Tim Rosenflanz
Tim Rosenflanz
From: https://github.com/keras-team/keras/issues/10768 by @hadaev8 Tokenizer will fit/transform the string into chars if a string is provided to `fit_on_texts`/`text_to_sequences` methods regardless of char_level setting. This is happening because the method expects...
When validation dataloader has more than one batch in it, the `output.y` of `model.predict` has an incorrect shape due to the incorrect concatenation axis. E.g. if I have 200 items...
**Is your feature request related to a current problem? Please describe.** Current torch model checkpointing logic is quite rigid. It only allows to track loss rather than other metrics and...
Using your implementation on the data that has been transposed previously causes a `RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans...
### Describe the bug Auto Arima keeps all model history during training/solving even when `return_valid_fits` is set to False. This causes memory consumption growth. Unless `return_valid_fits` is set to `True`,...
**Is your feature request related to a current problem? Please describe.** TSMixer Model with num_blocks higher than 4 aren't training well. It is somewhat nebulous to pinpoint but higher number...
**Describe the bug** TSMixer implementation overall uses `LayerNorm` and doesn't give a `BatchNorm` option. In the paper it was shortly mentioned that the `LayerNorm` gave better results on smaller batches....
**Is your feature request related to a current problem? Please describe.** Parallel timeseries.from_group_dataframe is currently passing sub_df around per group which can be slow when there are lot of groups...
Apologies if this is covered by some other feature request/issue already: Is there a way to nest global forecasting models that allows to properly train them with covariates? E.g. I...