Reconstruction Issue

[SanghoYum]

Hello, thank you for releasing the TinyMyo code and paper.

I have been running several experiments with TinyMyo for about two months. I am consistently observing an issue where the model reconstructs only a very weak waveform shape in the masked regions, but fails to recover the actual amplitude of the sEMG signal.

Problem

During masked EMG reconstruction, the reconstructed signal in the masked interval shows a very small waveform-like pattern, but the amplitude is almost collapsed near zero. The model seems to capture the rough temporal shape, but not the amplitude scale.

What I tried

To rule out implementation differences, I followed the TinyMyo setup as closely as possible:

• Used the original TinyMyo code without modifying the model architecture
• Matched the reported hyperparameters and library versions as much as possible
• Preprocessed the datasets according to the preprocessing methods described in the paper
• Ran pretraining on:
  • Ninapro DB6
  • Ninapro DB7
  • EMG2Pose

However, the same amplitude-collapse behavior appears across these datasets.

Observed behavior

The reconstruction loss decreases during training, but visually the reconstructed masked regions do not recover the original amplitude. The output remains much smaller than the target signal in amplitude.

Expected behavior

I expected the model to reconstruct both the temporal waveform shape and the amplitude scale of the masked EMG segment, similar to the reconstruction examples shown in the paper.

Questions

1. Is this amplitude-collapse behavior expected during TinyMyo pretraining?
2. Was any additional denormalization or post-processing used when visualizing the reconstruction results in the paper?
3. Could this issue be related to the normalization strategy, reconstruction target scale, or loss formulation?
4. Could you share the exact config/script used to generate the reconstruction visualization in the paper?
5. Are there any dataset-specific settings for Ninapro DB6, DB7, or EMG2Pose that are not included in the current repository?

I would be happy to provide training logs, reconstruction plots, and config files if needed.

Thank you.