Problem with Label Encoding

[anandh51295]

Thanks for this work i read you article on the nano nets site.
but i have this one doubt:
def onehot(idx, num_classes):
"""
1-hot encoding.
"""
encoding = np.zeros(num_classes)
encoding[1] = 1
return encoding

here you only set 1 to 1 position that only we got 1.0 accuracy but we need to set 1 based on the idx then we only get accuracy around 0.60. there is any way to improve the accuracy of the model

[adriansarno]

Thank you for your ticket Anand, I will address the first problem (onehot) in my next pull request.
The second problem (improve model accuracy) I will try to address in a future PR.
The validation accuracy (val_acc) with the fixed onehot is 0.86, so it is still reasonable. Please see last lines of the training log, pasted below:

Epoch: 0061 train_loss= 0.03845 train_acc= 0.86364 val_loss= 0.03895 val_acc= 0.85620 time= 0.61268 Validation cost is not improving. Early stopping... Optimization Finished! Test set results: cost= 0.03870 accuracy= 0.85879 time= 0.37214

[MotiBaadror]

Accuracy is calculating np.argmax(labels) which will give all values as [1,1,1,1,...] . Since we have encoded labels [[0,,4,0,0,0,],[0,3,0,0,0] ...]. So the model is learning to get higher probability for the first position. When I test the model np.argmax(pred[test_mask],1) will give 1000+ values as 1 and some rest are zerop out of 1200+.