mlr3torch

mlr3torch copied to clipboard

sebffischer

`nn("<id>")` shorcut for `po("nn_<id>")`

1

saves 3 characters per pipeop

sebffischer

Just like e.g. for callbacks, we should provide a convenience function to create a pipeop_torch. Also we should provide an `as_pipeop` converter for `nn_module_generator` objects.

sebffischer

We want to port all learners from torchvision: * [ ] [Resnet](https://github.com/mlverse/torchvision/blob/main/R/models-resnet.R) * [ ] [Alexnet](https://github.com/mlverse/torchvision/blob/main/R/models-alexnet.R) * [ ] [VGG](https://github.com/mlverse/torchvision/blob/main/R/models-vgg.R) * [ ] [mobilenet](https://github.com/mlverse/torchvision/blob/main/R/models-mobilenetv2.R)

sebffischer

bring learners back from attic and make them compatible with dataloader for `PipeOpTorch` / unify the relevant code The following architectures will be supported: * TabNet: * [ ] regr...

sebffischer

**Autotests** * [ ] Autotest for PipeOpTorch: * Should check that {input, output} correspond to the args / return type of the generated nn_module * Should check that the calculated...

sebffischer

forward methods with arguments

2

Some torch modules (like `nn_conv_transpose1d`) have an argument in their forward function (`output_size` in this case). We should have some mechanism to allow to set these values on the fly,...

sebffischer

HarutyunyanLiana

Se also here: https://github.com/mlr-org/mlr3torch/issues/13

sebffischer

See also: https://github.com/mlr-org/mlr3torch/issues/4

sebffischer