Adding doc, adding a timer, fixing bugs. #84

teytaud · 2019-09-26T09:18:25Z

Adding documentation.
Fixing a crash for DCGAN.
Adding a timer.

teytaud · 2019-09-26T12:00:44Z

README.md

@@ -64,8 +79,9 @@ And wait for a few days. Your checkpoints will be dumped in output_networks/cele
 For celebaHQ:

 ```
-python datasets.py celebaHQ $PATH_TO_CELEBAHQ -o $OUTPUT_DATASET - f
-python train.py PGAN -c config_celebaHQ.json --restart -n celebaHQ
+python datasets.py celebaHQ $PATH_TO_CELEBAHQ -o $OUTPUT_DATASET - f  # Prepare the dataset and build the configuration file.


not sure of this << - f >> ?

-f accelerate the training by saving intermediate images of smaller sizes.

teytaud · 2019-09-26T12:05:35Z

README.md

@@ -48,6 +48,21 @@ pip install -r requirements.txt
 - DTD: https://www.robots.ox.ac.uk/~vgg/data/dtd/
 - CIFAR10: http://www.cs.toronto.edu/~kriz/cifar.html

+For a quick start with CelebAHQ, you might:


I just wanted to facilitate people's life by copy-pasting the TLDR.

Good idea for celebaHQ, it was a pain in the *** to make the first time.

I would however add a section ## Quick download

README.md

teytaud · 2019-09-26T12:08:40Z

models/trainer/progressive_gan_trainer.py

@@ -22,6 +23,7 @@ def __init__(self,
                 pathdb,
                 miniBatchScheduler=None,
                 datasetProfile=None,
+                 max_time=0,


@brozi I added a timer, which will help for DFO or AS for GAN optimization.

teytaud · 2019-09-30T07:52:44Z

models/trainer/DCGAN_trainer.py

@@ -18,6 +19,7 @@ def getDefaultConfig(self):

    def __init__(self,
                 pathdb,
+                 miniBatchScheduler=None,                 


Adding this argument is necessary because dataset.py sets up miniBatchScheduler and we get a crash at runtime because init does not have such an argument ?

It shouldn't crash unless you add a scheduler in the configuration file.

Presumably there is a problem in the automatic generation of the configuration file.
@Molugan how should we proceed regarding this issue ? I would go for keeping my fix so that at least it does not crash.

The issue is DCGAN + celebaHQ in the dataset maker. I hadn'y considered the possibility that someone would want to use both.

Then the dataset maker should be changed, not this part: indeed the idea of the GANTrainer is to be as general as possible. Specific features like multi-scale are defined by child classes.

Molugan · 2019-10-23T11:10:54Z

README.md

@@ -3,7 +3,6 @@
 A GAN toolbox for researchers and developers with:
 - Progressive Growing of GAN(PGAN): https://arxiv.org/pdf/1710.10196.pdf
 - DCGAN: https://arxiv.org/pdf/1511.06434.pdf
- To come: StyleGAN https://arxiv.org/abs/1812.04948


So no StyleGAN implementation in the end, @Molugan ?😞

There is a PR for styleGAN !

styleGAN incoming (#95) so no need to remove this one

Molugan

I'm ok with most changes. Just a little stuff to correct and this can be merged.

was waiting for feedback around other possibilities for the bugfix; maybe time to go ahaead.

Molugan · 2020-01-07T12:18:36Z

README.md

-python train.py PGAN -c config_celebaHQ.json --restart -n celebaHQ
+python datasets.py celebaHQ $PATH_TO_CELEBAHQ -o $OUTPUT_DATASET  # Prepare the dataset and build the configuration file.
+python train.py PGAN -c config_celebaHQ.json --restart -n celebaHQ  # Train.
+python eval.py inception -n celebaHQ -m PGAN  # If you want to check the inception score.


I'd rather not add the inception part here.

Molugan · 2020-01-07T12:20:28Z

README.md

 ```
+You can compare choose for the optimization one of the optimizers in Nevergrad


You can compare choose for the optimization one of the optimizers in Nevergrad ->
You can also try out some optimizers from Nevergrad

