# Import to be able to import python package from src
import sys
sys.path.insert(0, '../src')

import pandas as pd
import numpy as np
import ontime as on

Models

This class implement a generic way to load models in onTime. It is compatible with Darts, Scikit-learn and PyTorch, with an aim to add other librairies and models soon. Let’s see how to use it.

Let’s generate random TimeSeries

# Start and end dates
start_date = pd.Timestamp('2022-01-01')
end_date = pd.Timestamp('2022-12-31')

# Make a random walk
ts = on.generators.random_walk().generate(start=start_date, end=end_date)
ts = ts.astype(np.float32)

# Make another random walk
new_ts = on.generators.random_walk().generate(start=start_date, end=end_date)
new_ts = ts.astype(np.float32)

Using a Darts model

The model can be loaded from the desired library

from darts.models import BlockRNNModel

Then, it can be used with this generic interface.

model = on.Model(BlockRNNModel,
                 input_chunk_length=12,
                 output_chunk_length=6,
                 n_rnn_layers=2,
                 n_epochs=50
                 )

Finally, the training and inference functions are the same than in Scikit Learn for instance :

# To train the model
model.fit(ts)

darts.models.forecasting.torch_forecasting_model INFO  Train dataset contains 348 samples.
darts.models.forecasting.torch_forecasting_model INFO  Time series values are 32-bits; casting model to float32.
INFO: GPU available: True (cuda), used: True
lightning.pytorch.utilities.rank_zero INFO  GPU available: True (cuda), used: True
INFO: TPU available: False, using: 0 TPU cores
lightning.pytorch.utilities.rank_zero INFO  TPU available: False, using: 0 TPU cores
INFO: HPU available: False, using: 0 HPUs
lightning.pytorch.utilities.rank_zero INFO  HPU available: False, using: 0 HPUs
INFO: You are using a CUDA device ('NVIDIA GeForce RTX 3070 Ti Laptop GPU') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
lightning.pytorch.utilities.rank_zero INFO  You are using a CUDA device ('NVIDIA GeForce RTX 3070 Ti Laptop GPU') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

  | Name            | Type             | Params | Mode
-------------------------------------------------------------
0 | criterion       | MSELoss          | 0      | train
1 | train_criterion | MSELoss          | 0      | train
2 | val_criterion   | MSELoss          | 0      | train
3 | train_metrics   | MetricCollection | 0      | train
4 | val_metrics     | MetricCollection | 0      | train
5 | rnn             | RNN              | 2.0 K  | train
6 | fc              | Sequential       | 156    | train
-------------------------------------------------------------
2.2 K     Trainable params
0         Non-trainable params
2.2 K     Total params
0.009     Total estimated model params size (MB)
8         Modules in train mode
0         Modules in eval mode

INFO: `Trainer.fit` stopped: `max_epochs=50` reached.
lightning.pytorch.utilities.rank_zero INFO  `Trainer.fit` stopped: `max_epochs=50` reached.

<ontime.core.modelling.model.Model at 0x7f7accf17460>

# To make a prediction from train series
darts_pred = model.predict(10)

INFO: GPU available: True (cuda), used: True
lightning.pytorch.utilities.rank_zero INFO  GPU available: True (cuda), used: True
INFO: TPU available: False, using: 0 TPU cores
lightning.pytorch.utilities.rank_zero INFO  TPU available: False, using: 0 TPU cores
INFO: HPU available: False, using: 0 HPUs
lightning.pytorch.utilities.rank_zero INFO  HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

Once trained, the model can also infer on a new series, or even on a list of series (batched inference)

# on a new single series
model.predict(n=10, ts=new_ts)
# on many series
model.predict(n=10, ts=[new_ts]*3)

INFO: GPU available: True (cuda), used: True
lightning.pytorch.utilities.rank_zero INFO  GPU available: True (cuda), used: True
INFO: TPU available: False, using: 0 TPU cores
lightning.pytorch.utilities.rank_zero INFO  TPU available: False, using: 0 TPU cores
INFO: HPU available: False, using: 0 HPUs
lightning.pytorch.utilities.rank_zero INFO  HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

