Closed chadcwilliams closed 6 months ago
That could be even helpful in some cases One case I can think of right now is AE-Training. Loss starts high and decreases quite fast. Then decrease-rate shrinks tremendously (in order of magnitudes) but it's still decreasing. That's due to the different stages of the training. At first it learns constant values (biases) to bring the curve into the correct value range and bring in the most general features. Then it learns to adapt that general curve to the given needs. So having the loss plot of the first training shows the first learning stage which has the biggest jump in loss. The latter learning stages would seem like constant lines even though there's still quite a lot happening there. We should discuss some solutions
When using the visualizer, it only plots the loss of the last training. We'll need to check and make sure it saves all losses across training and then adapt to ensure it plots all of it together.