search

NVIDIA-AI-IOT / jetracer

An autonomous AI racecar using NVIDIA Jetson Nano
MIT License
1.07k stars 321 forks source link

unable to train and run #22

Closed gbr1 closed 2 years ago

gbr1 commented 5 years ago

Hi! Just builded my Jetracer (latrax 1:18 -> youtube video ) I created a setup track: photo_2019-09-20_12-45-42
After following instructions: 1) basic motion works correctly; 2) interactive regression gives me always right-bottom blue point during "evaluation"; 3) road following: car always go right.

I always tried with a simple path: photo_2019-09-20_17-02-19 And car goes right.

Installed camera is: IMX219 with wide angle.

How do I choice correctly points? Which is correct sequence (create images -> train button -> evaluate -> save)? Is my track correct or are necessary other stuffs such as central line and uniform colours?

Thank you in advance!

jaybdub commented 5 years ago

Hi gbr1,

Thanks for reaching out and sharing your video! Very well done!

The track you created looks good to me, it should certainly be able to detect that. Hopefully we can get you up and running soon.

To answer your first question

How do I correctly choose points?

  1. Imagine the ideal path the robot would follow
  2. Click as far along that ideal path such that the robot could travel in a straight line towards the point without crossing the boundary

Could you provide the following information

  1. How many data samples did you collect / how was car placed? (data variance is very important)
  2. How many epochs did you train for? (it usually takes many, I'd guess at least 15+ to converge)

Please let me know if you have any questions or if anything is unclear.

Best, John

gbr1 commented 5 years ago

Hi @jaybdub! I took about 200 photos and 3 epochs maybe that was the problem. Then I moved on donkeycar software so I didn't try more.

Thank you.

RyanC1681 commented 4 years ago

Hi, John: I took >200 photos but epoch always = 0 and car always do circle, not steering straight. Could you give me a correct training sequence (live -> train button -> click left image -> count# inceasing -> evaluate -> save model)?
How to set another epoch or find out which point need more training/images?

Thanks

jaybdub commented 4 years ago

Hi RyanC1681,

The work flow is generally to

  1. Collect images
  2. Train for several epochs until loss converges
  3. Run live to visualization how well the network is performing
  4. Repeat 1->3 until you are happy with the performance, at this point save the model

Please let me know if this helps.

Best, John

ChrisW0rk commented 4 years ago

Hi jaybdub,

I know this is a bit old, but perhaps you could help me out with my current situation if you have time.

I can get to the point where I'm consistently training a model that performs well after 100+ images and 15 or so epochs. It tracks movements around the track with a high degree of accuracy.

I essentially do exactly the work flow you described above: collect images, train for several epochs, and run it live to make sure the model is performing well. Rinse and repeat, then save the model when I'm satisfied.

However, when I attempt to run it with the Road Following notebook, it just turns in a continuous circle to the right. It does seem to adjust the degree of turning at times, which indicates that some commands seem to be going through.

Do you have any idea why this may be happening, considering the perceived accurate performance running the Interactive Regression notebook?

Thanks!