Fix for loading pretrained models & storing models server side

[Nacho114]

Fix for loading pretrained models by conditional import of tfjs-node

When importing tfjs-node two things happen:

1. The appropriate tfjs engine is loaded
2. The appropriate functions are loaded

It seems that if we import tfjs-node the only 1. happens, since when loading and saving files the API still thinks it's tfjs (no node version); if we import 'tfjs-node' conditionally earlier then 2. seems to happen properly allowing us to save and load models again.

Note that due to some webpack dependency checks, we need to have tfjs-node inside a string to take advantage of lazy loading to avoid weback issues.

TODO:

• model.json was not added, check on this
• check if it is possible to save models in the server
• Added timeout for test, was running in infinite loop if server crashed right after build.
• find better way for model path (current one is relative, might not be ideal since it assumes CLI and server are at a same distance from model path).
• update TASK.md

Fixes #446
Also Fixes #363

[Nacho114]

Currently loading the model looks as follows:

export async function model (): Promise<tf.LayersModel> {
  const file = 'file://./../models/mobileNetV2_35_alpha_2_classes/model.json'
  return await tf.loadLayersModel(file)
}

the issue is that file is relative to where the code is executed from. Note that it is either run via the server, or the benchmark. In both cases the model is at ../ distance from the cwd. So it is not very robust. One solution is to store the path in process.env and define it once in benchmark and once in server, this way the stored models path is relative not to cwd, but to each env.

[Nacho114]

@tharvik @s314cy wdyt?

[Nacho114]

@tharvik @s314cy wdyt?

Found a clean solution by following the server config as suggested by @tharvik

[s314cy]