Hello,
Thank you for releasing the code and the 8-shot checkpoints. I am trying to reproduce the AcousticRooms pretraining run that produced xRIR_unseen.pth / xRIR_seen.pth (Table 1), and a few training details are not in the paper or README.
From train_xRIR_unseen.py I am assuming:
- optimizer AdamW,
lr=1e-3, weight_decay=1e-4
ExponentialLR with decay_epochs=50, gamma=0.1batch_size=64, num_shot=8, num_epoch=200- unseen-room split as in
treble_xRIR_dataset.py
Could you confirm whether the released checkpoints were trained with this script/schedule, and clarify the following?
-
Hardware. What GPU(s) did you use (type and count)? The seen checkpoint has a module. prefix, so it looks like DataParallel. Was that one node with several GPUs, and was the global batch still 64 (split across devices) or was the per-GPU batch 64?
-
Wall-clock / early stopping. Did you actually run all 200 epochs, or save the best-validation/test-loss checkpoint earlier? I noticed a leftover path in sim_to_real/eval_classroom.py pointing at .../ckpt/xRIR/20_best.pth. Was the AcousticRooms best checkpoint around epoch 20? Approximate time per epoch (or total training time) would also help.
-
ViT depth. §5.1 says $F_{vt}$ has 6 layers; model/xRIR.py defaults to depth=12, and the released weights match the 12-layer constructor. Which one was used for Table 1?
I would like to match your training budget rather than guess from the script defaults. Any of the above would be very helpful.
Thank you!
Hello,
Thank you for releasing the code and the 8-shot checkpoints. I am trying to reproduce the AcousticRooms pretraining run that produced
xRIR_unseen.pth/xRIR_seen.pth(Table 1), and a few training details are not in the paper or README.From
train_xRIR_unseen.pyI am assuming:lr=1e-3,weight_decay=1e-4ExponentialLRwithdecay_epochs=50,gamma=0.1batch_size=64,num_shot=8,num_epoch=200treble_xRIR_dataset.pyCould you confirm whether the released checkpoints were trained with this script/schedule, and clarify the following?
Hardware. What GPU(s) did you use (type and count)? The seen checkpoint has a
module.prefix, so it looks likeDataParallel. Was that one node with several GPUs, and was the global batch still 64 (split across devices) or was the per-GPU batch 64?Wall-clock / early stopping. Did you actually run all 200 epochs, or save the best-validation/test-loss checkpoint earlier? I noticed a leftover path in
sim_to_real/eval_classroom.pypointing at.../ckpt/xRIR/20_best.pth. Was the AcousticRooms best checkpoint around epoch 20? Approximate time per epoch (or total training time) would also help.ViT depth. §5.1 says$F_{vt}$ has 6 layers;
model/xRIR.pydefaults todepth=12, and the released weights match the 12-layer constructor. Which one was used for Table 1?I would like to match your training budget rather than guess from the script defaults. Any of the above would be very helpful.
Thank you!