TDA4VM: Supported TFLITE operator : GlobalAveragePooling2D

Azer Tiop

Part Number: TDA4VM

Tool/software:

Hi,

I am working on TDA4VM EVM for the first time. I'm using the latest version of SDK 10_00_00_05, latest version of EDGEAI_TIDL_TOOLS 10_00_08_00.

Here are all the supported Operators on TIDL-RT: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/10_00_08_00/docs/supported_ops_rts_versions.md

I want to run a custom model created and trained using TensorFlow and Keras on MMA. I noticed that the `Flatten` layer is not supported, so I would like to replace it with `GlobalAveragePooling2D`.

Here is my question: Will the Mean TFLite operator be called when I use the `GlobalAveragePooling2D` layer? If not, is `GlobalAveragePooling2D` supported, and which Keras layer is linked to the Mean TFLite operator?

Thanks,

Azer

over 1 year ago

0 Wen Li over 1 year ago

TI__Expert 8511 points

Hi Azer;

Thanks for the question. Our team will update you this week.

Best regards

Wen Li

+1 Chris Tsongas over 1 year ago

TI__Genius 16500 points

Hi Azer,

If you are new to the TDA4VM EVM I would recommend starting with a model from the Model Zoo. These models already have all their layers supported. Also, ONNX tends to be easier to start with as more tooling is available.

For your specific question, with TFLITE, the operator equivalent to ONNX GlobalAveragePool is AveragePool2d, and they bot map to TIDL_PoolingLayer. I am not 100% clear on your question. Here is the answer to what I think you asked. The AveragePool2d TFLITE and MEAN layers will be mapped to TIDL_PoolingLayer for running inference on the TDA4. The TIDL compiler will do this automatically and optimize it for the device. For any layer that is not supported, you can add this to the deny list, and it will run on the ARM core. Please see:

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/07_03_00_07/exports/docs/tidl_j7_02_00_00_07/ti_dl/docs/user_guide_html/md_tidl_osr_tflrt_tidl.html

For information about the deny list. All TIDL runtimes support this option.

Regards,
Chris

0 Azer Tiop over 1 year ago in reply to Chris Tsongas

Intellectual 350 points

Hi Chris,

Thank you for your reply. I also tried for the flatten layer that is supposed to not be supported by TFLITE operator.

However, the inference on C7x MMA is done without any problem :

I'm using edgeai_tidl_tool version 10_00_08_00. Is it supported now?

Regards,

Azer

0 Chris Tsongas over 1 year ago in reply to Azer Tiop

TI__Genius 16500 points

I'm sorry, I do not understand your final question. Is what supported now?

Chris

0 Azer Tiop over 1 year ago in reply to Chris Tsongas

Intellectual 350 points

Is the flatten layer supported by TFLITE operator now ? Because this is not written in the documentation of edgeai_tidl_tool version 10_00_08_00 here :

https://github.com/TexasInstruments/edgeai-tidl-tools/blob/10_00_08_00/docs/supported_ops_rts_versions.md

Regards,

Azer

0 Chris Tsongas over 1 year ago in reply to Azer Tiop

TI__Genius 16500 points

Hi Azer,

Can you please put this on a new thread? It is showing up as resolved, and I may miss it when I query. So, I will close it and look for a new thread on the TFLITE OP.

Thanks,

Chris

Processors

Processors forum

TDA4VM: Supported TFLITE operator : GlobalAveragePooling2D