MLflow modules are software components that deliver the core features that aid in the different phases of the ML life cycle. MLflow features are delivered through modules, extensible components that organize related features in the platform.
The following are the built-in modules in MLflow:
- MLflow Tracking: Provides a mechanism and UI to handle metrics and artifacts generated by ML executions (training and inference)
- Mlflow Projects: A package format to standardize ML projects
- Mlflow Models: A mechanism that deploys to different types of environments, both on-premises and in the cloud
- Mlflow Model Registry: A module that handles the management of models in MLflow and its life cycle, including state
In order to explore the different modules, we will install MLflow in your local environment using the following command:
pip install mlflow
Important note
It is crucial that the technical requirements are correctly installed on your local machine to allow you to follow along. You can also use the pip command with the required permissions.
Exploring MLflow projects
An MLflow project represents the basic unit of organization of ML projects. There are three different environments supported by MLflow projects: the Conda environment, Docker, and the local system.
Important note
Model details of the different parameters available on an MLProject file can be consulted in the official documentation available at https://www.mlflow.org/docs/latest/projects.html#running-projects.
The following is an example of an MLproject file of a conda environment:
name: condapred
conda_env:
image: conda.yaml
entry_points:
main:
command: "python mljob.py"
In the conda option, the assumption is that there is a conda.yaml file with the required dependencies. MLflow, when asked to run the project, will start the environment with the specified dependencies.
The system-based environment will look like the following; it's actually quite simple:
name: syspred
entry_points:
main:
command: "python mljob.py"
The preceding system variant will basically rely on the local environment dependencies, assuming that the underlying operating system contains all the dependencies. This approach is particularly prone to library conflicts with the underlying operating system; it might be valuable in contexts where there is already an existing operating system environment that fits the project.
The following is a Docker environment-based MLproject file:
name: syspred
docker_env:
image: stockpred-docker
entry_points:
main:
command: "python mljob.py"
Once you have your environment, the main file that defines how your project should look is the MLProject file. This file is used by MLflow to understand how it should run your project.
Developing your first end-to-end pipeline in MLflow
We will prototype a simple stock prediction project in this section with MLflow and will document the different files and phases of the solution. You will develop it in your local system using the MLflow and Docker installed locally.
Important note
In this section, we are assuming that MLflow and Docker are installed locally, as the steps in this section will be executed in your local environment.
The task in this illustrative project is to create a basic MLflow project and produce a working baseline ML model to predict, based on market signals over a certain number of days, whether the stock market will go up or down.
In this section, we will use a Yahoo Finance dataset available for quoting the BTC-USD pair in https://finance.yahoo.com/quote/BTC-USD/ over a period of 3 months. We will train a model to predict whether the quote will be going up or not on a given day. A REST API will be made available for predictions through MLflow.
We will illustrate, step by step, the creation of an MLflow project to train a classifier on stock data, using the Yahoo API for financial information retrieved using the package's pandas data reader:
- Add your
MLProject file:name: stockpred
docker_env:
image: stockpred-docker
entry_points:
main:
command: "python train.py"
The preceding MLProject file specifies that dependencies will be managed in Docker with a specific image name. MLflow will try to pull the image using the version of Docker installed on your system. If it doesn't find it, it will try to retrieve it from Docker Hub. For the goals of this chapter, it is completely fine to have MLflow running on your local machine.
The second configuration that we add to our project is the main entry point command. The command to be executed will invoke in the Docker environment the train.py Python file, which contains the code of our project.
- Add a Docker file to the project.
Additionally, you can specify the Docker registry URL of your image. The advantage of running Docker is that your project is not bound to the Python language, as we will see in the advanced section of this book. The MLflow API is available in a Rest interface alongside the official clients: Python, Java, and R:
FROM continuumio/miniconda:4.5.4
RUN pip install mlflow==1.11.0 \
&& pip install numpy==1.14.3 \
&& pip install scipy \
&& pip install pandas==0.22.0 \
&& pip install scikit-learn==0.20.4 \
&& pip install cloudpickle \
&& pip install pandas_datareader>=0.8.0
The preceding Docker image file is based on the open source package Miniconda, a free minimal installer with a minimal set of packages for data science that allow us to control the details of the packages that we need in our environment.
