Image Classification Benchmark

Classifies 803,580 rows (80,358 unique images, each repeated 10x) using ResNet18 model. Downloads images, applies preprocessing transforms, and runs inference to predict ImageNet labels across distributed GPU nodes.

Input Dataset: ImageNet benchmark dataset (S3 parquet format) Output Format: Parquet with image URLs and predicted labels Cluster: 8 worker nodes using g6.xlarge instances Benchmark Date: September 22, 2024 Framework Versions: Daft 0.6.2, Ray Data 2.49.2, AWS EMR Spark 7.10.0

Performance Results

Engine	Runtime
Daft	4m 23s
Ray Data	23m 30s
Spark	45m 7s