Transformers timenorm
The first time you run the code, the pre-trained model, clulab/roberta-timex-semeval, will be automatically downloaded in your computer. If you want to produce some predictions with this model, you need to pass as arguments the directory containing the input text and the target directory where the predictions will be stored. For example, to process the AQUAINT subset from TempEval-2013, just run:
python run_time.py -p /paht/to/anafora-annotation/TempEval-2013/AQUAINT -o /path/to/output/AQUAINT [--no_cuda]
Recall that the anafora-annotation folder includes both raw text and Anafora annotation files, but it could contain only the former since the latter are not needed to make predictions. This will be the case during the evaluation phase.
You can continue training the pre-trained model from its current checkpoint if you dispose of additional annotated data by running the following command:
python run_time.py -t /path/to/train-data/ -s /path/to/save-model/ [--no_cuda]
The train-data directory must follow a similar structure to the AQUAINT or TimeBank folders and include, for each document, both the raw text file (with no extension) and the Anafora annotation file (with .xml extension). After running the training, the save-model will contain two sub-folders, logs, with a set of log files that can be visualized with TensorBoard, and results, that contains all the checkpoints saved during the training and three files (pytorch_model.bin, training_args.bin and config.json) with the configuration and weights of the final model.
To use this new version of the model for predictions, you can run:
python run_time.py -p /paht/to/text/TempEval-2013/AQUAINT -o /path/to/output/AQUAINT -m /path/to/save-model/results/ [--no_cuda]
If you still want to continue the training from this point, just run:
python run_time.py -t /path/to/train-data/ -s /path/to/save-model-2/ -m /path/to/save-model/results/ [--no_cuda]
Use the --no_cuda option if you are going to run the commands above in the gpu. The /path/to/save-model/results/ value for the -m option can be also replaced by a model stored in the HuggingFace hub. E.g. -m clulab/fake-timex-model.
Run python run_time.py -h to explore additional options and arguments you can play with, like the hyperpameters of the model.