20 releases (stable)

1.0.14 Aug 16, 2023
1.0.8 Aug 15, 2023
1.0.7 Jun 27, 2023
0.1.4 Jun 10, 2023

#133 in Machine learning

Download history 2/week @ 2024-03-27 3/week @ 2024-04-03

208 downloads per month

Apache-2.0 and GPL-2.0-or-later

64KB
1.5K SLoC

🚧 Cog[-rust]: Containers for machine learning

Cog is an open-source tool that lets you package Rust ML models in a standard, production-ready container.

It's output should be interchangeable with Replicate's own Cog (for Python models).

Highlights

  • 📦 Docker containers without the pain. Writing your own Dockerfile can be a bewildering process. With Cog, you define your environment inside your Cargo.toml and it generates a Docker image with all the best practices: Nvidia base images, efficient caching of dependencies, minimal image sizes, sensible environment variable defaults, and so on.

  • 🤬️ No more CUDA hell. Cog knows which CUDA/cuDNN/tch/tensorflow combos are compatible and will set it all up correctly for you.

  • Define the inputs and outputs for your model in Rust. Then, Cog generates an OpenAPI schema and validates the inputs and outputs with JSONSchema.

  • 🎁 Automatic HTTP prediction server: Your model's types are used to dynamically generate a RESTful HTTP API using axum.

  • ☁️ Cloud storage. Files can be read and written directly to Amazon S3 and Google Cloud Storage. (Coming soon.)

  • 🚀 Ready for production. Deploy your model anywhere that Docker images run. Your own infrastructure, or Replicate.

How it works

Easily define your environment inside your Cargo.toml. Cog infers the rest:

[package]
name = "ml-model"

[package.metadata.cog]
cpu = true # optional, defaults to false
image = "docker-image-name" # optional, defaults to `cog-[package.name]`

Define how predictions are run on your model on your main.rs:

use anyhow::Result;
use async_trait::async_trait;
use cog_rust::Cog;
use schemars::JsonSchema;
use std::collections::HashMap;
use tch::{
  nn::{ModuleT, VarStore},
  vision::{imagenet, resnet::resnet50},
  Device,
};

#[derive(serde::Deserialize, schemars::JsonSchema)]
struct ModelRequest {
  /// Image to classify
  image: cog_rust::Path,
}

struct ResnetModel {
  model: Box<dyn ModuleT + Send>,
}

#[async_trait]
impl Cog for ResnetModel {
  type Request = ModelRequest;
  type Response = HashMap<String, f64>;

  async fn setup() -> Result<Self> {
    let mut vs = VarStore::new(Device::Cpu);
    vs.load("weights/model.safetensors")?;
    let model = Box::new(resnet50(&vs.root(), imagenet::CLASS_COUNT));

    Ok(Self { model })
  }

  fn predict(&self, input: Self::Request) -> Result<Self::Response> {
    let image = imagenet::load_image_and_resize224(&input.image)?;
    let output = self
      .model
      .forward_t(&image.unsqueeze(0), false)
      .softmax(-1, tch::Kind::Float);

    Ok(imagenet::top(&output, 5)
      .into_iter()
      .map(|(prob, class)| (class, 100.0 * prob))
      .collect())
  }
}

cog_rust::start!(ResnetModel);

Now, you can run predictions on this model:

$ cargo cog predict -i @input.jpg
--> Building Docker image...
--> Running Prediction...
--> Output written to output.jpg

Or, build a Docker image for deployment:

$ cargo cog build -t my-colorization-model
--> Building Docker image...
--> Built my-colorization-model:latest

$ docker run -d -p 5000:5000 --gpus all my-colorization-model

$ curl http://localhost:5000/predictions -X POST \
    -H 'Content-Type: application/json' \
    -d '{"input": {"image": "https://.../input.jpg"}}'

Why am I building this?

The Replicate team has done an amazing job building the simplest way to go from Python notebook to Docker image to API endpoint.

However, using Python as the base layer comes with its on share of challenges, like enormus image sizes or extra latency on model requests.

As the non-Python ML ecosystem slowly flourishes (see whisper.cpp and llama.cpp for example), cog-rust will provide that extra performance exposed on the same interfaces users and tools are already used to.

Prerequisites

  • macOS, Linux or Windows. Cog works anywhere Rust works.
  • Docker. Cog uses Docker to create a container for your model. You'll need to install Docker before you can run Cog.

Install

You can install Cog with Cargo:

cargo install cargo-cog

Dependencies

~26–40MB
~601K SLoC