PREPARE_PRETRAINED_MODELS.md

Prepare Pre-trained Models

Download pre-trained models and organize them as following:

code_root/
└── model/
    └── pretrained_model/
        ├── vl-bert-base-e2e.model
        ├── vl-bert-large-e2e.model
        ├── vl-bert-base-prec.model
        ├── vl-bert-large-prec.model
        ├── bert-base-uncased/
        │   ├── vocab.txt
        │   ├── bert_config.json
        │   └── pytorch_model.bin
        ├── bert-large-uncased/
        │   ├── vocab.txt
        │   ├── bert_config.json
        │   └── pytorch_model.bin
        └── resnet101-pt-vgbua-0000.model     

VL-BERT

Model Name	Download Link
vl-bert-base-e2e	GoogleDrive / BaiduPan
vl-bert-large-e2e	GoogleDrive / BaiduPan
vl-bert-base-prec	GoogleDrive / BaiduPan
vl-bert-large-prec	GoogleDrive / BaiduPan

Note: models with suffix "e2e" means parameters of Fast-RCNN is tuned during pre-training, while "prec" means Fast-RCNN is fixed during pre-training and for effeciency the visual features is precomputed using bottom-up-attention.

BERT & ResNet

Download following pre-trained BERT and ResNet and place them under this folder.

• BERT: GoogleDrive / BaiduPan
• ResNet101 pretrained on Visual Genome: GoogleDrive / BaiduPan (converted from caffe model)