Former-commit-id: ef93f456d6e14f061325063d7eee8f0ed9747da8

added back the ipynb README.ipynb and the example. I noticed what is wrong: the description fails to render. Going to use twine to check the validity.


Former-commit-id: 6a35cda9f714fa987f8eea5a01e308801d0235dc

Former-commit-id: 746e2a55af1c0457597488d002945f0968eb4743

Former-commit-id: 16f4fd7830f7fd429534a31e27db7c8bef510038

removed the files that can't be uploaded in pypi (only examples). Will need to put it back after. Saved on my PC


Former-commit-id: 676a37a8c6c268bd7eabe26daf2213a1edde60a2

Former-commit-id: 8c287144b92cbc67916d2f1c0be186ce336bb8a1

did some minor changes. I am trying to release it to Pypi. I removed the test_notebooks dir. It is now on my PC.


Former-commit-id: ddbb9e53447f838af9e15d230359400f63b8af4f

Implemented a PLN VAE that infers only the the mean and variance of the latent layer in the Test.ipynb. It seems to work since I retrieve the maximum of the ELBO. However, it takes many forward pass to retrieve it. Need to find a good optimizer.  One possible alternative of this would be to approximate with a non-diagonal gaussian. it is quite easy to implement. I also implemented a complete VAE where we don't suppose a PLN structure. It takes even much more time. I don't have found yet a good optimizer for this. I am going to launch many optimizers and compare them.


Former-commit-id: 3d1e28fcd6eeb18a9452ce075a82528e82d5c2a0

Former-commit-id: e82a63162e7a3e0979cd15ec77b893deb68d17e4

I linked the images of the numerical complexity that shows the time convergence of fastPLN and fastPLNPCA w.r.t p and w.r.t n. I also added the compelixty for each algorithm (i.e. linear for both IMPS_PLN and fastPLNPCA, quadratic for fastPLN).

Former-commit-id: bc1bf3d97b1ee0337c568153d28390f04dd7387c

Former-commit-id: 1cf70a85c00a2fa9fbb876d88251e1a20162bed5

Former-commit-id: 4066d4f19db86aaadd21e4791366e2edb66ae044

Former-commit-id: b8617da8aeeb164a83604501a3b6d423f7975efa

Former-commit-id: 7c228b34111380c964f1110fe908f7731bdde04a

added the README.ipynb that contains the formulas, and the README.md. The md will display formulas with dollars $, while the ipynb will render the formulas in math mode. However, we will see only the md first.


Former-commit-id: 1132eb17555fd4a5139b73cda57170b2acb41540

delete since useless. I will only display the ipynb that shows math in $$ $$ ...

Former-commit-id: d5201cefd9ffa1ceb11e7241bc8008ec47ea7834

Former-commit-id: a37bf0da43f78c087d149142d0ad5a1e62f2a522

Former-commit-id: f9b6a7caf01ac52b6e5ee07cca1f209939cb34ad

Former-commit-id: 5c174e2138452877e81fabbf0362e242600bf22d

Former-commit-id: a25c7acb2ccb9388305b1b20ed4f5b829ac150f7

Former-commit-id: 414f799b7f0e29c0691df7528067f9927f2f0960

Former-commit-id: 6e8fb82034579e9e10dd8477f4decf67124e91b6

Former-commit-id: 26bec74a6bd9baba6b6a4bcfb3f2326226b5b920

Tried to render the math in a jupyter notebook

Former-commit-id: a75078becd1a98a702099cc77afb3372f4b671bf

Changed the README.md so that it can displays the math (previously it was not showing the things in math mode)


Former-commit-id: 9630b88942b2d22fa2b6ae0b38fd948c74e78a49

Former-commit-id: 52b3108a7ca63be821ddcbc09cee952558506cc3

Former-commit-id: 87b00e922e8dc4f7b91590b28058290729de89f8

Former-commit-id: 9d59536dc9bcf21c4eb9c6b011d53677fca86e4c

Former-commit-id: 544b9bf9660ba05d15946e742e3a2be890a5ffc1

Fixed an error that occured when we wanted to take a GPU. Now we can use a gpu for every algorithm. I noticed a bug for IMS_PLN. The log likelihood is not the same when we take some batch size or others. We get the same result for batch sizes of 10 or 40, bur different when batch_size is 30 or 41 for example. I think this is due to the calculation of the likelihood  when batches have different size (200 is not divisible by 30 or 41 but divisible by 10 and 40)


Former-commit-id: 1e7fda3de3718b9a7bba88b9cee69001237bb7c9

Finished to review the code for the three .py files. Also run each file in a package that format your code to the right convention.


Former-commit-id: f58d49f0d5d4a108c402692099df6b10e6d09669

Added examples of code for each model (fastPLN, fastPLNPCA, IMPS_PLN). I added screenshots of the supposed results. I also added some formulas for the gradient in the mathematic description.


Former-commit-id: 61cea783be3f670cfc2fb726f0fa2e1cbade2785

Deleted some files that won't be needed in the package. I kept them in local if needed, but they're in git anyway. I added the constants on the ELBO on the PLN model to get a stop criteria.
I also modified the stop criterion of the IMPS_model : We add 1 to the criterion if the log likelihood had not be improved for nb_trigger epoch. We stop when the criterion reahces the parameter nb_plateau.
I also did some minor modifications in the utils.


Former-commit-id: 1f224df8314fa4a7c9818a21874afac17cfefbed

added tests datasets as examples for the package. Also updated models.py so that it takes pd.DataFrame as inputs instead of torch.tensor (in practice we extract data with pandas).


Former-commit-id: 410f9e4e91fd236845119e75b60abdb745b3f030

Former-commit-id: 290b6b4cd3887080f7e45cd09b85f7dc23db3f5c

changed the name SAGA to AdaSAGA vs AdaSVRG. I will use this file to compare both. I began the implementation of AdaSVRg but not finished. The parameter theta for adaptive termination is bad with 0.5 as said in the paper SVRG meets Adagrad. Given the examples, I found either 0.1 or even 0.001.
I also modified the sample_PLN funcion in the utils.py file to take only torch.tensors and not numpy arrays. All the others modifications are useless and minors.


Former-commit-id: 382a20ed3d75fc5cfdabad9db1ed034e87263b6c

Tried to see what was going wrong in the SAGARAD algorithm. Modified a little bug in the SAG algorithm.


Former-commit-id: 0b76edbd2891d728b0a1e13c1a4d4be93b6a80ce

Tried to figure out why we have convergence of the parameters in some cases and not in others cases for the PLNPCA variational and ZIPLN. I measured the SNR, but it is not really informative. I put the PLNPCA variational estimator in the file models.py. It only misses the ZIPLN file. I am currently trying to get the Graphical lasso estimator for the variational PLN. I also deleted some useless files.


Former-commit-id: cc4f9546add39a5790b9583473a514fceef8783a

Fixed the issues with the GPU. Now you can work with the GPU for fastPLN and IMPS PLN. It took me about 1h for n= 2000, p =1000 for fastPLN.


Former-commit-id: 86c8b99533bb0ae91704b39d3dacb778622374b6

modified the IMPS name in fastPLNmodels. I am also pushing to get it on my labtop and see a more accurate error, cuda is not telling me the error with gpu ...


Former-commit-id: a00164bbfc8dfb0e0ac96dc29421669a585c84fc