def cvae_loss_function(dec_en_batch_data: torch.tensor,
batch_data: torch.tensor, mu: torch.tensor,
log_var: torch.tensor) -> torch.tensor:
kld = -0.5 * torch.sum(1 + log_var - mu.pow(2) - log_var.exp())
bce = F.binary_cross_entropy(dec_en_batch_data,
batch_data,
reduction='sum')
return (bce + kld) / (batch_data.shape[0] * batch_data.shape[1])
def compute_loss(model: torch.nn.Module,
model_output: torch.tensor,
target_labels: torch.tensor,
is_normalize: bool = True) -> torch.tensor:
'''
Computes the loss between the model output and the target labels
Note: we have initialized the loss_criterion in the model with the sum
reduction.
Args:
- model: model (which inherits from nn.Module), and contains loss_criterion
- model_output: the raw scores output by the net [Dim: (N, 15)]
- target_labels: the ground truth class labels [Dim: (N, )]
- is_normalize: bool flag indicating that loss should be divided by the
batch size
Returns:
- the loss value
'''
loss = None
############################################################################
# Student code begin
############################################################################
loss_func = torch.nn.CrossEntropyLoss()
pred = model_output.exp() / (model_output.exp().sum(-1)).unsqueeze(-1)
loss = loss_func(model_output, target_labels)
if is_normalize == True:
loss /= model_output.shape[0]
############################################################################
# Student code end
############################################################################
return loss