We will specify the version of MLflow (our ML platform), numpy, and scipy for numerical calculations. Cloudpickle allows us to easily serialize objects. We will use pandas to manage data frames, and pandas_datareader to allow us to easily retrieve the data from public sources.
- Import the packages required for the project.
On the following listing, we explicitly import all the libraries that we will use during the execution of the training script: the library to read the data, and the different sklearn modules related to the chosen initial ML model:
import numpy as np
import datetime
import pandas_datareader.data as web
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report
from sklearn.metrics import precision_score
from sklearn.metrics import recall_score
from sklearn.metrics import f1_score
import mlflow.sklearn
We explicitly chose for the stock market movement detection problem a RandomForestClassifier, due to the fact that it's an extremely versatile and widely accepted baseline model for classification problems.
- Acquire your training data.
The component of the code that acquires the Yahoo Finance stock dataset is intentionally small, so we choose a specific interval of 3 months to train our classifier.
The acquire_training_data method returns a pandas data frame with the relevant dataset:
def acquire_training_data():
start = datetime.datetime(2019, 7, 1)
end = datetime.datetime(2019, 9, 30)
df = web.DataReader("BTC-USD", 'yahoo', start, end)
return dfThe format of the data acquired is the classic format for financial securities in exchange APIs. For every day of the period, we retrieve the following data: the highest value of the stock, the lowest, opening, and close values of the stock, as well as the volume. The final column represents the adjusted close value, the value after dividends, and splits:
Figure 1.1 – Excerpt from the acquired data
Figure 1.2 is illustrative of the target variable that we would like to achieve by means of the current data preparation process:
Figure 1.2 – Excerpt from the acquired data with the prediction column
- Make the data usable by scikit-learn.
The data acquired in the preceding step is clearly not directly usable by RandomForestAlgorithm, which thrives on categorical features. In order to facilitate the execution of this, we will transform the raw data into a feature vector using the rolling window technique.
Basically, the feature vector for each day becomes the deltas between the current and previous window days. In this case, we use the previous day's market movement (1 for a stock going up, 0 otherwise):
def digitize(n):
if n > 0:
return 1
return 0
def rolling_window(a, window):
"""
Takes np.array 'a' and size 'window' as parameters
Outputs an np.array with all the ordered sequences of values of 'a' of size 'window'
e.g. Input: ( np.array([1, 2, 3, 4, 5, 6]), 4 )
Output:
array([[1, 2, 3, 4],
[2, 3, 4, 5],
[3, 4, 5, 6]])
"""
shape = a.shape[:-1] + (a.shape[-1] - window + 1, window)
strides = a.strides + (a.strides[-1],)
return np.lib.stride_tricks.as_strided(a, shape=shape, strides=strides)
def prepare_training_data(data):
data['Delta'] = data['Close'] - data['Open']
data['to_predict'] = data['Delta'].apply(lambda d: digitize(d))
return data
The following example is illustrative of the data frame output produced with the binarized ups and downs of the previous days:
Figure 1.3 – Feature vector with binarized market ups and downs
- Train and store your model in MLflow.
This portion of the following code listing calls the data preparation methods declared previously and executes the prediction process.
The main execution also explicitly logs the ML model trained in the current execution in the MLflow environment.
if __name__ == "__main__":
with mlflow.start_run():
training_data = acquire_training_data()
prepared_training_data_df = prepare_training_data(training_data)
btc_mat = prepared_training_data_df.as_matrix()
WINDOW_SIZE = 14
X = rolling_window(btc_mat[:, 7], WINDOW_SIZE)[:-1, :]
Y = prepared_training_data_df['to_predict'].as_matrix()[WINDOW_SIZE:]
X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=0.25, random_state=4284, stratify=Y)
clf = RandomForestClassifier(bootstrap=True, criterion='gini', min_samples_split=2, min_weight_fraction_leaf=0.0, n_estimators=50, random_state=4284, verbose=0)
clf.fit(X_train, y_train)
predicted = clf.predict(X_test)
mlflow.sklearn.log_model(clf, "model_random_forest")
mlflow.log_metric("precision_label_0", precision_score(y_test, predicted, pos_label=0))
mlflow.log_metric("recall_label_0", recall_score(y_test, predicted, pos_label=0))
mlflow.log_metric("f1score_label_0", f1_score(y_test, predicted, pos_label=0))
mlflow.log_metric("precision_label_1", precision_score(y_test, predicted, pos_label=1))
mlflow.log_metric("recall_label_1", recall_score(y_test, predicted, pos_label=1))
mlflow.log_metric("f1score_label_1", f1_score(y_test, predicted, pos_label=1))The mlflow.sklearn.log_model(clf, "model_random_forest") method takes care of persisting the model upon training. In contrast to the previous example, we are explicitly asking MLflow to log the model and the metrics that we find relevant. This flexibility in the items to log allows one program to log multiple models into MLflow.
