Task 2.2: ML Model and Foundation Model (FM) Development - Train, fine-tune, and customize models for ML and AI solutions

Skill 2.2.1: Apply Amazon SageMaker AI built-in algorithms and common ML libraries

This skill is about knowing when and how to apply Amazon SageMaker AI built-in algorithms and common ML libraries/frameworks to train models. It is not only about memorizing algorithm names. The exam expects you to match a business problem to the correct SageMaker AI training approach, understand data input requirements, and know when to bring your own training code using a managed framework container.

---

1. Key Idea

SageMaker AI gives you three main ways to train:

Training route	Who supplies the learning logic?	Who supplies the container?	Best when
Built-in algorithm	AWS	AWS	The problem matches a standard algorithm and you want minimal code
Common ML library / Script Mode	You supply a training script	AWS supplies prebuilt framework container	You need a custom model, loss function, or architecture
Bring your own container	You supply both	You supply and maintain	You have special dependencies or inference requirements not covered by managed containers

For this skill, focus on the first two routes.

---

2. Amazon SageMaker AI Built-in Algorithms

Built-in algorithms are AWS-managed implementations for common ML tasks. You provide data and hyperparameters, and SageMaker AI runs the training job.

2.1 Built-in Algorithm Quick Reference

Problem type	SageMaker AI built-in algorithm	Typical use cases and notes
Tabular classification/regression	Linear Learner	Linear models; good baseline; supports regression and binary/multiclass classification
Tabular classification/regression/ranking	XGBoost	Strong default for structured data; handles nonlinear relationships and missing values
Sparse classification/regression/recommendation	Factorization Machines	Click prediction, recommendations, sparse categorical features
Clustering	K-Means	Customer segmentation, grouping; unsupervised
Dimensionality reduction	Principal Component Analysis (PCA)	Reduce features, noise reduction
Anomaly detection on structured data	Random Cut Forest (RCF)	Unsupervised anomaly detection; unusual metrics/events
IP- or identity-based anomaly detection	IP Insights	Suspicious IP, account, or network behavior patterns
Time-series forecasting	DeepAR	Probabilistic forecasting across multiple related time series; demand, energy, revenue
Text classification and word embeddings	BlazingText	Word2Vec-style embeddings and fastText-style text classification
Sequence-to-sequence tasks	Seq2Seq	Machine translation, summarization, conversational modeling
Topic modeling	Latent Dirichlet Allocation (LDA)	Statistical document-topic discovery
Topic modeling	Neural Topic Model (NTM)	Neural document-topic discovery
Object/entity embeddings	Object2Vec	Recommendations, matching, fraud detection, semantic search
Image classification	Image Classification	Assigns an image-level class label
Object detection	Object Detection	Detects objects and bounding boxes within images
Pixel-level image labeling	Semantic Segmentation	Assigns a class to every pixel

2.2 Choosing a Built-in Algorithm by Problem Pattern

Use this practical mapping:

If you have...	First built-in algorithm to consider
Labeled tabular data with regression/classification	XGBoost
Very large linear tabular model or interpretable baseline	Linear Learner
Sparse, high-cardinality categorical data	Factorization Machines
Unlabeled tabular data requiring segmentation	K-Means
Unlabeled tabular/metric data requiring anomaly detection	Random Cut Forest (RCF)
Time-series data with seasonality and many related series	DeepAR
Text classification or word representation	BlazingText
Image class label prediction	Image Classification
Image object localization	Object Detection
Pixel-level vision task	Semantic Segmentation

---

3. Common ML Libraries in SageMaker AI

Common ML libraries allow you to use custom training code while still running inside AWS-managed framework containers. In SageMaker AI, this is often called Script Mode.

