If you're referring to the oscillations in the training loss (and training PSNR), this is expected behavior for the model. The model is trained in a noisy-to-noisy fashion, meaning that clean images are not available and both inputs and targets are noisy. Therefore, the training loss is not expected to approach zero.

Let's denote:

$F(\cdot)$ as the network,
$x$ as a clean image,
$y = x + z$ as the noisy input (with noise $z$),
$\tilde{x} = x + w$ as the noisy label (with noise $w$).

The training goal is to learn a network $F(\cdot)$ such that, given a noisy input $y$, it outputs something close to the clean image $x$, i.e., $F(y) \approx x$.
However, since $x$ is unavailable during training, the network is trained to minimize the loss
$\tilde{l} = \left\lVert F(y) - \tilde{x}\right\rVert ^2_2$.

Even in an ideal (but unrealistic) scenario where $F(y) = x$, the loss would still be
$\tilde{l} = \left\lVert F(y) - \tilde{x}\right\rVert ^2_2 = \left\lVert x - \tilde{x}\right\rVert ^2_2 = \left\lVert w\right\rVert ^2_2 > 0$,
because the labels themselves are noisy.

As such, fluctuations in the training loss are expected, and the loss won't converge to zero. I recommend monitoring instead whether the expression
$\left\lVert F(y) - \pmb{x}\right\rVert ^2_2$
(on a validation set with known clean images, if available) decreases over the epochs. This would provide a more meaningful indicator of performance.

model seems not converge #1

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions