[FIXED] Local TFJS Model (Transfer Learning MobileNet) returns wrong predictions

Issue

I use transfer learning with MobileNet to solve an image problem. I loaded the images with ImageDataGenerator (rescale=1./127.5).

After training I converted it with:

tensorflowjs_converter --input_format=tf_saved_model --weight_shard_size_bytes 10000000000 model tmp

The next step is to load my model and my weight file in my Expo App:

const model = await tf.loadGraphModel(bundleResourceIO(modelJson,modelWeight));

And normalize my input image:

const normalized  = imageTensor.toFloat().sub(127.5).div(127.5);

The output is completely wrong (One class always 1 and all other ~0)

Does anybody know what the problem could be?

Solution

Solved it with this normalization:

const normalized  = imageTensor.toFloat().sub(127).div(128);

Answered By - Mare Seestern