Channel-specific NormalizeOp support [feature request?] about tflite-support HOT 4 CLOSED

As you've noticed, Codegen does support per channel normalization through metadata, same as TFLite Task library .

The comments have been updated. See here. Let me know if it makes sense to you.

from tflite-support.

Yes, I don't think tf.image support per channel normalization. Usually this is done simply by

VGG_MEAN = [123.68, 116.78, 103.94]
input_img_norm = (input_img - tf.constant(VGG_MEAN)) / 255.0

What does it mean: "except that the input is a UINT8 Tensor?

It means if the input tensor is UINT8, and if mean is 0 and stddev is 1, the returned TensorBuffer will not be a FLOAT32 tensor, but a UINT8 tensor, because the original input tensor is returned.

I'll update the comment shortly.

from tflite-support.

It looks like the codegen which uses support library already does this. After adding the appropriate metadata using Python, the relevant codegen is:

.add(new NormalizeOp(metadata.getImageMean(), metadata.getImageStddev()));

where the generated class Metadata contains

        public float[] getImageMean() {
            return Arrays.copyOf(imageMean, imageMean.length);
        }

I will leave it open for you to update the comment? Thanks for your reply.

from tflite-support.

Closing the issue. Feel free to reopen it if you have other questions.

from tflite-support.