Loss test - Learning the two-moons distribution with a normalizing flow

[Justinezgh]

Tests to combine the two losses : log likelihood and score.

[Justinezgh]

Notebook : https://colab.research.google.com/drive/1t4DaL02o31OCOFifDaQS2B1f_QN5-iRq?usp=sharing

[Justinezgh]

• 1000 updates
• loss = - log(p(x))

[Justinezgh]

• 1000 updates
• loss = norm(score_hat - score)**2

[Justinezgh]

• 1000 updates
• loss = -log(p(x)) + mu* norm(score_hat - score)**2

[Justinezgh]

• 10 000 updates
• loss = -log(p(x)) + mu* norm(score_hat - score)**2

[Justinezgh]

• 10 000 updates
• loss = -nu log(p(x)) + mu norm(score_hat - score)**2