Awesome
:houses: Airbert: In-domain Pretraining for Vision-and-Language Navigation :houses:
This repository stores some models trained in our experiments.
:keyboard: Downloading from the command line
We stored our models on Google Drive, as the service provides a storage of 15 GB for free.
You can use the great gdown script for downloading the models:
pip install gdown
gdown [link to Google Drive]
We also provide a Makefile to help you:
# Download everything
make all
# Download a specific model
make airbert-r2rRSA
# Get all commands
make help
:houses: Model pretrained on the BNB dataset
| Model | Description |
|---|---|
| airbert | Airbert model pretrained on the BNB dataset |
:alien: External models used in our scripts
| Model | Description |
|---|---|
| vilbert | ViLBERT model pretrained on Conceptual Captions |
| vlnbert | VLN-BERT: vilbert fine-tuned on R2R |
:robot: Finetuned models in discriminative setting
| Model | Description |
|---|---|
| airbert-r2rRS | airbert fine-tuned on R2R with the shuffling loss |
| airbert-r2rRSA | airbert fine-tuned on R2R with the shuffling loss + speaker data |
:robot: Finetuned models in generative setting
| Model | Description |
|---|---|
| REVERIE | Recurrent VLN-BERT for remote referring expression with pretrained Airbert as backbone |
| R2R | Recurrent VLN-BERT for vision-and-language navigation with pretrained Airbert as backbone |