retraining emulator for NMF with burst model
changhoonhahn opened this issue · comments
As I found in model_priors.ipynb, the 4 component NMF model introduces a strong prior on the SSFR and sSFH. Adding a burst ameliorates this issue:
Based on the priors, it seems that NFM with burst should be the fiducial model. I will therefore need to retrain the emulator.
As the first step I've constructed 5x10^6 training SEDs and 100,000 SEDs for validation. I've run PCA on the training SEDs split into 3 wavelength bins (same as before) with [50, 30, 30] components. Below is the fractional error of the SEDs reconstructed from the PCA components, which demonstrates that it's accurate enough.
fractional reconstruction error for N_train=5x10^6, n_pcas=[50, 30, 30], architecture=[4x256, 4x256, 4x256]
The emulator is nowhere close to accurate enough.
Here's a checklist to improve the emulator:
- increase PCA components for wave0. For wave0 there's already ~1% reconstruction error coming from the PCA.
- try wider architecture. We used 4x256. Try with 512 Runs out of memory
- try deeper architecture. Try more than 4 layers
- increase training set. Increasing the training data will improve the reconstruction errors
Increasing the number of PCA components for wave0 reduces the PCA reconstruction errors:
n_pca = [60, 30, 30]
n_pca = [70, 30, 30]
deeper architecture results in a better fit --- still not good enough
Focusing only on the central wavelength bin:
with a 5x256 architecture
with a 6x256 architecture
training PCA using 1000 batches does not improve the PCA validation error.
n_pca = [60, 30, 30]
n_pca = [70, 30, 30]
