search

XL2248 / SOV-MAS

The code and data for "Summary-Oriented Vision Modeling for Multimodal Abstractive Summarization"
7 stars 0 forks source link

code #2

Open mylovefanghui opened 1 year ago

mylovefanghui commented 1 year ago

Hello, could you please provide the directory setting for the following code to run? I don't know how to store the files inside, thank you very much

XL2248 commented 1 year ago

Hi, thanks for your attention.

I have reorganized the data dir using the high-resource, low-resource, few-shot, and zero-shot settings (have updated and you can also reorganize them according to the setting of our paper). The scripts are also changed to be consistent with the reorganized dir.

Now, you can use bash multimodal_dist_mmt5_32_ft.sh 4 11 high-resource 1.0 8 256 to train a multilingual model for reproducing the results of Table 1 of SOV-MAS. And you can change 'high-resource' to 'low-resource' (bash multimodal_dist_mmt5_32_ft.sh 4 11 low-resource 1.0 8 256) to get the results of multilingual training of Table 2 of SOV-MAS. It is similar to the few-shot setting (bash multimodal_dist_mmt5_32_ft.sh 4 11 few-shot 1.0 8 256).

For single-language training, you have to specify the datatype and the language, like this, bash single_lang_multimodal_train32.sh high-resource english; bash single_lang_multimodal_train32.sh low-resource bengali.

Hope it can clear this issue.