Getting low accuracy on the pokedex challenge

#1

I am getting only 82% accuracy so plz suggest me the ways to improve code and plz tell me how to improve model layers here is my code:- https://github.com/agarwalyash02/machine-learning/blob/master/pokedex_training_model_from_scratch.ipynb

#2

hey @Jan19LPN0013 ,
Accuracy in deep learning depends upon various factors like ,
what layers did you used , in what manner , with how many number of nodes , activation function , learning rate and many more things .
To give some suggestions for the above notebook :

  1. You can more augmentation techniques , to create more subsets of dataset.
  2. If classes are imbalanced , then you can assign custom class weights to your model for better performance.
  3. Instead of using directly the flatten layer you can use GlobalMaxPooling layer.
  4. Try using transfer learing , as those models are pre trained on really good and big dataset , hence they are more generalized and perform really good then our custom models.
    But you need to pay attention on your feed forward network while using transfer learning , as how you preprocess the information generated by these pre trained models is the actual task to perform.
  5. Try using Custom Callbacks , for scheduling learning rate. It really helps very much in making the model training a lot.
  6. When working on classification technique , always check the classification report , to see how the model perform with respect to each and individual class in dataset.
    Different models can have different performance measures , so if you have 3 or 4 different models performing differently on different classes , then you can ensemble there results, so that the final results are more generalized to the respective output classes.

I hope this would have helped you
Thank You and Happy creating Deep Learning Models :slightly_smiling_face: .

#3

I have added GlobalMaxPooling layer previously but it decreases the accuracy
and how to perform point 6 plz explain
and can you tell me what changes I have to make in layers as I don’t which will come and when in layers I have added

#4

OKAY

search about classification report sklearn.
you just need to provide true labels , and your predicted labels to check how is your model performing.
As accuracy is not just a single metric to check classification results.

adjustments to these layers are a bit to understand first.
what does a layer do ? Whether it is useful or not ? which activation layer to use ?
these are some questions you need to search first and understand.
if still there is some confusion in them , you can ask it. After this only i will be able to reply you.

#5

I am only asking check my layers of cnn model and give suggestion how to accurately use the layers or I have used layers properly or not

#6

hey @Jan19LPN0013 ,
there is not perfect rule of how to use the layer correctly, you just need to practice and check which performs much better.
but a bit suggestion.

  1. after flatten layer , you have reduced the number of nodes very heavily , from 18432 to directly 64.
    Try with 18432 to 1024 to 128 to 10 and try with leakyrelu activation between these layers.
  2. You can try taking higher kernel size in initial convolutional layers and then reducing it further.
  3. User Learning Rate scheduler , it will in much proper training.
  4. You can use more augmentations.

I hope these will help.

#7

ohk i will implement this change get back to you thankyou

#8

yeah sure. No problem.