Add ability to create Extractors 100% in the GUI

[KastanDay]

These are living notes, drawing from inspiration

GUI Extractor Requirements:

1. Launch-able in GUI.

   • Implementation details
     • CPU-only for now
     • Run on your Clowder instance
     • Unsafe code is okay, just have a big red warning.

2. Support inference Endpoints, can hit them with super simple curl.

curl -X POST \
    -H "Content-Type: application/json" \
    -d '{"prompt": "Software developers start with a Hello, [MASK]! script."}' \
    http://localhost:8000/

# result:
# {
#     "outputs": {
#         "score": 0.8281897306442261,
#         "sequence": "software developers start with a hello, world! script.",
#         "token": 2088,
#         "token_str": "world"
#     }
# }

3. Can connect via PyClowder SDK

# from Banana.dev as inspiration, but PyClowder already does this.
from banana_dev import Client

# Create a reference to your model on Banana
my_model = Client(
    api_key="YOUR_API_KEY", # Found in dashboard
    model_key="YOUR_MODEL_KEY", # Found in model view in dashboard
    url="http://localhost:8000",
)

# Specify the model's input JSON
inputs = {
    "prompt": "Software developers start with a Hello, [MASK]! script",
}

# Call your model's inference endpoint
result, meta = my_model.call("/", inputs)
print(result)

# result:
# {
#     "outputs": {
#         "score": 0.8281897306442261,
#         "sequence": "software developers start with a hello, world! script.",
#         "token": 2088,
#         "token_str": "world"
#     }
# }

4. Custom preprocessing setup() and predict() functions.
   1. Enter code via the GUI, 100% GUI.

# https://github.com/replicate/cog
from cog import BasePredictor, Input, Path
import torch

class Predictor(BasePredictor):
    def setup(self):
        """Load the model into memory to make running multiple predictions efficient"""
        self.model = torch.load("./weights.pth")

    # The arguments and types the model takes as input
    def predict(self,
          image: Path = Input(description="Grayscale input image")
    ) -> Path:
        """Run a single prediction on the model"""
        processed_image = preprocess(image)
        output = self.model(processed_image)
        return postprocess(output)

5. Warning & limits for Arbitrary code execution
   1. Limit outbound (network) requests from container.
   2. Copy the permissions from MacOS.
   3. Before trying to install an "arbitrary code execution" extractor you have to agree to a big flashy warning & "no guarantees" policy. Just so they understand the concern.

TODO: Investigate Cog as a Docker wrapper to automate job launching. It's designed for exactly this problem.

[KastanDay]

dw, I'll clean up these comments. Just keeping them for now as I do the initial implementation.

[KastanDay]

Maybe start with SegmentAnything model for Xiaoxia Liao to use.