search

cshizhe / asg2cap

Code accompanying the paper "Say As You Wish: Fine-grained Control of Image Caption Generation with Abstract Scene Graphs" (Chen et al., CVPR 2020, Oral).
MIT License
200 stars 29 forks source link

Can you tell me the directory format of the dataset #19

Open jiar1019 opened 3 years ago

jiar1019 commented 3 years ago

Can you tell me the directory format of the dataset

sailist commented 3 years ago

if you want to train MSCOCO, the directory construct will be formatted as follow:

.
├── annotation
│   ├── controllable
│   │   ├── cleaned_tst_region_descriptions.json
│   │   ├── int2word.npy
│   │   ├── public_split # 图片文件名，切分方式 
│   │   │   ├── trn_names.npy 
│   │   │   ├── tst_names.npy
│   │   │   └── val_names.npy
│   │   ├── regionfiles #  object ID  对应的图片，冗余信息，在 json 文件中完全包括了
│   │   │   ├── trn_names.npy
│   │   │   ├── tst_names.npy
│   │   │   └── val_names.npy
│   │   │   ├── train2014 # 每个json 是一个 dict，level1 每个key 是 object ID，对应的 value 是相应
│   │   │   │   ├── COCO_train2014_000000000009.json
│   │   │   │   ├── COCO_train2014_000000000025.json
│   │   │   │   ├── COCO_train2014_000000000030.json
│   │   │   ├── val2014
│   │   │   │   ├── COCO_val2014_000000000042.json
│   │   │   │   ├── COCO_val2014_000000000073.json
│   │   │   │   ├── COCO_val2014_000000000074.json
│   │   │   │   ├── COCO_val2014_000000000133.json
│   │   │   │   ├── COCO_val2014_000000000136.json
│   │   │   │   ├── COCO_val2014_000000040011.json
│   │   └── word2int.json
├── dir.txt
├── ordered_feature
│   ├── MP
│   │   └── resnet101.ctrl
│   │       ├── trn_ft.npy
│   │       ├── tst_ft.npy
│   │       └── val_ft.npy
│   └── SA
│       └── X_101_32x8d
│           └── objrels
│               ├── train2014_COCO_train2014_000000000009.jpg.hdf5
│               ├── train2014_COCO_train2014_000000581909.jpg.hdf5
│               ├── train2014_COCO_train2014_000000581921.jpg.hdf5
│               ├── val2014_COCO_val2014_000000000074.jpg.hdf5
│               ├── val2014_COCO_val2014_000000000139.jpg.hdf5
│               └── val2014_COCO_val2014_000000581929.jpg.hdf5
└── results
zxh19976666 commented 3 years ago

if you want to train MSCOCO, the directory construct will be formatted as follow:

.
├── annotation
│   ├── controllable
│   │   ├── cleaned_tst_region_descriptions.json
│   │   ├── int2word.npy
│   │   ├── public_split # 图片文件名，切分方式 
│   │   │   ├── trn_names.npy 
│   │   │   ├── tst_names.npy
│   │   │   └── val_names.npy
│   │   ├── regionfiles #  object ID  对应的图片，冗余信息，在 json 文件中完全包括了
│   │   │   ├── trn_names.npy
│   │   │   ├── tst_names.npy
│   │   │   └── val_names.npy
│   │   │   ├── train2014 # 每个json 是一个 dict，level1 每个key 是 object ID，对应的 value 是相应
│   │   │   │   ├── COCO_train2014_000000000009.json
│   │   │   │   ├── COCO_train2014_000000000025.json
│   │   │   │   ├── COCO_train2014_000000000030.json
│   │   │   ├── val2014
│   │   │   │   ├── COCO_val2014_000000000042.json
│   │   │   │   ├── COCO_val2014_000000000073.json
│   │   │   │   ├── COCO_val2014_000000000074.json
│   │   │   │   ├── COCO_val2014_000000000133.json
│   │   │   │   ├── COCO_val2014_000000000136.json
│   │   │   │   ├── COCO_val2014_000000040011.json
│   │   └── word2int.json
├── dir.txt
├── ordered_feature
│   ├── MP
│   │   └── resnet101.ctrl
│   │       ├── trn_ft.npy
│   │       ├── tst_ft.npy
│   │       └── val_ft.npy
│   └── SA
│       └── X_101_32x8d
│           └── objrels
│               ├── train2014_COCO_train2014_000000000009.jpg.hdf5
│               ├── train2014_COCO_train2014_000000581909.jpg.hdf5
│               ├── train2014_COCO_train2014_000000581921.jpg.hdf5
│               ├── val2014_COCO_val2014_000000000074.jpg.hdf5
│               ├── val2014_COCO_val2014_000000000139.jpg.hdf5
│               └── val2014_COCO_val2014_000000581929.jpg.hdf5
└── results

thank you for your reply but where are dir.txt and results?

sailist commented 3 years ago

dir.txt is not needed, it contains the output of the tree command, which is my reply content.

zxh19976666 commented 3 years ago

dir.txt is not needed, it contains the output of the tree command, which is my reply content.

ok thank you sir but when I run ‘python configs/prepare_coco_imgsg_config.py $mtype’ it show no module nemed caption how to fix it？

sailist commented 3 years ago

dir.txt is not needed, it contains the output of the tree command, which is my reply content.

ok thank you sir but when I run ‘python configs/prepare_coco_imgsg_config.py $mtype’ it show no module nemed caption how to fix it？

I'd suggest you reimplement this project, copy the core code, pad train logic and ignore this unimportant preprocess code. It won't take you much time. Here is my repo asg2cap implemented under my train framework