Considering that LSTM are good for time series sequences, I have trained LSTM for multiclass classification task in Pytorch environment, several parameters of LSTM model were tuned including number of layers, no of cells, adding dropout, regularization, training at different batch sizes and different number of epochs. But still I am not getting the best performance, the maximum accuracy I could get on my validation dataset is 80 percent while the training accuracy of 92% at last epoch. Training accuracy also saturates after certain no of epochs and does not improve. (Loss curve attached)
Not sure what's wrong either underfitting, overfitting or my dataset have too much variance.
I will be very thankful if someone can give some guidance.
Dear Hamza,
I think you should use another optimizer different from used in your experiment eg.(SGD, Adagrad, Adadelta, RMSprop, Adam).
Dear Hamza Azam
These links might be an asset, have a look:
1. https://machinelearningmastery.com/how-to-develop-rnn-models-for-human-activity-recognition-time-series-classification/
2. https://www.tesisenred.net/bitstream/handle/10803/670346/2020_Tesis_Sansano%20Sansano_Emilio.pdf?sequence=1&isAllowed=y
3. https://arxiv.org/pdf/2105.01753.pdf
Kind Regards
Qamar Ul Islam
Dear Hamza Azam ,
First of all try to change your hyperparameters such as Optimization, then despite the fact if your data set is cleaned then it shows not good variation of validation among your training data, then look for split data as well,
finally you can use other methods as well for time data such as recurrent networks or etc.
Best,
Thankyou everyone for very helpful advice, cleaning dataset and increasing its size helped me to achieve better results.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training