Welcome to Autopycoin

This is a deep learning package based on tensorflow. The models are unofficial implementations.

Available Models

from autopycoin.models import create_interpretable_nbeats, create_interpretable_nbeats, NBEATS, PoolNBEATS

Available Losses

from autopycoin.losses import QuantileLossError, SymetricMeanAbsolutePercentageError

Dataset maker

from autopycoin.dataset import WindowGenerator

How to use autopycoin

pip install autopycoin

How to use Autopycoin

This notebook is a quick start for autopycoin. We will demonstrate its efficiency through a toy example.

Import

As our package is based on tensorflow we need to import it.

import tensorflow as tf

Let's create our data thanks to random_ts which generate a fake time serie based on a trend and seasonality components.

from autopycoin.data import random_ts

tf.random.set_seed(0)

data = random_ts(n_steps=400, # Number of steps (second dimension)
                 trend_degree=2,
                 periods=[10], # We can combine multiple periods, period is the time length for a cyclical function to reproduce a similar output
                 fourier_orders=[10], # higher is this number, more complex is the output
                 trend_mean=0,
                 trend_std=1,
                 seasonality_mean=0,
                 seasonality_std=1, 
                 batch_size=1, # Generate a batch of data (first dimension)
                 n_variables=1, # Number of variables (last dimension)
                 noise=True, # add normal centered noise
                 seed=42)

denoised_data = random_ts(n_steps=400, # Number of steps (second dimension)
                 trend_degree=2,
                 periods=[10], # We can combine multiple periods, period is the time length for a cyclical function to reproduce a similar output
                 fourier_orders=[10], # higher is this number, more complex is the output
                 trend_mean=0,
                 trend_std=1,
                 seasonality_mean=0,
                 seasonality_std=1, 
                 batch_size=1, # Generate a batch of data (first dimension)
                 n_variables=1, # Number of variables (last dimension)
                 noise=False, # add normal centered noise
                 seed=42)

# Let's render it with matplotlib
import matplotlib.pyplot as plt

plt.rcParams['figure.figsize'] = (20, 10)
plt.plot(data[0], c='r', label='noised data')
plt.plot(denoised_data[0], label='denoised data')
plt.title('Generated data')
plt.legend()

<matplotlib.legend.Legend at 0x25a4597b040>

Tensorflow library

Tensorflow use dataset to feed its models. Hence, we create an object WindowGenerator to facilitate the creation of training, validation and training set.

from autopycoin.dataset import WindowGenerator

We first create an instance of Windowgenerator by defining important parameters value. By defining an input_width of 70, we set our input length.

• label_width set our label length.
• shift represents the shift between label and input data. Hence if you don't want label and input overlapping shift has to be equal to label_width. It can help for more complex model like lstm models which can be trained to reconstruct their inputs.
• test_size and valid_size are the numbers of examples to predict. If they are integer, they are the true number of examples, if they are float then it's a proportion of the entire data.

Note: We create lot of examples by shifting by one the window of prediction.

w = WindowGenerator(
        input_width=80,
        label_width=40,
        shift=40,
        test_size=50,
        valid_size=10,
        flat=True,
        batch_size=16,
        preprocessing = lambda x,y: (x, (x,y)) # NBEATS output
    )

We can't use it now as it needs to be initialized by an array, tensor or a dataframe

w = w.from_array(data=data[0], # Has to be 2D array
        input_columns=[0],
        known_columns=[],
        label_columns=[0],
        date_columns=[],)

