Open rajicon opened 3 years ago
Nevermind I found it. However, it is unclear that we need to seperately find it, perhaps this should be clarified in the README ?
What is the correct way to get the flicker related files? I don't know where to get flickr30k_test_resnet50_cnn_features.hdf5 , and so I figure I would ask on how to "correctly" get the files I need (as download_data.sh doesn't get these files). Is there a step I'm misunderstanding?
Hi @rajicon from here https://github.com/iacercalixto/MultimodalNMT
(1) Download the flickr30k From this paper: image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions, Transactions of the Association for Computational Linguistics
(2) Download the split from here (task 1) https://www.statmt.org/wmt16/multimodal-task.html
(3) run python extract_image_features.py --gpuid 0 --pretrained_cnn resnet50 --splits=train,valid,test --images_path /MultimodalNMT/flickr30k/images/ --train_fnames /MultimodalNMT/flickr30k/train_images.txt --valid_fnames /MultimodalNMT/flickr30k/val_images.txt --test_fnames /MultimodalNMT/flickr30k/test_images.txt
How do we get files like test_2016_flickr.lc.norm.tok.de ? I don't think the download_data script gets this. I also looked at the multi30k dataset (https://github.com/multi30k/dataset) , but can't seem to find the above file (the closest is test_2017_flickr.de).