Posts in category: Amazon AWS

Searchmetrics is a global provider of search data, software, and consulting solutions, helping customers turn search data into unique business insights. To date, Searchmetrics has helped more than 1,000 companies such as McKinsey & Company, Lowe’s, and AXA find an advantage in the hyper-competitive search landscape.

In 2021, Searchmetrics turned to AWS to help with artificial intelligence (AI) usage to further improve their search insights capabilities.

In this post, we share how Searchmetrics built an AI solution that increased the efficiency of its human workforce by 20% by automatically finding relevant search keywords for any given topic, using Amazon SageMaker and its native integration with Hugging Face.

“Amazon SageMaker made it a breeze to evaluate and integrate Hugging Face’s state-of-the-art NLP models into our systems. The solution we built makes us more efficient and greatly improves our user experience.”– Ioannis Foukarakis, Head of Data, Searchmetrics

Using AI to identify relevance from a list of keywords

A key part of Searchmetrics’ insights offering is its ability to identify the most relevant search keywords for a given topic or search intent.

To do this, Searchmetrics has a team of analysts assessing the potential relevance of certain keywords given a specific seed word. Analysts use an internal tool to review a keyword within a given topic and a generated list of potentially related keywords, and they must then select one or more related keywords that are relevant to that topic.

This manual filtering and selection process was time consuming and slowed down Searchmetrics’s ability to deliver insights to its customers.

To improve this process, Searchmetrics sought to build an AI solution that could use natural language processing (NLP) to understand the intent of a given search topic and automatically rank an unseen list of potential keywords by relevance.

Using SageMaker and Hugging Face to quickly build advanced NLP capabilities

To solve this, Searchmetrics’ engineering team turned to SageMaker, an end-to-end machine learning (ML) platform that helps developers and data scientists quickly and easily build, train, and deploy ML models.

SageMaker accelerates the deployment of ML workloads by simplifying the ML build process. It provides a broad set of ML capabilities on top of a fully managed infrastructure. This removes the undifferentiated heavy lifting that too-often hinders ML development.

Searchmetrics chose SageMaker because of the full range of capabilities it provided at every step of the ML development process:

• SageMaker notebooks enabled the Searchmetrics team to quickly spin up fully managed ML development environments, perform data preprocessing, and experiment with different approaches
• The batch transform capabilities in SageMaker enabled Searchmetrics to efficiently process its inference payloads in bulk, as well as easily integrate into its existing web service in production

Searchmetrics was also particularly interested in the native integration of SageMaker with Hugging Face, an exciting NLP startup that provides easy access to more than 7,000 pre-trained language models through its popular Tranformers library.

SageMaker provides a direct integration with Hugging Face through a dedicated Hugging Face estimator in the SageMaker SDK. This makes it easy to run Hugging Face models on the fully managed SageMaker infrastructure.

With this integration, Searchmetrics was able to test and experiment with a range of different models and approaches to find the best-performing approach to their use case.

The end solution uses a zero-shot classification pipeline to identify the most relevant keywords. Different pre-trained models and query strategies were evaluated, with facebook/bart-large-mnli providing the most promising results.

Using AWS to improve operational efficiency and find new innovation opportunities

With SageMaker and its native integration with Hugging Face, Searchmetrics was able to build, train, and deploy an NLP solution that could understand a given topic and accurately rank an unseen list of keywords based on their relevance. The toolset offered by SageMaker made it easier to experiment and deploy.

When integrated with Searchmetrics’s existing internal tool, this AI capability delivered an average reduction of 20% in the time taken for human analysts to complete their job. This resulted in higher throughput, improved user experience, and faster onboarding of new users.

This initial success has not only improved the operational performance of Searchmetrics’s search analysts, but has also helped Searchmetrics chart a clearer path to deploying more comprehensive automation solutions using AI in its business.

These exciting new innovation opportunities help Searchmetrics continue to improve their insights capabilities, and also help them ensure that customers continue to stay ahead in the hyper-competitive search landscape.

In addition, Hugging Face and AWS announced a partnership earlier in 2022 that makes it even easier to train Hugging Face models on SageMaker. This functionality is available through the development of Hugging Face AWS Deep Learning Containers (DLCs). These containers include Hugging Face Transformers, Tokenizers, and the Datasets library, which allows us to use these resources for training and inference jobs.

For a list of the available DLC images, see available Deep Learning Containers Images, which are maintained and regularly updated with security patches. You can find many examples of how to train Hugging Face models with these DLCs and the Hugging Face Python SDK in the following GitHub repo.

Learn more about how you can accelerate your ability to innovate with AI/ML by visiting Getting Started with Amazon SageMaker, getting hands-on learning content by reviewing the Amazon SageMaker developer resources, or visiting Hugging Face on Amazon SageMaker.

About the Author

Daniel Burke is the European lead for AI and ML in the Private Equity group at AWS. Daniel works directly with Private Equity funds and their portfolio companies, helping them accelerate their AI and ML adoption to improve innovation and increase enterprise value.

Read More

Identifying paraphrased text has business value in many use cases. For example, by identifying sentence paraphrases, a text summarization system could remove redundant information. Another application is to identify plagiarized documents. In this post, we fine-tune a Hugging Face transformer on Amazon SageMaker to identify paraphrased sentence pairs in a few steps.

A truly robust model can identify paraphrased text when the language used may be completely different, and also identify differences when the language used has high lexical overlap. In this post, we focus on the latter aspect. Specifically, we look at whether we can train a model that can identify the difference between two sentences that have high lexical overlap and very different or opposite meanings. For example, the following sentences have the exact same words but opposite meanings:

• I took a flight from New York to Paris
• I took a flight from Paris to New York

Solution overview

We walk you through the following high-level steps:

1. Set up the environment.
2. Prepare the data.
3. Tokenize the dataset.
4. Fine-tune the model.
5. Deploy the model and perform inference.
6. Evaluate model performance.

If you want to skip setting up the environment, you can use the following notebook on GitHub and run the code in SageMaker.

Hugging Face and AWS announced a partnership earlier in 2022 that makes it even easier to train Hugging Face models on SageMaker. This functionality is available through the development of Hugging Face AWS Deep Learning Containers (DLCs). These containers include Hugging Face Transformers, Tokenizers, and the Datasets library, which allows us to use these resources for training and inference jobs. For a list of the available DLC images, see Available Deep Learning Containers Images. They are maintained and regularly updated with security patches. You can find many examples of how to train Hugging Face models with these DLCs and the Hugging Face Python SDK in the following GitHub repo.

The PAWS dataset

Realizing the lack of efficient sentence pairs datasets that exhibit high lexical overlap without being paraphrases, the original PAWS dataset released in 2019 aimed to provide the natural language processing (NLP) community a new resource for training and evaluating paraphrase detection models. PAWS sentence pairs are generated in two steps using Wikipedia and the Quora Question Pairs (QQP) dataset. A language model first swaps words in a sentence pair with the same Bag of Words (BOW) to generate a sentence pair. A back translation step then generates paraphrases with high BOW overlap but using a different word order. The final PAWS dataset contains a total of 108,000 human-labeled and 656,000 noisily labeled pairs.

In this post, we use the PAWS-Wiki Labeled (Final) dataset from Hugging Face. Hugging Face has already performed the data split for us, which results in 49,000 sentence pairs in the training dataset, and 8,000 sentence pairs each for the validation and test datasets. Two sentence pair examples from the training dataset are shown in the following example. A label of 1 indicates that the two sentences are paraphrases of each other.

Prerequisites

You need to complete the following prerequisites:

1. Sign up for an AWS account if you don’t have one. For more information, see Set Up Amazon SageMaker Prerequisites.
2. Get started using SageMaker notebook instances.
3. Set up the right AWS Identity and Access Management (IAM) permissions. For more information, see SageMaker Roles.

Set up the environment

Before we begin examining and preparing our data for model fine-tuning, we need to set up our environment. Let’s start by spinning up a SageMaker notebook instance. Choose an AWS Region in your AWS account and follow the instructions to create a SageMaker notebook instance. The notebook instance may take a few minutes to spin up.

When the notebook instance is running, choose conda_pytorch_p38 as your kernel type. To use the Hugging Face dataset, we first need to install and import the Hugging Face library:

!pip --quiet install "sagemaker" "transformers==4.17.0" "datasets==1.18.4" --upgrade
!pip --quiet install sentence-transformers

import sagemaker.huggingface
import sagemaker
from datasets import load_dataset

Next, let’s establish a SageMaker session. We use the default Amazon Simple Storage Service (Amazon S3) bucket associated with the SageMaker session to store the PAWS dataset and model artifacts:

sess = sagemaker.Session()
role = sagemaker.get_execution_role()
bucket = sess.default_bucket()

Prepare the data

We can load the Hugging Face version of the PAWS dataset with its load_dataset() command. This call downloads and imports the PAWS Python processing script from the Hugging Face GitHub repository, which then downloads the PAWS dataset from the original URL stored in the script and caches the data as an Arrow table on the drive. See the following code:

dataset_train, dataset_val, dataset_test = load_dataset("paws", "labeled_final", split=['train', 'validation', 'test'])

Before we begin fine-tuning our pre-trained BERT model, let’s look at our target class distribution. For our use case, the PAWS dataset has binary labels (0 indicates the sentence pair is not a paraphrase, and 1 indicates it is). Let’s create a column chart to view the class distribution, as shown in the following code. We see that there is a slight class imbalance issue in our training set (56% negative samples vs. 44% positive samples). However, the imbalance is small enough to avoid employing class imbalance mitigation techniques.

import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

df = dataset_train.to_pandas()

ax = sns.countplot(x="label", data=df)
ax.set_title('Label Count for PAWS Dataset', fontsize=15)
for p in ax.patches:
    ax.annotate(f'n{p.get_height()}', (p.get_x()+0.4, p.get_height()), ha='center', va='top', color='white', size=13)

Tokenize the dataset

Before we can begin fine-tuning, we need to tokenize our dataset. As a starting point, let’s say we want to fine-tune and evaluate the roberta-base transformer. We selected roberta-base because it’s a general-purpose transformer that was pre-trained on a large corpus of English data and has frequently shown high performance on a variety of NLP tasks. The model was originally introduced in the paper RoBERTa: A Robustly Optimized BERT Pretraining Approach.