# Or using dataframe
import pandas as pd

data = pd.DataFrame(data.numpy()[0], columns=['test'])

w = w.from_array(data=data, # Has to be 2D array
        input_columns=['test'],
        known_columns=[],
        label_columns=['test'],
        date_columns=[],)

w.train

<BatchDataset element_spec=(TensorSpec(shape=(None, 80), dtype=tf.float32, name=None), (TensorSpec(shape=(None, 80), dtype=tf.float32, name=None), TensorSpec(shape=(None, 40), dtype=tf.float32, name=None)))>

w.valid

<BatchDataset element_spec=(TensorSpec(shape=(None, 80), dtype=tf.float32, name=None), (TensorSpec(shape=(None, 80), dtype=tf.float32, name=None), TensorSpec(shape=(None, 40), dtype=tf.float32, name=None)))>

w.test

<BatchDataset element_spec=(TensorSpec(shape=(None, 80), dtype=tf.float32, name=None), (TensorSpec(shape=(None, 80), dtype=tf.float32, name=None), TensorSpec(shape=(None, 40), dtype=tf.float32, name=None)))>

It's time for inference

We import the nbeats module where the nbdeats model is defined.

There are currently multiple version of nbeats implemented as create_interpretable_nbeats or create_generic_nbeats. But it is possible to customize it's own nbeats model by subclassing the nbeats Model and Layer.

from autopycoin.models import nbeats

model1 = nbeats.create_interpretable_nbeats(
            label_width=40,
            forecast_periods=[10],
            backcast_periods=[10],
            forecast_fourier_order=[10],
            backcast_fourier_order=[10],
            p_degree=1,
            trend_n_neurons=200,
            seasonality_n_neurons=200,
            drop_rate=0.,
            share=True)

model1.compile(tf.keras.optimizers.Adam(
    learning_rate=0.0015, beta_1=0.9, beta_2=0.999, epsilon=1e-07, amsgrad=True,
    name='Adam'), 
    loss='mse', 
    loss_weights=[1, 1], # In the paper = [0, 1]
    metrics=["mae"])

model1.fit(w.train, validation_data=w.valid, epochs=20)

