| Publisher | Alexander Petryaev |
|---|---|
| File size | 175.01MB |
| Number of files | 80 |
| Latest version | 2 |
| Latest release date | 2020-12-31 09:29:13 |
| First release date | 2020-10-26 09:42:16 |
| Supported Unity versions | 2018.4.2 or higher |
Deep Components allows you to integrate deep learning algorithms, structures, and methodology into your Unity3D application not only at the development stage but including the publishing stage as well. You won't need any other third-party software. Deep Components as a tool is completely integrated into Unity3D and has no extra software dependencies.
The Free version of Deep Components supports only the CPU backend. Note that when compiled to IL2CPP it works up to 5x faster than in editor.
The features provided are quite graphic and allows you to observe and supervise the training process of deep learning network.
Deep Components contains following layers with learnable parameters: Dense (aka Fully Connected, aka Linear), Convolutional, Deconvolutional.
Layers providing nonlinear activation: Leaky ReLU, Sigmoid, TanH, Softmax, CELU, Softsign.
Regularization layers: Dropout, Batch normalization.
Discretisation layers: Pooling, Upsampling.
Layer for variational inference: Stochastic.
Loss functions: Mean average Error, Mean squared error, Huber, Negative log likelihood, Binary cross entropy.
Optimization algoriths: Stochastic gradient descent with optional momentum, Root mean square propagation (RMSProp), Adaptive momentum estimation (Adam).
It allows arbitrary batch size (still, it has to fit the hardware limits).
Also, it includes demonstrative examples of how to use deep neural networks to solve classification, encoding/decoding problems, unsupervised learning, and generative models.
