Transformer-based models

[ChloeCarbonniere]

Hello,

Is it possible to implement transforme-based models on the STM32N6 ?

[RSERSTM]

Hello ChloeCarbonniere,
What do you mean by transforme-based models ? Transformer-based models ?
Attention layers ? Encoder-only computer vision models ? Generative models ?
Thank you,

[ChloeCarbonniere]

Mi final goal is to implement a DETR on the STM32N6.

[RSERSTM]

Ok I understand.
For info STM32N6-DK has 128MB of Flash to store weights and RAM (internal 4.2MB + external 32MB) to store activations(the data) https://www.st.com/en/evaluation-tools/stm32n6570-dk.html.
In other words regarding these kind of Transformers (DETR usually takes GBytes of space) for the moment we don't have these in the model zoo because they are often to big for STM32N6.
But it should be possible to implement very small Transformers with quantized (8bit) Attention layers if the weights + activations fit in memory.
Regards,