Server accepts file for pre-/post-processing functions

[rahul-tuli]

This PR makes code changes such that a custom pipeline, with custom pre-/post-processing functions can be created using the deepsparse.server directly. This enables a user to use our system without knowing anything about custom pipelines

Additionally this PR also makes documentation updates, to expose this pathway.

# custom-processing-functions-config.yaml

endpoints:
  - task: custom
    model: /home/rahul/models/resnet50.onnx
    kwargs:
      processing_file: /home/rahul/projects/deepsparse/processing.py

# processing.py

from torchvision import transforms
from PIL import Image
import torch

IMAGENET_RGB_MEANS = [0.485, 0.456, 0.406]
IMAGENET_RGB_STDS = [0.229, 0.224, 0.225]
preprocess_transforms = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=IMAGENET_RGB_MEANS, std=IMAGENET_RGB_STDS),
])

def preprocess(img_file):
    with open(img_file, "rb") as img_file:
        img = Image.open(img_file)
        img = img.convert("RGB")
    img = preprocess_transforms(img)
    batch = torch.stack([img])
    return [batch.numpy()]

def postprocess(outputs):
    return outputs

Invocation:

deepsparse.server --config_file custom-processing-functions-config.yaml

And Voila! Now the server creates a custom pipeline, with custom pre-/post- processing functions from the specified processing.py

Special thanks to @bfineran for thinking about this pathway
And @dbogunowicz for #821 from which the loading functions code is inspired from

Additionally this pathway will become highly useful for sparsify.package CLI

[dbogunowicz]

Oh so this is what you've been working on. I like that!

One thing which is missing are some unit tests. Any plans to add those as well?

[KSGulin]

This is a cool addition