I’ve replicated the steps Jeremy took on https://github.com/fastai/course-v3/blob/master/nbs/dl1/lesson7-superres-gan.ipynb
However, I did take a different approach where I am only using 20 images total of 2 types of dogs(10 each) as well as splitting the validation set to 30 percent of the ‘image_gen’ folder and increasing the lr to 1e-2 once I got to the pretraining of the critic part.
This gave similar results, but the below happened when I got to the pretrain the critic part. This might be more of a @jeremy question, but why is it that the critic’s loss doesn’t go down when the dataset is small? I would think it’ll be easier to distinguish between a fake and a real image, since there are less images to learn. It’s not like the MSEloss of the generator was lower than Jeremy’s notebook either. Shouldn’t this mean that the critic should have a easier time distinguishing the real from fakes?
def get_crit_data(classes, bs, size):
src = ImageList.from_folder(path, include=classes).split_by_rand_pct(0.3, seed=42)
ll = src.label_from_folder(classes=classes)
data = (ll.transform(get_transforms(max_zoom=2.), size=size)
.databunch(bs=bs).normalize(imagenet_stats))
data.c = 3
return data
data_crit = get_crit_data([name_gen, 'images'], bs=4, size=size)
data_crit.show_batch(rows=3, ds_type=DatasetType.Train, imgsize=3)
plt.show()
loss_critic = AdaptiveLoss(nn.BCEWithLogitsLoss())
def create_critic_learner(data, metrics):
return Learner(data, gan_critic(), metrics=metrics, loss_func=loss_critic, wd=wd)
learn_critic = create_critic_learner(data_crit, accuracy_thresh_expand)
learn_critic.fit_one_cycle(25, 1e-2)
learn_critic.save('critic-1')
data_crit = get_crit_data(['crappy', 'images'], bs=bs, size=size)
learn_crit = create_critic_learner(data_crit, metrics=None).load('critic-1')
the output
----------------------------------GAN--------------------------------------
epoch train_loss valid_loss accuracy_thresh_expand time
0 0.686626 0.702948 0.356364 00:10
1 0.691571 0.719672 0.367273 00:11
2 0.684327 0.766719 0.450909 00:11
3 0.694961 0.740039 0.374545 00:10
4 0.697699 0.698146 0.545455 00:10
5 0.692488 1.168247 0.545455 00:10
6 0.705878 0.690986 0.545455 00:10
7 0.705758 0.716483 0.483636 00:10
8 0.710638 0.705156 0.476364 00:10
9 0.710643 0.693130 0.443636 00:10
10 0.708336 0.695476 0.454545 00:10
11 0.706554 0.691266 0.545455 00:10
12 0.705258 0.690534 0.545455 00:10
13 0.704851 0.697592 0.454545 00:11
14 0.703753 0.696524 0.454545 00:11
15 0.702643 0.694397 0.454545 00:10
16 0.701649 0.693619 0.454545 00:10
17 0.700480 0.693081 0.450909 00:10
18 0.699535 0.693055 0.450909 00:11
19 0.698590 0.692627 0.538182 00:10
20 0.697668 0.692416 0.629091 00:10
21 0.697229 0.692438 0.600000 00:11
22 0.696679 0.692444 0.581818 00:10
23 0.696413 0.692449 0.581818 00:11
24 0.695998 0.692447 0.581818 00:10
