Hi, I experienced a problem in the behavior of tensorboard, when recording a scalar value with
In my experiment I use a set of loss criteria, whose values I record at each epoch, with
add_scalar('Lossname', loss.item(), epoch).
Two values are recorded correctly, so in the tensorboard viewer the scalar graphs are displayed correctly.
One value produces a graph with all infinite values, but if I print
loss3.item() at runtime, it is a simple float variable with values neither too large nor too small (between 0 and 10).
Why is it not logged correctly?
Also, since the total training loss is a weighted sum of the previous functions, I know that loss3 never goes to infinity because the training is successful.
The three loss functions are as follows:
- loss1: MSELoss() from PyTorch
- loss2: LogSTFTMagnitudeLoss() from ParallelWaveGAN
- loss3: SpectralConvergenceLoss() from ParallelWaveGAN
Why are the values returned by
loss3.item() not logged correctly?
I suspect that it is because loss1() and loss2() use pytorch’s default functions: loss1() is the simple MSELoss while loss2() contains L1Loss. In contrast, loss3() is completely hand-defined and perhaps its class is missing some method…