I am trying to enter the state farm competition. I am using the VGG16 model with batch norm.
I train the model, but my model always predicts the first class. I suspect that it may be because the training samples are not shuffled, and thus the solver always goes for the first one.
However, I see this is also the case in the lesson 1 cats and dogs, these are also not shuffled and the algorithm solves just fine.
Do I need to shuffle? Is the shuffle taken care of in vgg.fit?
I was using the get_data function but I saw it calls to a get_batches function that has shuffle=False. I changed that in utils.py and the samples are still not shuffling
Hi @josealb - It would be great if you can post some code snippet, to understand more about the issue. But in general, you may run into validation accuracy, if your data is not good and not shuffled. How did you create your train and valid data sets?
Thanks, I already solved the issue, it was indeed the shuffling, it didn’t work because I was taking the data from somwhere else in the code
The weird thing is the shuffle=True option doesnt show when you list the items before fitting, but works when you call the fit function.
Anyways it works now and lesson learned with get_batches always use shuffle=True
