pymlup 0.2.1

MLup framework, fast ml to production, easy to learn, easy to use.

PyMLup

Introduction

It's library for easy run ML in production.

All you need is to deliver the model file and config to the server (in fact, the config is not necessary) 🙃

PyMLup is a modern way to run machine learning models in production. The market time has been reduced to a minimum. This library eliminates the need to write your own web applications with machine learning models and copy application code. It is enough to have a machine learning model to launch a web application with one command.

• It's library learning only clean python;
• Use FastApi in web app backend;

Work tested with machine learning model frameworks (links to tests):

• scikit-learn>=1.2.0,<1.3.0
• tensorflow>=2.0.0,<3.0.0
• lightgbm>=4.0.0,<5.0.0
• torch>=2.0.0,<3.0.0
• onnx>=1.0.0,<2.0.0
• onnxruntime>=1.0.0,<2.0.0

Support and tested with machine learning libraries:

• numpy>=1.0.0,<2.0.0
• pandas>=2.0.0,<3.0.0
• joblib>=1.2.0,<1.3.0
• tf2onnx>=1.0.0,<2.0.0
• skl2onnx>=1.0.0,<2.0.0
• jupyter==1.0.0

The easiest way to poke:

pip install pymlup
mlup run -m /path/to/my/model.onnx

Useful links

• Docs under development;
• Examples;
• Tests models

How it's work

1. You are making your machine learning model. Optional: you are making mlup config for your model.
2. You deliver your model to server. Optional: you deliver your config to server.
3. Installing pymlup to your server and libraries for model.
4. Run web app from your model or your config 🙃

Requirements

Python 3.7+

• PyMLup stands on the shoulders of giants FastAPI for the web parts.
• Additionally, you need to install the libraries that your model uses.

Installation

pip install pymlup

You will also can install with ml backend library:

pip install "pymlup[scikit-learn]"  # For scikit-learn
pip install "pymlup[lightgbm]"      # For microsoft lightgbm
pip install "pymlup[tensorflow]"    # For tensorflow
pip install "pymlup[torch]"         # For torch
pip install "pymlup[onnx]"          # For onnx models: torch, tensorflow, sklearn, etc...

Examples

Examples code

import mlup

class MyAnyModelForExample:
    def predict(self, X):
        return X

ml_model = MyAnyModelForExample()


up = mlup.UP(ml_model=ml_model)
# Need call up.ml.load(), for analyze your model
up.ml.load()
# If you want testing your web app, you can run in daemon mode
# You can open browser http://localhost:8009/docs
up.run_web_app(daemon=True)

import requests
response = requests.post('http://0.0.0.0:8009/predict', json={'X': [[1, 2, 3], [4, 5, 6]]})
print(response.json())

up.stop_web_app()

You can check work model by config, without web application.

• predict - Get model predict as inner arguments as in web app.
• predict_from - As predict method, but not use data transformer before call model predict.
• async_predict - Asynchronous version of the predict method.

import mlup
import numpy

class MyAnyModelForExample:
    def predict(self, X):
        return X

ml_model = MyAnyModelForExample()
up = mlup.UP(ml_model=ml_model)
up.ml.load()

up.predict(X=[[1, 2, 3], [4, 5, 6]])
up.predict_from(X=numpy.array([[1, 2, 3], [4, 5, 6]]))
await up.async_predict(X=[[1, 2, 3], [4, 5, 6]])

Save ready application to disk

Make default config

If path endswith to json, make json config, else yaml config.

import mlup
mlup.generate_default_config('path_to_yaml_config.yaml')

From config

You can save ready config to disk, but you need set local storage and path to model file in server. In folder can there are many files, mask need for filter exactly our model file

import mlup
from mlup.ml.empty import EmptyModel  # This stub class
from mlup import constants

up = mlup.UP(ml_model=EmptyModel())
up.conf.storage_type = constants.StorageType.disk
up.conf.storage_kwargs = {
    'path_to_files': 'path/to/model/file/in/model_name.modelextension',
    'file_mask': 'model_name.modelextension',
}
up.to_yaml("path_to_yaml_config.yaml")
up.to_json("path_to_json_config.json")

# After in server
up = mlup.UP.load_from_yaml("path_to_yaml_config.yaml", load_model=True)
up.run_web_app()

From pickle

If you make pickle/joblib file your mlup with model, don't need to change storage type, because your model there is in your pickle/joblib file.

import pickle
import mlup
from mlup.ml.empty import EmptyModel  # This stub class

up = mlup.UP(ml_model=EmptyModel())

# You can create pickle file
with open('path_to_pickle_file.pckl', 'wb') as f:
    pickle.dump(up, f)

# After in server
with open('path_to_pickle_file.pckl', 'rb') as f:
    up = pickle.load(f)
up.ml.load()
up.run_web_app()

Change config

If you can change model settings (See config attributes docs), need call up.ml.load_model_settings().

import mlup

class MyAnyModelForExample:
    def predict(self, X):
        return X

ml_model = MyAnyModelForExample()

up = mlup.UP(
    ml_model=ml_model,
    conf=mlup.Config(port=8011)
)
up.ml.load()
up.conf.auto_detect_predict_params = False
up.ml.load_model_settings()

Examples server commands

mlup run

You can run web application from model, config, pickle up object. Bash command mlup run making this.

