Chapter 13 - RuntimeError: mat1 and mat2 shapes cannot be multiplied (1536x256 and 128x1) about modern-time-series-forecasting-with-python HOT 7 CLOSED

verified, pls close the issue

from modern-time-series-forecasting-with-python.

I suspect this is because of a bug fix I pushed recently... I'll take a look at this and get back.. Also, if you have faced some other problems (and even better solved it), please post it here so that others can benefit from that..

from modern-time-series-forecasting-with-python.

I suspect this is because of a bug fix I pushed recently... I'll take a look at this and get back.. Also, if you have faced some other problems (and even better solved it), please post it here so that others can benefit from that..

Oh thank you for your answer, I will try to remember most of the problems I faced which are the following:

• Need to add libraries to .yml file (tsfel, pymannkendall, window-ops, pytorch_tabular)

• Chapter02 01 -Pandas Refresher ... code block line 37 and 38:
  The data downloaded from "https://www.data.act.gov.au/resource/94a5-zqnn.csv" only contains the last month so you need to change the date values in line 38: (for March 2023)

df = df.loc["2023-03-01 7:00":"2023-03-14 16:00", "pm2_5_1_hr"].to_frame()

• In case you are running pytorch without gpu support:

Found all the TrainerConfig(). ... and change the following lines:
...
auto_select_gpus=True,
gpus=-1

change it for:
#auto_select_gpus=True,
accelerator = 'cpu',
gpus=None #-1

from modern-time-series-forecasting-with-python.

Same here. Pls help

I also confirm @nonpa-gt discovery with commenting gpus=-1 line.

from modern-time-series-forecasting-with-python.

This is on my radar... I'm currently on vacation, but plan to push a fix to this as soon as I get back!

from modern-time-series-forecasting-with-python.

I think the issue is that the hidden state size should be doubled when the bidirectional flag is turned on. Here is the change to /src/dl/models.py around line 210:

    rnn_output_size = self.hparams.hidden_size
    if self.hparams.bidirectional:
        rnn_output_size *= 2
    self.fc = nn.Linear(rnn_output_size, 1)
    # self.fc = nn.Linear(self.hparams.hidden_size, 1) # Removed

from modern-time-series-forecasting-with-python.

#21 Should have fixed this. It would be nice if someone can check and confirm

from modern-time-series-forecasting-with-python.