Open yonatanbitton opened 3 years ago
Update: Downgrading torch to 1.4.0 seems to solve the problem in 2b).
However, i'm afraid that after pre-training I'll still have the same problem as 1), meaning that the model architecture won't be suitable and I'll receive the same problem as in 1):
ans_weight = answer_state_dict['logit_fc.3.weight']
KeyError: 'logit_fc.3.weight'
In any case, if you have a suitable model (that wasn't pre-trained on the GQA) that I can use it would be best. Many thanks.
Hello First - thanks for this amazing repo.
I need a pretrained model that I can later fine-tune it on the GQA dataset, but I need a version that wasn't pretrained before on the GQA specifically.
I've tried 2 different approaches, both failed. 1) Finding a pretrained version that I can use: I found this post: https://github.com/airsplay/lxmert/issues/7 You provide a model that was pretrained on VQA. I've tried to use this version, by using the
gqa_finetune.bash
code and just replace the model in--loadLXMERTQA snap/pretrained/model
to the provided model (--loadLXMERTQA snap/pretrained/pretrained_vqa_model
), but I receive the following error why the model creates the GQA class:2) Pretraining it myself. I took the
lxmert_pretrain.bash
and removed thevgnococo
from --train
, to make it pretrain on the rest that are not GQA. 2a) when I try to pretrain it only a single GPU with this CMD: bashrun/lxmert_pretrain_no_gqa.bash 2
I receive out of memory error. 2b) When I try to pretrain it on > 1 GPUs, for example 2 GPUs using this CMD:bash run/lxmert_pretrain_no_gqa.bash 4,3 --multiGPU
I recieve this wierd stacktrace:Do you have an idea how can I solve it? Or is there any change that you have a pretrained model that is not familiar with the GQA data that I can use?