T2I_CL icon indicating copy to clipboard operation
T2I_CL copied to clipboard

Best DAMSM epoch ?

Open eoaiewo opened this issue 1 year ago • 9 comments

Can anyone help me, I train the DAMSM for 600 epochs for bird dataset, how to know which epoch is the best for DAMSM to choose in order to train the GAN model, does anyone know ?

eoaiewo avatar Nov 06 '23 21:11 eoaiewo

does anyone know ?

eoaiewo avatar Nov 07 '23 08:11 eoaiewo

It should be the epoch 200. You may also try other checkpoints.

huiyegit avatar Nov 07 '23 14:11 huiyegit

What do you mean try other checkpoints ?

What about valid loss, s_loss, w_loss outputs when I train the DAMSM, because after epoch 200, these stats seems decreasing should I choose based on these stats or something else ? Do you know ?

ghost avatar Nov 07 '23 17:11 ghost

The metric for DAMSM may be the IS score, FID for the image quality.

huiyegit avatar Nov 07 '23 18:11 huiyegit

What do you mean try other checkpoints ?

What about valid loss, s_loss, w_loss outputs when I train the DAMSM, because after epoch 200, these stats seems decreasing should I choose based on these stats or something else ? Do you know ?

Checkpoints refer to the models saved at some epochs.

huiyegit avatar Nov 07 '23 18:11 huiyegit

The process is to train the DAMSM first, then pick the best epoch to train the GAN model, the IS score or FID only used for GAN, for the DAMSM there are outputs like those I mention, I don't know how to choose

eoaiewo avatar Nov 07 '23 19:11 eoaiewo

You are right. The feedback of DAMSM model is after the GAN training.

huiyegit avatar Nov 07 '23 19:11 huiyegit

When I train DAMSM, in the line valid loss 6.02757 5.78502, why are there 2 numbers next to valid loss

and which number do I need to use ?

| epoch 100 |     0/  184 batches | ms/batch  3.33 | s_loss 0.01315 0.01326 | w_loss 0.01357 0.01273
-----------------------------------------------------------------------------------------
| end epoch 100 | valid loss 6.02757 5.78502 | lr 0.00026524|
-----------------------------------------------------------------------------------------

Do you know ?

eoaiewo avatar Nov 20 '23 19:11 eoaiewo

Do you know ?

eoaiewo avatar Nov 21 '23 17:11 eoaiewo