This paper presents a rigorous, comparative deep learning benchmark of modern architectures for facial expression recognition (FER) in the wild. We systematically implement and evaluate a representative set of architectures – including residual CNNs, hybrid CNN-Transformer models, Vision Transformers, and contrastive-pretrained ResNet variants – on three challenging datasets: FER2013, RAF-Basic, and RAF-Compound. Our study investigates overfitting, the effectiveness of strong data augmentations, validation/test performance, and architectural trends. We provide open, reproducible code and practical insights for real-world FER deployment. Extensive methodological descriptions, mathematical motivations, and ablation studies are included to make the benchmark both transparent and reusable.