EfficientNetV2 is the second generation of EfficientNet that optimizes for training speed in addition to inference efficiency — using a combination of Fused-MBConv blocks, progressive learning (increasing image size during training), and NAS optimized for training time.

What Is EfficientNetV2?

• Fused-MBConv: Replaces depthwise separable conv with regular conv in early stages (faster on modern hardware due to better utilization).
• Progressive Learning: Start training with small images and weak augmentation, gradually increase both.
• NAS Objective: Optimized for training speed (not just parameter count or FLOPs).
• Paper: Tan & Le (2021).

Why It Matters

• 5-11× Faster Training: EfficientNetV2-M trains 5× faster than EfficientNet-B7 with similar accuracy.
• Progressive Learning: Simple but effective — smaller images early = faster initial epochs.
• Hardware Aware: Recognizes that depthwise conv is slow on GPUs due to poor hardware utilization.

EfficientNetV2 is EfficientNet optimized for real-world speed — understanding that FLOPs don't equal training time and optimizing what actually matters.