search

JLREx / PAtt-Lite

Official implementation for PAtt-Lite: Lightweight Patch and Attention MobileNet for Challenging Facial Expression Recognition
MIT License
33 stars 3 forks source link

Lower Accuracy on RAF-DB Dataset #10

Open cyinen opened 5 months ago

cyinen commented 5 months ago

Hi， I attempted to replicate your results on the RAF-DB dataset using your notebook but achieved only about 88.07% accuracy, despite making no changes except for dataset replacement. This was consistent across tries, including on Kaggle with TensorBoard 2.12.

Could you offer any advice or adjustments to reach the expected accuracy?

Thanks for your work and help in advance.

Here is my notebook. result

JLREx commented 5 months ago

@cyinen Hi, have you tried enabling the sample_resizing? It could be due to this as the samples might not be using the full expression from the MobileNet with it still based on 224x224.

cyinen commented 5 months ago

Sorry, what's sample_resizing means? I have resized the samples to 224x224 before input to the network. Can you provide more details? Thanks!

kalpa61 commented 2 weeks ago

@JLREx Hello, my accuracy rate on the RAF-DB data set is also 88%, I don't know what the problem is, no code changes have been made

KrystianZielinski commented 2 days ago

try experimenting with different seed values for shuffling ur train dataset and see if the results vary. It was the case for me.