INFO: GPU available: True (cuda), used: True
lightning.pytorch.utilities.rank_zero INFO  GPU available: True (cuda), used: True
INFO: TPU available: False, using: 0 TPU cores
lightning.pytorch.utilities.rank_zero INFO  TPU available: False, using: 0 TPU cores
INFO: HPU available: False, using: 0 HPUs
lightning.pytorch.utilities.rank_zero INFO  HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

[<TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 40B
 array([[[14.379956]],

        [[13.368615]],

        [[13.446395]],

        [[15.118863]],

        [[14.379672]],

        [[13.368358]],

        [[13.446113]],

        [[15.118556]],

        [[13.832502]],

        [[14.525592]]], dtype=float32)
 Coordinates:
   * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
   * component  (component) object 8B 'random_walk'
 Dimensions without coordinates: sample
 Attributes:
     static_covariates:  None
     hierarchy:          None,
 <TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 40B
 array([[[14.379956]],

        [[13.368615]],

        [[13.446395]],

        [[15.118863]],

        [[14.379672]],

        [[13.368358]],

        [[13.446113]],

        [[15.118556]],

        [[13.832502]],

        [[14.525592]]], dtype=float32)
 Coordinates:
   * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
   * component  (component) object 8B 'random_walk'
 Dimensions without coordinates: sample
 Attributes:
     static_covariates:  None
     hierarchy:          None,
 <TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 40B
 array([[[14.379956]],

        [[13.368615]],

        [[13.446395]],

        [[15.118863]],

        [[14.379672]],

        [[13.368358]],

        [[13.446113]],

        [[15.118556]],

        [[13.832502]],

        [[14.525592]]], dtype=float32)
 Coordinates:
   * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
   * component  (component) object 8B 'random_walk'
 Dimensions without coordinates: sample
 Attributes:
     static_covariates:  None
     hierarchy:          None]

darts_pred = darts_pred.rename({'random_walk': 'darts prediction'}) # for plotting

Using a Scikit-learn API compatible model

from sklearn.neural_network import MLPRegressor

model = on.Model(MLPRegressor(),
                 lags=30)
model.fit(ts)
sk_pred = model.predict(10)
sk_pred = sk_pred.rename({'pred': 'sklearn prediction'})

/home/benjy/.cache/pypoetry/virtualenvs/ontime-2OQVvbNf-py3.10/lib/python3.10/site-packages/sklearn/neural_network/_multilayer_perceptron.py:691: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (200) reached and the optimization hasn't converged yet.
  warnings.warn(

# on a new single series (for now, inference can only be performed on single series)
model.predict(n=10, ts=new_ts)

<TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 80B
array([[[26.51311037]],

       [[26.79350356]],

       [[27.33226923]],

       [[26.90249441]],

       [[27.77318538]],

       [[27.31159983]],

       [[27.34664187]],

       [[26.14925155]],

       [[26.61025269]],

       [[25.49064079]]])
Coordinates:
  * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
  * component  (component) object 8B 'pred'
Dimensions without coordinates: sample
Attributes:
    static_covariates:  None
    hierarchy:          None

TimeSeries (DataArray)

• time: 10
• component: 1
• sample: 1

• 26.51 26.79 27.33 26.9 27.77 27.31 27.35 26.15 26.61 25.49

  array([[[26.51311037]],
  
         [[26.79350356]],
  
         [[27.33226923]],
  
         [[26.90249441]],
  
         [[27.77318538]],
  
         [[27.31159983]],
  
         [[27.34664187]],
  
         [[26.14925155]],
  
         [[26.61025269]],
  
         [[25.49064079]]])

