How Do Frameworks Like PyTorch and TensorFlow Handle Unrolled LSTM Layers from ONNX #6864
Labels
question
Questions about ONNX
Comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This above image represents an LSTM block in ONNX
I'm trying to understand how frameworks such as PyTorch and TensorFlow load layers that are stored in an unrolled state in ONNX.
For example, in ONNX, LSTM and RNN models are often represented in an unrolled form, consisting of multiple
Linear,Add, andConcatlayers. When converting an ONNX model to PyTorch, I have the following questions:Do these frameworks load the layers in their unrolled form, keeping them as separate operations?
Or do they use a pattern-matching engine to detect LSTM blocks in the ONNX graph?
If such an engine exists, does PyTorch then replace the unrolled structure with a native torch.nn.LSTM layer instead of preserving the individual operations?
The text was updated successfully, but these errors were encountered: