Epoch and Steps in Configs dont match Paper

[DominikVincent]

Hello,

I noticed a discrepancy between the training details in the paper and the provided config files, and I’d like to get clarification on how you trained your model:

On the one hand, you say:

But your config files all stages use 300 epochs.

On the other hand you say:

that your model is trained between 160k and 300k iterations, but according to the logging, that would indicate that you train for less than 12 epochs:

Could you clarify your training setup and update the config files to match the ones actually used for your results? Otherwise, it will be difficult for the community to fully benefit from and verify your contribution.

[BoifZ]

Thanks for the question! In our setup, each epoch contains ~800 batches. Therefore, 100 epochs = 80k iterations.

During the T2M pre-training, we train MotionGPT3 for 200 epochs (~160k iterations) and select the checkpoint with the lowest FID as the best model.

The training configuration reported in our paper reflects the evaluation setting, and the config file fields are left at their defaults. We recommend monitoring the validation curve during your first few runs and adjusting the total training based on the training dynamics and your specific needs.

[DominikVincent]

Thanks for the reply!

Just to double-check with you:

• The dataset contains 25,560 training samples and 1,530 validation samples (val dataset loading done 1530).
• In stage 1, I also observe ~800 batches/epoch with 1 GPU, and ~400 with 2 GPUs.
• In stage 2.1 (MoT_vae_stage2_all.yaml), I see ~2400/1200 batches per epoch for 1/2 GPUs respectively.
• In stage 2.2 (MoT_vae_stage2_instruct.yaml), I see 14398/7199 batches per epoch for 1/2 GPUs respectively.

Is that aligned with your setup? Do you explicitly adjust the batch size per GPU in your configs (e.g., 2 GPUs → batch size 16 per GPU → total batch size 32)?

---

For the training, could you provide exactly the config files that you use in your final experiments, as the available config parameters vastly exceed the ones that you mention in the paper?

Based on your README and paper, is the training procedure as follows (assuming a total batch size of 32, i.e., 32 on a single GPU or 16 per GPU when using 2 GPUs)?

Stage	Command	Epochs / Steps	Notes
1	python -m train --cfg configs/MoT_vae_stage1_t2m.yaml --nodebug	200 epochs (≈160k steps, 800 batches/epoch)	All defaults in config except epochs changed to 200
2 (part1)	python -m train --cfg configs/MoT_vae_stage2_all.yaml --nodebug	125 epochs (≈300k steps, 2400 batches/epoch)	All defaults in config except epochs changed to 125
2 (part2)	python -m train --cfg configs/MoT_vae_stage2_instruct.yaml --nodebug	≈80k steps, with (14398 batches per epoch -> 6 epochs)	Epochs and stage not explicitly mentioned in paper, assumed 80k iterations
3	python -m train --cfg configs/MoT_vae_stage3.yaml --nodebug	≈80k steps	Epochs not explicitly mentioned in paper, assumed 80k iterations as stated for all other training

All stages always take the last pretrained model from the previous stage.
Could you please confirm these settings, or share the exact config files you used to produce the final results in the paper, such that it is possible to recreate and verify your results?

Also, could you clarify why stage 2 is split into two separate trainings (multitask cross-modal learning vs. instruction finetuning)? Is this necessary?

Finally, monitoring validation is not straightforward: your code does not compute a validation loss, and the reported metrics don’t converge (see Issue #19). This makes it hard to follow your recommendation of “monitoring the validation curve.”

I appreciate your clarification; it will help a lot for reproducibility. Thanks!

[BoifZ]

Thanks for the detailed summary.

1. Training setup. Your breakdown of the training stages matches the configuration we used for the final training.

2. Batch size & iterations. We set batch_size=32 per GPU. With two GPUs, the effective batch size is 64. You can verify this by the iterations per epoch: with 1 GPU we observe ~800 batches/epoch, and with 2 GPUs ~400 batches/epoch—consistent with your note:

• In stage 1, I also observe ~800 batches/epoch with 1 GPU, and ~400 with 2 GPUs.

3. Stage 2 (cross-modal alignment). For Stage 2 we follow the multi-task training scheme proposed and validated in MotionGPT (task pretraining → instruct-tuning) for cross-modal alignment.

4. On No Validation Losses? #19 . We’ve replied there—please check the updated metric curves for the generated motions. In the paper we also report and plot R-Precision@3 and Multimodal Distance curves.

If anything looks off on your side, feel free to share the exact command/log so we can double-check.