heatmap generation process

[ybupengwang]

First of all, thank you very much for your contribution and for sharing your code — it has been incredibly helpful for my research and understanding.

While going through the implementation, I had a few questions regarding the heatmap generation process.

Specifically, I noticed that when using static heatmap regression and calling get_cepha_heatmap_datasets() to generate heatmaps in advance, the CPU usage spikes to 100% on my Windows 10 system. Additionally, when using get_cepha_landmark_datasets() with GaussianHeatmapGenerator(learnable=False), the heatmaps are generated repeatedly during the train_epoch phase, which significantly impacts training efficiency.

I was wondering if there might be a better or recommended way to handle this。Looking forward to your reply!

Best regards

[jefjonkers]

Hi, thank you so much for your kind words and for taking the time to dive into the code. I'm happy to hear it's been helpful for your research!

Regarding your questions:

• CPU spike when using get_cepha_heatmap_datasets(): You're absolutely right. The spike is due to the fact that the heatmap generation is performed at that exact moment, as the static heatmaps are being computed and cached ahead of time. Unfortunately, there's currently no built-in workaround for this — it's a trade-off between preprocessing upfront vs. doing it on-the-fly during training.

• Repeated heatmap generation with get_cepha_landmark_datasets() and GaussianHeatmapGenerator(learnable=False): This is expected behavior in the current setup, where the heatmaps are generated dynamically during each training epoch. If training efficiency is a concern, a recommended approach is to precompute the heatmaps once using get_cepha_heatmap_datasets() and store them in memory (RAM), allowing for quick access during training. This can help avoid redundant computation during each epoch. However, at the moment, we don't support saving the precomputed heatmaps to disk — but that's definitely a useful feature worth considering for future improvements. It would reduce memory pressure and avoid regeneration during each run.

Best regards