Corey Lowman

ONNX serialization/deserialization

16

Would be useful to export/import models to onnx https://github.com/onnx/onnx

5d should cover most use cases (maybe batched sequences of 3 channel images?), and then 6d to be extra sure

This would require some sort of mutable device that isn't just a type. Could be good prep for gpu devices

scalar arith can do in place modification

1

It would require something to change the id of the tensor though, since the input & result need to have different unique ids to have different gradients

Safe allocation of arrays on heap

1

This function would stop using `alloc_zeroed()` and `Box::from_raw()` https://github.com/coreylowman/dfdx/blob/main/src/devices/allocate.rs#L11: ```rust let layout = Layout::new::(); debug_assert_eq!(layout.size(), T::NUM_BYTES); unsafe { let ptr = alloc_zeroed(layout) as *mut T; Box::from_raw(ptr) } ```

Plot benchmark speed against pytorch

4

- [ ] Linear batched forward (matmul & broadcast add) - [ ] Backprop algorithm - [ ] Optimizer updates - [ ] forward with tape & without tape

Document differences from pytorch & other rust DL crates

2

E.g. - const generic sizes - clean/safe/hackable implementation - Gradients not stored directly on parameters, makes updating & writing optimizers much cleaner

Tensor from multi dimensional array

4

Hello, I wrote the following function for creating a tensor from any slice type, do you think it would be good to include? note I had to copy and paste...

- [x] #158 - [ ] #157 - [ ] #150 - [ ] #149 - [x] #147 - [x] #146 - [ ] #121 - [x] #108 - [...

CUDA kernels JIT vs compile time compilation

4

Should CUDA kernels be JIT compiled at runtime or somehow compiled when the program is built? Best case we can support both of these easily via a feature flag or...