Collecting Images For VAE [question]

araffin / learning-to-drive-in-5-minutes

Implementation of reinforcement learning approach to make a car learn to drive smoothly in minutes

https://towardsdatascience.com/learning-to-drive-smoothly-in-minutes-450a7cdb35f4

MIT License

284 stars 88 forks source link

Collecting Images For VAE [question] #31

Closed KarlXing closed 3 years ago

KarlXing commented 3 years ago

Hi, thank you for this interesting work. I'd like to collect images for the VAE training instead of using the pretrained VAE model. May I know how many samples did you collect for each level? Are there any suggestions about the manual control during data collection? I appreciate it for any advice on that. Thanks :)

araffin commented 3 years ago

Hello, The number of images does not really matter, what matter is that you cover most of the observation distribution, so it's good to do not only expert driving ;) Best is to check visually the trained autoencoder and to wait until convergence.

If you want number, I think I had something working with ~5k images or less (so ~5-10 minutes driving).

KarlXing commented 3 years ago

That makes sense. Thanks!