Feat discriminator unet mha #580
Conversation
wr0124
force-pushed
the
feat_discriminator_unet_mha
branch
from
439d06e
to
01134aa
Compare
models/base_gan_model.py
Outdated
| getattr(self, fake_name + "_with_context"), size=self.real_A.shape[2:] | ||
| ), | ||
| ) | ||
|
|
There was a problem hiding this comment.
All changes above are not from this PR, I believe they shouldn't be here, right ?
There was a problem hiding this comment.
Those codes were not written by me, and I have no clue where they came from. I had an accidental pull three weeks ago, which was fixed by Louis; I'm wondering if they were imported at that time. Can I delete them without destroying the work of others?"
There was a problem hiding this comment.
they are being removed
models/gan_networks.py
Outdated
| D_ndf (int) -- the number of filters in the first conv layer | ||
| D_ngf(int) -- the number of filters in the last cov layer |
There was a problem hiding this comment.
Don't change this, also I believe D_ngf doesn't exist.
There was a problem hiding this comment.
D_ngf is one parameter in the first Unet ( https://github.com/jolibrain/joliGEN/blob/98a09fe6360b396cd3d309227deef47d278cf3c3/models/modules/unet_architecture/unet_generator.py#L14C16-L14C16 ) , but since it is not relevant to this UNet architecture, I'll try to delete it from this pull request
There was a problem hiding this comment.
it is being removed
| @@ -1,11 +1,24 @@ | |||
| import functools | |||
|
|
|||
| import torch | |||
There was a problem hiding this comment.
OK, i'll check on it and delete it
There was a problem hiding this comment.
if i remove it, my code does not work.
models/modules/loss.py
Outdated
| self.gan_mode, target_real_label=target_real_label | ||
| ).to(self.device) | ||
|
|
||
| self.viz = visdom.Visdom( |
There was a problem hiding this comment.
Remove viz from here, viz is taken care of elsewhere.
There was a problem hiding this comment.
This can be left here for testing, but moved/removed afterwards.
There was a problem hiding this comment.
OK. I'll delete all the viz
models/modules/loss.py
Outdated
| mixed_images = masks * fake + (1 - masks) * real | ||
| return mixed_images, masks | ||
|
|
||
| train_use_cutmix = True # change this into True to use cutmix |
There was a problem hiding this comment.
We need to make this an option, in options/train_options.py and then pass it as argument to DualDiscriminatorGANLoss(DiscriminatorLoss) constructor.
There was a problem hiding this comment.
ok, I'll make this an option.
models/modules/loss.py
Outdated
| consistent_pred_pixel = torch.mul( | ||
| pred_fake_pixel, 1 - cutmix_pixel_label | ||
| ) + torch.mul(pred_real_pixel, cutmix_pixel_label) | ||
| loss_cutmix_pixel = ( |
There was a problem hiding this comment.
@wr0124 are you sure this is the loss from the paper ?
It should be D(cutmix(inputs)) - cutmix(D(inputs)) no ?
There was a problem hiding this comment.
yes, you are right. in the paper, it is || D(cutmix(inputs)) - cutmix(D(inputs)) ||² . So, D(cutmix(inputs)) is pred_cutmix_fake_pixel in my code, and cutmix(D(inputs)) is consistent_pred_pixel in my code @beniz
There was a problem hiding this comment.
OK cool, you can probably use MSELoss instead of the norm and **2, that´d be more efficient probably.
There was a problem hiding this comment.
yes, you are right. The MSELoss is being used here.
wr0124
force-pushed
the
feat_discriminator_unet_mha
branch
from
59bd27e
to
3f67ad8
Compare
wr0124
force-pushed
the
feat_discriminator_unet_mha
branch
from
3f67ad8
to
dc69f51
Compare
| @@ -379,6 +379,20 @@ def initialize(self, parser): | |||
| default=64, | |||
| help="# of discrim filters in the first conv layer", | |||
| ) | |||
| parser.add_argument( | |||
There was a problem hiding this comment.
I believe this is already done with --D_ndf.
options/train_options.py
Outdated
|
|
||
| parser.add_argument( | ||
| "--train_use_cutmix", | ||
| type=bool, |
There was a problem hiding this comment.
use action="store_true" instead.
Integrate CutMix into the unet_discriminator_mha which is original UNet.