The markdown parsing is broken/disabled for release notes. Sorry about that, I'm chasing the source of a crash that's been bringing this website down for the last couple of days.
- Fixed various bugs with `RNN` model and `LSTM` layer. Should be more consistent
- Moved auto calculation of biases to each layer so biases can more easily be controlled
- Standardized the `axis` definitions for mathematical operations on `Tensor`
```
+--------+
/ /|
/ Z |
+---X----+ |
| | |
| -1 Y +
| | /
| |/
+--------+
```
Along axis 0 the Tensor of shape AxBxC, where A is the columns, B is the rows, and C is the depth, would perform a mathematical function along the Y axis returning a (Ax1xC) Tensor
Along axis 1 the Tensor of shape AxBxC, where A is the columns, B is the rows, and C is the depth, would perform a mathematical function along the X axis returning a (1xBxC) Tensor
Along axis 2 the Tensor of shape AxBxC, where A is the columns, B is the rows, and C is the depth, would perform a mathematical function along the Z axis returning a (AxBx1) Tensor
Along axis -1 the Tensor of shape AxBxC, where A is the columns, B is the rows, and C is the depth, would perform a mathematical function along the Z axis returning a (1x1x1) Tensor Scalar
Added RNN and LSTM support with Embedding and LSTM layers.
- Instructions on how to use the layers can be found in the documentation of the layers themselves or in the FullModelTests file.
Added Vectorizer.
- This class vectorizes objects and returns it as a positional vector or a one hot vector. Instructions to use this can be found in the documentation on the class
Updated NumSwift to 2.0.1
Added dynamic member lookup to MetricsReporter
Remove dataset protocol and moved to NeuronDatasets
Drastically reduce size of the Neuron package by moving out unused datasets to a new repo NeuronDatasets.
Updated NumSwift dependency to tag 2.0.0
This is the big one! This is a complete re-write of Neuron! Changed effectively everything.
Please take a look at the README before updating to this version.
**It is INCOMPATIBLE with the older versions of Neuron**
Added activation to `addDenseNormal`
## What's Changed
* Conv by @wvabrinskas in https://github.com/wvabrinskas/Neuron/pull/17
- Added prototype `Convolution` support. **Use with caution as this can potentially use 600+% of the CPU. GPU support coming...**
- Refactored a bunch of functions to run concurrently. **You might see an increase in CPU usage depending on your training size.**
- Added a more robust metrics calculator and system.
- Added the MNIST dataset so the binary size of `Neuron` has increased due to this.
- Refactored the way `Brain` is constructed. This will make it a lot easier to create.
**Full Changelog**: https://github.com/wvabrinskas/Neuron/compare/1.5.1...1.6.0