@aiff22 Hi，
I try to train with my own data, but when level1 and level0, the MS-SSIM loss be Nan (actually MS-SSIM loss is not used at level1).
So I run the training code in debug mode, I found that, in some case, the SSIM could be negative.
I found the code of MS-SSIM implements at jorge-pessoa/pytorch-msssim and discussions about this problem

Following on from #44, there's a number of things that should be improved.

Easy enough:

• more documentation
• downsampling on activity and attenuation image should be made default in ScatterSimulation
• enabling of FBP recons (for scanners without gaps)
• rename some classes to include the PET name (as wouldn't work for SPECT etc)
• add a simple test for ScatterEstimation
• don't har

The next generation autograd is now called jax, and is built by the same guys who built autograd + more, with a somewhat nicer API. One other attractive feature is the ability to JIT compile functions to CPU, GPU and TPU!

As such, I'd like to propose switching over from autograd to jax in requirements.txt. Naturally, this is just a suggestion; I simply thought I'd check out the repo having

As tests are an essential part, it needs to be documented how to add them. Structure could be

1. Overview of testing strategy
   CMake, Travis
2. Adding new tests
   a. C++
   b. Python
   c. MATLAB
3. Brief info on Travis

I think some of the generic text can just be copied from the SIRF paper. In any case, keep this brief!

OPTICS (Ordering Points To Identify the Clustering Structure) is a clustering algorithm similar to DBSCAN. DBSCAN's major weakness is density tuning. OPTICS attempts to address this issue by ordering points and choosing the best epsilon.

We currently have an incomplete OPTICS implementation at [utils/clust