unit8co / darts Goto Github PK

View Code? Open in Web Editor NEW

7.8K 7.8K 842.0 181.53 MB

A python library for user-friendly forecasting and anomaly detection on time series.

Home Page: https://unit8co.github.io/darts/

License: Apache License 2.0

Python 99.98% Dockerfile 0.01% Shell 0.01%

anomaly-detection data-science deep-learning forecasting machine-learning python time-series

darts's People

Contributors

Stargazers

Watchers

Forkers

endrjuskr guillaumeraille wookai arpyt zwbjtu123 hivewang millerwu2014 a3digit jiapengwei manucob13 mhdella duanshuaimin czong tchigher nicolizamacorrea rkayastha flobrec cxz greg-lep lucasaos52 sbarman25 idahopotato1 mayvid9 ifv kn78 melsr patrick-dd nethask atul7cloudyuga seirdy droxef martins0n gbaf rajatdas amit2016-17 jahokas polybius12 hamid701 valeman muttdata 79212 ikedaosushi seanahmad natnaelt mhaupt63 zakopuro cammyblanc lewisbakkero masabhathini justinwhatley sudhang karthikbadari leotafti t-kota shubhsoni d3mi4n kusumy charlie5dh stephanesbizzera stardust0513 beesitech amitgupta223 jcassiojr mvahit krivokuca wl935 mohamedmassaoudi jinzhaolu luckyyangrun surajitdb hmz3192 imwildstone mveerara hrzn camilaagw tremamiguel maciejdomagala 2017wxyzwxyz jzjosuerivera71 seb-afk gyani91 sergio2526 vishalbelsare aesmin laiki mpetros accusenine yk-amarly-20 sczhai smilingjimmy xfw pennfranc rnogales94 rib4n chetanmehra alexcombessie cycle13 sam93katoch zerounnet kjappelbaum

darts's Issues

Installation failed: MAX_TORCH_SEED_VALUE too high and specific versions for pytorch and fbprophet needed

Install darts on windows

Following the install guide leads to "ERROR: No matching distribution found for torch==1.5.1; extra == "all" (from u8darts[all])"
You need to specify the versions:
conda install -c conda-forge -c pytorch pip fbprophet==0.6 pytorch==1.5.1 cpuonly
when fitting the model in "TCN-examples.ipynb" following error pops up:

File "C:\Users\JB\miniconda3\envs\darts2\lib\site-packages\darts\utils\torch.py", line 62, in decorator
    manual_seed(self._random_instance.randint(0, high=MAX_TORCH_SEED_VALUE))
  File "mtrand.pyx", line 743, in numpy.random.mtrand.RandomState.randint
  File "_bounded_integers.pyx", line 1343, in numpy.random._bounded_integers._rand_int32
ValueError: high is out of bounds for int32

I change coding in "...\site-packages\darts\utils\torch.py" in line 17:

#MAX_TORCH_SEED_VALUE = (1 << 63) - 1
MAX_TORCH_SEED_VALUE = 2147483647

After this change it works. It seems that results of model training is not affected.

System (please complete the following information):

Win 10
Python version: 3.7
darts version 0.4.0

Dockerfile to ease up dev setup

ModuleNotFoundError: No module named 'darts'

I installed all the dependencies packages on anaconda (as well as fbprophet and torch), and later I installed
darts, by "pip install u8darts", saying that all "Rerquirement alread satisfied".

However, When I try to import dart on python, it doesnt find the module.

Does Anybody knows what might be the problem?

Python version: [e.g. 3.6.5]

Update model?

Is your feature request related to a current problem? Please describe.
Is it possible to update an existing model after it has been trained? For example, a TCN model?

Additional context
As new training data arrives, I would like to be able to pickup the training from the last time.

Thanks for any advice

BTW, this package is really nice! So easy to use and intuitive :)

Transformer Forecasting Model

With the recent popularity of GPT3, it got me thinking about how a Transformer model could be applied to time-serires forecasting.
Could this sort of model be put on the 'wishlist' for Darts?

There was a discussion on Kaggle recently about this and it links to more information about the suject.
https://www.kaggle.com/c/m5-forecasting-accuracy/discussion/142833

I believe the Attention mechanisim of the model is capable of discovering certain patterns that other models struggle with and this could be a better model for certain domains of time-series forecasting.

There is a pytorch implementation here:
https://github.com/maxjcohen/transformer

a recent paper on the subject:
Deep Transformer Models for Time Series Forecasting:The Influenza Prevalence Case
https://arxiv.org/pdf/2001.08317

Fast Fourier Transform (FFT)

Add Fast Fourier Transform (FFT) to the library.

plot series without weekend data

How to plot , series.plot() if our series has no weekend data?

