* Vassilis – not many updates, may be small inconsistency in dataset given to amin – not the dataset given to paper. Will retrain model and find which results match paper.
* Patrick – in repository have a database of named models
* Vassilis – models may be quite big – 250mb – perhaps can save only weights
* Patrick – can we set seed so that model can be reproduced identically each time. Should be possible. Control random number generation.
* Patrick – in repository have a database of named models.
* Vassilis – models may be quite big – 250mb – perhaps can save only weights.
* Patrick – can we set seed so that model can be reproduced. identically each time. Should be possible. Control random number generation.
* Can save models on gitLab large file storage.
* Perhaps also, record scripts and setting for models that don’t work very well.
* Patrick: Performance poor with linear and circularized signal? * Not yet tested.
* Cusps linearly polarised.
* Cusps known really accurately – good to check why they don’t work with our method, but our method will not be the best.# Actions:
* Cusps known really accurately – good to check why they don’t work with our method, but our method will not be the best.
2.**Amin:**
* Amin – managed to train mixed model. Performance is poor. Perhaps model is bad. Accuracies around 90%.
3.**Michael:**
...
...
@@ -44,27 +44,27 @@
5. Report back to GraceDB
* Patrick - Need to get a review. – March pipeline deadline.
4.**Wasim:**
* In classification case look for pixel number – discreet
* In regression case look for RA and Dec
* In classification case look for pixel number – discreet.
* In regression case look for RA and Dec.
* Patrick – pixel map, in order to have probability.
* Michael - Map to distribution rather than single pixel.
* Perhaps try to reconstruct sky-map onto sphere.
* Wassim: Optimisation – perhaps have a model to predict ideal resolution.
* Feed input into another pipeline (?) to generate sky map and use as label.
* One way to do it – apply linear combination to get NULL stream. Make an image on sphere.
* Autoencoder from strain to latent space then to spherical image
* Autoencoder from strain to latent space then to spherical image.
# Actions
| Responsible | Action | Deadline |
|-------------|--------|----------|
| Vassilis | Store models on gitlab (?) | N/A |
| Michael | Set up trello | N/A |
| Vassilis | Store models on gitlab (?). | N/A |
| Michael | Set up trello. | N/A |
| Michael | Perform preliminary mass parameter estimation on CBCs. | N/A |
| Patrick | Set up draft of structure on wiki | N/A |
| Patrick | Set up draft of structure on wiki. | N/A |
| Patrick | Send glitch data and documentation.| N/A |
| Patrick | Dig up the requirements for a review readiness presentation. | N/A |
| Vassilis | Give people access to pipeline wiki/ | N/A |
| Wasim | Keep working toward single pixel case | N/A |
| Wasim |Look at deep sphere papers| N/A |
| Patrick | Dig up the requirements for a review readiness. presentation. | N/A |
| Vassilis | Give people access to pipeline wiki. | N/A |
| Wasim | Keep working toward single pixel case. | N/A |
| Wasim |Look at deep sphere papers. | N/A |
| Amin | Run false alarm rate and efficiency tests with current trained networks. | N/A |