We perform tokenization on the sentences with a roberta-base tokenizer from Hugging Face, which uses byte-level Byte Pair Encoding to split the document into tokens. For more details about the RoBERTa tokenizer, refer to RobertaTokenizer. Because our inputs are sentence pairs, we need to tokenize both sentences simultaneously. Because most BERT models require the input to have a fixed tokenized input length, we set the following parameters: max_len=128 and truncation=True. See the following code:

from transformers import AutoTokenizer
tokenizer_and_model_name = 'roberta-base'

# Download tokenizer
tokenizer = AutoTokenizer.from_pretrained(tokenizer_and_model_name)

# Tokenizer helper function
def tokenize(batch, max_len=128):
    return tokenizer(batch['sentence1'], batch['sentence2'], max_length=max_len, truncation=True)

dataset_train_tokenized = dataset_train.map(tokenize, batched=True, batch_size=len(dataset_train))
dataset_val_tokenized = dataset_val.map(tokenize, batched=True, batch_size=len(dataset_val))

The last preprocessing step for fine-tuning our BERT model is to convert the tokenized train and validation datasets into PyTorch tensors and upload them to our S3 bucket:

import botocore
from datasets.filesystems import S3FileSystem

s3 = S3FileSystem()
s3_prefix = 'sts-sbert-paws/sts-paws-datasets'

# convert and save train_dataset to s3
training_input_path = f's3://{sess.default_bucket()}/{s3_prefix}/train'
dataset_train_tokenized = dataset_train_tokenized.rename_column("label", "labels")
dataset_train_tokenized.set_format('torch', columns=['input_ids', 'attention_mask', 'labels'])
dataset_train_tokenized.save_to_disk(training_input_path,fs=s3)

# convert and save val_dataset to s3
val_input_path = f's3://{sess.default_bucket()}/{s3_prefix}/val'
dataset_val_tokenized = dataset_val_tokenized.rename_column("label", "labels")
dataset_val_tokenized.set_format('torch', columns=['input_ids', 'attention_mask', 'labels'])
dataset_val_tokenized.save_to_disk(val_input_path,fs=s3)

Fine-tune the model

Now that we’re done with data preparation, we’re ready to fine-tune our pre-trained roberta-base model on the paraphrase identification task. We can use the SageMaker Hugging Face Estimator class to initiate the fine-tuning process in two steps. The first step is to specify the training hyperparameters and metric definitions. The metric definitions variable tells the Hugging Face Estimator what types of metrics to extract from the model’s training logs. Here, we’re primarily interested in extracting validation set metrics at each training epoch.

# Step 1: specify training hyperparameters and metric definitions
hyperparameters = {'epochs': 4,
                   'train_batch_size': 16,
                   'model_name': tokenizer_and_model_name}
                   
metric_definitions=[
    {'Name': 'loss', 'Regex': "'loss': ([0-9]+(.|e-)[0-9]+),?"},
    {'Name': 'eval_loss', 'Regex': "'eval_loss': ([0-9]+(.|e-)[0-9]+),?"},
    {'Name': 'eval_accuracy', 'Regex': "'eval_accuracy': ([0-9]+(.|e-)[0-9]+),?"},
    {'Name': 'eval_f1', 'Regex': "'eval_f1': ([0-9]+(.|e-)[0-9]+),?"},
    {'Name': 'eval_precision', 'Regex': "'eval_precision': ([0-9]+(.|e-)[0-9]+),?"},
    {'Name': 'eval_recall', 'Regex': "'eval_recall': ([0-9]+(.|e-)[0-9]+),?"},
    {'Name': 'epoch', 'Regex': "'epoch': ([0-9]+(.|e-)[0-9]+),?"}
]           

The second step is to instantiate the Hugging Face Estimator and start the fine-tuning process with the .fit() method:

# Step 2: instantiate estimator and begin fine-tuning
from sagemaker.huggingface import HuggingFace

huggingface_estimator = HuggingFace(
                            entry_point='train.py',
                            source_dir='./scripts',
                            output_path=f's3://{sess.default_bucket()}',
                            base_job_name='huggingface-sdk-extension',
                            instance_type='ml.p3.8xlarge',
                            instance_count=1,
                            volume_size=100,
                            transformers_version='4.17.0',
                            pytorch_version='1.10.2',
                            py_version='py38',
                            role=role,
                            hyperparameters=hyperparameters,
                            metric_definitions=metric_definitions
                        )
                        
huggingface_estimator.fit({'train': training_input_path, 'test': val_input_path}, 
                          wait=True, 
                          job_name='sm-sts-blog-{}'.format(int(time.time())))

The fine-tuning process takes approximately 30 minutes using the specified hyperparameters.

Deploy the model and perform inference

SageMaker offers multiple deployment options depending on your use case. For persistent, real-time endpoints that make one prediction at a time, we recommend using SageMaker real-time hosting services. If you have workloads that have idle periods between traffic spurts and can tolerate cold starts, we recommend using Serverless Inference. Serverless endpoints automatically launch compute resources and scale them in and out depending on traffic, eliminating the need to choose instance types or manage scaling policies. We demonstrate how to deploy our fine-tuned Hugging Face model to both a real-time inference endpoint and a Serverless Inference endpoint.

Deploy to a real-time inference endpoint

You can deploy a training object onto real-time inference hosting within SageMaker using the .deploy() method. For a full list of the accepted parameters, refer to Hugging Face Model. To start, let’s deploy the model to one instance, by passing in the following parameters: initial_instance_count, instance_type, and endpoint_name. See the following code:

rt_predictor = huggingface_estimator.deploy(initial_instance_count=1,
instance_type="ml.g4dn.xlarge",
endpoint_name="sts-sbert-paws")

The model takes a few minutes to deploy. After the model is deployed, we can submit sample records from the unseen test dataset to the endpoint for inference.

Deploy to a Serverless Inference endpoint

To deploy our training object onto a serverless endpoint, we need to first specify a serverless config file with memory_size_in_mb and max_concurrency arguments:

from sagemaker.serverless.serverless_inference_config import ServerlessInferenceConfig

serverless_config = ServerlessInferenceConfig(
    memory_size_in_mb=6144,
    max_concurrency=1,
)

memory_size_in_mb defines the total RAM size of your serverless endpoint; the minimal RAM size is 1024 MB (1 GB) and it can scale up to 6144 MB (6 GB). Generally, you should aim to choose a memory size that is at least as large as your model size. max_concurrency defines the quota for how many concurrent invocations can be processed at the same time (up to 50 concurrent invocations) for a single endpoint.

We also need to supply the Hugging Face inference image URI, which you can retrieve using the following code:

image_uri = sagemaker.image_uris.retrieve(
    framework="huggingface",
    base_framework_version="pytorch1.10",
    region=sess.boto_region_name,
    version="4.17",
    py_version="py38",
    instance_type="ml.m5.large",
    image_scope="inference",
)

Now that we have the serverless config file, we can create a serverless endpoint in the same way as our real-time inference endpoint, using the .deploy() method:

sl_predictor = huggingface_estimator.deploy(
    serverless_inference_config=serverless_config, image_uri=image_uri
)

The endpoint should be created in a few minutes.

Perform model inference

To make predictions, we need to create the sentence pair by adding the [CLS] and [SEP] special tokens and subsequently submit the input to the model endpoints. The syntax for real-time inference and serverless inference is the same:

import random 

rand = random.randrange(0, 8000)

true_label = dataset_test[rand]['label']
sent_1 = dataset_test[rand]['sentence1']
sent_2 = dataset_test[rand]['sentence2']

sentence_pair = {"inputs": ['[CLS] ' + sent_1 + ' [SEP] ' + sent_2 + ' [SEP]']}


# real-time inference 
print('Sentence 1:', sent_1) 
print('Sentence 2:', sent_2)
print()
print('Inference Endpoint:', rt_predictor.endpoint_name)
print('True Label:', true_label)
print('Predicted Label:', rt_predictor.predict({"inputs": sentence_pair})[0]['label'])
print('Prediction Confidence:', rt_predictor.predict({"inputs": sentence_pair})[0]['score'])

# serverless inference
print('Sentence 1:', sent_1) 
print('Sentence 2:', sent_2)
print()
print('Inference Endpoint:', sl_predictor.endpoint_name)
print('True Label:', true_label)
print('Predicted Label:', sl_predictor.predict({"inputs": sentence_pair})[0]['label'])
print('Prediction Confidence:', sl_predictor.predict({"inputs": sentence_pair})[0]['score'])

In the following examples, we can see the model is capable of correctly classifying whether the input sentence pair contains paraphrased sentences.

The following is a real-time inference example.

The following is a Serverless Inference example.

Evaluate model performance

To evaluate the model, let’s expand the preceding code and submit all 8,000 unseen test records to the real-time endpoint:

from tqdm import tqdm

preds = []
labels = []

# Inference takes ~5 minutes for all test records using a fine-tuned roberta-base and ml.g4dn.xlarge instance

for i in tqdm(range(len(dataset_test))):
    true_label = dataset_test[i]['label']
    sent_1 = dataset_test[i]['sentence1']
    sent_2 = dataset_test[i]['sentence2']
    
    sentence_pair = {"inputs": ['[CLS] ' + sent_1 + ' [SEP] ' + sent_2 + ' [SEP]']}
    pred = rt_predictor.predict(sentence_pair)
    
    labels.append(true_label)
    preds.append(int(pred[0]['label'].split('_')[1]))

Next, we can create a classification report using the extracted predictions:

from sklearn.metrics import classification_report

print('Endpoint Name:', rt_predictor.endpoint_name)
class_names = ['paraphase', 'not paraphrase']
print(classification_report(labels, preds, target_names=class_names))

We get the following test scores.

We can observe that roberta-base has a combined macro-average F1 score of 92% and performs slightly better at detecting sentences that are paraphrases. The roberta-base model performs well, but it’s good practice to calculate model performance using at least one other model.

The following table compares roberta-base performance results on the same test set against another fine-tuned transformer called paraphrase-mpnet-base-v2, a sentence transformer pre-trained specifically for the paraphrase identification task. Both models were trained on an ml.p3.8xlarge instance.

The results show that roberta-base has a 1% higher F1 score with very similar training and inference times using real-time inference hosting on SageMaker. The performance difference between the models is relatively minor, however, roberta-base is ultimately the winner since it has marginally better performance metrics and almost identical training and inference times.

Clean up

When you’re done using the model endpoints, you can delete them to avoid incurring future charges:

rt_predictor.delete_endpoint()
sl_predictor.delete_endpoint()

Conclusion

In this post, we discussed how to rapidly build a paraphrase identification model using Hugging Face transformers on SageMaker. We fine-tuned two pre-trained transformers, roberta-base and paraphrase-mpnet-base-v2, using the PAWS dataset (which contains sentence pairs with high lexical overlap). We demonstrated and discussed the benefits of real-time inference vs. Serverless Inference deployment, the latter being a new feature that targets spiky workloads and eliminates the need to manage scaling policies. On an unseen test set with 8,000 records, we demonstrated that both models achieved an F1 score greater than 90%.

To expand on this solution, consider the following:

• Try fine-tuning with your own custom dataset. If you don’t have sufficient training labels, you could evaluate the performance of a fine-tuned model like the one demonstrated in this post on a custom test dataset.
• Integrate this fine-tuned model into a downstream application that requires information on whether two sentences (or blocks of text) are paraphrases of each other.

Happy building!

About the Authors

Bala Krishnamoorthy is a Data Scientist with AWS Professional Services, where he enjoys applying machine learning to solve customer business problems. He specializes in natural language processing use cases and has worked with customers in industries such as software, finance and healthcare. In his free time, he enjoys trying new food, watching comedies and documentaries, working out at Orange Theory, and being out on the water (paddle-boarding, snorkeling and hopefully diving soon).

