Loss functions

I’m lost too.

huber_loss

 huber_loss (preds, target, mask, use_mask=False, padding_mask=None,
             delta=1)

preds: [bs x num_patch x n_vars x patch_len] targets: [bs x num_patch x n_vars x patch_len] mask: [bs x num_patch x n_vars] padding_mask: [bs x num_patch]

source

cosine_similarity_loss

 cosine_similarity_loss (preds, target, mask, use_mask=False,
                         padding_mask=None)

preds: [bs x num_patch x n_vars x patch_len] targets: [bs x num_patch x n_vars x patch_len] mask: [bs x num_patch x n_vars]

source

cosine_similarity

 cosine_similarity (preds, target, mask, use_mask=False,
                    padding_mask=None)

source

mape

 mape (preds, target, mask, use_mask=False)

source

mae

 mae (preds, target, mask, use_mask=False, padding_mask=None)

source

rmse

 rmse (preds, target, mask, use_mask=False, padding_mask=None)

source

mse

 mse (preds, target, mask, use_mask=False, padding_mask=None)

source

r2_score

 r2_score (preds, target, mask, use_mask=False)

source

masked_mae_loss

 masked_mae_loss (preds, target, mask, use_mask=False, padding_mask=None)

preds: [bs x num_patch x n_vars x patch_len] targets: [bs x num_patch x n_vars x patch_len] mask: [bs x num_patch x n_vars] padding_mask: [bs x num_patch]

source

masked_mse_loss

 masked_mse_loss (preds, target, mask, use_mask=False, padding_mask=None)

preds: [bs x num_patch x n_vars x patch_len] targets: [bs x num_patch x n_vars x patch_len] mask: [bs x num_patch x n_vars] padding_mask: [bs x num_patch]

source

patch_continuity_loss

 patch_continuity_loss (preds)

preds: [bs x num_patch x n_vars x patch_len] targets: [bs x num_patch x n_vars x patch_len]

x = torch.randn(2,2, 2, 10)

patch_continuity_loss(x)

tensor(0.2936)

source

FocalLoss

 FocalLoss (weight=None, gamma=2.0, reduction='mean', ignore_index=-100)

adapted from tsai, weighted multiclass focal loss https://github.com/timeseriesAI/tsai/blob/bdff96cc8c4c8ea55bc20d7cffd6a72e402f4cb2/tsai/losses.py#L116C1-L140C20

criterion = FocalLoss(gamma=0.7, weight=None, ignore_index=0)
batch_size = 10

n_patch = 721
n_class = 5
#m = torch.nn.Softmax(dim=-1)
logits = torch.randn(batch_size, n_class, n_patch)
target = torch.randint(0, n_class, size=(batch_size, n_patch))
criterion(logits, target)

tensor(8.4217)

source

KLDivLoss

 KLDivLoss (weight=None, ignore_index=-100)

*Kullback-Leibler Divergence Loss with masking for ignore_index. Handles soft labels with ignore_index marked as -100.

Args: logits: [bs x n_classes x pred_labels] - model predictions targets: [bs x n_classes x soft_labels] - soft labels, with ignore_index positions marked as -100 or [bs x n_labels] - hard labels*

x = torch.randn(4,5,10)
y = torch.randint(0,5, size=(4,10))
y_og = y.clone()
y[0,0] = -100

KLDivLoss(ignore_index=-100)(x,y)

tensor(0.4065)