Predictions using Pytorch neural network does not seem right

Part 1 (2019)
1

After going through the first 5 lessons of fast.ai I decided I will try to create a neural network using Pytorch for tabular data. I was able to create a NN with 3 layers and added a LogSoftmax layer at the end along with NLLLoss. But when I use the network to make predictions, the predictions does not seem right to me. They are of very high values and If I am not mistaken the softmax layer should return me values that add upto 1.
Kaggle Kernel Link : https://www.kaggle.com/jinudaniel/reducing-fatalities-nn-approach-with-pytorch
Please let me know If i doing something wrong.

2

Did you come right with this? This is my issue:
Mismatch between FastAI prediction and Pytorch prediction - PyTorch Forums

3

Did you preprocess your data in the same way fastai does? (Normalize based on the training data, label-encode the categories, etc)

This may also be some help if you’re using tabular pandas: Exporting `TabularPandas` for Inference (Intermediate) | walkwithfastai

1 Like
4

Thank you. Absolutely right- it was normalization on the tabulardataloader, which is done prior to the model in the fastai predict method.