Remove feature but increase accuracy

I am new to PyTorch and have tried different methodl to increase the accuracy of my model (e.g. dropoff, data augmentation, normalization, batch normalization). However, most of the dones’t help much. Instead, when I remove some of the feature in my model (e.g. from 10 features to 2 feastures), the accuracy of the model increase a lot. Wondering if anyone got similar experience when developing / training the model. Appreicate if you can share your view / experience.