Part Number:TDA2EVM5777
Hi,
I was trying to convert a U-Net from Tensorflow to TIDL format using the TIDL conversion tool. However, it seems that Deconvolution is neither supported for Tensorflow Keras nor Slim.
Afterwards, I was looking aroung to see which layers of Tensorflow are currently supported. So far, I found this link
http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components_TIDL.html#neural-network-layers-supported-by-tidl
which does not state wether it's talking about Keras or Slim layers (I would assume it's Slim but am not really sure). And this link
https://e2e.ti.com/support/processors/f/791/t/735984?tisearch=e2e-sitesearch&keymatch=tensorflow
which shows an example of some Keras layers working but doesn't give info about all Keras layers supported (I assume that the same Keras layers are supported as Slim but am not really sure).
So my questions would be:
1. Can you give a summary of which TF Keras and which TF Slim layers are supported by the TIDL conversion tool including corresponding TF Version.
2. Is it true that you cannot convert any types of Autoencoder architectures trained in Tensorflow using the TIDL conversion tool without having to manually add the Deconv. layer into the conversion tool?
Best regards and thanks in advance,
Daniel