F1 score in PyTorch

f1_score.py

def f1_loss(y_true:torch.Tensor, y_pred:torch.Tensor, is_training=False) -> torch.Tensor:
    '''Calculate F1 score. Can work with gpu tensors
    
    The original implmentation is written by Michal Haltuf on Kaggle.
    
    Returns
    -------
    torch.Tensor
        `ndim` == 1. 0 <= val <= 1
    
    Reference
    ---------
    - https://www.kaggle.com/rejpalcz/best-loss-function-for-f1-score-metric
    - https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score
    - https://discuss.pytorch.org/t/calculating-precision-recall-and-f1-score-in-case-of-multi-label-classification/28265/6
    
    '''
    assert y_true.ndim == 1
    assert y_pred.ndim == 1 or y_pred.ndim == 2
    
    if y_pred.ndim == 2:
        y_pred = y_pred.argmax(dim=1)
        
    
    tp = (y_true * y_pred).sum().to(torch.float32)
    tn = ((1 - y_true) * (1 - y_pred)).sum().to(torch.float32)
    fp = ((1 - y_true) * y_pred).sum().to(torch.float32)
    fn = (y_true * (1 - y_pred)).sum().to(torch.float32)
    
    epsilon = 1e-7
    
    precision = tp / (tp + fp + epsilon)
    recall = tp / (tp + fn + epsilon)
    
    f1 = 2* (precision*recall) / (precision + recall + epsilon)
    f1.requires_grad = is_training
    return f1

It took some time to integrate this f1_loss into my solution but here is some notes:

1. results must be flat
2. predictions must be within 0..1 range (not logits)
3. you want to return negative f1 result to optimize NN

Here's what I came up with:

def f1_loss(y_true:torch.Tensor, y_pred:torch.Tensor, is_training=True) -> torch.Tensor:
    tp = (y_true * y_pred).sum().to(torch.float32)
    tn = ((1 - y_true) * (1 - y_pred)).sum().to(torch.float32)
    fp = ((1 - y_true) * y_pred).sum().to(torch.float32)
    fn = (y_true * (1 - y_pred)).sum().to(torch.float32)
    
    epsilon = 1e-7
    
    precision = tp / (tp + fp + epsilon)
    recall = tp / (tp + fn + epsilon)
    
    f1 = 2* (precision*recall) / (precision + recall + epsilon)
    return -f1

Usage:
f1_loss(target.to(torch.float32).flatten(), F.sigmoid(preds.flatten()))

Hope it will save you some time and nerves 😅