Pixels Predictions

Why SageMaker?

Before diving into actual project, it’s worth spending why I choose AWS SageMaker is the backbone of this project - Obviously Sagemaker been whole product compose other AWS service Glue together & Market Demand 😂

SageMaker is AWS’s fully managed machine learning platform — it handles the undifferentiated heavy lifting of ML infra

FYI, plain boto3+EC2 can do that too.The value is in what SageMaker automates around training: distributed data loading, experiment tracking, hyperparameter tuning jobs, and one-command deployment to a real-time inference endpoint.

Problem Statement.

I started as a weekend project, with 10-5 class subset, on the company I worked with given an opportunity to classify 50 different types of furniture on the high level, and it passes to specific model ( I always relies on ensambling btw ) to achieve good results.

A 10–15 class subset (like just the starters) trains fast and hits high accuracy easily — it doesn’t stress-test the pipeline or justify a managed platform like SageMaker. Going with the full 150-class Gen 1 Pokédex introduces real inter-class similarity, class imbalance, and a large enough dataset that distributed training, hyperparameter tuning, and hosted endpoints actually earn their place.

Criteria	SageMaker Studio (UI)	boto3 (General-Purpose SDK)	SageMaker Python SDK (sagemaker lib)
What it is	Web-based IDE for notebooks, pipelines, experiments	Low-level AWS SDK — same one used for S3, EC2, Lambda, etc.	High-level, ML-specific SDK built on top of boto3
Best for	Exploration, visual debugging, one-off experiments, demoing to non-engineers	Fine-grained control, scripting infrastructure alongside other AWS services	Actual training/deployment workflows — the “SageMaker way”
Learning curve	Low — click-and-run	Medium–high — raw API calls such as create_training_job() with verbose JSON configs	Low–medium — purpose-built abstractions such as Estimator, .fit(), .deploy()
Reproducibility	Poor — clicks aren’t code and are hard to version	Good — scriptable, but verbose and brittle	Good — scriptable and concise; designed for source control
Automation-friendly	No — manual by nature	Yes, but you rebuild ML-specific conveniences yourself	Yes — designed for CI/CD and SageMaker Pipelines
Abstraction level	Highest — little/no code required	Lowest — raw AWS API surface	Middle — hides boilerplate while keeping control of ML logic