TDA4VM: Need some clarity on CONV + BN + RELU fusion in TIDL

Hi,

Thomas Boulay said:

In TIDL doc, it is mentionned that CONV & RELU & BN will be merged to get better performance.

Yes, Relu and BN can be merged inside convolution.

And from the shared SS, i can see that relu is getting absorbed as part of convolutions as expected and supported.

However, we dont support absorption/merging of pooling layers into convolutions.

Hi, 

thanks for your answer. Sorry the title of the question is confusing. I wanted to talk about CONV + POOL + ReLU. In the tidl documentation, it is mentionned that Pooling could be absorbed as part of convolution too. Is it something obsolete?

Br,

Just some additional infos, 

I went though the documentation and found this section:

It seems that this mergepooling layer is checking the input of pooling layer itself. if this checking is done before CONV+RELU have been merged, it makes sense that pooling is not absorbed into conv when we have CONV+RELU+POOLING sequence.

If the sequence is CONV+POOLING +RELU, I think I could expect CONV+POOLING will be absorbed but we should get same issue with RELU?

Could you confirm these assumptions are valid or not please? Are there any constraint on conv/pooling kernel size or in the sequence order for make this merging possible?

Br,

I have assigned this thread to correct expert who can comment on this, please expect the response in coming week as team is busy in release cycle.

Thanks for your patience.

Thomas,
Can you let me know which SDK & documentation you're referring to? This option is currently unsupported. Our current most documentation is at : https://github.com/TexasInstruments/edgeai-tidl-tools 

Thanks,
Varun