Text emotion classification is one of the methods used to organize massively available texts into a predefined category to maximize the utilization of information in social media. From previous research, it is noted that a lot of work has been done for different languages; however, there is no work done for the Amharic language. This work aimed to develop the data and emotion classification models for Amharic social media text comments using deep learning approaches. We have annotated 14,716 emotion text comments into seven different emotion classes. LSTM, BiLSTM, CNN-BiLSTM, BiGRU, CNN, and BiLSTM using AmRoBERTa pre-trained embedding models are implemented. After experimentation and evaluation, the result showed that the CNN classifier outperforms better with 77.0% accuracy and F1-score. For the future, we recommend increasing the data and adapting different contextualized pre-trained embeddings for the effectiveness of the proposed deep learning approaches. The dataset is available publicly for further investigation.