Ivan Cui is a Data Scientist with AWS Professional Services, where he helps customers build and deploy solutions using machine learning on AWS. He has worked with customers across diverse industries, including software, finance, pharmaceutical, and healthcare. In his free time, he enjoys reading, spending time with his family, and maximizing his stock portfolio.

Read More

This is a guest post by Moovit’s Software and Cloud Architect, Sharon Dahan.

Moovit, an Intel company, is a leading Mobility as a Service (MaaS) solutions provider and creator of the top urban mobility app. Moovit serves over 1.3 billion riders in 3,500 cities around the world.

We help people everywhere get to their destination in the smoothest way possible, by combining all options for real-time trip planning and payment in one app. We provide governments, cities, transit agencies, operators, and all organizations with mobility challenges with AI-powered mobility solutions that cover planning, operations, and analytics.

In this post, we describe how Moovit built an automated pipeline to train and deploy BERT models which classify public transportation service alerts in multiple metropolitan areas using Apache Airflow and Amazon SageMaker.

The service alert challenge

One of the key features in Moovit’s urban mobility app is offering access to transit service alerts (sourced from local operators and agencies) to app users around the world.

A service alert is a text message that describes a change (which can be positive or negative) in public transit service. These alerts are typically communicated by the operator in a long textual format and need to be analyzed in order to classify their potential impact on the user’s trip plan. The service alert classification affects the way transit recommendations are shown in the app. An incorrect classification may cause users to ignore important service interruptions that may impact their trip plan.

Existing solution and classification challenges

Historically, Moovit applied both automated rule-based classification (which works well for simple logic) as well as manual human classification for more complex cases.

For example, the following alert “Line 46 will arrive 10 min later as a result of an accident with a deer.” Can be classified into one of the following categories:

1: "NO_SERVICE",
2: "REDUCED_SERVICE",
3: "SIGNIFICANT_DELAYS",
4: "DETOUR",
5: "ADDITIONAL_SERVICE",
6: "MODIFIED_SERVICE",
7: "OTHER_EFFECT",
9: "STOP_MOVED",

The above example should be classified as 3, which is SIGNIFICANT_DELAYS.

The existing rule-based classification solution searches the text for key phrases (for example delay or late) as illustrated in the following diagram.

While the rule-based classification engine offered accurate classifications, it was able to classify only 20% of the service alerts requiring the other 80% to be manually classified. This was not scalable and resulted in gaps in our service alerts coverage.

NLP based classification with a BERT framework

We decided to leverage a neural network that can learn to classify service alerts and selected the BERT model for this challenge.

BERT (Bidirectional Encoder Representations from Transformers) is an open-source machine learning (ML) framework for natural language processing (NLP). BERT is designed to help computers understand the meaning of ambiguous language in the text by using surrounding text to establish context. The BERT framework was pre-trained using text from the BooksCorpus with 800M words and English Wikipedia with 2,500M words, and can be fine-tuned with question-answer datasets.

We leveraged classified data from our rule-based classification engine as ground truth for the training job and explored two possible approaches:

• Approach 1: The first approach was to train using the BERT pre-trained model which meant adding our layers in the beginning and at the end of the pre-trained model.
• Approach 2: The second approach was to use the BERT tokenizer with a standard five-layer model.

Comparison tests showed that, due to the limited amount of available ground truth data, the BERT tokenizer approach yielded better results, was less time-consuming, and required minimal compute resources for training. The model was able to successfully classify service alerts that could not be classified with the existing rule-based classification engine.

The following diagram illustrates the model’s high-level architecture.

After we have the trained model, we deploy it to a SageMaker endpoint and expose it to the Moovit backend server (with request payload being the service alert’s raw text). See the following example code:

{
   "instances": [
       "Expect longer waits for </br> B4, B8, B11, B12, B14, B17, B24, B35, B38, B47, B48, B57, B60, B61, B65, B68, B82, and B83 buses.rnrnWe're working to provide as much service as possible."
   ]
}

The response is the classification and the level of confidence:

{
   "response": [
       {
           "id": 1,
           "prediction": "SIGNIFICANT_DELAYS",
           "confidance": 0.921
       }
   ]
}

From research to production – overcoming operational challenges

Once we trained an NLP model, we had to overcome several challenges in order to enable our app users to access service alerts at scale and in a timely manner:

• How do we deploy a model to our production environment?
• How do we serve the model at scale with low latency?
• How do we re-train the model in order to future proof our solution?
• How do we expand to other metropolitan areas (aka “metros”) in an efficient way?

Prior to using SageMaker, we used to take the trained ML models and manually integrate them into our backend environment. This created a dependency between the model deployment and a backend upgrade. As a result, our ability to deploy new models was very limited and resulted in extremely rare model updates.

In addition, serving an ML model can require substantial compute resources which are difficult to predict and need to be provisioned for in advance in order to ensure adherence to our strict latency requirements.  When the model is served within the backend this can cause unnecessary scaling of compute resources and erratic behavior.

The solution to both these challenges was to use SageMaker endpoints for our real time inference requirements. This enabled us to (1) de-couple the model serving and deployment cycle from the backend release schedule and (2) de-couple the resource provisioning required for model serving (also in peak periods) from the backend provisioning.

Because our group already had deep experience with Airflow, we decided to automate the entire pipeline using Airflow operators in conjunction with SageMaker. As you can see below, we built a full CI/CD pipeline to automate data collection, model re-training and to manage the deployment process. This pipeline can also be leveraged to make the entire process scalable to new metropolitan areas, as we continue to increase our coverage in additional cities worldwide.

AI Lake architecture

The architecture shown in the following diagram is based on SageMaker and Airflow; all endpoints exposed to developers use Amazon API Gateway. This implementation was dubbed “AI lake”.

SageMaker helps data scientists and developers to prepare, build, train, and deploy high-quality machine learning models quickly by bringing together a broad set of capabilities purpose-built for machine learning.

Moovit uses SageMaker to automate the training and deployment process. The trained models are saved to Amazon Simple Storage Service (Amazon S3) and cataloged.

SageMaker helps us significantly reduce the need for engineering time and lets us focus more on developing features for the business and less on the infrastructure required to support the model’s lifecycle. Below you can see Moovit’s SageMaker Training Jobs.

After we train the Metro’s model, we expose it using the SageMaker endpoint. SageMaker enables us to deploy a new version seamlessly to the app, without any downtime.

Moovit uses API Gateway to expose all models under the same domain, as shown in the following screenshot.

Moovit decided to use Airflow to schedule and create a holistic workflow. Each model has its own workflow, which includes the following steps:

• Dataset generation – The owner of this step is the BI team. This step automatically creates a fully balanced dataset with which to train the model. The final dataset is saved to an S3 bucket.
• Train – The owner of this step is the server team. This step fetches the dataset from the previous step and trains the model using SageMaker. SageMaker takes care of the whole training process, such as provisioning the instance, running the training code, saving the model, and saving the training job results and logs.
• Verify – This step is owned by the data science team. During the verification step, Moovit runs a confusion matrix and checks some of the parameters to make sure that the model is healthy and stands within proper thresholds. If the new model misses the criteria, the flow is canceled and the deploy step doesn’t run.
• Deploy – The owner of this step is the DevOps teams. This step triggers the deploy function for SageMaker (using Boto3) to update the existing endpoint or create a new one.

Results

With the AI lake solution and service alert classification model, Moovit accomplished two major achievements:

• Functional – In Metros where the service alert classification model was deployed, Moovit has achieved x3 growth in percentage of classified service alerts! (from 20% to over 60%)
• Operational – Moovit now has the ability to maintain and develop more ML models with less engineering effort, and with very clear and outlined best practices and responsibilities. This opens new opportunities for integrating AI and ML models into Moovit’s products and technologies.

The following charts illustrate the service alert classifications before (left) and after (right) implementing this solution – the turquoise area is the unclassified alerts (aka “modified service”).

Conclusion

In this post, we shared how Moovit used SageMaker with AirFlow to improve the number of classified service alerts by 200% (x3). Moovit is now able to maintain and develop more ML models with less engineering efforts and with very clear practices and responsibilities.

For further reading, refer to the following:

• Build end-to-end machine learning workflows with Amazon SageMaker and Apache Airflow
• Amazon SageMaker Operators in Apache Airflow
• SageMaker Workshop: AirFlow Integration

About the Authors

Sharon Dahan is a Software & Cloud Architect at Moovit. He is responsible for bringing innovative and creative solutions which can stand within Moovit’s tremendous scale. In his spare time, Sharon makes tasty hoppy beer.

Miron Perel is a Senior Machine Learning Business Development Manager with Amazon Web Services. Miron helps enterprise organizations harness the power of data and Machine Learning to innovate and grow their business.

Eitan Sela is a Machine Learning Specialist Solutions Architect with Amazon Web Services. He works with AWS customers to provide guidance and technical assistance, helping them build and operate machine learning solutions on AWS. In his spare time, Eitan enjoys jogging and reading the latest machine learning articles.

Read More

In this post, we demonstrate Kubeflow on AWS (an AWS-specific distribution of Kubeflow) and the value it adds over open-source Kubeflow through the integration of highly optimized, cloud-native, enterprise-ready AWS services.

Kubeflow is the open-source machine learning (ML) platform dedicated to making deployments of ML workflows on Kubernetes simple, portable and scalable. Kubeflow provides many components, including a central dashboard, multi-user Jupyter notebooks, Kubeflow Pipelines, KFServing, and Katib, as well as distributed training operators for TensorFlow, PyTorch, MXNet, and XGBoost, to build simple, scalable, and portable ML workflows.

AWS recently launched Kubeflow v1.4 as part of its own Kubeflow distribution (called Kubeflow on AWS), which streamlines data science tasks and helps build highly reliable, secure, portable, and scalable ML systems with reduced operational overheads through integrations with AWS managed services. You can use this Kubeflow distribution to build ML systems on top of Amazon Elastic Kubernetes Service (Amazon EKS) to build, train, tune, and deploy ML models for a wide variety of use cases, including computer vision, natural language processing, speech translation, and financial modeling.

Challenges with open-source Kubeflow

When you use an open-source Kubeflow project, it deploys all Kubeflow control plane and data plane components on Kubernetes worker nodes. Kubeflow component services are deployed as part of the Kubeflow control plane, and all resource deployments related to Jupyter, model training, tuning, and hosting are deployed on the Kubeflow data plane. The Kubeflow control plane and data plane can run on the same or different Kubernetes worker nodes. This post focuses on Kubeflow control plane components, as illustrated in the following diagram.

This deployment model may not provide an enterprise-ready experience due to the following reasons:

• All Kubeflow control plane heavy lifting infrastructure components, including database, storage, and authentication, are deployed in the Kubernetes cluster worker node itself. This makes it challenging to implement a highly available Kubeflow control plane design architecture with a persistent state in the event of worker node failure.
• Kubeflow control plane generated artifacts (such as MySQL instances, pod logs, or MinIO storage) grow over time and need resizable storage volumes with continuous monitoring capabilities to meet the growing storage demand. Because the Kubeflow control plane shares resources with Kubeflow data plane workloads (for example, for training jobs, pipelines, and deployments), right-sizing and scaling Kubernetes cluster and storage volumes can become challenging and result in increased operational cost.
• Kubernetes restricts the log file size, with most installations keeping the most recent limit of 10 MB. By default, the pod logs become inaccessible after they reach this upper limit. The logs could also become inaccessible if pods are evicted, crashed, deleted, or scheduled on a different node, which could impact your application log availability and monitoring capabilities.

