Hello,
I'm working on estimating the most probable parameter value to produce a specific observation using SNPE-C. For this, I'm using NSF as the density estimator, configured as follows: posterior_nn(model="nsf", hidden_features=50, num_transforms=1). My approach involves fitting on two descriptive statistics. I've set the learning rate at 1e-6 (too low). Other relevant settings include a validation_fraction of 0.25, a total of 80_000 simulations, a training batch size of 64, and stop_after_epochs=20. The prior set on the parameter is a HalfNormal. Here's the resulting PDF:
The train/validation loss curve:
Interestingly, the performance on the validation set is better than that on the training set in the same round and never overfits. This looks like an error in my pipeline but when I run it using slightly different simulators it disappears. Any insights into why this might be happening would be greatly appreciated.
Hello,
I'm working on estimating the most probable parameter value to produce a specific observation using SNPE-C. For this, I'm using NSF as the density estimator, configured as follows:
posterior_nn(model="nsf", hidden_features=50, num_transforms=1). My approach involves fitting on two descriptive statistics. I've set the learning rate at 1e-6 (too low). Other relevant settings include avalidation_fractionof 0.25, a total of 80_000 simulations, a training batch size of 64, andstop_after_epochs=20. The prior set on the parameter is aHalfNormal. Here's the resulting PDF:The train/validation loss curve:
Interestingly, the performance on the validation set is better than that on the training set in the same round and never overfits. This looks like an error in my pipeline but when I run it using slightly different simulators it disappears. Any insights into why this might be happening would be greatly appreciated.