46 lines
964 B
Markdown
46 lines
964 B
Markdown
# Watch And Help Dataset
|
|
|
|
Codes to reproduce results on WAH dataset[^1]
|
|
|
|
[^1]: Modified based on WAH train and test codes (https://github.com/xavierpuigf/watch_and_help)[https://github.com/xavierpuigf/watch_and_help].
|
|
|
|
## Data
|
|
|
|
Extact `dataset/watch_data.zip`
|
|
|
|
|
|
## Neural Network
|
|
|
|
Run `sh scripts/train_watch_strategy_full.sh` to train the model
|
|
|
|
To test model, either use trained model or extract checkpoints `checkpoints/train_strategy_full/lstmlast.zip`
|
|
|
|
Run `sh scripts/test_watch_strategy_full.sh` to test the model
|
|
|
|
|
|
|
|
## Prediction Split
|
|
|
|
Create artificial users and sample predictions from 10% to 90%
|
|
|
|
```
|
|
cd stan
|
|
sh split_user.sh
|
|
sh sampler_user.sh
|
|
```
|
|
|
|
|
|
## Bayesian Inference
|
|
|
|
|
|
Run inference to get results of user intention prediction and action length (0% to 100%) for all users
|
|
|
|
```
|
|
Rscript strategy_inference_test.R
|
|
```
|
|
|
|
Plot intention prediction results and 10% to 100% of actions results
|
|
|
|
```
|
|
sh plot_user_length.sh
|
|
sh plot_user_length_10_steps.sh
|