3.1 Common ML Library Reference

Library/framework	SageMaker AI usage	Typical scenarios
scikit-learn	SageMaker AI SKLearn container with script mode	Traditional tabular ML, feature engineering, proof of concept
XGBoost	Built-in algorithm or script mode	Structured data, ranking, custom objectives and metrics
TensorFlow/Keras	SageMaker AI TensorFlow container	Custom deep learning, computer vision, NLP, distributed training
PyTorch	SageMaker AI PyTorch container	Custom architectures, research workloads, transformer fine-tuning
Hugging Face Transformers	SageMaker AI Hugging Face container	Fine-tuning foundation models like BERT, GPT, T5, or embedding models
Apache MXNet	SageMaker AI MXNet container	Legacy deep learning workloads; historically used by SageMaker built-ins

3.2 Built-in Algorithm vs Common Library

Criterion	Built-in algorithm	Common library / Script Mode
Training code	Supplied by AWS	Supplied by you
Container	AWS-managed algorithm image	AWS-managed framework image
Customization	Hyperparameters only	Full model, data loading, loss, metrics
Setup effort	Very low	Medium
Use for custom architecture	No	Yes
Example	SageMaker AI XGBoost	PyTorch training script

---

4. Decision Flow for Selecting a Training Approach

flowchart TD
A[Identify problem and data] --> B{Data type?}
B -->|Structured tabular| C{Labeled?}
C -->|Yes supervised| D{High-cardinality sparse?}
D -->|Yes| E[Factorization Machines]
D -->|No| F[XGBoost or Linear Learner]
C -->|No clustering| G[K-Means]
C -->|No anomaly| H[Random Cut Forest or IP Insights]
C -->|No feature reduction| I[PCA]
B -->|Text| J{Modern transformer or custom NLP?}
J -->|Yes| K[Hugging Face or PyTorch Script Mode]
J -->|No| L[BlazingText]
B -->|Image| M{Need localization?}
M -->|Class only| N[Image Classification]
M -->|Bounding boxes| O[Object Detection]
M -->|Pixel masks| P[Semantic Segmentation]
B -->|Time series| Q[DeepAR]
B -->|Custom deep learning| R[PyTorch or TensorFlow Script Mode]

---

5. Important Input and Configuration Considerations

5.1 Data Input

• Training data is stored in Amazon S3.
• Built-in tabular algorithms commonly accept CSV.
• For SageMaker AI tabular supervised algorithms using CSV, the first column often contains the target label.
• Computer vision built-in algorithms typically use RecordIO or image files with a manifest file.
• Script mode gives you more flexibility: your script can read CSV, Parquet, JSON, images, or any supported format directly from S3.

5.2 Container and Script Selection

• Built-in algorithms require no training script.
• For common libraries, you must provide an entry point script containing your training logic.
• SageMaker AI supplies the framework container; you do not need to create a Docker image.

5.3 Hyperparameters

Built-in algorithms expose their own hyperparameters. Examples:

Algorithm	Common hyperparameters
XGBoost	num_round, eta, max_depth, min_child_weight, subsample, colsample_bytree, objective, eval_metric
Linear Learner	learning_rate, epochs, mini_batch_size, optimizer, L1, L2
DeepAR	prediction_length, context_length, epochs, learning_rate
Image Classification	epochs, learning_rate, batch_size, optimizer
K-Means	k, init_method
PCA	num_components, algorithm_mode

The exam generally tests whether you know which hyperparameters affect training behavior, not exact syntax.

---

6. Exam-Style Scenarios

Scenario 1: Customer Churn Prediction

A telecommunications company has a labeled tabular dataset containing customer demographics, billing history, and service usage. The data science team wants high predictive accuracy with minimal custom training code. The dataset has missing values and nonlinear feature relationships.

Which approach is most appropriate?

• A. SageMaker AI Linear Learner
• B. SageMaker AI Random Cut Forest
• C. SageMaker AI XGBoost built-in algorithm
• D. SageMaker AI Object Detection

Answer: C

Explanation: XGBoost is the strongest built-in choice for structured tabular classification with missing values and nonlinear relationships. Linear Learner is linear and often less accurate for complex tabular data. Random Cut Forest is unsupervised anomaly detection, and Object Detection is an image task.

---

Scenario 2: Unsupervised Login Anomaly Detection

A security team wants to detect unusual sign-in behavior based on IP address, account ID, and user-agent features. No fraud labels are available.

Which SageMaker AI built-in algorithm should the team choose?

• A. IP Insights
• B. XGBoost
• C. Linear Learner
• D. K-Means

Answer: A

Explanation: IP Insights is designed for unsupervised anomaly detection on IP-, account-, and entity-pattern data. XGBoost and Linear Learner are supervised and require labels. K-Means is for clustering, not anomaly scoring.