Multi-Quantile Recurrent Neural Network (MQRNN)

AWS Forecast Recipe:
https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-mqrnn.html

How to specify exogenous variables for TCN or Torch

Looking at the API for multivariate models, I am not sure how I can provide exogenous time series (input values known in the future) to the predict function.
Please advise.

[BUG] TqdmDeprecationWarning

When fitting a LSTM Model I get the folowing warning which I guess can be overcome by an update of the library. Nothing serious as everything runs so far. Just could be an issue in the 5.0 release of tqdm.

lib/python3.6/site-packages/darts/utils/utils.py:74: TqdmDeprecationWarning:

This function will be removed in tqdm==5.0.0
Please use tqdm.notebook.tqdm instead of tqdm.tqdm_notebook

Python version: [e.g. 3.6]
darts version [e.g. 0.2.1]

Fix dependency versions

[Documentation] TimeSeries parameter: frequency- what is it?

I tried to create a TimeSeries object with either constructor or from_dataframe method, but I get:

ValueError: Could not infer explicit frequency. Observed frequencies: {'D', 'B'}. Is Series too short (n=2)?

Could you please describe in more details the frequency parameter:

What are possible values?
How this parameter is used?
Example would be nice.
Is it possible to pass data with irregular time values?
Can I pass [1:N] for sequence of data, that are not bind to date? If so- what frequency should I use then?

ARIMA / AutoARIMA : Support exogenous variables

Is your feature request related to a current problem? Please describe.
Nope

Describe proposed solution
I LOVE this package and think it's very well-written! I'd love to add covariates for forecasting like in some other packages to go from ARIMA to ARIMAX. I'm assuming this would be an easy addition since I've seen it done elsewhere. Thanks!

Describe potential alternatives
A clear and concise description of any alternative solutions or existing features that might solve it.

Additional context
Add any other context or screenshots about the feature request here.

LSTM

Add LSTM to the library

seq2seq

Add seq2seq model to the library.

Use on data with regular breaks?

Hello - I recently discovered your library and it is very intriguing!

I frequently work with time series data which has regular breaks, during which there are no observed measurements (e.g., during the winter). However, when I try the darts library on this data, I get NaN MAPE error results, presumably due to these irregular intervals, as the forecasting methods don't know what to do with the gap periods.

I've tried using auto_fillna to fill the gaps for these time periods, which has successfully gotten the package to work, but I'm concerned that the errors and calibrations are biased, since the models are now being fit on inputed data, which makes up ~50% of observations in many cases. So when I try to compare models, half the resulting MAPE score is a function of which model is best at handling the inputed data.

Do you have a recommended workaround here? Any way to get the models to not score the time periods without actual observed data?

Thanks!

u8darts on Torch v1.7

Is possible to use u8darts with torch v1.7 ?

I have latest torch installed, and when i tried to install 'u8darts[all]', then error message shows :

ERROR: Could not find a version that satisfies the requirement torch==1.5.1;

Regards

Theta

Add Theta method to the library:
https://robjhyndman.com/papers/Theta.pdf

Transform function for TimeSeries

Describe proposed solution
New function in TimeSeries class named transform taking as parameter function and applying it to TimeSeries itself. For reference see pandas-pipe

signature:
transform(self, function, *args, **kwargs) -> TimeSeries

[BUG] Unittest may fail depending on the version of pmdarima and statsmodels

Describe the bug

Unittest may fail depending on the version of pmdarima and statsmodels.

System (please complete the following information)

Python version : 3.8.2
darts version : 0.2.3
OS : Ubuntu 20.04.1 LTS (Focal Fossa)

To Reproduce

I tried 3 combinations of the following.

pmdarima	statsmodels	results
1.6.1(not latest)	0.11.1(not latest)	OK
1.6.1(not latest)	0.12.0(latest)	`test_autoregression_models.py` fails because of Exponential smoothing model
1.7.1(latest)	0.11.1(not latest)	`test_autoregression_models.py` fails because of Auto-ARIMA model

I didn't try the case that both libraries are latest version (pmdarima 1.7.1, statsmodels 0.12) because pmdarima 1.7.1 requires statsmodels<0.12. Please see also #194.

