Facial emotion recognition (FER) plays a vital role in human-computer interaction, enabling applications starting from sentiment analysis to affective computing. In this study, we propose an innovative Multi-stage Hierarchical Attention Network model for Facial Emotion Recognition in Wild (MHAN-FERW). This model is designed to address the challenges encountered in uncontrolled environmental conditions when it comes to recognizing facial expressions. The MHAN-FERW architecture utilizes an EfficientNetB6 backbone network to extract features, followed by the application of an attention mechanism to capture spatial, channel-wise attention and temporal information, added with a feature pyramid network and fully connected layers to enhance robust feature representation. Through experimentation on the AffectNet dataset, MHAN-FERW achieves a notable accuracy of 67%, surpassing the state-of-the-art benchmarks by 3%. The proposed MHAN-FERW architecture demonstrates the efficacy of integrating attention mechanisms and feature pyramid networks for facial emotion recognition in challenging real-world scenarios. The insights gained from this study contributed in the advancement of FER which has potential applications in various fields of computing.