yanguanyue/DFTA-10000-steps

DFTA checkpoints

Full training checkpoint backup for the DFTA main method, legacy and complete ablation variants, and cross-dataset DFTA runs. Existing published checkpoint paths are preserved. This is an artifact backup, not a claim that the current local source tree equals a particular GitHub revision.

Source code: https://github.com/yanguanyue/DFTA- Baseline checkpoints: https://huggingface.co/yanguanyue/DFTA-baselines Generated images and masks (private; authorized access required): https://huggingface.co/datasets/yanguanyue/DFTA-generated-results

Main experiment: original HAM10000 10,000-step release

The trained checkpoint for the original DFTA main experiment is lightning_logs/version_4/checkpoints/last.ckpt. This is the original published 10,000-step DFTA model, as documented before the backup expansion. The version_4 directory is its training-run identifier.

File in this repository	Role in the main experiment	Original local path relative to EI/
lightning_logs/version_4/checkpoints/last.ckpt	Trained DFTA main-experiment checkpoint (10,000 steps); use this for the original main-method model	DFTA/checkpoint/flow/lightning_logs/version_4/checkpoints/last.ckpt
PRETRAINED/merged_pytorch_model.pth	Pretrained initialization from the public Siamese polyp-segmentation model; it is the starting weights, not the trained DFTA result	DFTA/checkpoint/flow/PRETRAINED/merged_pytorch_model.pth
lightning_logs/version_4/hparams.yaml	Supporting metadata for that same training run	DFTA/checkpoint/flow/lightning_logs/version_4/hparams.yaml

The original lightning_logs/version_4/events.out.tfevents.* and root loss_log.txt are training logs retained from the earlier publication, not additional checkpoints.

Published main checkpoint: 12,714,275,350 bytes; SHA-256: 4bad4bfb663ca646f347a511847f824501361cb82cacdbc836f2fa5ecb9e2421.

Published initialization: 5,719,202,593 bytes; SHA-256: 030933a07d210c9dad35bc2793b28fcc106e6b5866a3d29e825707a2303cb43a.

Distinguishing the added experiment backups

The original main-experiment paths above are preserved. The backup expansion adds the following separate categories; these folders are populated incrementally while uploading is in progress.

Path prefix	Experiment category
ablations/complete/	Complete ablation-study checkpoint variants
ablations/legacy/	Earlier ablation-study checkpoint variants
cross-dataset/isic2017/isic2017_compare_20260729/flow/	DFTA main method trained for the ISIC2017 cross-dataset experiment
cross-dataset/dermofit/dermofit_compare_20260805/FlowMain/	DFTA main method trained for the Dermofit cross-dataset experiment
backups/2026-10-23/ (only if needed)	A local backup that differs from an already published file; exact source-to-backup mappings are recorded in the completed manifest

For the original HAM10000 main experiment, select the version_4/checkpoints/last.ckpt linked above. Ablation variants and cross-dataset runs have their own prefixes. Comparison-method checkpoints are in the separate DFTA-baselines repository.

Layout and scope

• Existing root paths: the originally published DFTA checkpoint and pretrained initialization (reused only after matching SHA-256).
• ablations/complete/ and ablations/legacy/: distinct ablation runs, retaining original variant names.
• cross-dataset/: DFTA runs for ISIC2017 and Dermofit.
• backups/2026-10-23/: any artifact whose old published path has different content; existing files are not replaced.
• backup_manifest.json: the completed backup's original relative paths, byte sizes, SHA-256 checksums, and file metadata.

Full checkpoint states are preserved without loading, converting or stripping optimizer states. Additional model configuration files are included when present. Metrics, evaluation tables, logs, training previews, source code, raw datasets and credentials are outside this backup. Pre-existing repository files remain available.

License and use

DFTA-owned weights and outputs are CC BY-NC 4.0, with attribution to yanguanyue and the linked DFTA source repository. Any incorporated base models and other third-party material retain their upstream licenses; this card grants no additional rights over them. Outputs are research artifacts and are not intended for clinical use. Checkpoints may require the original software environment and upstream dependencies.

Backup status and restoration

Files are uploaded, committed and checked individually. A completed backup_manifest.json and the backup-2026-10-23 tag are published only after every selected artifact in this repository has been verified. Until then the repository may contain an incomplete backup. The download helper in restore/ restores missing files, verifies checksums and preserves conflicting existing files.