Kubeflow on AWS

Kubeflow on AWS provides a clear path to use Kubeflow, with the following AWS services:

• Application Load Balancer for secure external traffic management over HTTPS
• Amazon CloudWatch for persistent log management
• AWS Cognito for user authentication with Transport Layer Security (TLS)
• AWS Deep Learning Containers for highly optimized Jupyter notebook server images
• Amazon Elastic File System (Amazon EFS) or Amazon FSx for Lustre for a simple, scalable, and serverless file storage solution for increased training performance
• Amazon EKS for managed Kubernetes clusters
• Amazon Relational Database Service (Amazon RDS) for highly scalable pipelines and a metadata store
• AWS Secrets Manager to protect secrets needed to access your applications
• Amazon Simple Storage Service (Amazon S3) for an easy-to-use pipeline artifacts store

These AWS service integrations with Kubeflow (as shown in the following diagram) allow us to decouple critical parts of the Kubeflow control plane from Kubernetes, providing a secure, scalable, resilient, and cost-optimized design.

Let’s discuss the benefits of each service integration and their solutions around security, running ML pipelines, and storage.

Secure authentication of Kubeflow users with Amazon Cognito

Cloud security at AWS is the highest priority, and we’re investing in tightly integrating Kubeflow security directly into the AWS shared-responsibility security services, such as the following:

• Application Load Balancer (ALB) for external traffic management
• AWS Certificate Manager (ACM) to support TLS
• IAM roles for service accounts (IRSA) for fine-grained access control at the Kubernetes Pod level
• AWS Key Management Service (AWS KMS) for data encryption key management
• AWS Shield for DDoS protection

In this section, we focus on AWS Kubeflow control plane integration with Amazon Cognito. Amazon Cognito removes the need to manage and maintain a native Dex (open-source OpenID Connect (OIDC) provider backed by local LDAP) solution for user authentication and makes secret management easier.

You can also use Amazon Cognito to add user sign-up, sign-in, and access control to your Kubeflow UI quickly and easily. Amazon Cognito scales to millions of users and supports sign-in with social identity providers (IdPs), such as Facebook, Google, and Amazon, and enterprise IdPs via SAML 2.0. This reduces the complexity in your Kubeflow setup, making it operationally lean and easier to operate to achieve multi-user isolation.

Let’s look at a multi-user authentication flow with Amazon Cognito, ALB, and ACM integrations with Kubeflow on AWS. There are a number of key components as part of this integration. Amazon Cognito is configured as an IdP with an authentication callback configured to route the request to Kubeflow after user authentication. As part of the Kubeflow setup, a Kubernetes ingress resource is created to manage external traffic to the Istio Gateway service. The AWS ALB Ingress Controller provisions a load balancer for that ingress. We use Amazon Route 53 to configure a public DNS for the registered domain and create certificates using ACM to enable TLS authentication at the load balancer.

The following diagram shows the typical user workflow of logging in to Amazon Cognito and getting redirected to Kubeflow in their respective namespace.

The workflow contains the following steps:

1. The user sends an HTTPS request to the Kubeflow central dashboard hosted behind a load balancer. Route 53 resolves the FQDN to the ALB alias record.
2. If the cookie isn’t present, the load balancer redirects the user to the Amazon Cognito authorization endpoint so that Amazon Cognito can authenticate the user.
3. After the user is authenticated, Amazon Cognito sends the user back to the load balancer with an authorization grant code.
4. The load balancer presents the authorization grant code to the Amazon Cognito token endpoint.
5. Upon receiving a valid authorization grant code, Amazon Cognito provides the ID token and access token to load balancer.
6. After your load balancer authenticates a user successfully, it sends the access token to the Amazon Cognito user info endpoint and receives user claims. The load balancer signs and adds user claims to the HTTP header x-amzn-oidc-* in a JSON web token (JWT) request format.
7. The request from the load balancer is sent to the Istio Ingress Gateway’s pod.
8. Using an envoy filter, Istio Gateway decodes the x-amzn-oidc-data value, retrieves the email field, and adds the custom HTTP header kubeflow-userid, which is used by the Kubeflow authorization layer.
9. The Istio resource-based access control policies are applied to the incoming request to validate the access to the Kubeflow Dashboard. If either of those are inaccessible to the user, an error response is sent back. If the request is validated, it’s forwarded to the appropriate Kubeflow service and provides access to the Kubeflow Dashboard

Persisting Kubeflow component metadata and artifact storage with Amazon RDS and Amazon S3

Kubeflow on AWS provides integration with Amazon Relational Database Service (Amazon RDS) in Kubeflow Pipelines and AutoML (Katib) for persistent metadata storage, and Amazon S3 in Kubeflow Pipelines for persistent artifact storage. Let’s continue to discuss Kubeflow Pipelines in more detail.

Kubeflow Pipelines is a platform for building and deploying portable, scalable ML workflows. These workflows can help automate complex ML pipelines using built-in and custom Kubeflow components. Kubeflow Pipelines includes Python SDK, a DSL compiler to convert Python code into a static config, a Pipelines service that runs pipelines from the static configuration, and a set of controllers to run the containers within the Kubernetes Pods needed to complete the pipeline.

Kubeflow Pipelines metadata for pipeline experiments and runs are stored in MySQL, and artifacts including pipeline packages and metrics are stored in MinIO.

As shown in the following diagram, Kubeflow on AWS lets you store the following components with AWS managed services:

• Pipeline metadata in Amazon RDS – Amazon RDS provides a scalable, highly available, and reliable Multi-AZ deployment architecture with a built-in automated failover mechanism and resizable capacity for an industry-standard relational database like MySQL. It manages common database administration tasks without needing to provision infrastructure or maintain software.
• Pipeline artifacts in Amazon S3 – Amazon S3 offers industry-leading scalability, data availability, security, and performance, and could be used to meet your compliance requirements.

These integrations help offload the management and maintenance of the metadata and artifact storage from self-managed Kubeflow to AWS managed services, which is easier to set up, operate, and scale.

Support for distributed file systems with Amazon EFS and Amazon FSx

Kubeflow builds upon Kubernetes, which provides an infrastructure for large-scale, distributed data processing, including training and tuning large models with a deep network with millions or even billions of parameters. To support such distributed data processing ML systems, Kubeflow on AWS provides integration with the following storage services:

1. Amazon EFS – A high-performance, cloud-native, distributed file system, which you could manage through an Amazon EFS CSI driver. Amazon EFS provides ReadWriteMany access mode, and you can now use it to mount into pods (Jupyter, model training, model tuning) running in a Kubeflow data plane to provide a persistent, scalable, and shareable workspace that automatically grows and shrinks as you add and remove files with no need for management.
2. Amazon FSx for Lustre – An optimized file system for compute-intensive workloads, such as high-performance computing and ML, that you can manage through the Amazon FSx CSI driver. FSx for Lustre provides ReadWriteMany access mode as well, and you can use it to cache training data with direct connectivity to Amazon S3 as the backing store, which you can use to support Jupyter notebook servers or distributed training running in a Kubeflow data plane. With this configuration, you don’t need to transfer data to the file system before using the volume. FSx for Lustre provides consistent submillisecond latencies and high concurrency, and can scale to TB/s of throughput and millions of IOPS.

Kubeflow deployment options

AWS provides various Kubeflow deployment options:

• Deployment with Amazon Cognito
• Deployment with Amazon RDS and Amazon S3
• Deployment with Amazon Cognito, Amazon RDS, and Amazon S3
• Vanilla deployment

For details on service integration and available add-ons for each of these options, refer to Deployment Options. You can fit the option that best fits your use case.

In the following section, we walk through the steps to install AWS Kubeflow v1.4 distribution on Amazon EKS. Then we use the existing XGBoost pipeline example available on the Kubeflow central UI dashboard to demonstrate the integration and usage of AWS Kubeflow with Amazon Cognito, Amazon RDS, and Amazon S3, with Secrets Manager as an add-on.

Prerequisites

For this walkthrough, you should have the following prerequisites:

• An AWS account.
• An existing Amazon EKS cluster. It should be Kubernetes version 1.19 or higher. For automated cluster creation using eksctl, see Create an Amazon EKS Cluster and use the eksctl option.

Install the following tools on the client machine used to access your Kubernetes cluster. You can use AWS Cloud9, a cloud-based integrated development environment (IDE) for the Kubernetes cluster setup.

• AWS Command Line Interface (AWS CLI) – A command line tool for interacting with AWS services. For installation instructions, refer to Installing, updating, and uninstalling the AWS CLI.
• eksctl > 0.56 – A command line tool for working with Amazon EKS clusters that automates many individual tasks.
• kubectl – A command line tool for working with Kubernetes clusters.
• git – A distributed version control software.
• Python 3.8+ – The Python programming environment.
• pip – The package manager for Python.
• kustomize version 3.2.0 – A command line tool to customize Kubernetes objects through a kustomization file.

Install Kubeflow on AWS

Configure kubectl so that you can connect to an Amazon EKS cluster:

# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: MIT-0

# Set the cluster name, region where the cluster exists
export CLUSTER_NAME=<CLUSTER_NAME>
export CLUSTER_REGION=<CLUSTER_REGION>

aws eks update-kubeconfig --name $CLUSTER_NAME --region $CLUSTER_REGION
kubectl config current-context

Various controllers in Kubeflow deployment use IAM roles for service accounts (IRSA). An OIDC provider must exist for your cluster to use IRSA. Create an OIDC provider and associate it with for your Amazon EKS cluster by running the following command, if your cluster doesn’t already have one:

eksctl utils associate-iam-oidc-provider --cluster ${CLUSTER_NAME} 
--region ${CLUSTER_REGION} --approve

Clone the AWS manifests repo and Kubeflow manifests repo, and checkout the respective release branches:

git clone https://github.com/awslabs/kubeflow-manifests.git
cd kubeflow-manifests
git checkout v1.4.1-aws-b1.0.0
git clone --branch v1.4.1 https://github.com/kubeflow/manifests.git upstream

export kubeflow_manifest_dir=$PWD

For more information about these versions, refer to Releases and Versioning.

Set up Amazon RDS, Amazon S3, and Secrets Manager

You create Amazon RDS and Amazon S3 resources before you deploy the Kubeflow manifests. We use automated Python scripts that take care of creating the S3 bucket, RDS database, and required secrets in Secrets Manager. It also edits the required configuration files for the Kubeflow pipeline and AutoML to be properly configured for the RDS database and S3 bucket during Kubeflow installation.

