search

bcmi / CaGNet-Zero-Shot-Semantic-Segmentation

Code for our ACMMM2020 paper "Context-aware Feature Generation for Zero-shot Semantic Segmentation".
233 stars 14 forks source link

mean acc of unseen is NaN #10

Closed greathope closed 4 years ago

greathope commented 4 years ago

Hi, I checked the fasttext value and your pickled embedding value, it seems they does not match. I directly use word embeddings from fasttext, but the mean acc of unseen is NaN for my customed data, how to fix it? Thanks.

zhangxgu commented 4 years ago

Hi, @greathope .The word vectors of VOC and COCO-stuff is directly copied from SPNet, you can find them in their github. Also I have checked the word2vec of VOC, which is the same with the pretrained word files with normalization, except some differences between words 'airplane' and 'areoplane'. Besides, mean acc of unseen is not NaN when you are running the finetuning step, by loading the configs with ``finetune''.

greathope commented 4 years ago

@zhangxgu yes, I directly run your voc12 and context, the mean acc of unseen is ok, but when I run finetuning on my customed data, the mean acc of unseen is nan...

zhangxgu commented 4 years ago

@greathope You may need to check the path of val data.

qybing commented 4 years ago

I have the same problem.Have you solved it？