search

InternLM / InternLM-XComposer

InternLM-XComposer2 is a groundbreaking vision-language large model (VLLM) excelling in free-form text-image composition and comprehension.
1.91k stars 120 forks source link

vit.vision_tower.vision_model.post_layernorm.weight & bias NOT initialized? #296

Closed JasonLeeFdu closed 2 months ago

JasonLeeFdu commented 2 months ago

image

JasonLeeFdu commented 2 months ago

It says that there are 2 params which are not found in the pretrained weights. It is very unsettling since I copyed the code from huggingface-1.8B model card- quick start. image

JasonLeeFdu commented 2 months ago

Still image

yuhangzang commented 2 months ago

The post_layernorm is applied after the output of the ViT last layer, which is not used in our model and we delete them during training. So you can ignore this log, it does not affect the performance.