Is there any example custom steps for CNV network with non uniform input shape (specifically radio sample)? #1230
-
I have been following FINN tutorials on CNV network and tried to apply it to my model, but the tutorial used image input which is (1x3x32x32). My model used a radio sample as input (trained for Modulation Classification), which has (1x2x1024x1) as input shape. I managed to get to the hardware build step and got the driver zip file on FPGA, but I notice the input shape for folded and packed is not the same as the normal input. Also, when I perform execute_on_buffer(), the output remains the same regardless of input (0% accuracy compared to cppsim accuracy 86%). Is this because my folding step was wrong? (I used setfolding() transformation) |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 4 replies
-
Hi, maybe you can take inspiration from this finn-example for a 1D CNN on the RadioML18 dataset: https://github.com/Xilinx/finn-examples/tree/main/build/vgg10-radioml |
Beta Was this translation helpful? Give feedback.
Hi,
maybe you can take inspiration from this finn-example for a 1D CNN on the RadioML18 dataset: https://github.com/Xilinx/finn-examples/tree/main/build/vgg10-radioml