Create an IAM user with permissions to allow GetBucketLocation and read and write access to objects in an S3 bucket where you want to store the Kubeflow artifacts. Use the AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY of the IAM user in the following code:

cd ${kubeflow_manifest_dir}/tests/e2e/
export BUCKET_NAME=<S3_BUCKET_NAME>
export S3_ACCESS_KEY_ID=<YOUR_ACCESS_KEY_ID_FOR_S3>
export S3_SECRET_ACCESS_KEY=<YOUR_SECRET_ACCESS_KEY_FOR_S3>

#Install the dependencies for the script
pip install -r requirements.txt

#Replace YOUR_CLUSTER_REGION, YOUR_CLUSTER_NAME and YOUR_S3_BUCKET with your values.
PYTHONPATH=.. python utils/rds-s3/auto-rds-s3-setup.py --region ${CLUSTER_REGION} --cluster ${CLUSTER_NAME} --bucket ${BUCKET_NAME} --db_name kubeflow --db_root_user admin --db_root_password password --s3_aws_access_key_id ${S3_ACCESS_KEY_ID} --s3_aws_secret_access_key ${S3_SECRET_ACCESS_KEY}

Set up Amazon Cognito as the authentication provider

In this section, we create a custom domain in Route 53 and ALB to route external traffic to Kubeflow Istio Gateway. We use ACM to create a certificate to enable TLS authentication at ALB and Amazon Cognito to maintain the user pool and manage user authentication.

Substitute the following values in

${kubeflow_manifest_dir}/tests/e2e/utils/cognito_bootstrap/config.yaml:

• route53.rootDomain.name – The registered domain. Let’s assume this domain is example.com.
• route53.rootDomain.hostedZoneId – If your domain is managed in Route53, enter the hosted zone ID found under the hosted zone details. Skip this step if your domain is managed by another domain provider.
• route53.subDomain.name – The name of the subdomain where you want to host Kubeflow (for example, platform.example.com). For more information about subdomains, refer to Deploying Kubeflow with AWS Cognito as IdP.
• cluster.name – The cluster name and where Kubeflow is deployed.
• cluster.region – The cluster Region where Kubeflow is deployed (for example, us-west-2).
• cognitoUserpool.name – The name of the Amazon Cognito user pool (for example, kubeflow-users).

The config file looks something like the following code:

cognitoUserpool:
     name: kubeflow-users
 cluster:
     name: kube-eks-cluster
     region: us-west-2
 route53:
     rootDomain:
         hostedZoneId: XXXX
         name: example.com
     subDomain:
         name: platform.example.com

Run the script to create the resources:

cd ${kubeflow_manifest_dir}/tests/e2e/
PYTHONPATH=.. python utils/cognito_bootstrap/cognito_pre_deployment.py

The script updates the config.yaml file with the resource names, IDs, and ARNs it created. It looks something like the following code:

cognitoUserpool:
     ARN: arn:aws:cognito-idp:us-west-2:123456789012:userpool/us-west-2_yasI9dbxF
     appClientId: 5jmk7ljl2a74jk3n0a0fvj3l31
     domainAliasTarget: xxxxxxxxxx.cloudfront.net
     domain: auth.platform.example.com
     name: kubeflow-users
 kubeflow:
     alb:
         serviceAccount:
             name: alb-ingress-controller
             policyArn: arn:aws:iam::123456789012:policy/alb_ingress_controller_kube-eks-clusterxxx
 cluster:
     name: kube-eks-cluster
     region: us-west-2
 route53:
     rootDomain:
         certARN: arn:aws:acm:us-east-1:123456789012:certificate/9d8c4bbc-3b02-4a48-8c7d-d91441c6e5af
         hostedZoneId: XXXXX
         name: example.com
     subDomain:
         us-west-2-certARN: arn:aws:acm:us-west-2:123456789012:certificate/d1d7b641c238-4bc7-f525-b7bf-373cc726
         hostedZoneId: XXXXX
         name: platform.example.com
         us-east-1-certARN: arn:aws:acm:us-east-1:123456789012:certificate/373cc726-f525-4bc7-b7bf-d1d7b641c238

Build manifests and deploy Kubeflow

Deploy Kubeflow using the following command:

while ! kustomize build ${kubeflow_manifest_dir}/docs/deployment/cognito-rds-s3 | kubectl apply -f -; do echo "Retrying to apply resources"; sleep 10; done

Update the domain with the ALB address

The deployment creates an ingress-managed AWS application load balancer. We update the DNS entries for the subdomain in Route 53 with the DNS of the load balancer. Run the following command to check if the load balancer is provisioned (this takes around 3–5 minutes):

kubectl get ingress -n istio-system

NAME            CLASS    HOSTS   ADDRESS                                                                  PORTS   AGE
istio-ingress   <none>   *       ebde55ee-istiosystem-istio-2af2-1100502020.us-west-2.elb.amazonaws.com   80      15d

If the ADDRESS field is empty after a few minutes, check the logs of alb-ingress-controller. For instructions, refer to ALB fails to provision.

When the load balancer is provisioned, copy the DNS name of the load balancer and substitute the address for kubeflow.alb.dns in ${kubeflow_manifest_dir}/tests/e2e/utils/cognito_bootstrap/config.yaml. The Kubeflow section of the config file looks like the following code:

 kubeflow:
     alb:
         dns: ebde55ee-istiosystem-istio-2af2-1100502020.us-west-2.elb.amazonaws.com
         serviceAccount:
             name: alb-ingress-controller
             policyArn: arn:aws:iam::123456789012:policy/alb_ingress_controller_kube-eks-clusterxxx

Run the following script to update the DNS entries for the subdomain in Route 53 with the DNS of the provisioned load balancer:

cd ${kubeflow_manifest_dir}/tests/e2e/
PYTHONPATH=.. python utils/cognito_bootstrap/cognito_post_deployment.py

Troubleshooting

If you run into any issues during the installation, refer to the troubleshooting guide or start fresh by following the “Clean up” section in this blog.

Use case walkthrough

Now that we have completed installing the required Kubeflow components, let’s see them in action using one of the existing examples provided by Kubeflow Pipelines on the dashboard.

Access the Kubeflow Dashboard using Amazon Cognito

To get started, let’s get access to the Kubeflow Dashboard. Because we used Amazon Cognito as the IdP, use the information provided in the official README file. We first create some users on the Amazon Cognito console. These are the users who will log in to the central dashboard. Next, create a profile for the user you created. Then you should be able to access the dashboard through the login page at https://kubeflow.platform.example.com.

The following screenshot shows our Kubeflow Dashboard.

Run the pipeline

On the Kubeflow Dashboard, choose Pipelines in the navigation name. You should see four examples provided by Kubeflow Pipelines that you can run directly to explore various Pipelines features.

For this post, we use the XGBoost sample called [Demo] XGBoost – Iterative model training. You can find the source code on GitHub. This is a simple pipeline that uses the existing XGBoost/Train and XGBoost/Predict Kubeflow pipeline components to iteratively train a model until the metrics are considered good based on specified metrics.

To run the pipeline, complete the following steps:

1. Select the pipeline and choose Create experiment.
   
2. Under Experiment details, enter a name (for this post, demo-blog) and optional description.
3. Choose Next.
   
4. Under Run details¸ choose your pipeline and pipeline version.
5. For Run name, enter a name.
6. For Experiment, choose the experiment you created.
7. For Run type, select One-off.
8. Choose Start.
   

After the pipeline starts running, you should see components completing (within a few seconds). At this stage, you can choose any of the completed components to see more details.

Access the artifacts in Amazon S3

While deploying Kubeflow, we specified Kubeflow Pipelines should use Amazon S3 to store its artifacts. This includes all pipeline output artifacts, cached runs, and pipeline graphs—all of which can then be used for rich visualizations and performance evaluation.

When the pipeline run is complete, you should be able to see the artifacts in the S3 bucket you created during installation. To confirm this, choose any completed component of the pipeline and check the Input/Output section on the default Graph tab. The artifact URLs should point to the S3 bucket that you specified during deployment.

To confirm that the resources were added to Amazon S3, we can also check the S3 bucket in our AWS account via the Amazon S3 console.

The following screenshot shows our files.

Verify ML metadata in Amazon RDS

We also integrated Kubeflow Pipelines with Amazon RDS during deployment, which means that any pipeline metadata should be stored in Amazon RDS. This includes any runtime information such as the status of a task, availability of artifacts, custom properties associated with the run or artifacts, and more.

To verify the Amazon RDS integration, follow the steps provided in the official README file. Specifically, complete the following steps:

1. Get the Amazon RDS user name and password from the secret that was created during the installation:

   export CLUSTER_REGION=<region> 
   
   aws secretsmanager get-secret-value 
       --region $CLUSTER_REGION 
       --secret-id rds-secret 
       --query 'SecretString' 
       --output text

2. Use these credentials to connect to Amazon RDS from within the cluster:

   kubectl run -it --rm --image=mysql:5.7 --restart=Never mysql-client -- mysql -h <YOUR RDS ENDPOINT> -u admin -pKubefl0w

3. When the MySQL prompt opens, we can verify the mlpipelines database as follows:

   mysql> use mlpipeline; show tables;
   
   +----------------------+
   | Tables_in_mlpipeline |
   +----------------------+
   | db_statuses          |
   | default_experiments  |
   | experiments          |
   | jobs                 |
   | pipeline_versions    |
   | pipelines            |
   | resource_references  |
   | run_details          |
   | run_metrics          |
   +----------------------+

4. Now we can read the content of specific tables, to make sure that we can see metadata information about the experiments that ran the pipelines:

   mysql> select * from experiments;
   +--------------------------------------+---------+-------------------------------------------------------------------------+----------------+---------------------------+------------------------+
   | UUID                                 | Name    | Description                                                             | CreatedAtInSec | Namespace                 | StorageState           |
   +--------------------------------------+---------+-------------------------------------------------------------------------+----------------+---------------------------+------------------------+
   | 36ed05cf-e341-4ff4-917a-87c43be8afce | Default | All runs created without specifying an experiment will be grouped here. |     1647214692 |                           | STORAGESTATE_AVAILABLE |
   | 7a1d6b85-4c97-40dd-988b-b3b91cf31545 | run-1   |                                                                         |     1647216152 | kubeflow-user-example-com | STORAGESTATE_AVAILABLE |
   +--------------------------------------+---------+-------------------------------------------------------------------------+----------------+---------------------------+------------------------+
   2 rows in set (0.00 sec)

Clean up

To uninstall Kubeflow and delete the AWS resources you created, complete the following steps:

1. Delete the ingress and ingress-managed load balancer by running the following command:

   kubectl delete ingress -n istio-system istio-ingress

2. Delete the rest of the Kubeflow components:

   kustomize build ${kubeflow_manifest_dir}/docs/deployment/cognito-rds-s3 | kubectl delete -f -

3. Delete the AWS resources created by scripts:
   1. Resources created for Amazon RDS and Amazon S3 integration. Make sure you have the configuration file created by the script in ${kubeflow_manifest_dir}/tests/e2e/utils/rds-s3/metadata.yaml:

      cd ${kubeflow_manifest_dir}/tests/e2e/
      PYTHONPATH=.. python utils/rds-s3/auto-rds-s3-cleanup.py

   2. Resources created for Amazon Cognito integration. Make sure you have the configuration file created by the script in ${kubeflow_manifest_dir}/tests/e2e/utils/cognito_bootstrap/config.yaml:

      cd ${kubeflow_manifest_dir}/tests/e2e/
      PYTHONPATH=.. python utils/cognito_bootstrap/cognito_resources_cleanup.py

4. If you created a dedicated Amazon EKS cluster for Kubeflow using eksctl, you can delete it with the following command:

   eksctl delete cluster --region $CLUSTER_REGION --name $CLUSTER_NAME

Summary

In this post, we highlighted the value that Kubeflow on AWS provides through native AWS-managed service integrations for secure, scalable, and enterprise-ready AI and ML workloads. You can choose from several deployment options to install Kubeflow on AWS with various service integrations. The use case in this post demonstrated Kubeflow integration with Amazon Cognito, Secrets Manager, Amazon RDS, and Amazon S3. To get started with Kubeflow on AWS, refer to the available AWS-integrated deployment options in Kubeflow on AWS.