In the end, your project layout should look like the following, based on the files created previously:
├── Dockerfile
├── MLproject
├── README.md
└── train.py
- Build your project's Docker image.
In order to build your Docker image, you should run the following command:
docker build -t stockpred -f dockerfile
This will build the image specified previously with the stockpred tag. This image will be usable in MLflow in the subsequent steps as the model is now logged into your local registry.
Following execution of this command, you should expect a successful Docker build:
---> 268cb080fed2
Successfully built 268cb080fed2
Successfully tagged stockpred:latest
- Run your project.
In order to run your project, you can now run the MLflow project:
mlflow run .
Your output should look similar to the excerpt presented here:
MLFLOW_EXPERIMENT_ID=0 stockpred:3451a1f python train.py' in run with ID '442275f18d354564b6259a0188a12575' ===
precision recall f1-score support
0 0.61 1.00 0.76 11
1 1.00 0.22 0.36 9
accuracy 0.65 20
macro avg 0.81 0.61 0.56 20
weighted avg 0.79 0.65 0.58 20
2020/10/15 19:19:39 INFO mlflow.projects: === Run (ID '442275f18d354564b6259a0188a12575') succeeded ===
This contains a printout of your model, the ID of your experiment, and the metrics captured during the current run.
At this stage, you have a simple, reproducible baseline of a stock predictor pipeline using MLflow that you can improve on and easily share with others.
Re-running experiments
Another extremely useful feature of MLflow is the ability to re-run a specific experiment with the same parameters as it was run with originally.
For instance, you should be able to run your previous project by specifying the GitHub URL of the project:
mlflow run https://github.com/PacktPublishing/Machine-Learning-Engineering-with-MLflow/tree/master/Chapter01/stockpred
Basically, what happens with the previous command is that MLflow clones the repository to a temporary directory and executes it, according to the recipe on MLProject.
The ID of the experiment (or the name) allows you to run the project with the original parameters, thereby enabling complete reproducibility of the project.
The MLflow projects feature allows your project to run in advanced cloud environments such as Kubernetes and Databricks. Scaling your ML job seamlessly is one of the main selling points of a platform such as MLflow.
As you have seen from the current section, the MLflow project module allows the execution of a reproducible ML job that is treated as a self-contained project.
Exploring MLflow tracking
The MLflow tracking component is responsible for observability. The main features of this module are the logging of metrics, artifacts, and parameters of an MLflow execution. It provides vizualisations and artifact management features.
In a production setting, it is used as a centralized tracking server implemented in Python that can be shared by a group of ML practitioners in an organization. This enables improvements in ML models to be shared within the organization.
In Figure 1.4, you can see an interface that logs all the runs of your model and allows you to log your experiment's observables (metrics, files, models and artifacts). For each run, you can look and compare the different metrics and parameters of your module.
It addresses common pain points when model developers are comparing different iterations of their models on different parameters and settings.
The following screenshot presents the different metrics for our last run of the previous model:
Figure 1.4 – Sample of the MLFlow interface/UI
MLflow allows the inspection of arbitrary artifacts associated with each model and its associated metadata, allowing metrics of different runs to be compared. You can see the RUN IDs and the Git hash of the code that generated the specific run of your experiment:
Figure 1.5 – Inspecting logged model artifacts
In your current directory of stockpred, you can run the following command to have access to the results of your runs:
mlflow ui
Running the MLflow UI locally will make it available at the following URL: http://127.0.0.1:5000/.
