I am facing a strange issue while doing the Hindi Vowel Consonant problem. I have used a Resnet model, which gives the following train and val accuracies
Epoch: 5/5, Train Vowel acc: 63.14, Train Consonant acc: 71.97, Train Total acc: 55.08
Val Vowel acc : 60.70, Val Consonant acc : 70.40, Val Total acc : 52.70
Then I generate the output from the test_loader, parse the labels and get an output file as required. First few of 10000 rows as follows:
Then I upload this output in the Submit prediction section of the kaggle contest.
And I get a public score of 0.016… how is it possible that a model that generates 50% accurate scores (Vowel + Consonant), gets almost nothing right on the test ?
I have checked and double checked the code… any ideas would help ?