Suggention & Question

This error is likely caused by the threshold in this section.
Does anyone know how this threshold is determined?
Also, do you think it is reasonable to change the thresholds so that no errors occur?
I would like to get your opinions.

log details

pmdarima==1.6.1(not latest) statsmodels==0.11.1(not latest)

MY_DIR/darts$ python3 -m pip install -q pmdarima==1.6.1 statsmodels==0.11.1
MY_DIR/darts$ python3 -m unittest
./home/norihitoishida/.local/lib/python3.8/site-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)
loading checkpoint_3.pth.tar
....loading checkpoint_3.pth.tar
......../home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters.py:743: ConvergenceWarning: Optimization failed to converge. Check mle_retvals.
  warn("Optimization failed to converge. Check mle_retvals.",
/home/norihitoishida/.local/lib/python3.8/site-packages/Cython/Build/Inline.py:4: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/statespace/sarimax.py:975: UserWarning: Non-invertible starting MA parameters found. Using zeros as starting parameters.
  warn('Non-invertible starting MA parameters found.'
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:568: FutureWarning: fft=True will become the default in a future version of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
./home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters.py:743: ConvergenceWarning: Optimization failed to converge. Check mle_retvals.
  warn("Optimization failed to converge. Check mle_retvals.",
/home/norihitoishida/git/darts/darts/models/theta.py:108: RuntimeWarning: invalid value encountered in double_scalars
  drift = self.coef * np.array([i + (1 - (1 - self.alpha) ** self.length) / self.alpha for i in range(0, n)])
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:568: FutureWarning: fft=True will become the default in a future version of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
./home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters.py:743: ConvergenceWarning: Optimization failed to converge. Check mle_retvals.
  warn("Optimization failed to converge. Check mle_retvals.",
/home/norihitoishida/.local/lib/python3.8/site-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 201.50it/s]
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 179.40it/s]
.../home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:568: FutureWarning: fft=True will become the default in a future version of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
................................................/home/norihitoishida/git/darts/darts/tests/test_timeseries.py:410: DeprecationWarning: The default dtype for empty Series will be 'object' instead of 'float64' in a future version. Specify a dtype explicitly to silence this warning.
  TimeSeries.from_series(pd.Series(), freq='D')
................................................
----------------------------------------------------------------------
Ran 118 tests in 37.855s

OK

pmdarima==1.6.1(not latest) statsmodels==0.12.0(latest)

MY_DIR/darts$ python3 -m pip install -q pmdarima==1.6.1 statsmodels==0.12.0
MY_DIR/darts$ python3 -m unittest
./home/norihitoishida/.local/lib/python3.8/site-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)
loading checkpoint_3.pth.tar
....loading checkpoint_3.pth.tar
......../home/norihitoishida/git/darts/darts/models/exponential_smoothing.py:62: FutureWarning: the 'damped'' keyword is deprecated, use 'damped_trend' instead
  hw_model = hw.ExponentialSmoothing(series.values(),
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters/model.py:427: FutureWarning: After 0.13 initialization must be handled at model creation
  warnings.warn(
F/home/norihitoishida/.local/lib/python3.8/site-packages/Cython/Build/Inline.py:4: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/arima_model.py:472: FutureWarning: 
statsmodels.tsa.arima_model.ARMA and statsmodels.tsa.arima_model.ARIMA have
been deprecated in favor of statsmodels.tsa.arima.model.ARIMA (note the .
between arima and model) and
statsmodels.tsa.SARIMAX. These will be removed after the 0.12 release.

statsmodels.tsa.arima.model.ARIMA makes use of the statespace framework and
is both well tested and maintained.

To silence this warning and continue using ARMA and ARIMA until they are
removed, use:

import warnings
warnings.filterwarnings('ignore', 'statsmodels.tsa.arima_model.ARMA',
                        FutureWarning)
warnings.filterwarnings('ignore', 'statsmodels.tsa.arima_model.ARIMA',
                        FutureWarning)

  warnings.warn(ARIMA_DEPRECATION_WARN, FutureWarning)
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters/model.py:427: FutureWarning: After 0.13 initialization must be handled at model creation
  warnings.warn(
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:662: FutureWarning: fft=True will become the default after the release of the 0.12 release of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
./home/norihitoishida/git/darts/darts/models/exponential_smoothing.py:62: FutureWarning: the 'damped'' keyword is deprecated, use 'damped_trend' instead
  hw_model = hw.ExponentialSmoothing(series.values(),
/home/norihitoishida/.local/lib/python3.8/site-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 199.28it/s]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 196.29it/s]
.../home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:662: FutureWarning: fft=True will become the default after the release of the 0.12 release of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
................................................/home/norihitoishida/git/darts/darts/tests/test_timeseries.py:410: DeprecationWarning: The default dtype for empty Series will be 'object' instead of 'float64' in a future version. Specify a dtype explicitly to silence this warning.
  TimeSeries.from_series(pd.Series(), freq='D')
................................................
======================================================================
FAIL: test_models_performance (darts.tests.test_autoregression_models.AutoregressionModelsTestCase)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/norihitoishida/git/darts/darts/tests/test_autoregression_models.py", line 59, in test_models_performance
    self.assertTrue(current_mape < max_mape, "{} model exceeded the maximum MAPE of {}."
AssertionError: False is not true : Exponential smoothing model exceeded the maximum MAPE of 4.8.with a MAPE of 5.5900253625464575

