kuan2/snd-remask-dgle-remote-sensing-hpc

DGLE remote sensing HPC bundle

Research bundle for the two DGLE-protocol tasks:

Task	Labeled source	Unlabeled target train	Labeled target evaluation
vh2pd	Vaihingen IRRG, 256 patches	Potsdam RGB, 2,592 patches	Potsdam RGB, 864 patches
r2u	LoveDA Rural, 1,366 images	LoveDA Urban, 1,156 images	LoveDA Urban Val, 677 images

The bundle contains prepared PNG data in six ZIP64 archives, six CSV manifests, dataset metadata, the ImageNet ResNet101 initialization, and the Python/Shell code needed for source-only evaluation and the later SRMP experiment. Raw ISPRS/Zenodo download archives and local training results are excluded.

Restore on HPC

Download into a clean project directory:

hf download kuan2/snd-remask-dgle-remote-sensing-hpc \
  --repo-type dataset --local-dir /path/to/snd-remask-remote
cd /path/to/snd-remask-remote
python3 -m pip install -r requirements.txt
python3 scripts/restore_hf_remote_bundle.py

The restore script checks SHA-256 for every shipped artifact before extraction, then verifies all six dataset splits. Use a Python environment with CUDA-enabled PyTorch and the dependencies in requirements.txt.

If the HPC cannot reach Hugging Face, copy this whole folder from the local computer to the HPC, omitting only .cache/ and __pycache__/. The six ZIP archives and the ImageNet backbone are already inside it. From the copied folder, install the requirements in a CUDA-enabled Python environment and run python3 scripts/restore_hf_remote_bundle.py; no dataset download is needed. Alternatively, copy the already extracted data/remote_sensing/dgle/ directory, pretrain/resnet101-5d3b4d8f.pth, core/, datasets/, train_SRMP_remote.py, train_SRMP.py, train_ReMask.py, run_remote_sensing_hpc.sh, and requirements.txt. Also copy scripts/__init__.py, scripts/remote_sensing_transforms.py, scripts/remote_sensing_results.py, scripts/verify_dgle_remote_sensing.py, and scripts/compare_dgle_source_only.py. Keep the same relative paths. In that case, skip the restore step.

To reconstruct the same splits directly from the official ISPRS and Zenodo sources instead, run bash setup_dgle_remote_sensing_hpc.sh in this directory. The script prompts for the ISPRS share password shown on the official benchmark page and downloads only the needed source archives.

Inside an HPC GPU allocation, run the DGLE source-only validation first:

PYTHON=python3 bash run_remote_sensing_hpc.sh check
PYTHON=python3 SEED=1 bash run_remote_sensing_hpc.sh source vh2pd
PYTHON=python3 SEED=1 bash run_remote_sensing_hpc.sh source r2u

This does not run SRMP adaptation. It compares the trained source model's target-domain IoU against the Source-Only rows in DGLE Tables 1 and 2. The published DGLE source-training procedure leaves some details unspecified, so this is a documented attempt at replication, not the authors' official code.

For the later SRMP experiment, see REMOTE_SENSING_DGLE_SETUP.md; only start that after reviewing the source-only results.

The direct Bash launcher accepts adapt after source training. With the same seed and source settings, REL_WEIGHT=0 runs the pseudo-label baseline and REL_WEIGHT=0.1 runs SRMP. It needs a CUDA GPU allocation but does not use sbatch or Slurm directives.

Portable results (without weights)

After final evaluation, both source-only evaluation and adaptation write result_portable.json in their output directory. This one file contains per-class IoU, mIoU, confusion counts (GT rows / prediction columns), protocol, checkpoint hashes, training settings, and the source-only comparison when the source checkpoint and evaluation scope match. It contains no model or optimizer weights. Copy this JSON file alone to inspect or share a run.

To export a completed adaptation produced by an older version of the trainer, update train_SRMP_remote.py, run_remote_sensing_hpc.sh, and scripts/remote_sensing_results.py, then run:

PYTHON=python3 SEED=1 REL_WEIGHT=0 bash run_remote_sensing_hpc.sh export r2u

This evaluates the saved EMA teacher without training. The report is written to results_remote_sensing/r2u/rs_standard/crop512_batch4_accum1_source20000/seed1/rel0/srmp/result_portable.json. Use the same task, seed, and run-path settings as the original run. The method and iteration in the report come from the saved checkpoint, not eval CLI defaults.

Sources and usage

• ISPRS benchmark: Vaihingen and Potsdam imagery. The project owner states they have permission to distribute these prepared data. Publication of this bundle does not grant onward redistribution rights.
• LoveDA on Zenodo: Wuhan University RSIDEA; academic use only, noncommercial, CC BY-NC-SA 4.0 for the dataset, subject also to Google Earth terms.
• DGLE paper: experimental protocol and source-only reference scores.

Only source labels and target evaluation labels are present. Target training manifests contain no label paths and target train directories contain no masks.