Closed leonardoaraujosantos closed 7 years ago
I haven't tested it very much without the scaling operation, although it likely wouldn't make a massive difference in the result. The main bottlenecks at this point are training data. I did no data augmentation or preprocessing, though I really should have. This isn't a major project, so I didn't do too much work on it.
Absolutely phenomenal job with the simulation! It looks great. The simulation isn't performing well because there isn't a large amount of data on high curvature roads. You'll notice in the simulation that the car drives quite well on low curvature parts of the road (but not straight sections of course). This can be easily solved by a quick drive around the residential areas of my city though to get more data.
I might consider working more on this project now that you've made such a marvelous simulation tool! Let me know if you have any more questions!
-Sully Chen
On Tue, May 9, 2017 at 5:55 PM, Leonardo Araujo dos Santos < notifications@github.com> wrote:
Hi @SullyChen https://github.com/SullyChen did you test training your model without the, scaling operation (atan)? https://github.com/SullyChen/Autopilot-TensorFlow/blob/master/model.py#L88
Also did you do some sort of preprocessing to do image augmentation and to balance the dataset? I'm asking this because most of the time the data has no steering.
I've created another driving example based on your code (Including a game and a tcp/ip comm class) https://github.com/leonardoaraujosantos/DLMatFramework/tree/master/ virtual/tensorDriver
Here is a video of my current performance https://www.youtube.com/watch?v=PT8oToyV85w
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/SullyChen/Autopilot-TensorFlow/issues/20, or mute the thread https://github.com/notifications/unsubscribe-auth/ALMHR73KdCaPFYgdsH5IqtnuPyqPM6xmks5r4QsOgaJpZM4NWDnH .
Cool! Btw if you want to play (Create your own dataset) I'm trying now to play with the TFRecord and also try to balance the dataset by repeating steering angles that are not 0 Also check it out on drive_on_game.py that I did not need to filter the results: https://github.com/leonardoaraujosantos/DLMatFramework/blob/master/virtual/tensorDriver/drive_on_game.py#L94
Here is a python notebook showing how unbalance the dataset is right now: https://github.com/leonardoaraujosantos/DLMatFramework/blob/master/virtual/tensorDriver/ImproveDataset_LMDB.ipynb
Stuff related to the game (Code and builds) https://github.com/leonardoaraujosantos/DLMatFramework/releases/download/v0.2/WindowsBuild.zip https://github.com/leonardoaraujosantos/DLMatFramework/releases/download/v0.2/LinuxBuild.zip Source code of the game https://github.com/leonardoaraujosantos/DLMatFramework/releases/download/v0.2/UnityProject.zip
Thanks so much!
-Sully Chen
On Wed, May 10, 2017 at 3:16 AM, Leonardo Araujo dos Santos < notifications@github.com> wrote:
Cool! Btw if you want to play (Create your own dataset) I'm trying now to play with the TFRecord and also try to balance the dataset by repeating steering angles that are not 0 Also check it out on drive_on_game.py that I did not need to filter the results: https://github.com/leonardoaraujosantos/DLMatFramework/blob/master/ virtual/tensorDriver/drive_on_game.py#L94
Here is a python notebook showing how unbalance the dataset is right now: https://github.com/leonardoaraujosantos/DLMatFramework/blob/master/ virtual/tensorDriver/ImproveDataset_LMDB.ipynb
Stuff related to the game (Code and builds) https://github.com/leonardoaraujosantos/DLMatFramework/releases/ download/v0.2/WindowsBuild.zip https://github.com/leonardoaraujosantos/DLMatFramework/releases/ download/v0.2/LinuxBuild.zip Source code of the game https://github.com/leonardoaraujosantos/DLMatFramework/releases/ download/v0.2/UnityProject.zip
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/SullyChen/Autopilot-TensorFlow/issues/20#issuecomment-300439045, or mute the thread https://github.com/notifications/unsubscribe-auth/ALMHR1LvH7LVJr61vmw1AfR4mDZBOaE1ks5r4Y5ngaJpZM4NWDnH .
Sure no problem, I've learn a lot about tensorflow from your code, btw I balanced the dataset, just by repeating the samples that were not to frequent and it improved a lot (I also notice no difference taking out the atan scaling part) https://github.com/leonardoaraujosantos/DLMatFramework/blob/master/virtual/tensorDriver/ImproveDataset_LMDB.ipynb
Keep taking a look on the code because it seems that using feed_dict to train is also not the best way to use tensorflow anymore. https://blog.metaflow.fr/tensorflow-how-to-optimise-your-input-pipeline-with-queues-and-multi-threading-e7c3874157e0
Also all the stuff that I'm learning on "Deep learning" I'm adding on this book https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/
Wow! Looks great.
-Sully Chen
On Wed, May 10, 2017 at 1:26 PM, Leonardo Araujo dos Santos < notifications@github.com> wrote:
Sure no problem, I've learn a lot about tensorflow from your code, btw I balanced the dataset, just by repeating the samples that were not to frequent and it improved a lot (I also notice no difference taking out the atan scaling part) https://github.com/leonardoaraujosantos/DLMatFramework/blob/master/ virtual/tensorDriver/ImproveDataset_LMDB.ipynb
Keep taking a look on the code because it seems that using feed_dict to train is also not the best way to use tensorflow anymore. https://blog.metaflow.fr/tensorflow-how-to-optimise- your-input-pipeline-with-queues-and-multi-threading-e7c3874157e0
Also all the stuff that I'm learning on "Deep learning" I'm adding on this book https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/SullyChen/Autopilot-TensorFlow/issues/20#issuecomment-300602417, or mute the thread https://github.com/notifications/unsubscribe-auth/ALMHR9WK6jNiZ3W1L7wnkO_T86nwimcgks5r4h2DgaJpZM4NWDnH .
Hi @SullyChen did you test training your model without the, scaling operation (atan)? https://github.com/SullyChen/Autopilot-TensorFlow/blob/master/model.py#L88
Also did you do some sort of preprocessing to do image augmentation and to balance the dataset? I'm asking this because most of the time the data has no steering.
I've created another driving example based on your code (Including a game and a tcp/ip comm class) https://github.com/leonardoaraujosantos/DLMatFramework/tree/master/virtual/tensorDriver
Here is a video of my current performance https://www.youtube.com/watch?v=PT8oToyV85w