Epoch 1/20
9/9 [==============================] - 10s 264ms/step - loss: 107.7595 - output_1_loss: 48.8508 - output_2_loss: 58.9087 - output_1_mae: 4.6331 - output_2_mae: 5.6065 - val_loss: 28.2610 - val_output_1_loss: 6.3400 - val_output_2_loss: 21.9210 - val_output_1_mae: 2.0182 - val_output_2_mae: 3.4445
Epoch 2/20
9/9 [==============================] - 0s 28ms/step - loss: 17.0173 - output_1_loss: 5.0084 - output_2_loss: 12.0089 - output_1_mae: 1.7634 - output_2_mae: 2.6280 - val_loss: 9.0327 - val_output_1_loss: 3.2561 - val_output_2_loss: 5.7766 - val_output_1_mae: 1.4348 - val_output_2_mae: 1.9232
Epoch 3/20
9/9 [==============================] - 0s 31ms/step - loss: 7.7175 - output_1_loss: 2.6491 - output_2_loss: 5.0685 - output_1_mae: 1.2985 - output_2_mae: 1.7502 - val_loss: 4.7811 - val_output_1_loss: 1.9094 - val_output_2_loss: 2.8717 - val_output_1_mae: 1.1005 - val_output_2_mae: 1.3431
Epoch 4/20
9/9 [==============================] - 0s 31ms/step - loss: 4.4976 - output_1_loss: 1.9064 - output_2_loss: 2.5912 - output_1_mae: 1.1020 - output_2_mae: 1.2652 - val_loss: 4.1069 - val_output_1_loss: 1.5517 - val_output_2_loss: 2.5552 - val_output_1_mae: 0.9765 - val_output_2_mae: 1.3012
Epoch 5/20
9/9 [==============================] - 0s 30ms/step - loss: 3.3197 - output_1_loss: 1.4996 - output_2_loss: 1.8201 - output_1_mae: 0.9801 - output_2_mae: 1.0888 - val_loss: 3.4868 - val_output_1_loss: 1.3489 - val_output_2_loss: 2.1379 - val_output_1_mae: 0.9232 - val_output_2_mae: 1.1573
Epoch 6/20
9/9 [==============================] - 0s 31ms/step - loss: 2.8041 - output_1_loss: 1.3137 - output_2_loss: 1.4904 - output_1_mae: 0.9224 - output_2_mae: 0.9713 - val_loss: 3.0279 - val_output_1_loss: 1.1370 - val_output_2_loss: 1.8908 - val_output_1_mae: 0.8540 - val_output_2_mae: 1.0975
Epoch 7/20
9/9 [==============================] - 0s 31ms/step - loss: 2.5440 - output_1_loss: 1.2236 - output_2_loss: 1.3204 - output_1_mae: 0.8918 - output_2_mae: 0.9265 - val_loss: 2.9015 - val_output_1_loss: 1.0745 - val_output_2_loss: 1.8270 - val_output_1_mae: 0.8283 - val_output_2_mae: 1.0760
Epoch 8/20
9/9 [==============================] - 0s 30ms/step - loss: 2.4074 - output_1_loss: 1.1735 - output_2_loss: 1.2339 - output_1_mae: 0.8744 - output_2_mae: 0.8928 - val_loss: 2.8078 - val_output_1_loss: 1.0035 - val_output_2_loss: 1.8042 - val_output_1_mae: 0.8019 - val_output_2_mae: 1.0765
Epoch 9/20
9/9 [==============================] - 0s 33ms/step - loss: 2.2903 - output_1_loss: 1.1395 - output_2_loss: 1.1508 - output_1_mae: 0.8615 - output_2_mae: 0.8617 - val_loss: 2.7515 - val_output_1_loss: 0.9931 - val_output_2_loss: 1.7584 - val_output_1_mae: 0.7983 - val_output_2_mae: 1.0654
Epoch 10/20
9/9 [==============================] - 0s 42ms/step - loss: 2.2125 - output_1_loss: 1.1157 - output_2_loss: 1.0968 - output_1_mae: 0.8529 - output_2_mae: 0.8449 - val_loss: 2.7506 - val_output_1_loss: 0.9789 - val_output_2_loss: 1.7718 - val_output_1_mae: 0.7917 - val_output_2_mae: 1.0708
