autoencoders pretraining using RGB faces

Codes allowing to pre-train an autoencoder, using RGB facial images from the CelebA database.

added 1 commit

• 8f8b627f - removed os from requirements.txt

Compare with previous version

Hey,

Did you have a look at how things are organized in this package ? The idea is actually not to have everything in one place (hence the folders dataset, architectures, trainers ...). Your script is handling almost all of that !

I can understand that it may be easier for you, but that would be nice to comply to what has been done before, for consistency and reusability purposes.

Also, since neither your Dataset, nor your architecture seem to be defined in the script, I guess that you should at least provide the configuration file ...

Please have a look at the doc and try to follow the examples provided there, thanks !

@heusch,

This Merge Request is not meant for reviewing yet, it is explicit WIP :) Sure, will add datasets and configs to corresponding locations, and will let you know once it is ready for reviewing.

However, you are right, the code will be organized a bit differently. I have a generic training script, which is suitable for all DNN architectures I am planning to add to the package. Well, I think this is the best from re-usability point view, since doesn't require a separate trainer for each architecture.

Please let me know if you have different opinion. Thanks!

added 1 commit

• 720dd3ce - Removed standard lib packages from meta.yaml, added entry point for training script

Compare with previous version

This Merge Request is not meant for reviewing yet, it is explicit WIP :) Sure, will add datasets and configs to corresponding locations, and will let you know once it is ready for reviewing.

I made a comment early on just to avoid you unecessary and maybe redundant work ;)

However, you are right, the code will be organized a bit differently. I have a generic training script, > which is suitable for all DNN architectures I am planning to add to the package. Well, I think this is the best from re-usability point view, since doesn't require a separate trainer for each architecture.

I'm not familiar with autoencoders, but in my case (and in the package) different architectures (well, in the same "family" at least) could share a single trainer. This is the case for the CNNTrainer for instance: you can define whatever CNN (VGG, ResNet, LightCNN) to do classification, and the same trainer could be used as is. This does not apply to GAN, since a conditional GAN is quite different than a traditional GAN (in the problem definition I mean).

I don't know it what goes for CNN also applies to autoencoder, but if it does, it would be nice to have an AutoEncoderTrainer in trainers and the different network architectures in architectures.

Thanks !

Well, for me a single script worked for all nets :) Sure, will put to corresponding places. But in my case the trainer is not a class, just a script taking network configs as inputs. Will demonstrate how to run when ready. Thanks!

added 1 commit

• f979eb7d - Fixed the map_labels in the utils, fixing the unit tests

Compare with previous version

added 1 commit

• 06d5b2df - Added dataset class, Conv-AE model, config to train on CelebA, and train script

Compare with previous version

added 1 commit

• 89723900 - Added h5py to the meta.yaml

Compare with previous version

added 1 commit

• b2a8f933 - modified the init in datasets folder

Compare with previous version

added 1 commit

• 7762c4de - Added h5py to requirements

Compare with previous version

added 1 commit

• eca9f56b - Reverted changes in data_folder related to h5py import

Compare with previous version

added 1 commit

• 181b2fff - Added the unit test for ConvAutoencoder model

Compare with previous version

Hey @heusch ,

This is now ready, so I remove the WIP from MR. If interested to test the training you can run:

jman submit --queue gpu \
--name experiment_0_celeba_pretraining \
--log-dir <path_to_save_the_results>/logs/ \
--environment="PYTHONUNBUFFERED=1" -- \
./bin/pytorch-train-autoencoder-pad.py \
/idiap/temp/onikisins/project/ODIN/experiment_data/pad_experiments_bob4/batl_db/ae_refactor_test/celeba_preprocessing/experiment_0/small_copy/ \
<path_to_save_the_results>/ \
-c autoencoder/net1_celeba.py \
-cg bob.learn.pytorch.config \
-gpu \
-vv

This will train an autoencoder on a small subset of CelebA faces. Thanks!

unmarked as a Work In Progress

Could you rename that to train_autoencoder.py ?

yes, will do so.

Apart from that, it looks alright .. That's a bit different than how I did it, but this should be ok ! I will check more in details when I'll have some more time, since it may actually be a better way to handle data for instance (although less generic I guess).

Could you also update the doc ? That would be great to have this explained in plain text as well.

Thanks

added 1 commit

• 1876beb7 - Renamed the training script to more specific, now train_autoencoder

Compare with previous version

Thanks for the review! Yes, I will create a section on AEs in the docs.

added 1 commit

• 61202ef7 - Updated the documentation on AEs, trained on CelebA

Compare with previous version

Hello @heusch,

The documentation on AEs is now updated, ready to be merged :), unless you have more comments. Thanks!

Thanks for the documentation. Just don't forget to update the name of the training script (as previously mentioned above)

Already updated.

Sure ? Just build the package (and the doc), and I still have a ./bin/pytorch-train-autoencoder-pad.py script ...

Ah, sorry, I guess you mentioned different place, Will rename now. Thanks!

Here as well

changed this line in version 13 of the diff

Also, make sure that docstrings are right (within bob, it has been agreed to use numpy style: https://sphinxcontrib-napoleon.readthedocs.io/en/latest/example_numpy.html)

For instance, your DataFolder class is not displayed correctly in the Python API section of the doc. You may want to have a look at other classes in the packages.

added 1 commit

• 2da69255 - Renamed the script used for AE training, both in setup.py and docs

Compare with previous version

added 1 commit

• 18e00bbf - Updated the doc in data_folder to match the numpy format

Compare with previous version

added 1 commit

• 842c9bbd - Updated the inint in datasets folder, making the source link available in docs

Compare with previous version

added 1 commit

• 5e44dfde - Added the doc to the __init__ in data_folder, to fix the sphinx warning

Compare with previous version

added 1 commit

• 7e723289 - Renamed callable to object, sphinx warning

Compare with previous version

@heusch , addressed your comments. The script is renamed, and the class documentation is in numpy format. Thanks!

All good thanks !

merged

mentioned in commit c75f9631