In the particular case of the runs shown in the following screenshot, we have a named experiment where the parameter of the size of the window in the previous example was tweaked. Clear differences can be seen between the performance of the algorithms in terms of F1 score:
Figure 1.6 – Listing of MLflow runs
Another very useful feature of MLFlow tracking is the ability to compare between different runs of jobs:
Figure 1.7 – Comparison of F1 metrics of job runs
This preceding visualization allows a practitioner to make a decision as to which model to use in production or whether to iterate further.
Exploring MLflow Models
MLflow Models is the core component that handles the different model flavors that are supported in MLflow and intermediates the deployment into different execution environments.
We will now delve into the different models supported in the latest version of MLflow.
As shown in the Getting started with MLflow section, MLflow models have a specific serialization approach for when the model is persisted in its internal format. For example, the serialized folder of the model implemented on the stockpred project would look like the following:
├── MLmodel
├── conda.yaml
└── model.pkl
Internally, MLflow sklearn models are persisted with the conda files with their dependencies at the moment of being run and a pickled model as logged by the source code:
artifact_path: model_random_forest
flavors:
python_function:
env: conda.yaml
loader_module: mlflow.sklearn
model_path: model.pkl
python_version: 3.7.6
sklearn:
pickled_model: model.pkl
serialization_format: cloudpickle
sklearn_version: 0.23.2
run_id: 22c91480dc2641b88131c50209073113
utc_time_created: '2020-10-15 20:16:26.619071'
~
MLflow, by default, supports serving models in two flavors, namely, as a python_function or in sklearn format. The flavors are basically a format to be used by tools or environments serving models.
A good example of using the preceding is being able to serve your model without any extra code by executing the following command:
mlflow models serve -m ./mlruns/0/b9ee36e80a934cef9cac3a0513db515c/artifacts/model_random_forest/
You have access to a very simple web server that can run your model. Your model prediction interface can be executed by running the following command:
curl http://127.0.0.1:5000/invocations -H 'Content-Type: application/json' -d '{"data":[[1,1,1,1,0,1,1,1,0,1,1,1,0,0]]}' [1]%
The response to the API call to our model was 1; as defined in our predicted variable, this means that in the next reading, the stock will move up.
The final few steps outline how powerful MLflow is as an end-to-end tool for model development, including for the prototyping of REST-based APIs for ML services.
The MLflow Models component allows the creation of custom-made Python modules that will have the same benefits as the built-in models, as long as a prediction interface is followed.
Some of the notable model types supported will be explored in upcoming chapters, including the following:
- XGBoost model format
- R functions
- H2O model
- Keras
- PyTorch
- Sklearn
- Spark MLib
- TensorFlow
- Fastai
Support for the most prevalent ML types of models, combined with its built-in capability for on-premises and cloud deployment, is one of the strongest features of MLflow Models. We will explore this in more detail in the deployment-related chapters.
Exploring MLflow Model Registry
The model registry component in MLflow gives the ML developer an abstraction for model life cycle management. It is a centralized store for an organization or function that allows models in the organization to be shared, created, and archived collaboratively.
The management of the model can be made with the different APIs of MLflow and with the UI. Figure 1.7 demonstrates the Artifacts UI in the tracking server that can be used to register a model:
Figure 1.8 – Registering a model as an artifact
Upon registering the model, you can annotate the registered model with the relevant metadata and manage its life cycle. One example is to have models in a staging pre-production environment and manage the life cycle by sending the model to production:
Figure 1.9 – Managing different model versions and stages
The model registry module will be explored further in the book, with details on how to set up a centralized server and manage ML model life cycles, from conception through to phasing out a model.
United States
Great Britain
India
Germany
France
Canada
Russia
Spain
Brazil
Australia
Singapore
Hungary
Ukraine
Luxembourg
Estonia
Lithuania
South Korea
Turkey
Switzerland
Colombia
Taiwan
Chile
Norway
Ecuador
Indonesia
New Zealand
Cyprus
Denmark
Finland
Poland
Malta
Czechia
Austria
Sweden
Italy
Egypt
Belgium
Portugal
Slovenia
Ireland
Romania
Greece
Argentina
Netherlands
Bulgaria
Latvia
South Africa
Malaysia
Japan
Slovakia
Philippines
Mexico
Thailand
Filter
