sample_embedding_folder2/0768712.txt

Ticket Name: TDA2: Pytorch network implementation on TDA2x/4x

Query Text:
Part Number: TDA2 I am going to implement deep learning network ESPNet on TDA2x and TDA4x(in future). ESPNet source was implemented in Pytorch framework. I found there are Caffe/Tensorflow contents in TIDL document. I have two questions. 1. how to use pytorch outputs in TIDL. Should I export pytorch outputs in ONNX format and import it to Tensoflow or Caffe, after then follow the TIDL guid? Is there any related documents? 2. If some core operations are not supported in TIDL because of TIDL limitations, Should I implement it by myself? ex)Convolution Dialition 1,2,4 only tested / Deconvolution only supported stride value is 2 / ... Then where should I start? Thanks.

Responses:
1. ONNX mdoels are not supported in TDA2x/3x TIDL. 2. We can split caffe/tensorflow in to multiple sub TIDL models. Use TIDL for processing these mdoels and user can implements unsuported models and run them between the sub models. Please get in touch with TI representative for TDA4x related questions.

Thanks for your answer! Then will ONNX be supported in TDA4x? Currently, What approach should I follow for using pytorch framework in TDA2x? Thanks.

Please get in touch with your local TI support for TDA4x related questions. For TDA2x, you may try to convert the model to caffe format.