Training CNN on PYNQ z2

Support
1

Hello,
I want to use pynq z2 for a CNN task by training the CNN on PC/google colab and then loading the trained(saved) network h5 file for the final classification task but it had some version issues.
So, I switched to training the CNN on pynq itself. But, it seems to ignore the batch size and take the whole train set for 10 epochs.

Training on pynq:

image
image1362Ɨ741 30.1 KB

Training on Colab:

image
image975Ɨ585 165 KB

This makes the training time so long.
Please help

1 Like
2

Hi there,

What are the version differences? Your best bet is probably to sort that out rather than training on the Z2 if you want speed, it will be much slower than a dedicated PC/server machine regardless. You could try to upgrade/downgrade the tensorflow version on one of your environments, or look into saving it with onnx, Iā€™m sure tensorflow ought to have hooks for that.

On the epochs issue you might have to add a steps_per_epoch parameter to your fit() function.

Thanks
Shawn

1 Like