• Coordinates: (2)
  • time
    (time)
    datetime64[ns]
    2023-01-01 ... 2023-01-10

    array(['2023-01-01T00:00:00.000000000', '2023-01-02T00:00:00.000000000',
           '2023-01-03T00:00:00.000000000', '2023-01-04T00:00:00.000000000',
           '2023-01-05T00:00:00.000000000', '2023-01-06T00:00:00.000000000',
           '2023-01-07T00:00:00.000000000', '2023-01-08T00:00:00.000000000',
           '2023-01-09T00:00:00.000000000', '2023-01-10T00:00:00.000000000'],
          dtype='datetime64[ns]')

  • component
    (component)
    object
    'pred'

    array(['pred'], dtype=object)

• Indexes: (2)
  • time
    PandasIndex

    PandasIndex(DatetimeIndex(['2023-01-01', '2023-01-02', '2023-01-03', '2023-01-04',
                   '2023-01-05', '2023-01-06', '2023-01-07', '2023-01-08',
                   '2023-01-09', '2023-01-10'],
                  dtype='datetime64[ns]', name='time', freq='D'))

  • component
    PandasIndex

    PandasIndex(Index(['pred'], dtype='object', name='component'))

• Attributes: (2)

  static_covariates :

  None

  hierarchy :

  None

Using a PyTorch model

You can create and use custom PyTorch models and wrap them in our wrapper. It allows to quickly test its PyTorch model.

For more complex behaviors, you can also create a new Darts model by inheriting one of the existing class in the library. You can find existing model implementations here.

import torch.nn as nn

class SimpleGRU(nn.Module):
    def __init__(self, input_dim=1, hidden_dim=32, output_dim=1, num_layers=1):
        super(SimpleGRU, self).__init__()
        self.hidden_dim = hidden_dim
        self.output_size = output_dim

        self.gru = nn.GRU(input_dim, hidden_dim, num_layers, batch_first=True)
        self.fc = nn.Linear(hidden_dim, output_dim)  # Map hidden state to output dimension

    def forward(self, x):
        # Pass through GRU
        out, _ = self.gru(x)
        out = self.fc(out)
        out = out[:, -6:, :] # the model output 6 features
        return out

model = on.Model(
    SimpleGRU,
    input_chunk_length=12,
    output_chunk_length=6, # should be equal what the output size the model was trained with
    num_layers=2,
    n_epochs=50,
        train_data_module_params={
        "val_split":0.2
    }
)

model.fit(ts)

INFO: GPU available: True (cuda), used: True
lightning.pytorch.utilities.rank_zero INFO  GPU available: True (cuda), used: True
INFO: TPU available: False, using: 0 TPU cores
lightning.pytorch.utilities.rank_zero INFO  TPU available: False, using: 0 TPU cores
INFO: HPU available: False, using: 0 HPUs
lightning.pytorch.utilities.rank_zero INFO  HPU available: False, using: 0 HPUs
INFO: You are using a CUDA device ('NVIDIA GeForce RTX 3070 Ti Laptop GPU') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
lightning.pytorch.utilities.rank_zero INFO  You are using a CUDA device ('NVIDIA GeForce RTX 3070 Ti Laptop GPU') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
2025-01-06 16:26:59.828777: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2025-01-06 16:27:00.179475: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:477] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
E0000 00:00:1736177220.320421  284635 cuda_dnn.cc:8310] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
E0000 00:00:1736177220.361063  284635 cuda_blas.cc:1418] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2025-01-06 16:27:00.703629: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
INFO: LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
lightning.pytorch.accelerators.cuda INFO  LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
INFO:
  | Name    | Type      | Params | Mode
----------------------------------------------
0 | model   | SimpleGRU | 9.7 K  | train
1 | loss_fn | MSELoss   | 0      | train
----------------------------------------------
9.7 K     Trainable params
0         Non-trainable params
9.7 K     Total params
0.039     Total estimated model params size (MB)
4         Modules in train mode
0         Modules in eval mode
lightning.pytorch.callbacks.model_summary INFO
  | Name    | Type      | Params | Mode
----------------------------------------------
0 | model   | SimpleGRU | 9.7 K  | train
1 | loss_fn | MSELoss   | 0      | train
----------------------------------------------
9.7 K     Trainable params
0         Non-trainable params
9.7 K     Total params
0.039     Total estimated model params size (MB)
4         Modules in train mode
0         Modules in eval mode

/home/benjy/.cache/pypoetry/virtualenvs/ontime-2OQVvbNf-py3.10/lib/python3.10/site-packages/lightning/pytorch/loops/fit_loop.py:298: The number of training batches (9) is smaller than the logging interval Trainer(log_every_n_steps=50). Set a lower value for log_every_n_steps if you want to see logs for the training epoch.