See mlup run --help for full docs.

From model

mlup run -m /path/to/your/model.extension

This will run code something like this:

import mlup
from mlup import constants

up = mlup.UP(
    conf=mlup.Config(
        storage_type=constants.StorageType.disk,
        storage_kwargs={
            'path_to_files': '/path/to/your/model.extension',
            'files_mask': r'.+',
        },
    )
)
up.ml.load()
up.run_web_app()

You change config attributes in this mode. For this, you can add arguments --up.<config_attribute_name>=new_value. (For more examples see mlup run --help).

From config

mlup run -c /path/to/your/config.yaml
# or mlup run -ct json -c /path/to/your/config.json

This will run code something like this:

import mlup

up = mlup.UP.load_from_yaml(conf_path='/path/to/your/config.yaml', load_model=True)
up.run_web_app()

From mlup.UP pickle/joblib object

mlup run -b /path/to/your/up_object.pckl
# or mlup run -bt joblib -b /path/to/your/up_object.joblib

This will run code something like this:

import pickle

with open('/path/to/your/up_object.pckl', 'rb') as f:
    up = pickle.load(f)
up.run_web_app()

mlup make-app

This command making .py file with mlup web application and your model, config, pickle up object or with default settings.

See mlup make-app --help for full docs.

With default settings

mlup make-app example_without_data_app.py

This command is making something like this:

# example_without_data_app.py
import mlup


# You can load the model yourself and pass it to the "ml_model" argument.
# up = mlup.UP(ml_model=my_model, conf=mlup.Config())
up = mlup.UP(
    conf=mlup.Config(
        # Set your config, for work model and get model.
        # You can use storage_type and storage_kwargs for auto_load model from storage.
    )
)
up.ml.load()
up.web.load()

# If you want to run the application yourself, or add something else to it, use this variable.
# Example with uvicorn: uvicorn example_app:app --host 0.0.0.0 --port 80
app = up.web.app

if __name__ == '__main__':
    up.run_web_app()

And you can write your settings and run web application:

python3 example_without_data_app.py

With only model

mlup make-app -ms /path/to/my/model.onnx example_without_data_app.py

This command is making something like this:

# example_without_data_app.py
import mlup
from mlup import constants


up = mlup.UP(
    conf=mlup.Config(
        # Set your config, for work model and get model.
        storage_type=constants.StorageType.disk,
        storage_kwargs={
            'path_to_files': '/path/to/my/model.onnx',
            'files_mask': 'model.onnx',
        },
    )
)
up.ml.load()
up.web.load()

# If you want to run the application yourself, or add something else to it, use this variable.
# Example with uvicorn: uvicorn example_app:app --host 0.0.0.0 --port 80
app = up.web.app

if __name__ == '__main__':
    up.run_web_app()

And you can run web application:

python3 example_without_data_app.py

With only config

mlup make-app -cs /path/to/my/config.yaml example_without_data_app.py

This command is making something like this:

# example_without_data_app.py
import mlup


up = mlup.UP.load_from_yaml('/path/to/my/config.yaml', load_model=False)
up.ml.load()
up.web.load()

# If you want to run the application yourself, or add something else to it, use this variable.
# Example with uvicorn: uvicorn example_app:app --host 0.0.0.0 --port 80
app = up.web.app

if __name__ == '__main__':
    up.run_web_app()

And you can run web application:

python3 example_without_data_app.py

With only binary UP object

mlup make-app -bs /path/to/my/up.pickle example_without_data_app.py

This command is making something like this:

# example_without_data_app.py
import pickle


with open('/path/to/my/up.pickle', 'rb') as f:
    up = pickle.load(f)

if not up.ml.loaded:
    up.ml.load()
up.web.load()

# If you want to run the application yourself, or add something else to it, use this variable.
# Example with uvicorn: uvicorn example_app:app --host 0.0.0.0 --port 80
app = up.web.app

if __name__ == '__main__':
    up.run_web_app()

And you can run web application:

python3 example_without_data_app.py

mlup validate-config

This command use for validation your config. This command have alpha version and need finalize.

See mlup validate-config --help for full docs.

mlup validate-config /path/to/my/conf.yaml

Web application interface

By default, web application starting on http://localhost:8009 and have api docs.

Interactive API docs

Now go to http://localhost:8009/docs.

You will see the automatic interactive API documentation (provided by Swagger UI):

Api points

/health

Use for check health web application.

