search

the-full-stack / fsdl-text-recognizer-2021-labs

Complete deep learning project developed in Full Stack Deep Learning, Spring 2021
https://bit.ly/berkeleyfsdl
MIT License
452 stars 281 forks source link

Lab 3: Can't achieve 90% accuracy #31

Closed oOXpycTOo closed 3 years ago

oOXpycTOo commented 3 years ago

Hi, thanks for the course. I faced a problem, that I can't achieve 90% accuracy on EMNISTLines dataset, no matter what I try to do. I even tried to simply execute the initial architecture without any changes, but I get no more than 82% of accuracy. For other experiments in this lab, the accuracy is even worse than it's written in the corresponding readme. Could you please, point out, what I'm doing wrong? I simply clone the repo and executed this line: python training/run_experiment.py --max_epochs=10 --gpus=1 --num_workers=4 --data_class=EMNISTLines --min_overlap=0 --max_overlap=0 --model_class=LineCNNSimple --window_width=28 --window_stride=28

just-eoghan commented 3 years ago

Did you ever try increasing the max_epochs?

oOXpycTOo commented 3 years ago

I found the solution and it was just to change the dimension in softmax function for accuracy prediction, as described here: https://github.com/full-stack-deep-learning/fsdl-text-recognizer-2021-labs/issues/27#issue-869823325