Model load_to_fb

Terell_Dale · March 8, 2021, 6:28pm

Hello,

I saved a .tflite and I am trying to store model in the cam’s frame buffer stack. However, the syntax in the documentation, tf.load(path[, load_to_fb=False]), returns a syntax error. When I use the person_detection network in the syntax, tf.load(path, load_to_fb=True), there is no error. However, with my model used as, tf.load(“fmnist.tflite”, load_to_fb=True), I get a strange error, “OSError: Input dimension [c] should be 1 or 3!” I don’t think this is an issue with my network because it was made from a google tutorial. Is there any other explanation for this?
Screenshot from 2020-07-27 17-27-26.png

kwagyeman · March 8, 2021, 6:28pm

It’s an error with your network.

That error is generated by this line of code here:

github.com

openmv/tensorflow-lib/blob/master/libtf.cc#L81


if (model_input->dims->size == 2) {

    *input_height = model_input->dims->data[0];
    *input_width = model_input->dims->data[1];
    *input_channels = 1;
    *signed_or_unsigned = (model_input->type == kTfLiteInt8) || (model_input->type == kTfLiteFloat32);
    *is_float = model_input->type == kTfLiteFloat32;

    return 0;

} else if (model_input->dims->size == 3) {

    if ((model_input->dims->data[2] != 1) && (model_input->dims->data[2] != 3)) {
        error_reporter->Report("Input dimension [c] should be 1 or 3!");
        return 1;
    }

    *input_height = model_input->dims->data[0];
    *input_width = model_input->dims->data[1];
    *input_channels = model_input->dims->data[2];
    *signed_or_unsigned = (model_input->type == kTfLiteInt8) || (model_input->type == kTfLiteFloat32);

So, our loader noticed your model has 3 dims. Then we say the first and second channels are w and h. So, the third must be the number of channels. 1 for grayscale and 3 for rgb.

You’re network doesn’t have 1 or 3 channels. So, we don’t know how to process it then.

…

Note, we are using Edge Impulse to generate models for the OpenMV Cam now: https://www.edgeimpulse.com/

Please try using this service. It will actually generate useful networks for you that run on the camera.

MNIST is just a test example for running on the PC. It’s not really useful in real world situations.