plain `python script.py` is enough:
- **Experiment 1** orchestrates one worker subprocess per GPU and
schedules the 5 folds across them (4 folds in parallel, then fold 4
on the first free GPU). Total wall time ≈ 2 × (single-fold time)
on 4 GPUs.
- **Experiment 2** shards the 1.3 M Bavaria inference grid into
equal-sized contiguous slices, one per GPU, and concatenates the
shards back into a single GeoTIFF/parquet at the end. Total wall
time ≈ (single-GPU time) / N_GPUs.
Override with `--gpus 0,1` (subset) or `--sequential` (debug). Per-GPU
stdout lands in `{spatial_kfold,uncertainty}/worker_logs/<id>_gpu_<g>.log`.
If you really want to run both experiments concurrently on disjoint
GPU subsets:
```bash
python rebuttal/gpu_experiments/spatial_kfold/run_kfold.py --gpus 0,1 &
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py --gpus 2,3 &
wait
```
The k-fold checkpoints are saved per fold so a crash on fold N does
not invalidate folds 0..N-1.
## What to do with the outputs
- **`kfold_results.md`** → paste the table into manuscript §2.5 and cite
the row in the responses to R1.3 and R3.6.
- **`figure_kfold.png`** → new Figure in the supplement (or main text if
the editor agrees).
- **`figure_uncertainty_3panel.png`** → new Figure in manuscript §4.6.
- **All four** → referenced in the response letter with explicit page /
line numbers.
## Important assumptions flagged in the scripts
Every `# ASSUMPTION:` comment in the two scripts marks a place where the
spec the user provided and the actual codebase disagree, or where a default
needed to be chosen. Skim them before launching — they are the only
places where the project lead might want to override the default.
The four notable ones:
1. **Optimizer / lr.** Spec said Adam @ 1e-3 + exponential decay.
`train.py` uses Adam @ 2e-4 with no scheduler. Scripts follow `train.py`
(the spec's overriding instruction was "match the original exactly").
2. **Log transform.** Spec said `log1p` / `expm1`. `train.py` uses
`torch.log(OC + 1e-10)` / `np.exp(pred)`. Scripts follow `train.py`.
3. **Model class.** Spec referred to "the SGT 1.1M". `train.py` imports
`SimpleSGT` but the saved 1.1 M checkpoints are `EnhancedSGT` —
`running.py` has `from EnhancedSGT import EnhancedSGT as SimpleSGT`.
Scripts use `EnhancedSGT` (verified by state-dict shape and parameter
count = 1,121,637).
4. **Normalisation in MC dropout.** The MC accumulator is in *original
SOC units* (the log transform is inverted inside the MC loop). The
spec said this explicitly; flagging here only to note it.
## Sanity checks before running
```bash
# 1. confirm Model B exists at the documented path
ls -la /home/valerian/SGTPublication/Weights-ResidualsModels-MappingInference-SOCmapping/TemporalFusionTransformer/finalResults2023_1milVersion_TRANSFORM_log_LOSS_l1/
# 2. confirm model_ready_dataset.parquet exists
ls -la /home/valerian/SGTPublication/rebuttal/model_ready_dataset.parquet
# 3. confirm rasterio + pyproj are importable (for Experiment 2 GeoTIFFs)
python -c "import rasterio, pyproj; print(rasterio.__version__, pyproj.__version__)"
# 4. confirm the EnhancedSGT module imports and parameter count matches
python -c "import sys; sys.path.insert(0, '/home/valerian/SGTPublication/SOCmapping/SpatiotemporalGatedTransformer'); from EnhancedSGT import EnhancedSGT; m = EnhancedSGT(input_channels=6, height=5, width=5, time_steps=5, d_model=128); print(sum(p.numel() for p in m.parameters() if p.requires_grad))"
# expected: 1121637
```
|