The task of automatic misogyny identifica-tion and categorization has not received asmuch attention as other natural language taskshave, even though it is crucial for identify-ing hate speech in social Internet interactions.In this work, we address this sentence clas-sification task from a representation learningperspective, using both a bidirectional LSTMand BERT optimized with the following met-ric learning loss functions: contrastive loss,triplet loss, center loss, congenerous cosineloss and additive angular margin loss. We setnew state-of-the-art for the task with our fine-tuned BERT, whose sentence embeddings canbe compared with a simple cosine distance,and we release all our code as open source foreasy reproducibility. Moreover, we find that al-most every loss function performs equally wellin this setting, matching the regular cross en-tropy loss.