Converting Keras model, Conv2d error

I have followed the tutorial to convert MobileNet to a k210 for a custom model defined in Keras. At the last step (to generate the model for the k210l), I get the following error

Layer Conv2d is not supported

any idea on how to fix this ? Can’t we use conv2d layers in our models for the k210 ?

Thanks for your help

can you post your h5 or pb file? also with some pic dataset

You can download a sample set with h5 file and some image samples.

Perhaps you used wrong parameters? https://github.com/kendryte/nncase#supported-layers
When using TensorFlow Conv2d/DepthwiseConv2d kernel=3x3 stride=2 padding=same, you must first use tf.pad([[0,0],[1,1],[1,1],[0,0]]) to pad the input and then use Conv2d/DepthwiseConv2d with valid padding.

Hi,

i managed to get it to work by making sure that every conv2d layer have the ‘same’ padding. Do you have any tip so that the conversion from h5 to kmodel will result in a small memory footprint ? When i convert for now, i have a kmodel that is as large as the .pb graph.

Regards,

you need use ncc’s 8bit mode to decrase size

I have created a Colab notebook using the flower dataset for transfer learning using mobilenet v1. The end result is a kmodel file which you can upload to Maixpy

Transfer Learning