Starting with v1.3, you can follow the AWS Labs repository to track all AWS contributions to Kubeflow. You can also find us on the Kubeflow #AWS Slack Channel; your feedback there will help us prioritize the next features to contribute to the Kubeflow project.

About the Authors

Kanwaljit Khurmi is an AI/ML Specialist Solutions Architect at Amazon Web Services. He works with the AWS product, engineering and customers to provide guidance and technical assistance helping them improve the value of their hybrid ML solutions when using AWS. Kanwaljit specializes in helping customers with containerized and machine learning applications.

Meghna Baijal is a Software Engineer with AWS AI making it easier for users to onboard their Machine Learning workloads onto AWS by building ML products and platforms such as the Deep Learning Containers, the Deep Learning AMIs, the AWS Controllers for Kubernetes (ACK) and Kubeflow on AWS. Outside of work she enjoys reading, traveling and dabbling in painting.

Suraj Kota is a Software Engineer specialized in Machine Learning infrastructure. He builds tools to easily get started and scale machine learning workload on AWS. He worked on the AWS Deep Learning Containers, Deep Learning AMI, SageMaker Operators for Kubernetes, and other open source integrations like Kubeflow.

Read More

In this post, we walk you through two sampling techniques in Amazon SageMaker Data Wrangler so you can quickly create processing workflows for your data. We cover both random sampling and stratified sampling techniques to help you sample your data based on your specific requirements.

Data Wrangler reduces the time it takes to aggregate and prepare data for machine learning (ML) from weeks to minutes. You can simplify the process of data preparation and feature engineering, and complete each step of the data preparation workflow, including data selection, cleansing, exploration, and visualization, from a single visual interface. With Data Wrangler’s data selection tool, you can choose the data you want from various data sources and import it with a single click. Data Wrangler contains over 300 built-in data transformations so you can quickly normalize, transform, and combine features without having to write any code. With Data Wrangler’s visualization templates, you can quickly preview and inspect that these transformations are completed as you intended by viewing them in Amazon SageMaker Studio, the first fully integrated development environment (IDE) for ML. After your data is prepared, you can build fully automated ML workflows with Amazon SageMaker Pipelines and save them for reuse in Amazon SageMaker Feature Store.

What is sampling and how can it help

In statistical analysis, the total set of observations is known as the population. When working with data, it’s often not computationally feasible to measure every observation from the population. Statistical sampling is a procedure that allows you to understand your data by selecting subsets from the population.

Sampling offers a practical solution that sacrifices some accuracy for the sake of practicality and ease. To ensure your sample is a good representation of overall population, you can employ sampling strategies. Data Wrangler supports two of the most common strategies: random sampling and stratified sampling.

Random sampling

If you have a large dataset, experimentation on that dataset may be time-consuming. Data Wrangler provides random sampling so you can efficiently process and visualize your data. For example, you may want to compute the average number of purchases for a customer within a time frame, or you may want to compute the attrition rate of a subscriber. You can use a random sample to visualize approximations to these metrics.

A random sample from your dataset is chosen so that each element has an equal probability of being selected. This operation is performed in an efficient manner suitable for large datasets, so the sample size returned is approximately the size requested, and not necessarily equal to the size requested.

You can use random sampling if you want to do quick approximate calculations to understand your dataset. As the sample size gets larger, the random sample can better approximate the entire dataset, but unless you include all data points, your random sample may not include all outliers and edge cases. If you want to prepare your entire dataset interactively, you can also switch to a larger instance type.

As a general rule, the sampling error in computing the population mean using a random sample tends to 0 as the sample gets larger. As the sample size increases, the error decreases as the inverse of the square root of the sample size. The takeaway being, the larger the sample, the better the approximation.

Stratified sampling

In some cases, your population can be divided into strata, or mutually exclusive buckets, such as geographic location for addresses, publication year for songs, or tax brackets for incomes. Random sampling is the most popular sampling technique, but if some strata are uncommon in your population, you can use stratified sampling in Data Wrangler to ensure that each strata is proportionally represented in your sample. This may be useful to reduce sampling errors as well as to ensure you’re capturing edge cases during your experimentation.

In the real world, fraudulent credit card transactions are rare events and typically make up less than 1% of your data. If we were to sample randomly, it’s not uncommon for the sample to contain very few or no fraudulent transactions. As a result, when training a model, we would have too few fraudulent examples to learn an accurate model. We can use stratified sampling to make sure we have proportional representation of fraudulent transactions.

In stratified sampling, the size of each strata in the sample is proportional to the size of the strata in the population. This works by dividing your data into strata based on your specified column, selecting random samples from each strata with the correct proportion, and combining those samples into a stratified sample of the population.

Stratified sampling is a useful technique when you want to understand how different groups in your data compare with each other, and you want to ensure you have appropriate representation from each group.

Random sampling when importing from Amazon S3

In this section, we use random sampling with a dataset consisting of both fraudulent and non-fraudulent events from our fraud detection system. You can download the dataset to follow along with this post (CC 4.0 international attribution license).

At the time of this writing, you can import datasets from Amazon Simple Storage Service (Amazon S3), Amazon Athena, Amazon Redshift, and Snowflake. Our dataset is very large, containing 1 million rows. In this case, we want to sample 1,0000 rows on import from Amazon S3 for some interactive experimentation within Data Wrangler.

1. Open SageMaker Studio and create a new Data Wrangler flow.
2. Under Import data, choose Amazon S3.
   
3. Choose the dataset to import.
4. In the Details pane, provide your dataset name and file type.
5. For Sampling, choose Random.
6. For Sample size, enter 10000.
7. Choose Import to load the dataset into Data Wrangler.
   

You can visualize two distinct steps on the data flow page in Data Wrangler. The first step indicates the loading of the sample dataset based on the sampling strategy you defined. After the data is loaded, Data Wrangler performs auto detection of the data types for each of the columns in the dataset. This step is added by default for all datasets.

You can now review the random sampled data in Data Wrangler by adding an analysis.

8. Choose the plus sign next to Data types and choose Analysis.
   
9. For Analysis type¸ choose Scatter Plot.
10. Choose feat_1 and feat_2 as for X axis and Y axis, respectively.
11. For Color by, choose is_fraud.

When you’re comfortable with the dataset, proceed to do further data transformations as per your business requirement to prepare your data for ML.

In the following screenshot, we can observe the fraudulent (dark blue) and non-fraudulent (light blue) transactions in our analysis.

In the next section, we discuss using stratified sampling to ensure the fraudulent cases are chosen proportionally.

Stratified sampling with a transform

Data Wrangler allows you to sample on import, as well as sampling via a transform. In this section, we discuss using stratified sampling via a transform after you have imported your dataset into Data Wrangler.

1. To initiate sampling, on the Data Flow tab, choose the plus sign next to the imported dataset and choose Add Transform.
   

At the time of this writing, Data Wrangler provides more than 300 built-in transformations. In addition to the built-in transforms, you can write your own custom transforms in Pandas or PySpark.

2. From the Add transform list, choose Sampling.
   

You can now use three distinct sampling strategies: limit, random, and stratified.

3. For Sampling method, choose Stratified.
4. Use the is_fraud column as the stratify column.
5. Choose Preview to preview the transformation, then choose Add to add this transformation as a step to your transformation recipe.
   

Your data flow now reflects the added sampling step.

Now we can review the random sampled data by adding an analysis.

6. Choose the plus sign and choose Analysis.
7. For Analysis type¸ choose Histogram.
8. Choose is_fraud for both X axis and Color by.
9. Choose Preview.

In the following screenshot, we can observe the breakdown of fraudulent (dark blue) and non-fraudulent (light blue) cases chosen via stratified sampling in the correct proportions of 20% fraudulent and 80% non-fraudulent.

Conclusion

It is essential to sample data correctly when working with extremely large datasets and to choose the right sampling strategy to meet your business requirements. The effectiveness of your sampling relies on various factors, including business outcome, data availability, and distribution. In this post, we covered how to use Data Wrangler and its built-in sampling strategies to prepare your data.

You can start using this capability today in all Regions where SageMaker Studio is available. To get started, visit Prepare ML Data with Amazon SageMaker Data Wrangler.

Acknowledgements

The authors would like to thank Jonathan Chung (Applied Scientist) for his review and valuable feedback on this article.

About the Authors

Ben Harris is a software engineer with experience designing, deploying, and maintaining scalable data pipelines and machine learning solutions across a variety of domains.

Vishaal Kapoor is a Senior Applied Scientist with AWS AI. He is passionate about helping customers understand their data in Data Wrangler. In his spare time, he mountain bikes, snowboards, and spends time with his family.

Meenakshisundaram Thandavarayan is a Senior AI/ML specialist with AWS. He helps Hi-Tech strategic accounts on their AI and ML journey. He is very passionate about data-driven AI.

Ajai Sharma is a Principal Product Manager for Amazon SageMaker where he focuses on Data Wrangler, a visual data preparation tool for data scientists. Prior to AWS, Ajai was a Data Science Expert at McKinsey and Company, where he led ML-focused engagements for leading finance and insurance firms worldwide. Ajai is passionate about data science and loves to explore the latest algorithms and machine learning techniques.

Read More

The adoption of AWS cloud technology at NatWest Group means moving our machine learning (ML) workloads to a more robust and scalable solution, while reducing our time-to-live to deliver the best products and services for our customers.

In this cloud adoption journey, we selected the Customer Lifetime Value (CLV) model to migrate to AWS. The model enables us to understand our customers better and deliver personalized solutions. The CLV model consists of a series of separate ML models that are brought together into a single pipeline. This requires a scalable solution, given the amount of data it processes.

This is the final post in a four-part series detailing how NatWest Group partnered with AWS Professional Services to build a scalable, secure, and sustainable MLOps platform. This post is intended for ML engineers, data scientists, and C-suites executives who want to understand how to build complex solutions using Amazon SageMaker. It proves the platform’s flexibility by demonstrating how, given some starter code templates, you can deliver a complex, scalable use case quickly and repeatably.