Epoch 11/20
9/9 [==============================] - 0s 34ms/step - loss: 2.1483 - output_1_loss: 1.0973 - output_2_loss: 1.0510 - output_1_mae: 0.8459 - output_2_mae: 0.8261 - val_loss: 2.6904 - val_output_1_loss: 0.9565 - val_output_2_loss: 1.7339 - val_output_1_mae: 0.7839 - val_output_2_mae: 1.0580
Epoch 12/20
9/9 [==============================] - 0s 31ms/step - loss: 2.0970 - output_1_loss: 1.0808 - output_2_loss: 1.0162 - output_1_mae: 0.8401 - output_2_mae: 0.8130 - val_loss: 2.6640 - val_output_1_loss: 0.9480 - val_output_2_loss: 1.7161 - val_output_1_mae: 0.7820 - val_output_2_mae: 1.0537
Epoch 13/20
9/9 [==============================] - 0s 29ms/step - loss: 2.0540 - output_1_loss: 1.0680 - output_2_loss: 0.9860 - output_1_mae: 0.8356 - output_2_mae: 0.8008 - val_loss: 2.6468 - val_output_1_loss: 0.9326 - val_output_2_loss: 1.7142 - val_output_1_mae: 0.7754 - val_output_2_mae: 1.0538
Epoch 14/20
9/9 [==============================] - 0s 29ms/step - loss: 2.0171 - output_1_loss: 1.0562 - output_2_loss: 0.9609 - output_1_mae: 0.8310 - output_2_mae: 0.7910 - val_loss: 2.6439 - val_output_1_loss: 0.9350 - val_output_2_loss: 1.7090 - val_output_1_mae: 0.7752 - val_output_2_mae: 1.0473
Epoch 15/20
9/9 [==============================] - 0s 29ms/step - loss: 1.9864 - output_1_loss: 1.0468 - output_2_loss: 0.9396 - output_1_mae: 0.8273 - output_2_mae: 0.7809 - val_loss: 2.6282 - val_output_1_loss: 0.9284 - val_output_2_loss: 1.6998 - val_output_1_mae: 0.7728 - val_output_2_mae: 1.0454
Epoch 16/20
9/9 [==============================] - 0s 30ms/step - loss: 1.9587 - output_1_loss: 1.0372 - output_2_loss: 0.9215 - output_1_mae: 0.8237 - output_2_mae: 0.7729 - val_loss: 2.6171 - val_output_1_loss: 0.9228 - val_output_2_loss: 1.6943 - val_output_1_mae: 0.7705 - val_output_2_mae: 1.0443
Epoch 17/20
9/9 [==============================] - 0s 36ms/step - loss: 1.9331 - output_1_loss: 1.0287 - output_2_loss: 0.9045 - output_1_mae: 0.8206 - output_2_mae: 0.7672 - val_loss: 2.6150 - val_output_1_loss: 0.9207 - val_output_2_loss: 1.6943 - val_output_1_mae: 0.7690 - val_output_2_mae: 1.0435
Epoch 18/20
9/9 [==============================] - 0s 31ms/step - loss: 1.9092 - output_1_loss: 1.0210 - output_2_loss: 0.8882 - output_1_mae: 0.8180 - output_2_mae: 0.7597 - val_loss: 2.5927 - val_output_1_loss: 0.9081 - val_output_2_loss: 1.6846 - val_output_1_mae: 0.7642 - val_output_2_mae: 1.0409
Epoch 19/20
9/9 [==============================] - 0s 30ms/step - loss: 1.8908 - output_1_loss: 1.0147 - output_2_loss: 0.8761 - output_1_mae: 0.8155 - output_2_mae: 0.7552 - val_loss: 2.6052 - val_output_1_loss: 0.9144 - val_output_2_loss: 1.6908 - val_output_1_mae: 0.7666 - val_output_2_mae: 1.0435
Epoch 20/20
9/9 [==============================] - 0s 31ms/step - loss: 1.8719 - output_1_loss: 1.0081 - output_2_loss: 0.8638 - output_1_mae: 0.8130 - output_2_mae: 0.7485 - val_loss: 2.6011 - val_output_1_loss: 0.9084 - val_output_2_loss: 1.6927 - val_output_1_mae: 0.7635 - val_output_2_mae: 1.0453