----------------------------------------------------------------------
Ran 118 tests in 39.151s

FAILED (failures=1)

pmdarima==1.7.1(latest) statsmodels==0.11.1(not latest)

MY_DIR/darts$ python3 -m pip install -q pmdarima==1.7.1 statsmodels==0.11.1
MY_DIR/darts$ python3 -m unittest
./home/norihitoishida/.local/lib/python3.8/site-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)
loading checkpoint_3.pth.tar
....loading checkpoint_3.pth.tar
......../home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters.py:743: ConvergenceWarning: Optimization failed to converge. Check mle_retvals.
  warn("Optimization failed to converge. Check mle_retvals.",
/home/norihitoishida/.local/lib/python3.8/site-packages/Cython/Build/Inline.py:4: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/base/model.py:567: ConvergenceWarning: Maximum Likelihood optimization failed to converge. Check mle_retvals
  warn("Maximum Likelihood optimization failed to converge. "
F/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters.py:743: ConvergenceWarning: Optimization failed to converge. Check mle_retvals.
  warn("Optimization failed to converge. Check mle_retvals.",
/home/norihitoishida/git/darts/darts/models/theta.py:108: RuntimeWarning: invalid value encountered in double_scalars
  drift = self.coef * np.array([i + (1 - (1 - self.alpha) ** self.length) / self.alpha for i in range(0, n)])
/home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:568: FutureWarning: fft=True will become the default in a future version of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
./home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/holtwinters.py:743: ConvergenceWarning: Optimization failed to converge. Check mle_retvals.
  warn("Optimization failed to converge. Check mle_retvals.",
/home/norihitoishida/.local/lib/python3.8/site-packages/torch/storage.py:34: FutureWarning: pickle support for Storage will be removed in 1.5. Use `torch.save` instead
  warnings.warn("pickle support for Storage will be removed in 1.5. Use `torch.save` instead", FutureWarning)
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 200.28it/s]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 17/17 [00:00<00:00, 198.92it/s]
.../home/norihitoishida/.local/lib/python3.8/site-packages/statsmodels/tsa/stattools.py:568: FutureWarning: fft=True will become the default in a future version of statsmodels. To suppress this warning, explicitly set fft=False.
  warnings.warn(
................................................/home/norihitoishida/git/darts/darts/tests/test_timeseries.py:410: DeprecationWarning: The default dtype for empty Series will be 'object' instead of 'float64' in a future version. Specify a dtype explicitly to silence this warning.
  TimeSeries.from_series(pd.Series(), freq='D')
................................................
======================================================================
FAIL: test_models_performance (darts.tests.test_autoregression_models.AutoregressionModelsTestCase)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/norihitoishida/git/darts/darts/tests/test_autoregression_models.py", line 59, in test_models_performance
    self.assertTrue(current_mape < max_mape, "{} model exceeded the maximum MAPE of {}."
AssertionError: False is not true : Auto-ARIMA model exceeded the maximum MAPE of 13.7.with a MAPE of 30.32851040186142

----------------------------------------------------------------------
Ran 118 tests in 41.443s

FAILED (failures=1)

DeepAR+

AWS Forecast Recipe:
https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-deeparplus.html

Research reference dataset for comparison

Documentation: comparison to sktime and tslearn

Is your feature request related to a current problem? Please describe.
There are many libraries for time-series forecasting. They have strong and weak points. It's hard to understand strong points and weak points of Darts if you are not a contributor of Darts.

Describe proposed solution
It would be good if authors describe use-cases where this library is the best choice. It would be good to compare strong and week points of Darts with two popular python libraries for time-series forecasting: https://github.com/alan-turing-institute/sktime and https://github.com/tslearn-team/tslearn (around 1.5k stars each).

Add an example notebook for the FB Prophet model

Is your feature request related to a current problem? Please describe.
Currently, I couldn't see an example to demonstrate fbprophet usage through darts. I am opening this issue to create an example notebook for prophet usage.

Describe proposed solution
Adding a new jupyter notebook to examples folder to show the usage of fbprophet with different seasonal frequencies & holidays.

