The results of using FP model for MMNIST are really bad

edenton / svg

186 stars 56 forks source link

The results of using FP model for MMNIST are really bad #14

Open zyong812 opened 4 years ago

zyong812 commented 4 years ago

My results of using fix-prior model for MMNIST generation are really bad. As the following gif shows, after 25 epochs. I have tried the code with over 100 epochs, but the results don't improve much.

I don't find the comparison of FP and LP over MMNIST in the paper. What's your results in such case? Anybody get good results using FP on MMNIST? Thanks!

sample_25

xjgaocs commented 4 years ago

I got a similiar results as yours using fp setting. Actually, I haven't got a better results of fp than lp.

DanielTakeshi commented 3 years ago

SVG-LP seems to be working well, maybe try that over FP.

From the command in the README, I have this after 25 epochs:

sample_25

Here's the result after 300 epochs, i.e., this is the sample_299.gif produced from the code:

sample_299