Closed jbueltemeier closed 2 years ago
Merging #248 (432ad21) into master (7ba28c6) will decrease coverage by
0.5%
. The diff coverage is86.5%
.
@@ Coverage Diff @@
## master #248 +/- ##
========================================
- Coverage 98.6% 98.1% -0.6%
========================================
Files 39 39
Lines 1558 1610 +52
========================================
+ Hits 1537 1580 +43
- Misses 21 30 +9
Impacted Files | Coverage Δ | |
---|---|---|
pystiche_papers/ulyanov_et_al_2016/_utils.py | 100.0% <ø> (ø) |
|
pystiche_papers/ulyanov_et_al_2016/_loss.py | 91.5% <83.6%> (-8.5%) |
:arrow_down: |
pystiche_papers/ulyanov_et_al_2016/_data.py | 100.0% <100.0%> (ø) |
|
pystiche_papers/ulyanov_et_al_2016/_modules.py | 100.0% <100.0%> (ø) |
|
pystiche_papers/ulyanov_et_al_2016/_nst.py | 95.4% <100.0%> (+0.1%) |
:arrow_up: |
pystiche_papers/li_wand_2016/_utils.py | 96.1% <0.0%> (ø) |
|
pystiche_papers/gatys_et_al_2017/_utils.py | 100.0% <0.0%> (ø) |
|
pystiche_papers/gatys_ecker_bethge_2016/_utils.py | 100.0% <0.0%> (ø) |
Continue to review full report at Codecov.
Legend - Click here to learn more
Δ = absolute <relative> (impact)
,ø = not affected
,? = missing data
Powered by Codecov. Last update 7ba28c6...432ad21. Read the comment docs.
This PR has been divided into several smaller PRs: #250 Manipulate gradients in ulyanov_et_al_2016 #251 Change Ulyanov stylization pretransformation #252 Remove unused texture images in ulyanov_et_al_2016 #253 Add missing permalinks in ulyanov_et_al_2016
The handling of the image pre-processing of the content images has still been left open. This currently leads to a failure of the training. This will be addressed in a new PR.
So far, this implementation has not produced good results. This was because in the reference implementation the backward pass of the gradient is normalized and this is missing in this implementation. The results are now better, but still with measures, probably due to different hyperparameter that are too high or too low. Therefore, similar results as in the paper can now be created if the
style_weight
is changed.Therefore in the following two points which should be discussed in advance:
Due to the changed hyperparameter setting from #244 it is now easier to adjust the hyperparameter. Therefore i would adjust the
style_weight
for this replication so that besides the results with the default parameters similar results to the paper are created.One problem with the replication of the master branch is that the images are too small and have to be enlarged for training. I currently added an additional
resize
in thecontent_image_transform
before theValidRandomCrop
, but this is not implemented in the reference implementation.