Molugan · 2020-01-07T12:22:19Z

README.md

 ```
+for the SWD score, to be maximized, or for the inception score:


No, the SWD should be minimized

Molugan · 2020-01-07T12:24:04Z

README.md

-### Inspirational generation
+### Inspirational generation (https://arxiv.org/abs/1906.11661)
+
+You might want to generate clothese (or faces, or whatever) using an inspirational image, e.g.:


typo: clothese -> clothes

I don't find the description very clear :(.

Molugan · 2020-01-07T12:27:01Z

models/trainer/gan_trainer.py

@@ -50,6 +51,7 @@ def __init__(self,
            - modelLabel (string): name of the model
            - config (dictionary): configuration dictionnary.
            for all the possible options
+            - max_time (int): max number of seconds for training (0 = infinity).


NIT: working with second is not very practical, hours looks like a better unit to me.

Molugan

The README should be updated. I'm pushing a fix for the dataset maker.

Update README.md

3950b0c

teytaud changed the title ~~Update README.md~~ [WIP] Update README.md Sep 26, 2019

facebook-github-bot added the CLA Signed Do not delete this pull request or issue due to inactivity. label Sep 26, 2019

teytaud added 8 commits September 26, 2019 11:25

Update README.md

c4c4b87

Update README.md

d5f5841

Update README.md

8d2cd8f

Update README.md

c6379f9

Update train.py

1a07faf

Update inspirational_generation.py

5599bc2

Update README.md

da128a1

Update README.md

39e7ee9

teytaud commented Sep 26, 2019

View reviewed changes

teytaud added 2 commits September 26, 2019 14:01

Update train.py

05a8a1c

Update progressive_gan_trainer.py

b15a162

teytaud changed the title ~~[WIP] Update README.md~~ [WIP] Adding doc, adding a timer. Sep 26, 2019

teytaud commented Sep 26, 2019

View reviewed changes

README.md Show resolved Hide resolved

teytaud changed the title ~~[WIP] Adding doc, adding a timer.~~ Adding doc, adding a timer. Sep 26, 2019

teytaud requested review from Molugan, coupriec and baptisteroziere September 26, 2019 12:07

teytaud commented Sep 26, 2019

View reviewed changes

teytaud added 8 commits September 30, 2019 06:54

Update README.md

2b13f24

Update README.md

b1f0983

Update gan_trainer.py

1805b95

Update progressive_gan_trainer.py

b3d62b0

Update README.md

d6577ba

Update README.md

dcdaa66

Update DCGAN_trainer.py

57fdfaf

Update progressive_gan_trainer.py

d2a11a9

teytaud added 2 commits September 30, 2019 07:21

Update gan_trainer.py

fa54887

Update DCGAN_trainer.py

8374753

teytaud commented Sep 30, 2019

View reviewed changes

teytaud and others added 9 commits October 1, 2019 13:53

Update README.md

88eca90

Update README.md

d0a40a6

add inspirational images example

10e2247

Update README.md

dcb4f4b

Update README.md

b43210c

Update README.md

4451d70

Update README.md

4fc893d

Update README.md

3dccf5e

Update train.py

d812b0a

Molugan reviewed Oct 23, 2019

View reviewed changes

Molugan previously requested changes Oct 23, 2019

View reviewed changes

teytaud requested a review from Molugan December 25, 2019 08:03

teytaud changed the title ~~Adding doc, adding a timer.~~ Adding doc, adding a timer, fixing bugs. Jan 2, 2020

Molugan reviewed Jan 7, 2020

View reviewed changes

Molugan suggested changes Jan 7, 2020

View reviewed changes

BUGFIX: minibatchsceduler remaining with DCGAN

6f0cec7

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adding doc, adding a timer, fixing bugs. #84

Adding doc, adding a timer, fixing bugs. #84

teytaud commented Sep 26, 2019 •

edited

Loading

teytaud Sep 26, 2019

Molugan Jan 7, 2020

teytaud Sep 26, 2019

Molugan Oct 23, 2019

teytaud Sep 26, 2019

teytaud Sep 30, 2019

Molugan Oct 23, 2019

teytaud Dec 25, 2019

Molugan Jan 7, 2020

Molugan Jan 7, 2020 •

edited

Loading

Molugan Oct 23, 2019

kondela Oct 23, 2019

Molugan Jan 7, 2020

Molugan Jan 7, 2020

Molugan left a comment

Molugan Jan 7, 2020

Molugan Jan 7, 2020

Molugan Jan 7, 2020 •

edited

Loading

Molugan Jan 7, 2020 •

edited

Loading

Molugan Jan 7, 2020

Molugan left a comment

		```
		for the SWD score, to be maximized, or for the inception score:

Adding doc, adding a timer, fixing bugs. #84

Are you sure you want to change the base?

Adding doc, adding a timer, fixing bugs. #84

Conversation

teytaud commented Sep 26, 2019 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Molugan Jan 7, 2020 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Molugan left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Molugan Jan 7, 2020 • edited Loading

Choose a reason for hiding this comment

Molugan Jan 7, 2020 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Molugan left a comment

Choose a reason for hiding this comment

teytaud commented Sep 26, 2019 •

edited

Loading

Molugan Jan 7, 2020 •

edited

Loading

Molugan Jan 7, 2020 •

edited

Loading

Molugan Jan 7, 2020 •

edited

Loading