Alexia Jolicoeur-Martineau

icon indicating a website link https://ajolicoeur.wordpress.com/ icon indicating an email [email protected]

icon company Samsung SAIT icon location Montréal, Qc, Canada icon location AI researcher at the Samsung SAIT AI Lab Montréal. I build generative models for images, videos, text, tabular data, NN weights, molecules, and video games.

Results 32 comments of Alexia Jolicoeur-Martineau

SELU weight init

Thanks for raising this point! You're right, it's said in the paper on page 6 (https://arxiv.org/pdf/1706.02515.pdf) that one should use var = 1/n_weights. It's still an attracting point so not...

SELU weight init

Just saw this earlier today: http://cs231n.github.io/neural-networks-2/#init. It shows why use fan_in.

GBDTMO crash in Google colab (see simple example)

It still crash, see the colab. It now installs numpy==1.18.0 before everything else.

GBDTMO crash in Google colab (see simple example)

I added 'pip list' to show all installed Python packages in the colab and it doesn't seem to have anaconda installed. It's native python. There's no way to log the...

GBDTMO crash in Google colab (see simple example)

Outside colab, the error still happens and it returns a Segmentation fault.

Adding FID statistics calculation as an option (can now do "train", "eval", or "fid_stats")

Alright, will change this.

Adding FID statistics calculation as an option (can now do "train", "eval", or "fid_stats")

Will now test to see if I can get the same FID on a model.

Adding FID statistics calculation as an option (can now do "train", "eval", or "fid_stats")

Very close results, but not exactly the same, sadly! FID: 534.9150390625 vs FID: 535.936279296875 Let me know if you figure out anything to be changed. At least it seems very...

Adding FID statistics calculation as an option (can now do "train", "eval", or "fid_stats")

Then, that might be fine, floating points errors are acceptable. It's trained for 8 iterations on a tiny batch 😂.

Adding FID statistics calculation as an option (can now do "train", "eval", or "fid_stats")

Good news, I tried with your pre-trained model (cifar10_continuous_ve) on chkpt 24 on a FID on only 2k samples (so it finishes quickly enough). The results from two runs with...