@ -111,7 +111,7 @@ Save the edited file and `source` it
source set_env.sh
```
Run the below line to unset the environment variables
Run the below line to unset the environment variables (whendone with experiments)
```shell
sh unset_env.sh
```
@ -139,7 +139,7 @@ Pass the data split name to `-d` flag. Possible values are: `5-dpd`, `10-dpd`, `
Example training command: `sh train_baseline.sh -d 50-dpd`
### Belief State Prediction
Choose a checkpoint of the saved baseline model to generate belief state predictions.
Choose a checkpoint of the saved baseline model to generate belief states.
Set the `MODEL_CHECKPOINT` environment variable with the path to the chosen model checkpoint. It should only contain the path from the "experiment-{datetime}" folder.
```shell
@ -172,7 +172,7 @@ python evaluate.py
| 50-dpd | 28.64 |
| 100-dpd | 33.11 |
| 125-dpd | 35.79 |
| 250-dpd | 40.38 |
| 250-dpd | **40.38** |
## Prompt Learning Experiments
@ -181,8 +181,9 @@ The data for training the prompt learning model is available under [data/prompt-
`create_dataset.py` ([link](utils/create_dataset.py)) has the scripts for converting/creating the data for training the prompt-based model.
> **Note:**
> Running `create_dataset.py` can take some time as it needs to download, install and run Stanford CoreNLP `stanza` package.
> All the data required for training the prompt-based model is available under [data](data) directory of this repo.
> Running `create_dataset.py` can take some time as it needs to download, install and run Stanford CoreNLP `stanza` package. This scripts downloads coreNLP files of size `~1GB` and requires significant amount of RAM and processor capabilities to run efficiently.
>
> All the data required for training the prompt-based model is already available under the [data](data) directory of this repo.
### Install the requirements
After following the environment setup steps in the previous [section](#environment-setup), install the required python modules for prompt model training.
@ -206,17 +207,17 @@ Example training command: `sh train_baseline.sh -d 50-dpd`
**Some `train_prompting.sh` flags**:
`--num_epochs 10` - Number of epochs
`--num_epochs` - Number of epochs
`--learning_rate 5e-5` - Initial learning rate for Optimizer
`--learning_rate` - Initial learning rate for Optimizer
`--with_inverse_prompt` - Use Inverse Prompt while training
`--with_inverse_prompt` - Use Inverse Prompt while training**(recommended)**
`--inverse_prompt_weight 0.1` - Weight of the inverse prompt for loss function
`--inverse_prompt_weight` - Weight of the inverse prompt for loss function
**Note:** The defaults in `train_prompting.sh` are the best performing values.
### Belief State Generations (Prompt Generation)
### Belief State Generations (Prompt-based slot generation)
Now, the belief states can be generated by prompting. Choose a prompt fine-tuned model from the saved epochs and run the below script to generate belief states.
Generate belief states by running the below script:
@ -228,9 +229,9 @@ The argument `-m` takes the relative path of saved model from `SAVED_MODELS_PROM
The generated belief states (outputs) are saved under `OUTPUTS_DIR_PROMPT` folder. Some of the best outputs are uploaded to this repository and can be found under [outputs](outputs) folder.
The generated belief states (outputs) are saved under `OUTPUTS_DIR_PROMPT` folder. Some of the output files are uploaded to this repository and can be found under [outputs](outputs/prompt-learning) folder.
### Prompting Evaluation
### Evaluation ofprompt-based generations
The standard Joint Goal Accuracy (**JGA**) is used to evaluate the belief state predictions. In order to exclude the influence of wrongly extracted values, **JGA*** is computed only for values that are extracted correctly at each turn.
The [evaluate.py](prompt-learning/evaluate.py) file can be used to verify the below JGA scores.