<keras.callbacks.History at 0x25a4268de20>

from autopycoin.models import nbeats

model2 = nbeats.create_generic_nbeats(
            label_width=40,
            g_forecast_neurons=16, 
            g_backcast_neurons=16, 
            n_neurons=16, 
            n_blocks=3,
            n_stacks=2,
            drop_rate=0.1,
            share=True)

model2.compile(tf.keras.optimizers.Adam(
    learning_rate=0.0015, beta_1=0.9, beta_2=0.999, epsilon=1e-07, amsgrad=True,
    name='Adam'), loss='mse', 
    loss_weights=[1, 1], # In the paper = [0, 1]
    metrics=["mae"])

model2.fit(w.train, validation_data=w.valid, epochs=20)

Epoch 1/20
9/9 [==============================] - 10s 273ms/step - loss: 42.6128 - output_1_loss: 20.5989 - output_2_loss: 22.0138 - output_1_mae: 2.9896 - output_2_mae: 3.3102 - val_loss: 35.9356 - val_output_1_loss: 17.8830 - val_output_2_loss: 18.0526 - val_output_1_mae: 2.7465 - val_output_2_mae: 2.8812
Epoch 2/20
9/9 [==============================] - 0s 34ms/step - loss: 33.6867 - output_1_loss: 16.9436 - output_2_loss: 16.7431 - output_1_mae: 2.5379 - output_2_mae: 2.6609 - val_loss: 32.3713 - val_output_1_loss: 16.0248 - val_output_2_loss: 16.3464 - val_output_1_mae: 2.5090 - val_output_2_mae: 2.4740
Epoch 3/20
9/9 [==============================] - 0s 35ms/step - loss: 31.6279 - output_1_loss: 16.0887 - output_2_loss: 15.5393 - output_1_mae: 2.4137 - output_2_mae: 2.4602 - val_loss: 31.0422 - val_output_1_loss: 15.9003 - val_output_2_loss: 15.1419 - val_output_1_mae: 2.4889 - val_output_2_mae: 2.3757
Epoch 4/20
9/9 [==============================] - 0s 38ms/step - loss: 30.7158 - output_1_loss: 15.7325 - output_2_loss: 14.9833 - output_1_mae: 2.3807 - output_2_mae: 2.4034 - val_loss: 30.2672 - val_output_1_loss: 15.5197 - val_output_2_loss: 14.7476 - val_output_1_mae: 2.4714 - val_output_2_mae: 2.3704
Epoch 5/20
9/9 [==============================] - 0s 36ms/step - loss: 29.9507 - output_1_loss: 15.4581 - output_2_loss: 14.4926 - output_1_mae: 2.3601 - output_2_mae: 2.3899 - val_loss: 29.1492 - val_output_1_loss: 15.3752 - val_output_2_loss: 13.7740 - val_output_1_mae: 2.4536 - val_output_2_mae: 2.3348
Epoch 6/20
9/9 [==============================] - 0s 37ms/step - loss: 28.9890 - output_1_loss: 15.2142 - output_2_loss: 13.7748 - output_1_mae: 2.3620 - output_2_mae: 2.3576 - val_loss: 27.7325 - val_output_1_loss: 14.7231 - val_output_2_loss: 13.0094 - val_output_1_mae: 2.4348 - val_output_2_mae: 2.3628
Epoch 7/20
9/9 [==============================] - 0s 37ms/step - loss: 27.9566 - output_1_loss: 14.8224 - output_2_loss: 13.1342 - output_1_mae: 2.3521 - output_2_mae: 2.3442 - val_loss: 27.0617 - val_output_1_loss: 14.4031 - val_output_2_loss: 12.6587 - val_output_1_mae: 2.4278 - val_output_2_mae: 2.3942
Epoch 8/20
9/9 [==============================] - 0s 38ms/step - loss: 26.8509 - output_1_loss: 14.4216 - output_2_loss: 12.4294 - output_1_mae: 2.3607 - output_2_mae: 2.3420 - val_loss: 26.5709 - val_output_1_loss: 14.1705 - val_output_2_loss: 12.4004 - val_output_1_mae: 2.4366 - val_output_2_mae: 2.4046
Epoch 9/20
9/9 [==============================] - 0s 45ms/step - loss: 25.5537 - output_1_loss: 13.9055 - output_2_loss: 11.6482 - output_1_mae: 2.3517 - output_2_mae: 2.2923 - val_loss: 25.4198 - val_output_1_loss: 13.6695 - val_output_2_loss: 11.7503 - val_output_1_mae: 2.4041 - val_output_2_mae: 2.2838
Epoch 10/20
