Loads of small things, all of which broke everything before getting to a better place
- Max: working autoencoder as dataredux and interpolation up
- Max: tutorial on simple regression feature comparison
- SC: working autoencoder as dataredux (the other part will come soon here)
- outputLayer is now tapOut. Its numbering is now different: 0 counting from the input layer, so 1 is the first hidden layer and -1 is the default (the last, like all our interface).
- it has a new friend, tapIn, which allows you to feed the data for predict and predictpoint in the middle somewhere. 0 is the input, and 1 is the first hidden layer
Normalize and Standardize
- now have an ‘inverse’ parameters for transform and transformPoint to allow query from the transformed space to the original.
Now returns/passes the variance, aka the fidelity of the new representation for a given number of dimensions.
Most TB1 objects have a blocking mode in KR which allow to keep them on the main buffer thread of the server (faster for small jobs avoiding large memory copying)
- all json load/save/states bugs/oddities reported
- Max: cluttering when buffer resizing
- bufnmf parameter check order
- general buf resize sanity check
- Max: buf resize of dataset is done just in low priority thread (faking mode 2 otherwise)
- SC: most dataset objects in KR are much more efficient
- SC: kr bus assignation
Should mostly be edge cases. Feel free to test against your reports and update the issues.