Model predicting one class using Saved_Model

zulyang · September 29, 2020, 4:41am

Hi, has anyone successfully loaded their Saved_Model generated from EdgeImpulse into their own projects? Tried with a bunch of test images and am consistently getting the same classification. Am i missing something here? Do i need to add an activation function?

Printing Predictions:

Model Summary:
Screenshot 2020-09-29 at 12.34.07 PM

janjongboom · September 29, 2020, 7:42am

@zulyang How are you feeding the input data in?

zulyang · September 29, 2020, 10:22am

As such! Im reading a video frame by frame, resizing it, and using the model to predict the classification of the image.

janjongboom · September 29, 2020, 11:04am

@zulyang You’ll need to match the expected input of the model if you use just the TFLite model and not the rest of the Edge Impulse SDK (as we do it automatically), see an example in Python here:

github.com

edgeimpulse/processing-blocks/blob/master/image/dsp.py#L9


import argparse
import json
import math
import numpy as np
import sys
import io, base64
from PIL import Image


def generate_features(draw_graphs, raw_data, axes, sampling_freq, channels):


    graphs = []
    pixels = []
    width = raw_data[0]
    height = raw_data[1]
    raw_data = raw_data[2:].astype(dtype=np.uint32).view(dtype=np.uint8)
    bs = raw_data.tobytes()
    ix = 0


    if channels == 'Grayscale':

I guess normalization is not the same. You can do a quick sanity check by looking at the Raw features and the Processed features on ‘Live classification’.

Here ‘raw features’ is the RGB data, and processed features is the output of the DSP step (which should be passed into the NN):