Autoregressive forecasting

Nice work, guys!

It's a common place to use features generated from timeseries itself like lags, rolling statistics and etc.
I've seen some approaches like using sklearn transformers and pipelines for fitting, and forecasting via recursive procedure( iteratively generating one step ahead predictions )

Have you any ideas about it?
I think it would be nice feature.

Latest pmdarima (1.7.1) requires statsmodels (>=0.11,<0.12)

Latest pmdarima (1.7.1) requires statsmodels>=0.11,<0.12(latest), because latest statsmodels seems to have bugs.

If users follow the official instructions for setting up Docker, they will end up with the latest statsmodels(0.12) installed.
So, should we fix requirements/main.txt like this?

- statsmodels>=0.11.1
+ statsmodels>=0.11.1, <0.12

https://github.com/unit8co/darts/blob/master/requirements/main.txt

I would like to get your opinion. Thanks!

Visualisation with plotly / plotly express

Is there any way, how to do a visualisation of a forecast (and original data) with plotly or plotly express instead of matplotlib?

from darts.models import ExponentialSmoothing
train, val = series.split_before(pd.Timestamp('2020-05-01'))
model = ExponentialSmoothing()
model.fit(train)
prediction = model.predict(len(val))

Is there any way, how to store the forecast values to dataframe?

LSTM Autoencoder with attention mechanisms

Problem
Adding LSTM-autoencoder with attention mechanisms for forecasting.

Solution
The solution is an implementation of the following paper. It will be possible to activate and deactivate both attention mechanisms.

`backtest()`: Retrain every n steps

Recently found your library and it is really great and handy!
The evaluation of the forecasting performance is best done using a cross-validation approach. The backtest_forecasting()-function does that - although it currently iterates and re-trains the model on every single time step. In my application, I am training ten-thousands of different time series and it becomes computationally unfeasible to retrain on every time step. Another approach already implemented in scikit-learn is TimeSeriesSplit(), which generates a k-fold cross-validation split designed for time series.
https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.TimeSeriesSplit.html

One solution to the issue would be to add support for different cross-validation techniques implemented in scikit-learn (such as TimeSeriesSplit()). Another solution I see could be to add a stride parameter in backtest_forecasting() that specifies how many time steps it should skip while backtesting.

[BUG] ForecastingModel.backtest: Can bypass sanity checks

Describe the bug
Invalid optional arguments to ForecastingModel.backtest can bypass sanity checks when provided as positional arguments

To Reproduce

NaiveDrift().backtest(linear_series, None, start=0.7, forecast_horizon=-1)
NaiveDrift().backtest(linear_series, None, 0.7, -1)

Expected behavior
The first line above is caught by the sanity checks and raise ValueError: 'The provided forecasting horizon must be a positive integer.' as expected

For the second line however the negative forecast_horizon param clears the sanity checks and produces an error further down the line:

NaiveDrift().backtest(linear_series, None, 0.7, -1)
  File ".../darts/utils/utils.py", line 120, in sanitized_method
    return method_to_sanitize(self, *args, **kwargs)
  File ".../darts/models/forecasting_model.py", line 217, in backtest
    pred = self.predict(forecast_horizon, **predict_kwargs)
  File ".../darts/models/baselines.py", line 98, in predict
    forecast = np.linspace(last, last_value, num=n)
  File "<__array_function__ internals>", line 5, in linspace
  File ".../lib/python3.8/site-packages/numpy/core/function_base.py", line 115, in linspace
    raise ValueError("Number of samples, %s, must be non-negative." % num)
ValueError: Number of samples, -1, must be non-negative.

System (please complete the following information):

Python version: 3.8.5
darts version: develop branch, commit #2977f4f

numpy version incompatibility using Darts and Tensorflow

DARTS depends on numpy 1.19.0 and Tensorflow requires numpy<1.19.0

If DARTS will require numpy<1.19.0 it will be possible to have DARTS and Tensorflow in the same environment.

Plot in-sample forecast?

Hello, is there a way to plot in-sample forecast after having fitted a model?
The backtest() and historical_forecasts() are of course more rigorous for a proper backtesting but I would simply like sometimes to see how the model performs in sample before moving on to more sophisticate analysis.

Non-Parametric Time Series (NPTS)

AWS Forecast Recipe:
https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-npts.html

No matching distribution found for torch>=1.4.0 (from u8darts)

When I use pip install u8darts

Python version: 3.7.1

In sample predictions

I would like to have in sample prediction as an option.

This will help collect model errors so users can diagnose models.

