Model Overfiting with LSTM layers

General Discussion
,
#1

I am trying to train a model over human skeleton data and was able to achieve good accuracy overtraining but the validation loss reaches a point and starts to increase again. Model validation accuracy doesn’t decrease over time. Clearly, it overfitting and I totally understand. For reducing this I tried most of the techniques but was not able to decrease the validation_loss. I had tried dropouts, reducing model capacity, and adding loss for layers but no luck.
Log graph would be seen below

Any ideas to improve the model??

logs
logs1920×953 81.1 KB

#3

Have you tried to augment the dataset?

#4

Yap, I have tried that too. But I fell for this model parameter it’s more data. Maybe I could be wrong.

#5

Have you tried to overfit also the train+validation?

#6

No when i did data augmentation, train seems to be fine but the valdication loss only increases back

#7

Are you handling a 2d or 3d pose dataset?

#8

its 3d pose but it has been reshaped to 75 (i.e. 25 key points * 3)

#9

I don’t know your dataset but If you cannot collect more train data to cover your validation distribtuion you can try with some interesting augmentation approach like:

https://arxiv.org/abs/2105.02465

#10

I am using the NTU-RGBD dataset for training. According to your idea what should be validation distribution. My dataset size is around 18000 samples and split 80:10:10. Also model parameters is around 210,864.

#12

Are you traning on NTU-RGBD and evaluating on your own custom dataset?

#13

I am using NTU-RGBD for both training, and validation.

#14

Was in your graph the loss/accurancy on the Action recognition or on the keypoints?

#15

it was on action recognition because i have considered 20 classes for prediction

#16

Have you corrently sampled/balanced all the classes in training set?

#17

yes, I have considered while preparing dataset

#18

Have you tried to build the confusion Matrix or the classification error for each class to check how It is distributed?

#20

This is data distribution, numbers respond to classes.
Counter({14: 850, 16: 849, 9: 848, 18: 848, 5: 848, 4: 847, 6: 847, 17: 846, 1: 845, 19: 845, 3: 845, 15: 845, 12: 844, 8: 844, 10: 844, 2: 843, 0: 841, 11: 840, 13: 837, 7: 834})

#21

Yes but I meant how the validation error is distributed over classes.

Before debugging your custom model have you tried to reproduce approximate results with any well known model on this dataset?

#23

how can i get validation error over distributed classes. I mean how to visualise loss over classes.

#24

You can play the confusion matrix preparing the validation GT label and predictions