DimplesL

结合代码 # Truncate sequences to max length as image embeddings can make the sequence longer tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None) if tokenizer_model_max_length is not None: new_input_embeds = [x[:tokenizer_model_max_length] for x...

看了一下训练配置，是在stage2修改model_max_length为4096来实现的吗？这样修改对于llama没有其他影响吗？

> Hi, the max token size of [Vicuna-1.5](https://huggingface.co/lmsys/vicuna-7b-v1.5/blob/main/config.json) is 4K. So, it still within the limit. I'm not very sure, but as the training set in [train_vicuna_7b.sh,](https://github.com/lm-sys/FastChat/blob/main/scripts/train_vicuna_7b.sh) max length is...

> 你好， @DimplesL image token个数为1024，而每个区域对应的mask token和position token分别都只有1个，详见 > > https://github.com/CircleRadon/Osprey/blob/ca9f26dbd9a0907d8ff686588a394fa897b60828/osprey/model/osprey_arch.py#L184-L187 > > > 正常训练和推理一般是不会超过2048的。 感谢指正，看了一下特征的变换，确实如此。 关于这部分还有个问题： 在mask extractor这个模块，是有线性层等参数的，独立于projector层参数，想确认下这部分的参数在训练过程保存是怎么设置的