Describe proposed solution
An extra method to the model. predict_in_sample

in_sample_prediction = model.predict_in_sample(training_data)

Describe potential alternatives
Modifying predict method to be forecast and generalising predict like in Statsmodels ARIMA

Make setup lighter

Is your feature request related to a current problem? Please describe.
Currently the default setup of has ~2.4 GB of dependencies. This make it unusable for AWS lambdas or other resource constrained environments.

Describe proposed solution
Change the setup so that the heaviest dependencies can be optionally installed:
Eg:

pip install u8darts[core] # Bare minimun
pip install u8darts[torch] # ~1.5G (1,4G torch +74M caffe2 + (14+3.4)M tensorboard)
pip install u8darts[fbprophet] # ~500M (31M  fbprophet + 474M pystan)
pip install u8darts[statsmodels] # 42M

`inverse_transform` and multivariate time series

DataTransformer.inverse_transform isn’t easy to use with multivariate time series in a “real” use case situation. I think a very common use case would be to:

transform the multivariate time series (e.g. re-scale using scaler)
make a prediction for a single main target variable: say forecast holds that prediction
inverse transform back to the original scale

For 3. I'd like to be able to call scaler.inverse_transform(forecast), but there's a dimension mismatch: inverse_transform() expects a multivariate time series here (in particular a multivariate time series of the same width it was fitted with). The scaler cannot know which variable(s) the forecast corresponds to.

There are workarounds, but I couldn't come up with a clean and easy way… Can someone think of one?

Prophet return every prediction value

Is your feature request related to a current problem? Please describe.
In order to create an outlier detection with Prophet, i need the full dataframe that's return Prophet

Describe proposed solution
Remove the hardcoded ["yhat"] from Prophet.predict add a variable asking to return just yhat or all the predictions: 'yhat_lower', 'yhat_upper', etc..
https://unit8co.github.io/darts/_modules/darts/models/prophet.html#Prophet.predict

Thanks very much! very useful framework

ForecastingModel.backtest() fails if forecast_horizon isn't provided

darts/darts/models/forecasting_model.py

Lines 148 to 157 in 2977f4f

 if hasattr(n, 'forecast_horizon'): 

 raise_if_not(n.forecast_horizon > 0, 'The provided forecasting horizon must be a positive integer.', logger) 

 # check start parameter 

 if hasattr(n, 'start'): 

 if isinstance(n.start, float): 

 raise_if_not(n.start >= 0.0 and n.start < 1.0, '`start` should be between 0.0 and 1.0.', logger) 

 elif isinstance(n.start, pd.Timestamp): 

 if (hasattr(n, 'trim_to_series') and n.trim_to_series) or not hasattr(n, 'trim_to_series'): 

 raise_if_not(n.start + training_series.freq() * n.forecast_horizon in training_series,

n.forecast_horizon at line 157 fails when hasattr(n, 'forecast_horizon') == false, ie. every time no explicit value is passed when calling backtest()

Error produced:

AttributeError: 'types.SimpleNamespace' object has no attribute 'forecast_horizon'

[BUG]

Outlier detection

Problem.
Currently there is no functionality to spot outliers of specific entries within a given TimeSeries instance.

Proposed solution
One way to spot outliers could be to use one or more of the existing forecasting models (possibly as an ensemble) to produce a historic forecast, and look at the deviations of the forecasts from the actual data. Large deviations might be indicators for outliers.
Inspiration: https://towardsdatascience.com/anomaly-detection-with-time-series-forecasting-c34c6d04b24a

Conda-forge recipe request

Is your feature request related to a current problem? Please describe.
I'd like to be able to conda install darts

Describe proposed solution
Create a conda-forge recipe and link it to the PyPi package or GitHub release.

Describe potential alternatives
Pip installing to a conda environment works but PyPi and conda packages don't always play nice together when upgrading.

Implement `map` method for TimeSeries class

Missing functionality
Although adding or multiplying a scalar value to every entry in a TimeSeries instance is very easy, when it comes to more complex element-wise operations (e.g. taking the sine function value of every entry in the series) there is no easy way to create a new TimeSeries instance, given a TimeSeries instance and operation.

