how to use ts_learner + TSUnwindowedDataset properly for TS classification?

[bellerofonte]

Hi everyone!

First of all I have to notice the huge amount of features this project implements. That is really great! But on another hand - it is sometimes unclear how to use those features properly.
I am trying to solve standard classification problem and facing to following issue:

I've installed tsai 0.3.0 in separate env with python 3.9.10 under miniforge's conda.
I have a pandas dataframe with ~130 features and ~500K samples. I've labeled a target for each sample which is either 0 or 1.
So I'm doing this:

X, y = df2xy(df_train_val, target_col='target', steps_in_rows=False, to3d=False)
X.shape, y.shape
# ---> ((482659, 134), (482659,))
# ~70K samples left for test dataset

np.unique(y)
# ---> array([0, 1])

next, I have to create TSUnwindowedDataset to be able to fit target using sliding window

dset = TSUnwindowedDataset(X, y, y_func=y_func, window_size=10, stride=1, drop_start=0, drop_end=0, seq_first=True)
len(dset)
# ---> 482650 # correct length for window = 10

y_func is something to notice here. I as far a I understand, then using sliding window, y (target) also becomes a slide, so I have to transform it into one value (last value of the slide)

def y_func(o): 
    return o[:,-1]

next, I make splits and datasets:

splits = TimeSplitter(valid_size=TEST_FRAC)(range_of(len(dset)))
splits
# ---> ((#410252) [0,1,2,3,4,5,6,7,8,9...],
#       (#72398) [410252,410253,410254,410255,410256,410257,410258,410259,410260,410261...])

dsets = TSUnwindowedDatasets(dset, splits=splits)
len(dsets.train), len(dsets.valid)
# ---> (410252, 72398)

at this moment everything seems to be ok.

finally, I create loader and learner:

batch_tfms = TSNormalize(by_sample=True, by_var=True, range=(0,1))
dls = TSDataLoaders.from_dsets(dsets.train, dsets.valid, bs=32, batch_tfms=batch_tfms, device='cpu')
clf2 = ts_learner(dls, arch=TST, loss_func=BCELossFlat(), metrics=RocAucBinary(), cbs=ShowGraph())
clf2.fit_one_cycle(10, 3e-4)
clf2.export("save/clf2.pkl") 

and i fails at fit_one_cycle method:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Input In [131], in <module>
      1 clf2 = ts_learner(dls, arch=TST, loss_func=BCELossFlat(), metrics=RocAucBinary(), cbs=ShowGraph())
----> 2 clf2.fit_one_cycle(10, 3e-4)
      3 clf2.export("save/clf2.pkl")

...

File ~/.local/lib/python3.9/site-packages/torch/nn/functional.py:2915, in binary_cross_entropy(input, target, weight, size_average, reduce, reduction)
   2912     new_size = _infer_size(target.size(), weight.size())
   2913     weight = weight.expand(new_size)
-> 2915 return torch._C._nn.binary_cross_entropy(input, target, weight, reduction_enum)

RuntimeError: all elements of input should be between 0 and 1

I have also tried TSStandardize batch transformer before using TSNormalize but got the same error.
So can anybody explain what am I doing wrong here?