Retraining of the model on new dataset about patchtst

[vivekugale9887]

I have a model which is trained on stock data of one company and I want to retrain that model again on stock data of another company? Is it possible using PatchTST & if yes, please guide me how to do it.

[ikvision]

@vivekugale9887 I adopted PatchTST on my dataset, most of the work is adding a dataloader.

1. You would need to create a new class of type Dataset.
2. Implement a read_data method like

   PatchTST/PatchTST_supervised/data_provider/data_loader.py

   Line 221 in 2c68d0d

   def __read_data__(self):

3. implement a getitem method like
   

   PatchTST/PatchTST_supervised/data_provider/data_loader.py

   Line 82 in 2c68d0d

   def __getitem__(self, index):

4. Add your new class to the data factory

   PatchTST/PatchTST_supervised/data_provider/data_factory.py

   Line 4 in 2c68d0d

   data_dict = {

[yuqinie98]

Hi @ikvision Thanks very much for answering this in details! Besides I would like to address a few more thoughts:

1. Make your dataset csv in the same format and location of the sample datasets as described in readme file.
2. For your setting I supposed that it is similar to transfer learning which is in table 5 of the paper. PatchTST can do it, and what you need to do is a 2-step process: self-supervised training on one dataset and fine-tune on the other one.
3. Although it is doable, we do not recommend going with financial dataset at this time point because of the reason we mentioned in appendix A.1.1.

[vivekugale9887]

hi @ikvision , thank you for the quick reply. I am yet to try what you suggested to me. While searching for the incremental learning, I came across this notebook and am using the way it has mentioned the incremental learning.

https://github.com/timeseriesAI/tsai/blob/main/tutorial_nbs/14_Inference_Partial_Fit_and_Fine_Tune.ipynb