Hello :) Today is Day 22!
A quick summary of today:
- Card image classifier using neural networks on Kaggle
I learned about CNN during Course 1, 2 of the TensorFlow developer certificate preparation course and wanted to apply it to new data today. I searched Kaggle and found the following card image dataset.
Following the course, first I used ImageDataGenerator
Then made the model
Wrote the below to compare train and valid results
And finally, evaluated the model on unseen data
The results… not that good haha but it was very good practice overall.
Various models were constructed to obtain good performance, by changing layers, learning rates, data augmentation variables, filters of layers… Still, more than 65% accuracy was not obtained with the valid dataset. The loss of some models increased significantly after 5-10 epochs, but the loss of any model did not reach 50% or more accuracy after 100 epochs.
Below are pics of some of the models that I ran.
- model 1
- model 2
-
model 3
-
model 4
- model 5
- model 6
stopped the training
- model 7
stopped the training
- model 8
- model 9
- model 10
- model 11
- model 12
I don’t include all the information in the model picture, it’s Model 1 - base model, and then I just changed the variables. It took quite a while to train. It took about an hour per 20 epochs.
Tomorrow I will try to make some kind of NLP model.
That is all for today!
See you tomorrow :)