Some general questions


#1

Hello rajendra & others,

I’ve made some nice progress with the help of DLS. Meanwhile, some questions came up which I have compiled below, which you can hopefully answer without much trouble.

  1. Which variable is plotted on the X-axis in the results tab plots? I figured it is most likely ‘batch’ for training plots and ‘epoch’ for validation plots? Is this right?

  2. Should I randomize my train.csv, or is the ‘shuffle data’ function in the data tab enough?

  3. Do you have a hint on how to deal with unbalanced data? (I have one class, where samples are very hard to make, and for the second class I can make virtually unlimited samples).

  4. Does DLS reduce learning rate automatically? Is there a way for reducing LR according to loss in DLS?

  5. How can I decide whether to save weights on lowest loss or best accuracy?

  6. With my dataset (2 classes, 11k samples) my results usually have quite extreme spikes every time a new epoch starts. is this normal behaviour? (see image below).

  7. In general, does this look like a sucessfull training run? Can I be sure it isn’t overfitting?

  8. This was archived with inception preprocessing and a model created by autoML function. Does it make sense to try to use an actual inception or resNet layer inside the model, like this now?

  9. I heard that DLS is not going to get another free version, instead there will be a professional version? Is that correct?

  10. Is there an official documentation on DLS?

  11. How can I stay up to date on DLS development? Blog on deep cognition seems dead to me (last update there was for DLS 2.2.0).

Thanks you for your answers.

As a feature request: I would like to be able to pan and zoom all four plots simultaneously in the result tab, instead of just one at a time.


#2

It would be very helpful to me, if someone had an answer to at least some questions above. Thank you.


#3

Hi

  1. Yes you are right
  2. Shuffle data function is enough
  3. Yes there are techniques that will help you to deal with unbalanced dataset.
    you can refer to this link https://www.analyticsvidhya.com/blog/2017/03/imbalanced-classification-problem/
    You can’t deal with unbalanced data using DLS GUI
  4. DLS reduces learning rate acc to the hyperparameters and optimizer you have selected. You can change the LR from the hyper parameter tab.
  5. It’s depend on the model to model and the dataset. You can train your model and analyze the result based on which you can select. Like If your model is not over fitting then you can select the best acc or if you acc graph is not constant i.e. you can see accuracy getting exp higher at some point but your loss is reducing then you can select the lowest loss.
  6. Extreme spike may be due to unbalanced data or random shuffling and yeah i do also get spikes but as I can see in your case spikes are little higher than expectations but its alright.
  7. No i don’t think so that your model is over fitting bcz there is not that much diff btw training and validation acc and your loss is also low. If you want to check then split your data into train-val-test data and then run the inference on the test dataset and download the CSV to check the accuracy in test data.
  8. I didn’t get this point.
  9. No DLS Desktop version is free and it will get updated with time.
  10. Yes you can find it on the website and moreover for usecase you can find articles on Medium on diff probs solved using DLS.
  11. DLS latest version is DLS 2.5.0. Whenever there is an update DeepCognition sends email to all the registered users.

Thanks


#4

Thank you this helps me a lot.