Accelerate image processing on the client with GPU.js

[tddough98]

Here, we improve the image processing and display for the DeepCell Label client with GPU acceleration. I revamped existing ____Canvascomponents to use a kernel compiled with GPU.js instead of processing image data on the CPU, resulting in large speedups for image processing, especially for larger images. For a 4.5 megapixel image, changing the range of a channel now take around 6ms to complete the render compared to about 80ms in our existing deployment.

Each Canvas is now explicitly composed into a single canvas, instead of implicitly composed with a stack of canvases in the same position on the page. This address issue #274.

I've also addressed #273 by sending some additional ADD_LAYER and REMOVE_LAYER events to parent actors to ensure that the relevant hooks update when the event occurs.

I removed routes and logic from the backend that are now exclusively handled by the client including

• creating a label image from the label array with a colormap
• creating outlines from the label array

[willgraf]

As discussed, I think this is a good opportunity to add some tests to the Canvas components, especially as we have moved some more complicated logic into them.

It looks as though there is a failure on the 17.x node version for tests as well.

[willgraf]

LGTM. Some unit tests for the individual canvas components would be nice to include though.

[tddough98]

I added a test that Canvas sends events to the canvasMachine and that BrushCanvas has a canvas of the right size. I spent some time trying to test the canvas contents of BrushCanvas, but I hit some walls. As Jest runs tests in a node environment, some browser APIs, including the canvas, are not available. I installed canvas as a dev dependency which recreates the canvas API for node, but this does not include webgl functionality, which is a key need here. One solution seemed to be running tests with jest-electron so that webgl is available, but that will require ejecting create-react-app, which I'm not inclined to do in this PR. I'll explore that more another time, but for now, the two components tested should provide at least some foundation and models for testing more canvas functionality in the future.

[tddough98]

Unfortunately another bamboozling update for testing components with GPU.js. The standard tool for javascript code coverage is istanbul, which turns out is incompatible with GPU.js gpujs/gpu.js#663. The recommendation is to use v8 code coverage instead, but this requires using the coverageProvider option for jest, which is not supported with create-react-app...

I'm taking out the BrushCanvas test