Commit graph

108 commits

Author SHA1 Message Date
Yin Li
1963530984 Add gradient norms to tensorboard 2020-03-02 11:31:38 -05:00
Yin Li
9567db7332 Add weight initialization 2020-03-01 22:31:30 -05:00
Yin Li
75b1c19dcd Combine input and output figures 2020-03-01 15:40:53 -05:00
Yin Li
a46746287a Fix bug when env SLURM_STEP_NUM_NODES is not defined 2020-02-14 15:51:36 -05:00
Yin Li
5cb4a1bbae Remove label smoothing for generator 2020-02-14 09:22:57 -05:00
Yin Li
b67079bf72 Add runtime address and port determination and share them via file
Together with slurm step node counts, make it possible to launch
multiple training in one job
2020-02-13 19:56:54 -06:00
Yin Li
1818e11265 Add simple label smoothing by --adv-real-label 2020-02-13 16:44:56 -05:00
Yin Li
d2840f01b0 Add --adv-start as epoch to start adversarial training
This is similar to the deprecated --adv-delay, but specify
absolute epoch (--adv-delay counts from start_epoch in resumed training)
2020-02-13 16:00:50 -05:00
Yin Li
e8039dcccc Move map loss computation forward in training 2020-02-13 14:40:51 -05:00
Yin Li
53ed5a91f4 Change to nearest interpolation from linear in model super-resolution
To be consistent with the data super-resolution in data/fields.py
2020-02-13 12:01:04 -05:00
Yin Li
db3414e11c Disable D grad when updating G; Revert D and G order
The reason for the latter is to follow most code but I haven't
found the reason
2020-02-13 11:33:40 -05:00
Yin Li
cc0efd28ec Fix bug: pprint -> pformat 2020-02-11 12:55:08 -05:00
Yin Li
2e2adc761d Add spectral normalization 2020-02-11 11:58:05 -05:00
Yin Li
b41e85eda5 Fix stupid merging bug 2020-02-09 23:07:40 -05:00
Yin Li
16b82fcc56 Merge branch 'anneal_loss'
Conflicts:
	map2map/models/patchgan.py
	map2map/train.py
2020-02-09 21:25:39 -05:00
Yin Li
2aed00d97e Add argument to disable ReduceLROnPlateau by default 2020-02-09 20:15:23 -05:00
Yin Li
09493ad4ec Add non-strict model state loading 2020-02-09 17:32:09 -05:00
Yin Li
f831afbccf Add PatchGAN42 similar to the PatchGAN in pix2pix 2020-02-07 15:26:16 -05:00
Yin Li
45c1d57e72 Add pretty-print argument to job stdout 2020-02-07 11:40:08 -05:00
Yin Li
3fb9708575 Remove saving / loading optimizer & scheduler states
Conflicts:
	map2map/train.py
2020-02-07 10:15:14 -05:00
Yin Li
23b903f81b Remove saving / loading optimizer & scheduler states 2020-02-07 10:02:32 -05:00
Yin Li
1cd34c2eed Remove lazy argument feeding from field dataset 2020-02-07 09:49:27 -05:00
Yin Li
890a459363 Fix bug 2020-02-06 19:34:43 -06:00
Yin Li
cd63324724 Change delayed adversary to a smooth annealing scheme 2020-02-06 19:04:30 -05:00
Yin Li
291dfb24b3 Add diff between out and tgt to figure 2020-02-05 21:19:03 -05:00
Yin Li
a5c48e71b0 Change noise channel removal from input 2020-02-05 14:40:45 -05:00
Yin Li
ef0235f97b Fix upsampling crop bug 2020-02-05 14:10:58 -05:00
Yin Li
b609797e27 Add per field colormap and Fix scope bug 2020-02-05 11:58:05 -05:00
Yin Li
db69e9f953 Add figures with tensorboard 2020-02-03 22:18:08 -05:00
Yin Li
7f6578c63e Add delay to scheduler by adv_delay 2020-02-03 14:44:14 -05:00
Yin Li
0a2fc9a9e9 Add wrappers of adversary model and adversarial loss
commit c0dafec94bb7d131938650027f84e5308bf16ffd
Author: Yin Li <eelregit@gmail.com>
Date:   Mon Feb 3 11:18:08 2020 -0600

    Fix bug

commit b470b873649515f4b8a1cac7b4b33181eac51199
Author: Yin Li <eelregit@gmail.com>
Date:   Mon Feb 3 09:39:08 2020 -0600

    Fix bug

commit 9f8f64b3510c72bfcf2a1236ba5285edf280701c
Author: Yin Li <eelregit@gmail.com>
Date:   Mon Feb 3 10:20:37 2020 -0500

    Add wrappers of adversary model and adversarial loss
2020-02-03 12:05:47 -06:00
Yin Li
0721301113 Add noise channels to the input 2020-01-29 15:57:10 -05:00
Yin Li
862c9e75a0 Swap generator & discriminator order in training
The reasoning is that updating generator first will free up the memory
taken by the graph of the model
2020-01-22 19:00:13 -05:00
Yin Li
f99fd6b177 Add delay epoch before updating generator with adversarial loss 2020-01-22 18:44:09 -05:00
Yin Li
c68b9928ee Add separate input and target normalization 2020-01-21 18:51:32 -05:00
Yin Li
2b7e559910 Change adversary optim scheduler to sync with model optim scheduler 2020-01-20 16:25:15 -05:00
Yin Li
e7d2435a96 Fix adversary model optim scheduler 2020-01-19 19:24:10 -05:00
Yin Li
1b1e0e82fa Fix (possibly) multiple DistributedDataParallel reduce interference 2020-01-10 14:39:16 -05:00
Yin Li
15384dc9bd Add optional adversary model and make validation optional 2020-01-09 20:24:46 -05:00
Yin Li
9cf97b3ac1 Fix seeding bug introduced in the completely wrong commit f64b1e4 2020-01-06 20:20:05 -05:00
Yin Li
848dc87169 Change __dict__ to getattr 2019-12-23 16:11:43 -05:00
Yin Li
77710bc8a3 Cache data during the first epoch 2019-12-18 17:51:11 -05:00
Yin Li
01b0c8b514 Add data division, good with data caching 2019-12-18 17:13:40 -05:00
Yin Li
843fe09a92 Add data caching, and new pad and crop features 2019-12-17 12:00:13 -05:00
Yin Li
d03bcb59a1 Fix bug from 0533150 2019-12-12 19:26:57 -05:00
Yin Li
0533150194 Save past best models by not overwriting them 2019-12-12 18:05:16 -05:00
Yin Li
341bdbff84 Fix issue that pytorch 1.1 does not flush tensorboard 2019-12-12 15:25:45 -05:00
Yin Li
bd3798222a Add weight decay 2019-12-12 12:04:39 -05:00
Yin Li
7b6ff73be1 Move UNet with ResBlock to VNet and Revert UNet to the previous simple version 2019-12-09 21:53:27 -05:00
Yin Li
0764a1006e Remove unnecessary arguments --in-channels and --out-channels 2019-12-09 10:19:21 -05:00