Saving models with TSUnwindowedDataset Dataloaders: AttributeError: 'TSUnwindowedDataset' object has no attribute 'new_empty'

[GardarGardarsson]

Having successfully trained a regression model on a large datafile, that required me to use the TSUnwindowedDataset to circumvent memory and storage requirements, upon trying to save the model I received the following error:

~/anaconda3/envs/TSAI/lib/python3.9/site-packages/fastai/data/core.py in <listcomp>(.0)
    144     def __getitem__(self, i): return self.loaders[i]
    145     def new_empty(self):
--> 146         loaders = [dl.new(dl.dataset.new_empty()) for dl in self.loaders]
    147         return type(self)(*loaders, path=self.path, device=self.device)
    148 

`AttributeError: 'TSUnwindowedDataset' object has no attribute 'new_empty'

Following are glimpses from my code.
I'm loading the regression features and targets as Pandas DataFrames and converting them as:

X, y = df2xy(data, sample_col=None, feat_col=None, data_cols=feature_names, target_col=target_names, to3d=False)

In setting up TSUnwindowedDataset() I use the following:

window_size = 24
stride      = 1
drop_start  = 0
drop_end    = 0
seq_first   = True
horizon     = 1

dset        = TSUnwindowedDataset(X, y, 
                                  y_func      = None       , 
                                  window_size = window_size, 
                                  stride      = stride     , 
                                  horizon     = horizon    ,
                                  drop_start  = drop_start , 
                                  drop_end    = drop_end   , 
                                  seq_first   = seq_first  )
len(dset)

Then get the splits, all the while watching out for not overreaching on the returned indices given the window size, and then pass the splits along with the dataset to TSUnwindowedDatasets() as per below:

splits = get_splits(y[:-window_size], n_splits=1, valid_size=.2, stratify=True, random_state=41, shuffle=True)
dsets  = TSUnwindowedDatasets(dset, splits=splits)

Standardising the data on a pre-calculated mean and std. deviation, and constructing the TSDataLoader:

batch_tfms = TSStandardize(mean=x_avg, std=x_std) # Standardise the batch on the pre-calculated mean and std.dev 
dls        = TSDataLoaders.from_dsets(dsets.train, dsets.valid, batch_tfms=batch_tfms, bs=[64])

I then find the learning rate as demonstrated in many of your tutorials, and when it comes to training I do:

learn = ts_learner(dls, 
                   InceptionTimePlus, 
                   metrics =  [mae ,
                               rmse], 
                   cbs     =  [ShowGraph(),
                               SaveModelCallback(monitor      = 'valid_loss', 
                                                 comp         = None        ,
                                                 min_delta    = 0.0         ,  
                                                 fname        = 'model'     , 
                                                 every_epoch  = False       , 
                                                 at_end       = False       , 
                                                 with_opt     = False       , 
                                                 reset_on_fit = True        ) ] )

learn.fit_one_cycle(10, lr_max=lr)

I then used the standard line to save the trained model as per:

PATH = Path('./models/Regression.pkl')
PATH.parent.mkdir(parents=True, exist_ok=True)
learn.export(PATH)

But ultimately got:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_3698523/2688106701.py in <module>
      1 PATH = Path('./models/Regression.pkl')
      2 PATH.parent.mkdir(parents=True, exist_ok=True)
----> 3 learn.export(PATH)

~/anaconda3/envs/TSAI/lib/python3.9/site-packages/fastai/learner.py in export(self, fname, pickle_module, pickle_protocol)
    367     self._end_cleanup()
    368     old_dbunch = self.dls
--> 369     self.dls = self.dls.new_empty()
    370     state = self.opt.state_dict() if self.opt is not None else None
    371     self.opt = None

~/anaconda3/envs/TSAI/lib/python3.9/site-packages/fastai/data/core.py in new_empty(self)
    144     def __getitem__(self, i): return self.loaders[i]
    145     def new_empty(self):
--> 146         loaders = [dl.new(dl.dataset.new_empty()) for dl in self.loaders]
    147         return type(self)(*loaders, path=self.path, device=self.device)
    148 

~/anaconda3/envs/TSAI/lib/python3.9/site-packages/fastai/data/core.py in <listcomp>(.0)
    144     def __getitem__(self, i): return self.loaders[i]
    145     def new_empty(self):
--> 146         loaders = [dl.new(dl.dataset.new_empty()) for dl in self.loaders]
    147         return type(self)(*loaders, path=self.path, device=self.device)
    148 

`AttributeError: 'TSUnwindowedDataset' object has no attribute 'new_empty'

[GardarGardarsson]

@oguiza any updates on this?

[oguiza]

Hi @GardarGardarsson,
I'm sorry I didn't provide an update before.
To fix this issue I'll need to update a core element of tsai. What I mean is that it's not a small change. Anyway, I think I'll have it fixed sometime next week. I'll provide an update as soon as it'd done.
Thanks for your patience!

[GardarGardarsson]

Thanks for the update and there's absolutely no rush on my end, take all the time you need.

[oguiza]

Hi @GardarGardarsson,
First of all, sorry for the long delay.
I've added a new_empty attribute to TSUnwindowedDataset. It'd be good if you can try it.
However, I think at some point I'll need to integrate the SlidingWindow functionality with TSUndindowedDataset to make the solution more flexible and robust. But I don't have the timing yet.

[oguiza]

I'll close this issue due to a lack of response. Please, reopen it if necessary.

[AbdullahJirjees]

Hello @GardarGardarsson
I am having exactly your issue, any idea how did you solve it?

[GardarGardarsson]

Hey @AbdullahJirjees, no unfortunately I have not revisited this following @oguiza 's update.
I hope you figure it out.

[oguiza]

I'll close this issue due to the lack of response.