How much the loss has to be reduced before the model is feasible?
Opened this issue · 3 comments
leonzhu211 commented
As I trained it with my own dataset and got the following loss output
| item | value |
|---|---|
| grad_norm | 0.0203 |
| loss | 0.00342 |
| loss_q0 | 0.0074 |
| loss_q1 | 0.00366 |
| loss_q2 | 0.00215 |
| loss_q3 | 0.000243 |
| mse | 0.00342 |
| mse_q0 | 0.0074 |
| mse_q1 | 0.00366 |
| mse_q2 | 0.00215 |
| mse_q3 | 0.000243 |
| param_norm | 1.02e+03 |
| samples | 3.36e+06 |
| step | 4.2e+05 |
the test result is totally unacceptable.
How much the loss has to be reduced before the model is feasible?
qqpann commented
I'm curious too, @Hxyz-123 ?
The MSE seems pretty low compared to 0.27 in the paper? Correct me if I'm wrong!
MEIZU16 commented
Do you have any idea now ?
leonzhu211 commented
NO