docs/gpu_experiments/QUICKSTART.md

SOCmapping rebuttal — Runpod quickstart (copy-paste)

Single-page, command-only. For background / troubleshooting see RUNPOD_SETUP.md (longer doc) and checkpoint_note.md.

Pod requirement: ≥ 60 GB /workspace volume, any NVIDIA GPU with ≥ 12 GB VRAM (RTX 3090 / 4090 / A100 / L40S all work).

---

1 — System prep

apt update && apt install -y \
    git git-lfs curl unzip build-essential \
    python3.10-venv python3-pip
git lfs install
nvidia-smi | head -5

python3.10-venv is mandatory — Ubuntu's stock python3.10 doesn't include venv and bash setup_venv.sh will fail with No module named ensurepip if it's missing.

2 — Clone code

mkdir -p /workspace/SOC && cd /workspace/SOC
git clone https://github.com/ValerianFourel/SOCmapping.git
cd SOCmapping && git log --oneline -5

3 — Install HF CLI

pip install --upgrade "huggingface_hub[cli,hf_transfer]"
export HF_HUB_ENABLE_HF_TRANSFER=1
echo 'export HF_HUB_ENABLE_HF_TRANSFER=1' >> ~/.bashrc

4 — Download dataset (≈ 17 GB zip → ≈ 25 GB unzipped)

mkdir -p /workspace/SOC/Data && cd /workspace/SOC/Data
hf download ValerianFourel/SOCmappingRastersAndSoilSamples \
    SOCmappingData.zip --repo-type dataset --local-dir .
df -h /workspace                              # confirm ≥ 30 GB free
unzip -q SOCmappingData.zip
rm SOCmappingData.zip
ls /workspace/SOC/Data/Data                   # the zip nests under Data/

5 — Download model weights (≈ 10 MB)

mkdir -p /workspace/SOC/Weights && cd /workspace/SOC/Weights

# Model A (eval / Table 2 baseline)
hf download ValerianFourel/Weights-ResidualsModels-MappingInference-SOCmapping \
    --include "TemporalFusionTransformer/residualModels1mil_normalize_composite_l2_v2/*" \
    --local-dir .

# Model B (operational mapping — input to Experiment 2)
hf download ValerianFourel/Weights-ResidualsModels-MappingInference-SOCmapping \
    "TemporalFusionTransformer/finalResults2023_1milVersion_TRANSFORM_log_LOSS_l1/TFT_model_BEST_OVERALL_from_run_1_MAX_OC_150_TIME_BEGINNING_2007_TIME_END_2023_TRANSFORM_log_LOSS_l1_R2_1.0000.pth" \
    --local-dir .

find . -name "*.pth" -size +1M                # expect 2 .pth files

6 — Tell the codebase where things live (no more symlinks)

config.py and the GPU scripts read the four abstract roots from env vars + a walk-up fallback (SOCmapping/_paths.py). On Runpod where the clone lives at /workspace/SOC/SOCmapping/, the walk-up resolves the sibling Data/ and Weights-…/ automatically — usually no env vars needed.

If your layout differs (e.g. SOC_DATA_DIR on a separate volume), set the relevant overrides. Add them to ~/.bashrc so they persist:

# Project root containing Data/, Weights-…/, SOCmapping/ as siblings
export SOC_PROJECT_ROOT=/workspace/SOC

# Or per-component (overrides PROJECT_ROOT for that one)
# export SOC_DATA_DIR=/workspace/SOC/Data
# export SOC_WEIGHTS_DIR=/workspace/SOC/Weights
# export SOC_REBUTTAL_DIR=/workspace/SOC/SOCmapping/rebuttal
# export SOC_COORDS_1MIL_CSV=/workspace/SOC/Data/Coordinates1Mil/coordinates_Bavaria_1mil.csv

echo 'export SOC_PROJECT_ROOT=/workspace/SOC' >> ~/.bashrc

The dataset zip wraps everything in a top-level Data/, so on Runpod the actual data sits at /workspace/SOC/Data/Data/. Either flatten it (cd /workspace/SOC/Data && mv Data/* Data/.[!.]* . && rmdir Data) or set SOC_DATA_DIR=/workspace/SOC/Data/Data explicitly.

Verify:

python /workspace/SOC/SOCmapping/_paths.py
# Should print all four roots with ✓ markers (no ✗ MISSING)

# And from the SGT config, the actual file paths should resolve:
python -c "
import sys; sys.path.insert(0, '/workspace/SOC/SOCmapping/SpatiotemporalGatedTransformer')
import config
import os
for k in ('file_path_LUCAS_LFU_Lfl_00to23_Bavaria_OC',
          'file_path_coordinates_Bavaria_1mil'):
    v = getattr(config, k)
    print(f'{k}\n  → {v}\n  exists? {os.path.exists(v)}')"

7 — Build venv (auto-detects CUDA, ≈ 3 GB)

cd /workspace/SOC/SOCmapping/rebuttal/gpu_experiments
bash setup_venv.sh

Expected last line: forward pass OK, output shape: (2,) device=cuda:0.

8 — Sanity-load Model B

source /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/.venv/bin/activate
python - <<'PY'
import sys, torch
sys.path.insert(0, '/workspace/SOC/SOCmapping/SpatiotemporalGatedTransformer')
from EnhancedSGT import EnhancedSGT
ck = torch.load(
    '/home/valerian/SGTPublication/Weights-ResidualsModels-MappingInference-SOCmapping/'
    'TemporalFusionTransformer/finalResults2023_1milVersion_TRANSFORM_log_LOSS_l1/'
    'TFT_model_BEST_OVERALL_from_run_1_MAX_OC_150_TIME_BEGINNING_2007_'
    'TIME_END_2023_TRANSFORM_log_LOSS_l1_R2_1.0000.pth',
    map_location='cuda', weights_only=False)
sd = {k.replace('module.', ''): v for k, v in ck['model_state_dict'].items()}
m = EnhancedSGT(input_channels=6, height=5, width=5, time_steps=5, d_model=128,
                num_heads=4, dropout=0.3, num_encoder_layers=3,
                expansion_factor=4).cuda()
miss, extra = m.load_state_dict(sd, strict=False)
print(f'params={sum(p.numel() for p in m.parameters() if p.requires_grad):,}  '
      f'missing={len(miss)}  unexpected={len(extra)}')
print('forward OK', m(torch.randn(4, 6, 5, 5, 5, device="cuda")).shape)
PY

Expected: params=1,120,546 missing=0 unexpected=0 forward OK torch.Size([4]).

9 — Launch experiments (auto-uses every visible GPU)

Both experiments shard themselves across all CUDA devices in the pod with no extra flags — no accelerate launch, no torchrun. Internally each spawns one subprocess per GPU and orchestrates the work.

source /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/.venv/bin/activate
cd /workspace/SOC/SOCmapping

# Experiment 2 — MC dropout uncertainty map
#   Full 1.3 M grid on 4 GPUs ≈ 45 min (vs ≈ 3 h single-GPU)
#   400 k uniform sub-sample on 4 GPUs ≈ 14 min (recommended for a draft)
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py --max-points 400000
python rebuttal/gpu_experiments/uncertainty/plot_uncertainty.py        # CPU only

# Experiment 1 — spatial 5-fold CV
#   4 GPUs → folds 0-3 in parallel, then fold 4 → ≈ 2 × 1-fold time
#   ≈ 6 h on 4 × 4090 (vs ≈ 15 h single-GPU)
python rebuttal/gpu_experiments/spatial_kfold/run_kfold.py

Useful overrides:

# Restrict to specific GPUs
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py --gpus 0,1
python rebuttal/gpu_experiments/spatial_kfold/run_kfold.py --gpus 0,2

# Force the legacy single-GPU sequential mode (for debugging)
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py --sequential
python rebuttal/gpu_experiments/spatial_kfold/run_kfold.py --sequential

# Cap the Experiment 2 inference grid (uniform stride sub-sample)
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py --max-points 400000
# or equivalently
SOC_MAX_INFERENCE_POINTS=400000 python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py

# Take every k-th row (alternative phrasing of the same sub-sampling)
python rebuttal/gpu_experiments/uncertainty/mc_dropout_inference.py --stride 3   # ~ 433k points

Per-GPU worker logs land in:

rebuttal/gpu_experiments/spatial_kfold/worker_logs/fold_<i>_gpu_<g>.log
rebuttal/gpu_experiments/uncertainty/worker_logs/shard_<i>_gpu_<g>.log

nvidia-smi -l 5 should show ≈ 100% utilisation across all visible GPUs once both experiments are running.

Live progress in % from another terminal

# Snapshot
python /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/progress.py

# Auto-refresh every 5 s (Ctrl+C to stop)
python /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/progress.py --watch

# Only MC dropout or only k-fold
python /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/progress.py --watch --mc
python /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/progress.py --watch --kfold

The script parses each worker's log tail for batch N/M (MC dropout) or Fold N | Epoch E/270 (k-fold), shows per-shard / per-fold status, and prints an aggregate %.

10 — Push outputs to the SOCrebuttal HF dataset

export HF_TOKEN="hf_xxx_paste_a_write_token_from_huggingface.co/settings/tokens"
python /workspace/SOC/SOCmapping/rebuttal/gpu_experiments/upload_to_hf.py
# → https://huggingface.co/datasets/ValerianFourel/SOCrebuttal

11 — (Optional) rsync outputs to your laptop

# From your LAPTOP, not the SSH session:
rsync -avzP --exclude=.venv \
  -e "ssh -i ~/.ssh/id_ed25519" \
  "<USER>@ssh.runpod.io:/workspace/SOC/SOCmapping/rebuttal/gpu_experiments/" \
  ~/SGTPublication/SOCmapping/rebuttal/gpu_experiments/

---

If something breaks

symptom	fix
bash: unzip: command not found	apt install -y unzip
No module named ensurepip from setup_venv.sh	apt install -y python3.10-venv
nvidia-smi: command not found	pod has no GPU — switch template
hf not found	pip install --upgrade "huggingface_hub[cli,hf_transfer]"
HF download stalls	unset HF_HUB_ENABLE_HF_TRANSFER and retry
Coordinates (lon,lat) not found in Elevation mid-training	symlink in step 6 is wrong — check it resolves to the inner Data/Data/ not the outer Data/
torch.cuda.OutOfMemoryError in MC dropout	drop BATCH_SIZE in mc_dropout_inference.py from 256 → 128
huggingface-cli: command is deprecated	use hf (CLI was renamed in huggingface_hub ≥ 0.27)
Disk full during unzip	df -h /workspace, resize pod volume to ≥ 60 GB, restart, re-run step 4