WU Chen

@antimora I just need a small value to prevent log(0), so now I think it may not be necessary to use 1e-15; 1e-5 should be small enough. However, I think...

I think the current implementation of bidirectional lstm can work during inference, but the implementation cannot update the parameters of the model in backward propagation due to #1098. Using `Tensor::cat`...

@antimora Could you reopen this pull request? The issue regarding Autodiff has been resolved in #1575, and I think it's time to proceed with implementing the Bidirectional LSTM.

Wgpu test failed, but I don't know why...

The outputs of both `Lstm` and `BiLstm` are now aligned with PyTorch. The script I used to generate tests for Bidirectional LSTM can be found [here](https://github.com/wcshds/with-burn-rs/blob/main/gen_bidiretional_lstm_test.py). The tests for the...

@nathanielsimard @louisfd I changed the batch size to 2, and the test passed. This seems to be a bug with the Wgpu backend, so I've opened an issue #1656.

Is it possible to manually define differentials during the activation function backpropagation? So we don't have to automatically differentiate `log` and `exp`.

> If you could provide a minimal code, we will look into this. PyTorch screen shot appears to store weights as an array, which is supported by PyTorchFileRecorder. Sure. Rust...

@antimora @laggui Therefore, importing LSTM's weights from PyTorch should do some remap like this lstm.weight_ih_l0 => lstm.input_gate.input_transform.weight => lstm.forget_gate.input_transform.weight => lstm.cell_gate.input_transform.weight => lstm.output_gate.input_transform.weight lstm.bias_ih_l0 => lstm.input_gate.input_transform.bias => lstm.forget_gate.input_transform.bias => lstm.cell_gate.input_transform.bias...

@nathanielsimard Hello, I am always interested in the implementation of lstm in burn. I still think lstm is buggy right now. If a linear layer is added after the lstm,...