Hello, I'm working on a speech classification model built on Pytorch's nn.Conv1d. PyTorch expects the conv1d input shape to be (batch, channel, length), in my case (time_stamps, 1, feat_dimension).
After converting to TFLite, MLIA analysis shows that Conv ops only run on the NPU if batch_size=1.
If batch_size>1, it runs on the CPU with the message: "IFM Tensor batch size must be 1."
Since my user case requires time_stamps > 1 (a.k.a batch_size > 1), how can I solve to ensure Conv ops run on NPU?
Thank you,Hao
Hi, we might need a bit more info about especially your NPU (which Ethos?). TFLite or TFLite micro?
I use the ethos-u55-128, with command like mlia check model_int8.tflite --target-profile ethos-u55-128. tflite model.
Just wonder if I can use batch_size > 1 as well as Conv ops run on NPU. Thanks.
Hi Hao, can you provide more info on the version of MLIA you are using, please? Many thanks!
Hi Isabella, I'm using the mlia 0.9.2 via pip install. Thanks.
Ok, thanks for confirming.There will be an update on pypi soon that will fix this. I'll notify when it's ready to use.
That would be great, thanks!
FYI - if you build mlia and it's ethos-u plugin yourself you can have the fix now.