TRAINING...

[SaiSurya9999]

-- I have few doubts in training all set hardware is assembled and at last came to the training phase.
-- I have uploaded an image where you can see my training track.
-- Should we do the total training session in one execution of collect_training.py..??
-- For each execution there is an npz file created if we use one npz file the previous npz file which we trained is dumped..??
-- Could you give me your 300 images npz file or collection of those so that i can use them and train my model fast with more data.
-- When i place my npz file in model_test pynb file then i get some size error i have uploaded that below.

[hamuchiwa]

@SaiSurya9999 Make sure to slow down the car before collecting data, if it runs too fast, the video streaming couldn't keep up.
You can collect multiple times, in training it will loads all files in training_data folder.
This is an old project, I don't even have the car anymore.
Check the input size, in notebook it's 240320, your image may be 120320.

[SaiSurya9999]

How to increase the validation accuracy..??
I have uploaded the image of my track where i have drove it from both sides many times approx 350 frames were recorded..But there is no increase in validation accuracy of training..
It is biased to forward please help me @hamuchiwa

[SaiSurya9999]

I have slowed the vehicle so no problem about speed or stream rate cope up..
Please do help me in time @hamuchiwa i have to complete this project by the end of this month..
Should we delete npz files after training the model i mean they are being stored in folder even after training. so next session of training file will be created and when we train the older files will be combined and trained again...!!

[hamuchiwa]

@SaiSurya9999 You can collect multiple times and then train together. If you are satisfied what you have collected, yes, then you need to delete that file before training again.

[Siraj95Abudaber]

Hi
I have the same problem i can't increase the accuracy of my training i tried a lot but i can't get above 70%
could you please upload your .npz so i can try it.
Thanx @hamuchiwa

[hamuchiwa]

@Sirajabudabe Since it's duplicated with another thread of yours, I'll close the other one. Can you upload your data?

[SaiSurya9999]

How to reset the neural network model as a baby to train it again with nice balanced data..??
Thank you @hamuchiwa..!!

[hamuchiwa]

@SaiSurya9999 All model related stuff is in model.py, you can create your own architecture there. Every time you run model_training.py, it will overwrite the old one with the same file name nn_model.xml.

[SaiSurya9999]

@hamuchiwa What is your training accuracy when you successfully made the car play with corners..??
(I mean car taking successful turns at right point.)

[Siraj95Abudaber]

@hamuchiwa https://drive.google.com/drive/folders/1D2WbSvCy1LnNEoIAA69qk_IvJFVrXvgO
or if you can send me your E-mail!

[hamuchiwa]

@Sirajabudabe Can you remove the permission, I'll let you know after downloading, then you can delete the file.

[Siraj95Abudaber]

https://drive.google.com/file/d/1kcK2YEr8PdTN50MhcPSfWqxI0GHSEC-a/view?usp=sharing
could you please run it on your neural network and tell me what you got @hamuchiwa

[hamuchiwa]

@Sirajabudabe I have uploaded a jupyter notebook using keras. After a quick run, as you can see from the plots, the model is overfitting, I'll leave it to you for tuning. Good luck!
Untitled.ipynb.zip

[Siraj95Abudaber]

@hamuchiwa sorry but am not really familiar to deep learning and machine learning techniques , how can i run this model
is the problem in cv2 neural network ?
sorry again for my many questions