pe `(batch_size, output_dim`): The image embeddings obtained by
            applying the projection layer to the pooled output of vision model.
        Nr[