INFO: `Trainer.fit` stopped: `max_epochs=50` reached.
lightning.pytorch.utilities.rank_zero INFO  `Trainer.fit` stopped: `max_epochs=50` reached.

<ontime.core.modelling.model.Model at 0x7f7aca538310>

torch_pred = model.predict(10)
torch_pred = torch_pred.rename({'0': 'torch prediction'})

/home/benjy/projects_dev/ontime/src/ontime/core/modelling/libs/pytorch/pytorch_forecasting_model.py:97: UserWarning: The requested prediction horizon (n=10) exceeds the model's output_chunk_length (6). The model will use an iterative forecasting approach, which may result in reduced accuracy due to error propagation.
  warnings.warn(

# on a new single series
model.predict(n=10, ts=new_ts)
# on many series
model.predict(n=10, ts=[new_ts]*3)

/home/benjy/projects_dev/ontime/src/ontime/core/modelling/libs/pytorch/pytorch_forecasting_model.py:97: UserWarning: The requested prediction horizon (n=10) exceeds the model's output_chunk_length (6). The model will use an iterative forecasting approach, which may result in reduced accuracy due to error propagation.
  warnings.warn(
/home/benjy/projects_dev/ontime/src/ontime/core/modelling/libs/pytorch/pytorch_forecasting_model.py:97: UserWarning: The requested prediction horizon (n=10) exceeds the model's output_chunk_length (6). The model will use an iterative forecasting approach, which may result in reduced accuracy due to error propagation.
  warnings.warn(

[<TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 40B
 array([[[15.901117 ]],

        [[15.901123 ]],

        [[15.9012165]],

        [[15.901222 ]],

        [[15.90128  ]],

        [[15.901265 ]],

        [[15.900374 ]],

        [[15.900061 ]],

        [[15.899954 ]],

        [[15.899914 ]]], dtype=float32)
 Coordinates:
   * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
   * component  (component) <U1 4B '0'
 Dimensions without coordinates: sample
 Attributes:
     static_covariates:  None
     hierarchy:          None,
 <TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 40B
 array([[[15.901117 ]],

        [[15.901123 ]],

        [[15.9012165]],

        [[15.901222 ]],

        [[15.90128  ]],

        [[15.901265 ]],

        [[15.900374 ]],

        [[15.900061 ]],

        [[15.899954 ]],

        [[15.899914 ]]], dtype=float32)
 Coordinates:
   * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
   * component  (component) <U1 4B '0'
 Dimensions without coordinates: sample
 Attributes:
     static_covariates:  None
     hierarchy:          None,
 <TimeSeries (DataArray) (time: 10, component: 1, sample: 1)> Size: 40B
 array([[[15.901117 ]],

        [[15.901123 ]],

        [[15.9012165]],

        [[15.901222 ]],

        [[15.90128  ]],

        [[15.901265 ]],

        [[15.900374 ]],

        [[15.900061 ]],

        [[15.899954 ]],

        [[15.899914 ]]], dtype=float32)
 Coordinates:
   * time       (time) datetime64[ns] 80B 2023-01-01 2023-01-02 ... 2023-01-10
   * component  (component) <U1 4B '0'
 Dimensions without coordinates: sample
 Attributes:
     static_covariates:  None
     hierarchy:          None]

Comparing predictions

(on.Plot()
    .add(on.marks.line, ts[-50:])
    .add(on.marks.line, darts_pred)
    .add(on.marks.line, sk_pred)
    .add(on.marks.line, torch_pred)
    .properties(width=600, height=300)
    .show()
)