训练在16张NVIDIA A100 GPU上进行,每张GPU配备80GB内存。在推理阶段,所有答案生成都使用确定性贪婪解码,确保结果的可重现性和可比较性。对于伪问答生成,团队选择使用核采样技术来为每个视觉输入生成多样化的问答对。
Abstract: Causal inference is an important field in data science and cognitive artificial intelligence. It requires the construction of complex probabilistic models to describe the causal ...