Test In Colab

User Guide Tutorial 04: Prediction

This tutorial shows how to use TemporAI prediction plugins.

All prediction plugins

To see all the relevant plugins:

[ ]:

from tempor import plugin_loader
from rich.pretty import pprint

all_prediction_plugins = plugin_loader.list()["prediction"]

pprint(all_prediction_plugins, indent_guides=False)

{
    'one_off': {
        'classification': ['cde_classifier', 'ode_classifier', 'nn_classifier', 'laplace_ode_classifier'],
        'regression': ['laplace_ode_regressor', 'nn_regressor', 'ode_regressor', 'cde_regressor']
    },
    'temporal': {'classification': ['seq2seq_classifier'], 'regression': ['seq2seq_regressor']}
}

Now also load data source(s) we will use:

[ ]:

SineDataSource = plugin_loader.get_class("prediction.one_off.sine", plugin_type="datasource")
DummyTemporalPredictionDataSource = plugin_loader.get_class(
    "prediction.temporal.dummy_prediction", plugin_type="datasource"
)

Using a one-off prediction plugin

One-off prediction is the task of predicting a single value for each sample (may be classification or regression).

[ ]:

from tempor import plugin_loader

dataset = SineDataSource(random_state=42).load()
print(dataset)

model = plugin_loader.get("prediction.one_off.classification.nn_classifier", n_iter=50)
print(model)

OneOffPredictionDataset(
    time_series=TimeSeriesSamples([100, *, 5]),
    static=StaticSamples([100, 4]),
    predictive=OneOffPredictionTaskData(targets=StaticSamples([100, 1]))
)
NeuralNetClassifier(
    name='nn_classifier',
    category='prediction.one_off.classification',
    plugin_type='method',
    params={
        'n_static_units_hidden': 100,
        'n_static_layers_hidden': 2,
        'n_temporal_units_hidden': 102,
        'n_temporal_layers_hidden': 2,
        'n_iter': 50,
        'mode': 'RNN',
        'n_iter_print': 10,
        'batch_size': 100,
        'lr': 0.001,
        'weight_decay': 0.001,
        'window_size': 1,
        'device': None,
        'dataloader_sampler': None,
        'dropout': 0,
        'nonlin': 'relu',
        'random_state': 0,
        'clipping_value': 1,
        'patience': 20,
        'train_ratio': 0.8
    }
)

[ ]:

# Targets:
dataset.predictive.targets

StaticSamples with data:

0
sample_idx
0 0
1 1
2 0
3 0
4 1
... ...
95 1
96 1
97 1
98 0
99 0

100 rows × 1 columns

[ ]:

# Train.
model.fit(dataset)

2023-11-16 22:18:45 | INFO     | tempor.models.ts_model:_train:388 | Epoch:0| train loss: 0.68824702501297, validation loss: 0.6883962154388428
2023-11-16 22:18:45 | INFO     | tempor.models.ts_model:_train:388 | Epoch:10| train loss: 0.6999298930168152, validation loss: 0.6879696249961853
2023-11-16 22:18:45 | INFO     | tempor.models.ts_model:_train:388 | Epoch:20| train loss: 0.6818333268165588, validation loss: 0.6867876052856445
2023-11-16 22:18:45 | INFO     | tempor.models.ts_model:_train:388 | Epoch:30| train loss: 0.694374680519104, validation loss: 0.6851812601089478
2023-11-16 22:18:45 | INFO     | tempor.models.ts_model:_train:388 | Epoch:40| train loss: 0.6282318234443665, validation loss: 0.6936542391777039

NeuralNetClassifier(
    name='nn_classifier',
    category='prediction.one_off.classification',
    plugin_type='method',
    params={
        'n_static_units_hidden': 100,
        'n_static_layers_hidden': 2,
        'n_temporal_units_hidden': 102,
        'n_temporal_layers_hidden': 2,
        'n_iter': 50,
        'mode': 'RNN',
        'n_iter_print': 10,
        'batch_size': 100,
        'lr': 0.001,
        'weight_decay': 0.001,
        'window_size': 1,
        'device': None,
        'dataloader_sampler': None,
        'dropout': 0,
        'nonlin': 'relu',
        'random_state': 0,
        'clipping_value': 1,
        'patience': 20,
        'train_ratio': 0.8
    }
)

[ ]:

# Predict:

model.predict(dataset)

StaticSamples with data:

feat_0
sample_idx
0 0.0
1 1.0
2 1.0
3 0.0
4 0.0
... ...
95 1.0
96 0.0
97 1.0
98 0.0
99 1.0

100 rows × 1 columns

Using a temporal prediction plugin

Temporal prediction is the task of predicting a time series for each sample (may be classification or regression).