---

Scenario 3: Custom PyTorch Model Without Container Maintenance

A team has written a custom PyTorch model for product recommendation. The model uses a specialized neural network architecture and a custom loss function. The team does not want to build or maintain its own Docker container.

Which SageMaker AI option should the team use?

• A. SageMaker AI built-in Factorization Machines
• B. SageMaker AI PyTorch Script Mode
• C. SageMaker AI built-in PCA
• D. Amazon RDS with SQL analytics

Answer: B

Explanation: Script Mode allows the team to provide its custom PyTorch training script while using an AWS-managed PyTorch container. Built-in algorithms do not support custom architectures or custom loss functions.

---

Scenario 4: Multi-SKU Demand Forecasting

A retailer needs to forecast weekly demand for thousands of SKUs. The data contains multiple related time series with seasonal patterns and occasional promotions.

Which built-in algorithm is designed for this task?

• A. DeepAR
• B. Sequence-to-Sequence
• C. Linear Learner
• D. Image Classification

Answer: A

Explanation: DeepAR is SageMaker AI's built-in probabilistic forecasting algorithm for one-dimensional time series. It is designed for multiple related series and can incorporate covariates such as promotions.

---

Scenario 5: Image Class Labels vs Object Bounding Boxes

A warehouse uses computer vision to identify whether an image contains a pallet, box, or conveyor belt. The team initially only needs class labels for each image.

Later, the team also needs bounding boxes around each detected object.

Which built-in algorithm pair should be used?

• A. Image Classification first, then Object Detection
• B. Object Detection first, then Semantic Segmentation
• C. Semantic Segmentation first, then PCA
• D. Image Classification first, then BlazingText

Answer: A

Explanation: Image Classification assigns an image-level class label. Object Detection localizes objects by returning bounding boxes. Semantic Segmentation is for pixel-level masks, not bounding boxes.

---

7. Exam Tips and Traps

7.1 Strong Signal Patterns

If the question says...	Think...
Labeled tabular data, high accuracy, not custom code	XGBoost built-in
Sparse categorical data, click prediction, recommendations	Factorization Machines
Time-series forecasting	DeepAR
No labels, detect anomalies in metrics	Random Cut Forest
No labels, detect suspicious IP/account behavior	IP Insights
Image class only	Image Classification
Image with bounding boxes	Object Detection
Pixel-level masks	Semantic Segmentation
Word embeddings or basic text classification	BlazingText
Custom PyTorch/TensorFlow model	Script Mode with managed framework container
Transformer fine-tuning	Hugging Face or PyTorch Script Mode

7.2 Common Traps

• Do not choose a supervised algorithm for an unsupervised task. For anomaly detection without labels, XGBoost or Linear Learner is usually wrong.
• Do not confuse Image Classification with Object Detection. Image Classification gives class labels; Object Detection gives bounding boxes and class labels.
• Do not choose built-in algorithms for custom architectures. If the team has custom neural network code, use Script Mode with PyTorch, TensorFlow, or another framework.
• Do not assume BlazingText is the best option for transformer-based NLP. Modern foundation model tasks usually require Hugging Face or PyTorch Script Mode.
• Do not select K-Means for classification. K-Means is clustering and does not use labels for training.
• Do not treat SageMaker AI Autopilot as a built-in algorithm. Autopilot is an AutoML service that explores data and models automatically.
• Do not ignore data format. Built-in tabular algorithms commonly expect CSV; vision built-ins commonly expect RecordIO or manifest-based image data.
• Do not confuse DeepAR with general regression. DeepAR is specifically for time-series forecasting.

7.3 Cost and Operational Traps

• Built-in algorithms are often simpler and faster to deploy than custom script-mode jobs.
• Script Mode still requires a training script; cannot run without an entry point.
• Managed framework containers reduce infrastructure work but do not replace the need for correct training code.
• For standard tabular problems, rewriting as a custom PyTorch model may add complexity without enough benefit.

---

8. Quick Memory Map

flowchart LR
A[Standard problem] --> B[Built-in algorithm]
A2[Custom code] --> C[Script Mode with common ML library]
B --> D[XGBoost, Linear Learner, DeepAR, Image Classification]
C --> E[PyTorch, TensorFlow, scikit-learn, Hugging Face]

One-Line Exam Reminder

Standard problem + no custom code = built-in algorithm.

Custom model + team-owned script = SageMaker AI Script Mode with a common ML library.