HTTP's methods: HEAD, OPTIONS, GET

Return JSON

{'status': 200} and status code is 200.

/info

Use for get model and application information. If set debug=True in config, return full config.

HTTP's methods: GET

Return JSON:

{
  "model_info": {
    "name": "MyFirstMLupModel",
    "version": "1.0.0.0",
    "type": "sklearn",
    "columns": null
  },
  "web_app_info": {
    "version": "1.0.0.0",
    "column_validation": false,
    "debug": false
  }
}

If set in config debug=True, return another json, almost complete config. But no sensitive data.

{
  "web_app_config": {
    "host": "localhost",
    "port": 8009,
    "web_app_version": "1.0.0.0",
    "column_validation": false,
    "custom_column_pydantic_model": null,
    "mode": "mlup.web.architecture.directly_to_predict.DirectlyToPredictArchitecture",
    "max_queue_size": 100,
    "ttl_predicted_data": 60,
    "ttl_client_wait": 30.0,
    "min_batch_len": 10,
    "batch_worker_timeout": 1.0,
    "is_long_predict": false,
    "show_docs": true,
    "debug": true,
    "throttling_max_requests": null,
    "throttling_max_request_len": null,
    "timeout_for_shutdown_daemon": 3.0,
    "item_id_col_name": "mlup_item_id"
  },
  "model_config": {
    "name": "MyFirstMLupModel",
    "version": "1.0.0.0",
    "type": "sklearn",
    "columns": null,
    "predict_method_name": "predict",
    "auto_detect_predict_params": true,
    "storage_type": "mlup.ml.storage.memory.MemoryStorage",
    "binarization_type": "auto",
    "use_thread_loop": true,
    "max_thread_loop_workers": true,
    "data_transformer_for_predict": "mlup.ml.data_transformers.numpy_data_transformer.NumpyDataTransformer",
    "data_transformer_for_predicted": "mlup.ml.data_transformers.numpy_data_transformer.NumpyDataTransformer",
    "dtype_for_predict": null
  }
}

/predict

Use for call predict in model.

HTTP's methods: POST

Requests body data:

{
  "data_for_predict": [
    "input_data_for_obj_1",
    "input_data_for_obj_2",
    "input_data_for_obj_3"
  ]
}

Key data_for_predict is default key for inner data. In config by default set param auto_detect_predict_params is True. This param activate analyze model predict method, get arguments from and generate API by params. If auto_detect_predict_params found params, he changes data_for_predict to finding keys and change API docs.

Example for scikit-learn models:

{
  "X": [
    "input_data_for_obj_1",
    "input_data_for_obj_2",
    "input_data_for_obj_3"
  ]
}

input_data_for_obj_1 maybe any valid JSON data. These data are run through data transformers from config data_transformer_for_predict.

By default, this param is mlup.ml.data_transformers.numpy_data_transformer.NumpyDataTransformer.

Return JSON:

{
  "predict_result": [
    "predict_result_for_obj_1",
    "predict_result_for_obj_2",
    "predict_result_for_obj_3"
  ]
}

predict_result_for_obj_1 will be valid JSON data. These data, after being predicted by the model, are run through data transformers from config data_transformer_for_predicted.

By default, this param is mlup.ml.data_transformers.numpy_data_transformer.NumpyDataTransformer.

Validation

This method have validation for inner request data. It's making from config columns and flag column_validation.

Web application modes

Web application have three works modes:

• directly_to_predict - is Default. User request send directly to model.
• worker_and_queue - ml model starts in thread worker and take data for predict from queue. Web application new user requests send to queue and wait result from results queue.
• batching - ml model start in thread worker and take data for predict from queue. But not for one request, but combines data from several requests and sends it in one large array to the model. Web application new user requests send to queue and wait result from results queue.

This param is naming mode.

import mlup
from mlup.ml.empty import EmptyModel
from mlup import constants

up = mlup.UP(
    ml_model=EmptyModel(),
    conf=mlup.Config(
        mode=constants.WebAppArchitecture.worker_and_queue,
    )
)

If your model is light, or you hae many CPU/GPU/RAM, you can run many processes:

import mlup
from mlup.ml.empty import EmptyModel
from mlup import constants

up = mlup.UP(
    ml_model=EmptyModel(),
    conf=mlup.Config(
        mode=constants.WebAppArchitecture.worker_and_queue,
        uvicorn_kwargs={'workers': 4},
    )
)

Metrics

MlUp PyPI download statistics: https://pepy.tech/project/pymlup

TODO

• Full docs by all pymlup library.
• Added length inner array validation.
• Modern mlup validate-config command.
• Auto search params by model and test data.
• Support multi models pipelines. Now support is only single file model and single model predict method.
• Add C++ implementation this application by config or model. Support everything that is possible according to the mlup config.
• Add hub for save configs.