[ ]:

from tempor import plugin_loader

dataset = DummyTemporalPredictionDataSource(random_state=42, temporal_covariates_missing_prob=0.0).load()
print(dataset)

model = plugin_loader.get("prediction.temporal.regression.seq2seq_regressor", epochs=10)
print(model)

TemporalPredictionDataset(
    time_series=TimeSeriesSamples([100, *, 5]),
    static=StaticSamples([100, 3]),
    predictive=TemporalPredictionTaskData(
        targets=TimeSeriesSamples([100, *, 2])
    )
)
Seq2seqRegressor(
    name='seq2seq_regressor',
    category='prediction.temporal.regression',
    plugin_type='method',
    params={
        'encoder_rnn_type': 'LSTM',
        'encoder_hidden_size': 100,
        'encoder_num_layers': 1,
        'encoder_bias': True,
        'encoder_dropout': 0.0,
        'encoder_bidirectional': False,
        'encoder_nonlinearity': None,
        'encoder_proj_size': None,
        'decoder_rnn_type': 'LSTM',
        'decoder_hidden_size': 100,
        'decoder_num_layers': 1,
        'decoder_bias': True,
        'decoder_dropout': 0.0,
        'decoder_bidirectional': False,
        'decoder_nonlinearity': None,
        'decoder_proj_size': None,
        'adapter_hidden_dims': [50],
        'adapter_out_activation': 'Tanh',
        'predictor_hidden_dims': [],
        'predictor_out_activation': None,
        'max_len': None,
        'optimizer_str': 'Adam',
        'optimizer_kwargs': {'lr': 0.01, 'weight_decay': 1e-05},
        'batch_size': 32,
        'epochs': 10,
        'padding_indicator': -999.0
    }
)

[ ]:

# Targets:
dataset.predictive.targets

TimeSeriesSamples with data:

0 1
sample_idx time_idx
0 0 -3.110475 -3.566948
1 1.528495 -0.653673
2 2.275307 -0.695371
3 4.844060 3.469371
4 4.420301 5.147500
... ... ... ...
99 7 5.994185 6.225290
8 10.913662 5.346697
9 9.558824 7.585175
10 10.194430 5.795619
11 13.774189 8.457336

1573 rows × 2 columns

[ ]:

# Train.
model.fit(dataset);

Preparing data for decoder training...
Preparing data for decoder training DONE.
=== Training stage: 1. Train encoder ===
Epoch: 0, Loss: 68.64124755859375
Epoch: 1, Loss: 35.376345291137696
Epoch: 2, Loss: 21.269977149963378
Epoch: 3, Loss: 17.418056030273437
Epoch: 4, Loss: 13.851958045959472
Epoch: 5, Loss: 8.335217628479004
Epoch: 6, Loss: 6.385376873016358
Epoch: 7, Loss: 5.1125048828125
Epoch: 8, Loss: 4.690582141876221
Epoch: 9, Loss: 4.15880220413208
=== Training stage: 2. Train decoder ===
Epoch: 0, Loss: 29.110622051453316
Epoch: 1, Loss: 4.4235033275153635
Epoch: 2, Loss: 3.954428958957566
Epoch: 3, Loss: 3.8989672662446226
Epoch: 4, Loss: 3.8317033024119103
Epoch: 5, Loss: 3.8341090269664964
Epoch: 6, Loss: 3.717077842797864
Epoch: 7, Loss: 3.7477998285170564
Epoch: 8, Loss: 3.859872080239
Epoch: 9, Loss: 3.8217773019903305

[ ]:

# Predict:

model.predict(dataset, n_future_steps=10)

TimeSeriesSamples with data:

0 1
sample_idx time_idx
0 11 10.618200 8.973257
12 10.728621 9.108867
13 10.735380 9.118745
14 10.728010 9.093669
15 10.738855 9.097306
... ... ... ...
99 17 12.021331 10.417010
18 12.091549 10.484824
19 12.144796 10.523143
20 12.133350 10.510198
21 12.097664 10.484853

1000 rows × 2 columns