Challenges

NatWest Group, on its mission to delight its customers while remaining in line with regulatory obligations, has been working with AWS to build a standardized and secure solution to implement and productionize ML workloads. Previous implementations in the organization led to data silos and long lead times for environments spin up and spin down. This has also led to underutilized compute resources. To help improve this, AWS and NatWest collaborated to develop a series of ML project and environment templates using AWS services.

NatWest Group is using ML to derive new insights so we can predict and adapt to our customers’ future banking needs across the bank’s retail, wealth, and commercial operations. When developing a model and deploying it to production, many considerations need to be addressed to ensure compliance with the bank’s standards. These include requirements on model explainability, bias, data quality, and drift monitoring. The templates developed through this collaboration incorporate features to assess these points, and are now used by NatWest teams to onboard and productionize their use cases in a secure multi-account setup using SageMaker.

The templates embed standards for production-ready ML workflows by incorporating AWS best practices using MLOps capabilities in SageMaker. They also include a set of self-service, secure, multi-account infrastructure deployments for AWS ML services and data services via Amazon Simple Storage Service (Amazon S3).

The templates help use case teams within NatWest to do the following:

• Building capabilities in line with the NatWest MLOps maturity model
• Implement AWS best practices for MLOps capabilities in SageMaker services
• Create and use templated standards to develop production-ready ML workflows
• Provide self-service, secure infrastructure deployment for AWS ML services
• Reduce time to spin up and spin down environments for projects using a managed approach
• Reduce the cost of maintenance due to standardization of processes, automation, and on-demand compute
• Productionize code, which allows for the migration of existing models where code is functionally decomposed to take advantage of on-demand cloud architectures while improving code readability and remove technical debt
• Use continuous integration, deployment, and training for proof of concept (PoC) and use case development, as well as enable additional MLOps features (model monitoring, explainability, and condition steps)

The following sections discuss how NatWest uses these templates to migrate existing projects or create new ones.

Custom SageMaker templates and architecture

NatWest and AWS built custom project templates with the capabilities of existing SageMaker project templates, and integrated these with infrastructure templates to deploy to production environments. This setup already contains example pipelines for training and inference, and allows NatWest users to immediately use multi-account deployment.

The multi-account setup keeps development environments secure while allowing testing in a production-like environment. This enables project teams to focus on the ML workloads. The project templates take care of infrastructure, resource provisioning, security, auditability, reproducibility, and explainability. They also allow for flexibility so that users can extend the template based on their specific use case requirements.

Solution overview

To get started as a use case team at NatWest, the following steps are required based on the templates built by NatWest and provided via AWS Service Catalog:

1. The project owner provisions new AWS accounts, and the operations team sets up the MLOps prerequisite infrastructure to these new accounts.
2. The project owner creates an Amazon SageMaker Studio data science environment using AWS Service Catalog.
3. The data scientist lead starts a new project and creates SageMaker Studio users via the templates provided in AWS Service Catalog.
4. The project team works on the project, updating the template’s folder structure to aid their use case.

The following figure shows how we adapted the template architecture to fit our CLV use case needs. It showcases how multiple training pipelines are developed and run within a development environment. The deployed models can then be tested and productionized within pre-production and production environments.

The CLV model is a sequence of different models, which includes different tree-based models and an inference setup that combines all the outputs from these. The data used in this ML workload is collected from various sources and amounts to more than half a billion rows and nearly 1,900 features. All processing, feature engineering, model training, and inference tasks on premises were done using PySpark or Python.

SageMaker pipeline templates

CLV is composed of multiple models that are built in a sequence. Each model feeds into another, so we require a dedicated training pipeline for each one. Amazon SageMaker Pipelines allows us to train and register multiple models using the SageMaker model registry. For each pipeline, existing NatWest code was refactored to fit into SageMaker Pipelines, while ensuring the logical components of processing, feature engineering, and model training stayed the same.

In addition, the code for applying the models to perform inference consecutively was refactored into one pipeline dedicated for inference. Therefore, the use case required multiple training pipelines but only one inference pipeline.

Each training pipeline results in the creation of an ML model. To deploy this model, approval from the model approver (a role defined for ML use cases in NatWest) is required. After deployment, the model is available to the SageMaker inference pipeline. The inference pipeline applies the trained models in sequence to the new data. This inference is performed in batches, and the predictions are combined to deliver the final customer lifetime value for each customer.

The template contains a standard MLOps example with the following steps:

• PySpark Processing
• Data Quality Check and Data Bias Check
• Model Training
• Condition
• Create Model
• Model Registry
• Transform
• Model Bias Check, Explainability Check, and Quality Check

This pipeline is described in more detail in Part 3: How NatWest Group built audible, reproducible, and explainable ML models with Amazon SageMaker.

The following figure shows the design of the example pipeline provided by the template.

Data is preprocessed and split into train, test, and validation sets (Processing step) before training the model (Training). After the model is deployed (Create Model), it is used to create batch predictions (Transform). The predictions are then postprocessed and stored in Amazon S3 (Processing).

Data quality and data bias checks provide baseline information on the data used for training. Model bias, explainability, and quality checks use predictions on the test data to investigate the model behavior further. This information as well as model metrics from the model evaluation (Processing) are later displayed within the model registry. The model is only registered when a certain condition with regards to previous best models is met (Condition step).

Any artifacts and datasets created when a pipeline is run are saved in Amazon S3 using the buckets that are automatically created when provisioning the template.

Customizing the pipelines

We needed to customize the template to ensure the logical components of the existing codebase were migrated.

We used the LightGBM framework in building our models. The first of the models built in this use case is a simple decision tree to enforce certain feature splits during model training. Furthermore, we split the data based on a given feature. This procedure effectively resulted in training two separate decision tree models.

We can use trees to apply two kinds of predictions: value and leaf. In our case, we use value predictions on the test set to calculate custom metrics for model evaluation, and we use leaf prediction for inference.

Therefore, we needed to add specifications to the training and inference pipeline provided by the template. To account for the complexity provided by the use case model as well as to showcase the flexibility of the template, we added additional steps to the pipeline and customized the steps to apply the required code.

The following figure shows our updated template. It also demonstrates how you can have as many deployed models as you need, which you can pass to the inference pipeline.

Our first model enforces business knowledge and splits the data on a given feature. For this training pipeline, we had to register two different LightGBM models for each dataset.

The steps involved in training them were almost identical. Therefore, all steps were carried out twice for each data split compared to the standard template—except the first processing steps. A second processing step is applied to cater to different models with unique datasets.

In the following sections, we discuss the customized steps in more detail.

Processing

Each processing step is provided with a processor instance, which handles the SageMaker processing tasks. This use case uses two kinds of processors to run Python code:

• PySpark processor – Each processing job with a PySpark processor uses its own set of Spark configurations to optimize the job.
• Script processor – We use the existing Python code from the use case to generate custom metrics based on the predictions and actual data provided, and format the output to be viewed later in the model registry (JSON). We use a custom image created by the template.

In each case, we can modify the instance type, instance count, and size in GB of the Amazon Elastic Block Store (Amazon EBS) volume to adjust to the training pipeline needs. In addition, we apply custom container images created by the template, and we can extend the default libraries if needed.

Model training

SageMaker training steps support managed training and inference for a variety of ML frameworks. In our case, we use the custom Scikit-learn image provided using Amazon ECR in combination with the Scikit-learn estimator to train our two LightGBM models.

The previous design for the decision tree model involved a custom class, which was complicated and had to be simplified by fitting separate decision trees to filtered slices of the data. We accomplished this by implementing two training steps with different parameter and data inputs required for each model.

Condition

Conditional steps in SageMaker Pipelines support conditional branching in the running of steps. If all the conditions in the condition list evaluate to True, the “if” steps are marked as ready to run. Otherwise, the “else” steps are marked as ready to run. For our use case, we use two condition steps (ConditionLessThan) to determine if the current custom metrics for each model (for example the mean absolute error calculated on a test set) are below a performance threshold. This way, we only register models when the model quality is acceptable.

Create model

The create model step helps make models available for inference tasks and also inherits the code to be used to create predictions. Because our use case needs value and leaf predictions for two different models, the training pipelines have four different create model steps.

Each step is given an entry point file depending whether the consecutive transform step is to predict a leaf or value on given data.

Transform

Each transform step uses a model (created by the create model step) to return predictions. In our case, we have one for each created model, resulting in a total of four transform steps. Each of these steps returns either leaf or value predictions for each of the models. The output is customized depending on the next step in the pipeline. For example, a transform step for value predictions has different output filters to create a specific input required for the following model check steps in the pipeline.

Model registry

Finally, both model branches within the training pipeline have a unique model registry step. Like the standard setup, each model registry step takes the model-specific information from all the check steps (quality, explainability, and bias) as well as custom metrics and model artefacts. Each model is registered with a unique model package group.

While applying the use case-specific changes to the example code base, pipelines ran using the cache configurations for each pipeline step to enhance the debugging experience. Caching steps are useful when developing SageMaker pipelines because it reduces cost and saves time when testing pipelines.

We can optimize the settings of each step that requires an instance type and count (and if applicable EBS volume) for on-demand instances according to the use case needs.

Benefits

The AWS-NatWest collaboration has brought innovation to the implementation of ML models via SageMaker pipelines and MLOps best practices. NatWest Group now uses a standardized and secure solution to implement and productionize ML workloads on AWS, via flexible custom templates. Additionally, the model migration detailed in this post created multiple immediate and ongoing business benefits.

Firstly, we reduced complexity via standardization:

• We can integrate custom third-party models to SageMaker in a highly modular way to build a reusable solution
• We can create non-standard pipelines using templates, and these templates are supported in a standardized MLOps production environment

We saw the following benefits in software and infrastructure engineering:

• We can customize templates based on individual use cases
• Functionally decomposing and refactoring the original model code allows us to rearchitect, remove a significant technical debt caused by legacy code, and move our pipelines to a managed on-demand execution environment
• We can use the added capabilities available via the standardized templates to upgrade and standardize model explainability, as well as check data and model quality and bias
• The NatWest project team is empowered to provision environments, implement models, and productionize models in an almost completely self-service environment

Finally, we achieved improved productivity and lower cost of maintenance:

• Reduced time-to-live for ML projects – We can now start projects with generic templates that implement code standards, unit tests, and CI/CD pipelines for production-ready use case development from the beginning of a project. The new standardization means that we can expect reduced time to onboard future use cases.
• Reduced cost for running ML workflows in the cloud – Now we can run data processing and ML workflows with a managed architecture and adapt the underlying infrastructure to use case requirements using standardized templates.
• Increased collaboration between project teams – The standardized development of models means that we’ve increased the reusability of model functions developed by individual teams. This creates an opportunity to implement continuous improvement strategies in both development and operations.
• Boundaryless delivery – This solution allows us to make our elastic on-demand infrastructure available 24/7.
• Continuous integration, delivery, and testing – Reduced deployment times between development completion and production availability lead to the following benefits:
  • We can optimize the configuration of our infrastructure and software for model training and inference runtime.
  • The ongoing cost of maintenance is reduced due to standardization of processes, automation, and on-demand compute. Therefore, in production, we only incur runtime costs once a month during retraining and inference.

