Loss json in download.sh

[dljzx]

captions ./data/MSCOCO/MSCOCO/mscoco_caption_train2014_cn_processed.json file is not downloaded by the shell command file.Or it could be instead by captions_train2014_cn_translation.json ?

[dljzx]

and these's no ..._processed_dic.json in download.sh which make my training pro lots of TypeError.

[apple2373]

you can generate these files based on the script in the code

cd ./code/
python preprocess_MSCOCO_captions.py \
--input ../data/MSCOCO/captions_train2014_cn_translation.json \
--output ../data/MSCOCO/captions_train2014_cn_translation_processed.json \
--outdic ../data/MSCOCO/captions_train2014_cn_translation_processed_dic.json \
--outfreq ../data/MSCOCO/captions_train2014_cn_translation_processed_freq.json \
--cut 5 \
--char True \
cd ../

Then

python train_caption_model.py --savedir ./experiment1cn --epoch 50 --batch 120 --gpu 0 \
--vocab ./data/MSCOCO/captions_train2014_cn_translation_processed_dic.json \
--captions ./data/MSCOCO/captions_train2014_cn_translation_processed.json\

[apple2373]

Well, in general, this codebase is outdated and I don't recommend for training. I put a note on the readme.

[dljzx]

Oh thank you so much for your reply. I am just little confused that many json files you used were not downloaded by download.sh. For example no json files suffixed with dlc. Could you tell me where you get em?

[apple2373]

The jsons suffixed with _dic are at data/MSCOCO. I don't get them from somewhere else but I generate by myself. You can generate with code/preprocess_MSCOCO_captions.py .

[dljzx]

Oh, That's my fault. Thanku so much