Missing S6 (new→old) config/checkpoint and mismatch between released S5 YAML and checkpoint config

[tzcinhust]

Hi, thank you for releasing SADA.

I am trying to reproduce the EPIC-Kitchens S6 result in Table 1 (new → old):

• tIoU 0.1–0.5: 23.94 / 22.95 / 21.47 / 19.16 / 15.24
• Average mAP: 20.55

I used repository commit:

a96a15b0d0c905a553c7a2b14f607c9485ebe758

However, I could not find an author-released S6 (new_vs_old) configuration or checkpoint. The released checkpoint package appears to contain S1, S2, S4, S5, and CharadesEgo, but not S3 or S6. The repository contains configs/old_vs_new/version_final_ckpt.yaml, but no corresponding new_vs_old configuration.

I also found that the configuration embedded in the official S5 checkpoint:

best_ckpt_old_vs_new/saved_ckpt/best_ckpt/checkpoint-epoch10.pt

differs substantially from the released:

configs/old_vs_new/version_final_ckpt.yaml

Parameter	S5 checkpoint config	Released YAML
dim_proj_adv	128	128
pl_threshold	0.55	0.3
conditional_adv_drop_rate	0.5	0.1
loss_weights.weight_adv	3	2
loss_weights.weight_downstream	3	2
disc_learning_rate	0.02	0.004
beta	1	0.5
weight_source_cond_adv	0.6	0.75
weight_target_cond_adv	0.6	0.75
Level weights	[0.3, 0.3, 0.2, 0.4, 0.3, 0.9]	[0.6, 0.1, 1.0, 0.6, 0.2, 0.4]

The tensor shapes in the official S5 checkpoint also confirm that its discriminator width is 128:

• Standard discriminator: (128, 512) → (128, 128) → (1, 128)
• Conditional discriminator: (128, 1024) → (128, 128) → (1, 128)

For S6, I reconstructed the direction {EK100 \ EK55} → EK55 and used the supplementary Table A parameters:

• Level weights: [0.6, 0.1, 1.0, 0.6, 0.2, 0.4]
• alpha / pl_threshold = 0.3
• weight_adv = 2.0
• Seed 6
• Checkpoint selection based only on source-validation downstream loss

Results so far:

• Width 128: target-training Avg mAP was 18.30 at epoch 5 and 17.91 at the source-validation-selected epoch 10 checkpoint. Later epochs declined.
• Width 512 controlled run: target-training Avg mAP was 18.19 at epoch 5.

I also observe differences between the released annotation package and the S6 split counts reported in the paper appendix:

• Paper source: train 150 / validation 44
• Released package source: train 157 / validation 37
• Paper target: train 342 / validation 90
• Released package target: train 345 / validation 87

Could you please clarify or share:

1. The exact original S6 new_vs_old YAML/configuration used for the paper.
2. The official S6 checkpoint, if available.
3. The annotation/split version used to obtain the S6 train/validation counts reported in the paper.
4. Whether configs/old_vs_new/version_final_ckpt.yaml is intended to reproduce S5, given that it differs from the configuration embedded in the official S5 checkpoint.
5. The exact dim_proj_adv used for S6. The paper architecture description mentions width 512, while the official S5 checkpoint uses 128.
6. Whether additional scenario-specific parameters not listed in Table A are required to reproduce S6.

Any guidance or released artifacts would be greatly appreciated. Thank you!