Proposed solution
The most straight-forward way to implement this functionality would be implement a map function that takes as arguments a TimeSeries and a function, and returns a new TimeSeries instance (pretty much like the built-in python function).

backtests' example: AttributeError: 'numpy.float64' object has no attribute 'width'

When I was trying to follow the article backtests' example

https://medium.com/unit8-machine-learning-publication/darts-time-series-made-easy-in-python-5ac2947a8878

I got this error

AttributeError                            Traceback (most recent call last)
<ipython-input-14-9213dab113d8> in <module>
      2 series.plot(label='data')
      3 for i, m in enumerate(models):
----> 4     err = mape(backtests[i], series)
      5     backtests[i].plot(lw=3, label='{}, MAPE={:.2f}%'.format(m, err))
      6 plt.title('Backtests with 3-months forecast horizon')

~/.pyenv/versions/3.8.3/envs/keras/lib/python3.8/site-packages/darts/metrics/metrics.py in wrapper_multivariate_support(*args, **kwargs)
     31         series2 = kwargs['series2'] if 'series2' in kwargs else args[0] if 'series1' in kwargs else args[1]
     32 
---> 33         raise_if_not(series1.width == series2.width, "The two TimeSeries instances must have the same width.", logger)
     34 
     35         num_series_in_args = int('series1' not in kwargs) + int('series2' not in kwargs)

AttributeError: 'numpy.float64' object has no attribute 'width'

To Reproduce

I used this part of the code

from darts.models import Prophet
models = [ExponentialSmoothing(), Prophet()]
backtests = [model.backtest(series,
                            start=pd.Timestamp('19550101'),
                            forecast_horizon=3)
             for model in models]

from darts.metrics import mape
series.plot(label='data')
for i, m in enumerate(models):
    err = mape(backtests[i], series)
    backtests[i].plot(lw=3, label='{}, MAPE={:.2f}%'.format(m, err))
plt.title('Backtests with 3-months forecast horizon')
plt.legend()

Expected behavior
I would like to get the article's plot and results.

System (please complete the following information):
u8darts==0.5.0
fbprophet==0.6
Python 3.8.3

Additional context

I don't know if this message is expected but in this part of the code

backtests = [model.backtest(series,
                            start=pd.Timestamp('19550101'),
                            forecast_horizon=3)
             for model in models]

I get this message many times

INFO:fbprophet:Disabling weekly seasonality. Run prophet with weekly_seasonality=True to override this.

Accept Prophet with `growth='logistic'`

Proplem
To use Prophet with growth='logistic', the dataframe provided must have a column named 'cap'.

Describe proposed solution
Accept 'cap' and 'floor' as arrays provided as parametes on the model initialization.

Describe potential alternatives
Do not consider TimeSeries to be always multivariate and rename the columns in order.

ConvergenceWarning: Optimization failed to converge. Check mle_retvals.

Describe the bug
Simple Exponential Smoothing does not converge for the air passenger data as shown in the demo.

To Reproduce
Please check my Colab notebook: https://colab.research.google.com/drive/1H8Dqyx6lfe818PeHy_qdVgmoN3Btoiam?usp=sharing

Expected behavior
Exponential smoothing should converge for this very simple dataset.

System (please complete the following information):

Python version: 3.8
darts version 0.5.0

Additional context
The autocorr() function doesn't seem to exist.

Unified API : `model.params` attribute

Hello! Firstly, I wanted to thank you for this wonderful library.

Unified API to get model's params might make darts more convenient.
For example, to get params of Exponential Smoothing model, We have to code like this:

model = ExponentialSmoothing()
model.fit(train)

print(model.model.model.params) # {'smoothing_level': 0.5789473661331209, 'smoothing_slope': ...

So, I want to implement ~~model.get_params() method like sklearn.~~ model.params attribute with property like statsmodels and fbprophet.

I would like to get your opinion before send PR. Thanks!

[BUG] Inverse transform does not work on series of length < 3

First off, I want to say that I'm not an ML engineer at all. We have our own custom timeseries forecasting code that our ML engineers wrote. I stumbled upon darts, and we are very excited by the prospect of using it, given that we can use so many different kinds of models to generate our predictions. Our problem statement is as follows: We know the weekly consumption of our widgets by customers, and want to predict how many they will use next week. We only need 1 prediction. To that end, I wrote some code that cleans and preps the data, and is then passed to darts. Everything works as expected, until we hit the inverse_transform() function.

