Artificial-Judge

Artifical Judge attempts to use the BERT NLP pre-trained model to predict the winners of supreme court cases. This notebook uses this kaggle dataset which can be found in the justice.csv file in this repo.

Setup

To begin, install conda or miniconda for your machine: https://docs.conda.io/projects/conda/en/latest/user-guide/install/index.html
Install jupyter lab as well:

conda install jupyterlab -c conda-forge

Next, create a conda env to run this notebook in:

conda create --name myenv python=3.7

run conda activate myenv to enter in to the env. Note: run conda deactivate to go back to base

Prepare a kernel for jupyter lab in your environment:

• Activate the environment: source activate myEnv
• Install ipykernel: conda install ipykernel
• Run this command: python -m ipykernel install --user --name myEnv --display-name “my_environment”

Check import in the notebook for all necessary packages. You can check to see what packages are installed in your current conda env with conda list. If you are missing a package, simply run conda install packagename

Fine Tuning

Vailla Pytorch

For Vanilla Pytorch see section 3.1.1 and see the following code to tune some hyerparameters

train_dataloader = DataLoader(
            train_dataset,
            sampler = RandomSampler(train_dataset),
            batch_size = 16
        )
        
...

model = BertForSequenceClassification.from_pretrained(
    'bert-large-uncased',
    num_labels = 2,
    output_attentions = False,
    output_hidden_states = False,
)

optimizer = torch.optim.AdamW(model.parameters(), 
                              lr = 5e-5,
                              eps = 1e-08
                              )

or edit num_epochs before the training loop in 3.1.2

Transformers Trainer

See section 3.2.3 "Fine Tuning":

#Load pretrained model
model = BertForSequenceClassification.from_pretrained('bert-base-uncased', 
                                                      num_labels=2, 
                                                      #hidden_dropout_prob=0.4,
                                                      #attention_probs_dropout_prob=0.4
                                                      )

training_args = TrainingArguments(
    output_dir="test_trainer", 
    logging_dir='logs', 
    evaluation_strategy="epoch",
    per_device_train_batch_size=16,  
    per_device_eval_batch_size=16,
    num_train_epochs=4,
    gradient_accumulation_steps=1,
    learning_rate=5e-5,
    logging_steps=50,
)
metric = evaluate.load("accuracy")
def compute_metrics(eval_pred):
    logits, labels = eval_pred
    predictions = np.argmax(logits, axis=1)
    return metric.compute(predictions=predictions, references=labels)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    compute_metrics=compute_metrics,
)