"Why Aidge" in Get Started
I propose a new section in the "Get Started" page of the documentation, to highlight the main differentiating factors of the framework. I think it is something very important and could also form the basis of future presentations.
New features
- supervised learning (Optimizer, lr scheduler, loss, backpropagation)
- initializer (filler)
- New attribute (Constant) for Producers to match ONNX parameters. A constant Producer cannot be updated by the learning process.
- better GraphView display with Tensors sizes, Operator type, unique indexes
- create custom implementation for any operator
- Tensor CPU implementation in core (easier prototyping)
- Pytorch interoperability
- parallel Scheduler
- support of RNN
- enhance
Tensor- Tensor view to access without updating them
- scalar Tensor
- operator+,-,*,/
- introduction of dataloader to handle your own datasets
- better user interface
__get_attr__,__set_attr__ - a complete logging mechanism with adjustable level
- introduction of PTQ
- many new tutorials
- update with operators:
| operator | core | cpu / forward | cpu / backward | cuda / forward | ONNX import | ONNX export |
|---|---|---|---|---|---|---|
| AvgPooling | ||||||
| Erf | ||||||
| Concat | ||||||
| FC |
|
|
||||
| Gather | ||||||
| GlobalAveragePooling | ||||||
| Identity | ||||||
| LeakyReLU | ||||||
| LSTM | ||||||
| MatMul | ||||||
| MaxPooling | ||||||
| Memorize | ||||||
| Pop | ||||||
| ReduceMean | ||||||
| ReLU | ||||||
| Reshape | ||||||
| Sigmoid | ||||||
| Slice | ||||||
| Sqrt | ||||||
| Sub | ||||||
| Tanh | ||||||
| Transpose |