WU Chen

After using `slice_assign`, gradient descent cannot track the model parameters

6

I found that after using `slice_assign` in the loss function, gradient descent cannot track the model parameters. I believe this is the main reason why the loss becomes NaN after...

implement ctc loss function

10

I need ctc loss function in CRNN model. I tried to implement it based on PyTorch implementation, but the results obtained after calling forward() are somewhat different from PyTorch's. I...

implement bidirectional lstm

2

I need bidirectional lstm in CRNN model. ### Checklist - [x] Confirmed that `run-checks all` script has been executed. - [x] Made sure the book is up to date with...

Small negative values cause the gradient of `sigmoid` to become `NaN`

1

Here is the code: ```rust use burn::{ backend::{Autodiff, NdArray}, tensor::{activation, Data, Tensor}, }; fn main() { let data = Data::::from([-90.0]); let device = Default::default(); let tensor_1 = Tensor::::from_data(data, &device).require_grad(); let...

fix the problem of sigmoid gradient generating NaN

1

## Pull Request Template ### Checklist - [x] Confirmed that `run-checks all` script has been executed. - [x] Made sure the book is up to date with changes in this...

Provide a method to chunk the weights before importing using PyTorchFileRecorder

5

### Description I encountered difficulties when importing the weights of an LSTM from PyTorch. This is because PyTorch often tends to concatenate different weights together, resulting in significant differences between...

Wgpu backend of image-classification-web example cannot work

13

When I choose the wgpu backend, I get errors in the console.  After I disable the `autotune` feature of [burn-wgpu](https://github.com/tracel-ai/burn/blob/a53f459f205889a22ecea3713bbae12d3de7eb0c/examples/image-classification-web/Cargo.toml#L21-L23), the wgpu backend still cannot work. The live demo...

Bitwise operations are very common, and most deep learning frameworks support these operations. However, Burn seems to lack support for them. For me, if bitwise operations were available, I could...