Allow tensor as input image

[lindapaiste]

We want any model which accepts an image input to also accept a 3D tensor of image pixels. This means that we should not call tf.browser.fromPixels directly with the input as it will be an error if called with a tensor. I am instead calling a new utility toTensor which is a simple wrapper around tf.browser.fromPixels that will first check if the input is already a tensor and just return it.

Change tf.browser.fromPixels to toTensor in:

• BodyPix
• CartoonGAN
• Pix2Pix
• StyleTransfer
• UNET

---

There are chunks of logic in Darknet and Doodlenet for checking if the image is a tensor and conditionally converting it. These can be deleted entirely because the ImageClassifier handles converting the input to a tensor before calling the specific model's classify() function.

---

Two other minor changes:

• I realized that await this.ready and await mediaReady should be awaited in parallel rather than sequentially so I changed that in a few places.
• The existing imgToTensor function, which rescales and crops in addition to converting, should use 127.5 as the divisor rather than 127.

[lindapaiste]

• The existing imgToTensor function, which rescales and crops in addition to converting, should use 127.5 as the divisor rather than 127.

In hindsight I think that I was wrong about this. I got confused thinking that it should be half of 255, but I was forgetting that 0 is part of the range. 0 to 127 is 128 values and 0 to 255 is 256 values so it is already half.