Conclusion

With SageMaker and the assistance of AWS innovation leadership, NatWest has created an ML productivity virtuous circle. The templates, modular and reusable code, and standardization freed up the project team from existing delivery constraints. Furthermore, the MLOps automation freed up support effort to enable the team to work on other use cases and projects, or to improve existing models and processes.

This is the fourth post of a four-part series on the strategic collaboration between NatWest Group and AWS Professional Services. Check out the rest of the series for the following topics:

• Part 1 explains how NatWest Group partnered with AWS Professional Services to build a scalable, secure, and sustainable MLOps platform
• Part 2 describes how NatWest Group used AWS Service Catalog and SageMaker to deploy their compliant and self-service MLOps platform
• Part 3 provides an overview of how NatWest Group uses SageMaker services to build auditable, reproducible, and explainable ML models

About the Authors

Pauline Ting is Data Scientist in the AWS Professional Services team. She supports customers in the financial and sports & media industries in achieving and accelerating their business outcome by developing AI/ML solutions. In her spare time, Pauline enjoys traveling, surfing, and trying new dessert places.

Maren Suilmann is a data scientist at AWS Professional Services. She works with customers across industries unveiling the power of AI/ML to achieve their business outcomes. In her spare time, she enjoys kickboxing, hiking to great views and board game nights.

Craig Sim is a Senior Data Scientist at NatWest Group with a passion for data science research, particularly within graph machine learning domain, and model development process optimisation using MLOps best practices. Craig additionally has extensive experience in software engineering and technical program management within financial services. Craig has an MSc in Data Science, PGDip in Software Engineering and a BEng(Hons) in Mechanical and Electrical Engineering. Outside of work Craig is a keen golfer, having played at junior and senior county level. Craig was fortunate to be able to incorporate his golf interest, with data science, by collaborating with the PGA European Tour and World Golf Rankings for his MSc Data Science thesis. Craig also enjoys tennis and skiing and is a married father of three, now adult, children.

Shoaib Khan is a Data Scientist at NatWest Group. He is passionate for solving business problems particularly in the domain of customer lifetime value and pricing using MLOps best practices. He is well equipped with managing a large machine learning codebase and is always excited to test new tools and packages. Loves to educate others around him as he runs a YouTube channel called convergeML. In his spare time, enjoys taking long walks and travel.

Read More

This is the third post of a four-part series detailing how NatWest Group, a major financial services institution, partnered with AWS Professional Services to build a new machine learning operations (MLOps) platform. This post is intended for data scientists, MLOps engineers, and data engineers who are interested in building ML pipeline templates with Amazon SageMaker. We explain how NatWest Group used SageMaker to create standardized end-to-end MLOps processes. This solution reduced the time-to-value for ML solutions from 12 months to less than 3 months, and reduced costs while maintaining NatWest Group’s high security and auditability requirements.

NatWest Group chose to collaborate with AWS Professional Services given their expertise in building secure and scalable technology platforms. The joint team worked to produce a sustainable long-term solution that supports NatWest’s purpose of helping families, people, and businesses thrive by offering the best financial products and services. We aim to do this while using AWS managed architectures to minimize compute and reduce carbon emissions.

SageMaker project templates

One of the exciting aspects of using ML to answer business challenges is that there are many ways to build a solution that include not only the ML code itself, but also connecting to data, preprocessing, postprocessing, performing data quality checks, and monitoring, to name a few. However, this flexibility can create its own challenges for enterprises aiming to scale the deployment of ML-based solutions. It can lead to a lack of consistency, standardization, and reusability that limits the time to create new business propositions. This is where the concept of a template for your ML code comes in. It allows you to define a standard way to develop an ML pipeline that you can reuse across multiple projects, teams, and use cases, while still allowing the flexibility needed for key components. This makes the solutions we build more consistent, robust, and easier to test. This also makes development go much faster.

Through a recent collaboration between NatWest Group and AWS, we developed a set of SageMaker project templates that embody all the standard processes and steps that form part of a robust ML pipeline. These are used within the data science platform built on infrastructure templates that were described in Part 2 of this series.

The following figure shows the architecture of our training pipeline template.

The following screenshot shows what is displayed in Amazon SageMaker Studio when this pipeline is successfully run.

All of this helps improve our ML development practices in multiple ways, which we discuss in the following sections.

Reusable

ML projects at NatWest often require cross-disciplinary teams of data scientists and engineers to collaborate on the same project. This means that shareability and reproducibility of code, models, and pipelines are two important features teams should be able to rely on.

To meet this requirement, the team set up Amazon SageMaker Pipelines templates for training and inference as dual starting points for new ML projects. The templates are deployed by the data science team via a single click into new development accounts via AWS Service Catalog as SageMaker projects. These AWS Service Catalog products are managed and developed by a central platform team so that best practice updates are propagated across all business units, but consuming teams can contribute their own developments to the common codebase.

To achieve easy replication of templates, NatWest utilized AWS Systems Manager Parameter Store. The parameters were referenced in the generic templates, where we followed a naming convention in which each project’s set of parameters uses the same prefix. This means templates can be started up quickly and still access all necessary platform resources, without users needing to spend time configuring them. This works because AWS Service Catalog replaces any parameters in the use case accounts with the correct value.

Not only do these templates enable a much quicker startup of new projects, but they also provide more standardized model development practices in teams aligned to different business areas. This means that model developers across the organization can more easily understand code written by other teams, facilitating collaboration, knowledge sharing, and easier auditing.

Efficient

The template approach encouraged model developers to write modular and decoupled pipelines, allowing them to take advantage of the flexibility of SageMaker managed instances. Decoupled steps in an ML pipeline mean that smaller instances can be used where possible to save costs, so larger instances are only used when needed (for example, for larger data preprocessing workloads).

SageMaker Pipelines also offers a step-caching capability. Caching means that steps that were already successfully run aren’t rerun in case of a failed pipeline—the pipeline skips those, reuses their cached outputs, and starts only from the failed step. This approach means that the organization is only charged for the higher-cost instances for a minimal amount of time, which also aligns ML workloads with NatWest’s climate goal to reduce carbon emissions.

Finally, we used SageMaker integration with Amazon EventBridge and Amazon Simple Notification Service (Amazon SNS) for custom email notifications on key pipeline updates to ensure that data scientists and engineers are kept up to date with the status of their development.

Auditable

The template ML pipelines contain steps that generate artifacts, including source code, serialized model objects, scalers used to transform data during preprocessing, and transformed data. We use the Amazon Simple Storage service (Amazon S3) versioning feature to securely keep multiple variants of objects like these in the same bucket. This means objects that are deleted or overwritten can be restored. All versions are available, so we can keep track of all the objects generated in each run of the pipeline. When an object is deleted, instead of removing it permanently, Amazon S3 inserts a delete marker, which becomes the current object version. If an object is overwritten, it results in a new object version in the bucket. If necessary, it can be restored to the previous version. The versioning feature provides a transparent view for any governance process, so the artifacts generated are always auditable.

To properly audit and trust our ML models, we need to keep a record of the experiments we ran during development. SageMaker allows for metrics to be automatically collected by a training job and displayed in SageMaker experiments and the model registry. This service is also integrated with the SageMaker Studio UI, providing a visual interface to browse active and past experiments, visually compare trials on key performance metrics, and identify the best performing models.

Secure

Due to the unique requirements of NatWest Group and the financial services industry, with an emphasis on security in a banking context, the ML templates we created required additional features compared to standard SageMaker pipeline templates. For example, container images used for processing and training jobs need to download approved packages from AWS CodeArtifact because they can’t do so from the internet directly. The templates need to work with a multi-account deployment model, which allows secure testing in a production-like environment.

SageMaker Model Monitor and SageMaker Clarify

A core part of NatWest’s vision is to help our customers thrive, which we can only do if we’re making decisions in a fair, equitable, and transparent manner. For ML models, this leads directly to the requirement for model explainability, bias reporting, and performance monitoring. This means that the business can track and understand how and why models make specific decisions.

Given the lack of standardization across teams, no consistent model monitoring capabilities were in place, leading to teams creating a lot of bespoke solutions. Two vital components of the new SageMaker project templates were bias detection on the training data and monitoring the trained model. Amazon SageMaker Model Monitor and Amazon SageMaker Clarify suit these purposes, and both operate in a scalable and repeatable manner.

Model Monitor continuously monitors the quality of SageMaker ML models in production against predefined baseline metrics. Clarify offers bias checks on training data based on the Deequ open-source library. After the model is trained, Clarify offers model bias and explainability checks using various metrics, such as Shapley values. The following screenshot shows an explainability report viewed in SageMaker Studio and produced by Clarify.

The following screenshot shows a similarly produced example model bias report.

Conclusion

NatWest’s ML templates align our MLOps solution to our need for auditable, reusable, and explainable MLOps assets across the organization. With this blueprint for instantiating ML pipelines in a standardized manner using SageMaker features, data science and engineering teams at NatWest feel empowered to operate and collaborate with best practices, with the control and visibility needed to manage their ML workloads effectively.

In addition to the quick-start ML templates and account infrastructure, several existing NatWest ML use cases were developed using these capabilities. The development of these use cases informed the requirements and direction of the overall collaboration. This meant that the project templates we set up are highly relevant for the organization’s business needs.

This is the third post of a four-part series on the strategic collaboration between NatWest Group and AWS Professional Services. Check out the rest of the series for the following topics:

• Part 1 explains how NatWest Group partnered with AWS Professional Services to build a scalable, secure, and sustainable MLOps platform
• Part 2 describes how NatWest Group used AWS Service Catalog and SageMaker to deploy their compliant and self-service MLOps platform
• Part 4 details how NatWest data science teams migrate their existing models to SageMaker architectures

About the Authors

Ariadna Blanca Romero is a Data Scientist/ML Engineer at NatWest with a background in computational material science. She is passionate about creating tools for automation that take advantage of the reproducibility of results from models and optimize the process to ensure prompt decision-making to deliver the best customer experience. Ariadna volunteers as a STEM virtual mentor for a non-profit organization helping the professional development of Latin-American students and young professionals, particularly for empowering women. She has a passion for practicing yoga, cooking, and playing her electric guitar.

Lucy Thomas is a data engineer at NatWest Group, working in the Data Science and Innovation team on ML solutions used by data science teams across NatWest. Her background is in physics, with a master’s degree from Lancaster University. In her spare time, Lucy enjoys bouldering, board games, and spending time with family and friends.

Sarah Bourial is a Data Scientist at AWS Professional Services. She is passionate about MLOps and supports global customers in the financial services industry in building optimized AI/ML solutions to accelerate their cloud journey. Prior to AWS, Sarah worked as a Data Science consultant for various customers, always looking to achieve impactful business outcomes. Outside work, you will find her wandering around art galleries in exotic destinations.

Pauline Ting is Data Scientist in the AWS Professional Services team. She supports financial, sports and media customers in achieving and accelerating their business outcome by developing AI/ML solutions. In her spare time, Pauline enjoys traveling, surfing, and trying new dessert places.

Read More