Describe the bug
We only care about 1 prediction step, so we invoke the prediction by saying: pred_series = my_model.predict(n=1).
This works, and I am able to get the scaled version of what the usage is going to be next week. However, I want the human-readable/understandable number, and so I do: print(transformer.inverse_transform(pred_series)). I then get the following error:

  File "train/run_models.py", line 124, in get_model
    print(transformer.inverse_transform(pred_series))
  File "/home/ec2-user/darts/lib/python3.7/site-packages/darts/preprocessing/scaler_wrapper.py", line 102, in inverse_transform
    reshape((-1, series.width))))
  File "/home/ec2-user/darts/lib/python3.7/site-packages/darts/timeseries.py", line 571, in from_times_and_values
    return TimeSeries(df, freq, fill_missing_dates)
  File "/home/ec2-user/darts/lib/python3.7/site-packages/darts/timeseries.py", line 58, in __init__
    'is not passed', logger)
  File "/home/ec2-user/darts/lib/python3.7/site-packages/darts/logging.py", line 54, in raise_if_not
    raise ValueError(message)

I read through the source code and didn't see any place in the inverse_transform() function where the frequency is being passed. Merely:

return TimeSeries.from_times_and_values(series.time_index(),
                                                self.transformer.inverse_transform(series.values().
                                                                                   reshape((-1, series.width))))

I then tried to pass the frequency argument as follows: print(transformer.inverse_transform(pred_series), "W-SUN"), and got this error instead:

  File "train/run_models.py", line 124, in get_model
    print(transformer.inverse_transform(pred_series, "W-SUN"))
TypeError: inverse_transform() takes 2 positional arguments but 3 were given

I put "W" instead of "W-SUN" too, with similar results (same error, that is).

So my question is, how to use darts to get precisely 1 prediction? And for it to be not scaled? I could generate 3, pick the first one and that'd be that, but I'm not sure if that's the best/right approach to this problem.

To Reproduce
Here's part of my code. Prepare any dataset of your choice ahead of this snippet, please.

    ......
    ......

    # Number of previous time stamps taken into account.
    SEQ_LENGTH = 2
    # Number of features in last hidden state
    HIDDEN_SIZE = 15 * SEQ_LENGTH
    # number of output time-steps to predict
    OUTPUT_LEN = 1
    # Number of stacked rnn layers.
    NUM_LAYERS = 2

    my_model = RNNModel(
        model='LSTM',
        output_length=OUTPUT_LEN,
        hidden_size=HIDDEN_SIZE,
        n_rnn_layers=NUM_LAYERS,
        input_length=SEQ_LENGTH,
        batch_size=100,
        n_epochs=150,
        model_name='Air_RNN', log_tensorboard=True
    )

    my_model.fit(train_transformed, val_transformed, verbose=True)

    pred_series = my_model.predict(n=3)

    backtest_series = backtest_forecasting(series_transformed, my_model, pd.Timestamp('20200621'),
                                       fcast_horizon_n=1, verbose=True)

    print('RMSE: {:.4f}'.format(rmse(transformer.inverse_transform(series_transformed),
                                 transformer.inverse_transform(backtest_series))))

    my_model.fit(series_transformed, verbose=True)

    pred_series = my_model.predict(n=1)

    print(pred_series)

    # Error is in this next line. Everything above this works like a charm
    print(transformer.inverse_transform(pred_series))

Expected behavior
I expect to see a single record with an inverse transformed value.

System (please complete the following information):

Python version: 3.7.7
darts version: 0.2.1

Additional context
Add any other context about the problem here.

Add interpolate to TimeSeries class

Problem
dataframe with missing dates.
TimeSeries fill missing dates, but add nan value to the filled date
And almost all model will fail when fit

Solution
Add interpolate function from pandas when this occurs. Let the user define the interpolate method.
https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.Series.interpolate.html#pandas.Series.interpolate

By the way: pretty cool lib!

How to save underlying data forecast?

Is it possible within Darts to not only save the graphs, but also save the underlying data of the forecast?

	if hasattr(n, 'forecast_horizon'):
	raise_if_not(n.forecast_horizon > 0, 'The provided forecasting horizon must be a positive integer.', logger)

	# check start parameter
	if hasattr(n, 'start'):
	if isinstance(n.start, float):
	raise_if_not(n.start >= 0.0 and n.start < 1.0, '`start` should be between 0.0 and 1.0.', logger)
	elif isinstance(n.start, pd.Timestamp):
	if (hasattr(n, 'trim_to_series') and n.trim_to_series) or not hasattr(n, 'trim_to_series'):
	raise_if_not(n.start + training_series.freq() * n.forecast_horizon in training_series,

unit8co / darts Goto Github PK

darts's People

Contributors

Stargazers

Watchers

Forkers

darts's Issues

Describe the bug

System (please complete the following information)

To Reproduce

Suggention & Question

log details

pmdarima==1.6.1(not latest) statsmodels==0.11.1(not latest)

pmdarima==1.6.1(not latest) statsmodels==0.12.0(latest)

pmdarima==1.7.1(latest) statsmodels==0.11.1(not latest)

Recommend Projects

Recommend Topics

Recommend Org