mrq/DL-Art-School
1
3
Fork 5
Code Issues Pull Requests Packages Projects Releases Wiki Activity
aa0305def9
DL-Art-School/codes/options/train/train_ESRGAN_blacked.yml

86 lines
1.6 KiB
YAML
Raw Normal View History

Add my own configs
2020-04-22 06:37:54 +00:00
#### general settings
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
name: blacked_fix_and_upconv
Add my own configs
2020-04-22 06:37:54 +00:00
use_tb_logger: true
model: srgan
distortion: sr
scale: 4
gpu_ids: [0]
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
amp_opt_level: O1
Add my own configs
2020-04-22 06:37:54 +00:00
#### datasets
datasets:
train:
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
name: vixcloseup
Add my own configs
2020-04-22 06:37:54 +00:00
mode: LQGT
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
dataroot_GT: K:\4k6k\4k_closeup\hr
dataroot_LQ: K:\4k6k\4k_closeup\lr_corrupted
Enable megabatching
2020-05-02 23:46:59 +00:00
doCrop: false
Add my own configs
2020-04-22 06:37:54 +00:00
use_shuffle: true
Add support for passthrough disc/gen Add RRDBNetXL, which performs processing at multiple image sizes. Add DiscResnet_passthrough, which allows passthrough of image at different sizes for discrimination. Adjust the rest of the repo to allow generators that return more than just a single image.
2020-05-04 20:01:43 +00:00
n_workers: 0 # per GPU
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
batch_size: 40
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
target_size: 256
Add my own configs
2020-04-22 06:37:54 +00:00
color: RGB
val:
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
name: adrianna_val
Add my own configs
2020-04-22 06:37:54 +00:00
mode: LQGT
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
dataroot_GT: E:\4k6k\datasets\adrianna\val\hhq
dataroot_LQ: E:\4k6k\datasets\adrianna\val\hr
Add my own configs
2020-04-22 06:37:54 +00:00
#### network structures
network_G:
which_model_G: RRDBNet
in_nc: 3
out_nc: 3
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
nf: 48
Add my own configs
2020-04-22 06:37:54 +00:00
nb: 23
network_D:
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
which_model_D: discriminator_resnet
Add my own configs
2020-04-22 06:37:54 +00:00
in_nc: 3
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
nf: 48
Add my own configs
2020-04-22 06:37:54 +00:00
#### path
path:
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
pretrain_model_G: ../experiments/rrdb_blacked_gan_g.pth
pretrain_model_D: ~
Add my own configs
2020-04-22 06:37:54 +00:00
strict_load: true
Add support for passthrough disc/gen Add RRDBNetXL, which performs processing at multiple image sizes. Add DiscResnet_passthrough, which allows passthrough of image at different sizes for discrimination. Adjust the rest of the repo to allow generators that return more than just a single image.
2020-05-04 20:01:43 +00:00
resume_state: ../experiments/blacked_fix_and_upconv/training_state/16500.state
Add my own configs
2020-04-22 06:37:54 +00:00
#### training settings: learning rate scheme, loss
train:
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
lr_G: !!float 1e-5
Add my own configs
2020-04-22 06:37:54 +00:00
weight_decay_G: 0
beta1_G: 0.9
beta2_G: 0.99
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
lr_D: !!float 4e-5
Add my own configs
2020-04-22 06:37:54 +00:00
weight_decay_D: 0
beta1_D: 0.9
beta2_D: 0.99
lr_scheme: MultiStepLR
niter: 400000
warmup_iter: -1 # no warm up
Enable megabatching
2020-05-02 23:46:59 +00:00
lr_steps: [5000, 20000, 40000, 60000]
Add my own configs
2020-04-22 06:37:54 +00:00
lr_gamma: 0.5
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
mega_batch_factor: 4
Add my own configs
2020-04-22 06:37:54 +00:00
pixel_criterion: l1
pixel_weight: !!float 1e-2
feature_criterion: l1
Enable megabatching
2020-05-02 23:46:59 +00:00
feature_weight: 0
feature_weight_decay: .9
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
feature_weight_decay_steps: 501
feature_weight_minimum: 0
Fixup upconv for the next attempt!
2020-05-02 01:56:14 +00:00
gan_type: gan # gan | ragan
Enable megabatching
2020-05-02 23:46:59 +00:00
gan_weight: 1
Add my own configs
2020-04-22 06:37:54 +00:00
D_update_ratio: 1
Support inference across batches, support inference on cpu, checkpoint This is a checkpoint of a set of long tests with reduced-complexity networks. Some takeaways: 1) A full GAN using the resnet discriminator does appear to converge, but the quality is capped. 2) Likewise, a combination GAN/feature loss does not converge. The feature loss is optimized but the model appears unable to fight the discriminator, so the G-loss steadily increases. Going forwards, I want to try some bigger models. In particular, I want to change the generator to increase complexity and capacity. I also want to add skip connections between the disc and generator.
2020-05-04 14:48:25 +00:00
D_init_iters: 997
Add my own configs
2020-04-22 06:37:54 +00:00
manual_seed: 10
val_freq: !!float 5e2
#### logger
logger:
print_freq: 50
save_checkpoint_freq: !!float 5e2
Reference in New Issue Copy Permalink