9/9 [==============================] - 0s 39ms/step - loss: 24.2401 - output_1_loss: 13.3442 - output_2_loss: 10.8959 - output_1_mae: 2.3349 - output_2_mae: 2.2453 - val_loss: 24.0776 - val_output_1_loss: 12.8328 - val_output_2_loss: 11.2448 - val_output_1_mae: 2.4019 - val_output_2_mae: 2.3412
Epoch 11/20
9/9 [==============================] - 0s 39ms/step - loss: 22.8726 - output_1_loss: 12.5804 - output_2_loss: 10.2922 - output_1_mae: 2.3036 - output_2_mae: 2.2185 - val_loss: 22.0797 - val_output_1_loss: 11.8361 - val_output_2_loss: 10.2435 - val_output_1_mae: 2.3351 - val_output_2_mae: 2.2578
Epoch 12/20
9/9 [==============================] - 0s 37ms/step - loss: 21.3568 - output_1_loss: 11.8595 - output_2_loss: 9.4973 - output_1_mae: 2.2715 - output_2_mae: 2.1294 - val_loss: 20.7301 - val_output_1_loss: 11.2053 - val_output_2_loss: 9.5248 - val_output_1_mae: 2.2682 - val_output_2_mae: 2.1940
Epoch 13/20
9/9 [==============================] - 0s 37ms/step - loss: 19.9969 - output_1_loss: 10.9618 - output_2_loss: 9.0351 - output_1_mae: 2.2187 - output_2_mae: 2.1088 - val_loss: 19.5411 - val_output_1_loss: 10.4076 - val_output_2_loss: 9.1335 - val_output_1_mae: 2.2347 - val_output_2_mae: 2.1147
Epoch 14/20
9/9 [==============================] - 0s 39ms/step - loss: 18.2909 - output_1_loss: 10.0314 - output_2_loss: 8.2595 - output_1_mae: 2.1757 - output_2_mae: 2.0099 - val_loss: 17.4031 - val_output_1_loss: 9.2104 - val_output_2_loss: 8.1927 - val_output_1_mae: 2.1068 - val_output_2_mae: 1.9743
Epoch 15/20
9/9 [==============================] - 0s 38ms/step - loss: 16.9036 - output_1_loss: 9.1755 - output_2_loss: 7.7281 - output_1_mae: 2.1134 - output_2_mae: 1.9526 - val_loss: 15.8572 - val_output_1_loss: 8.1932 - val_output_2_loss: 7.6641 - val_output_1_mae: 2.0244 - val_output_2_mae: 1.9575
Epoch 16/20
9/9 [==============================] - 0s 37ms/step - loss: 15.6807 - output_1_loss: 8.4435 - output_2_loss: 7.2372 - output_1_mae: 2.0389 - output_2_mae: 1.8990 - val_loss: 15.5640 - val_output_1_loss: 8.1138 - val_output_2_loss: 7.4503 - val_output_1_mae: 2.0440 - val_output_2_mae: 1.9175
Epoch 17/20
9/9 [==============================] - 0s 39ms/step - loss: 14.5375 - output_1_loss: 7.7209 - output_2_loss: 6.8166 - output_1_mae: 1.9561 - output_2_mae: 1.8343 - val_loss: 14.9077 - val_output_1_loss: 7.8515 - val_output_2_loss: 7.0563 - val_output_1_mae: 2.0159 - val_output_2_mae: 1.8616
Epoch 18/20
9/9 [==============================] - 0s 39ms/step - loss: 13.4057 - output_1_loss: 7.0658 - output_2_loss: 6.3399 - output_1_mae: 1.9002 - output_2_mae: 1.7726 - val_loss: 12.9685 - val_output_1_loss: 6.5909 - val_output_2_loss: 6.3776 - val_output_1_mae: 1.8825 - val_output_2_mae: 1.8605
Epoch 19/20
9/9 [==============================] - 0s 37ms/step - loss: 11.6256 - output_1_loss: 6.0771 - output_2_loss: 5.5485 - output_1_mae: 1.7945 - output_2_mae: 1.6762 - val_loss: 11.7937 - val_output_1_loss: 6.1221 - val_output_2_loss: 5.6716 - val_output_1_mae: 1.8922 - val_output_2_mae: 1.7786
Epoch 20/20
9/9 [==============================] - 0s 39ms/step - loss: 11.2368 - output_1_loss: 5.8081 - output_2_loss: 5.4287 - output_1_mae: 1.7738 - output_2_mae: 1.6614 - val_loss: 10.8702 - val_output_1_loss: 5.4575 - val_output_2_loss: 5.4126 - val_output_1_mae: 1.7811 - val_output_2_mae: 1.7010





<keras.callbacks.History at 0x25a4c7eb3d0>

Evaluation

model1.evaluate(w.test)

4/4 [==============================] - 0s 15ms/step - loss: 2.6927 - output_1_loss: 1.2548 - output_2_loss: 1.4379 - output_1_mae: 0.8934 - output_2_mae: 0.9630





[2.692697048187256,
 1.2548267841339111,
 1.4378703832626343,
 0.8934133052825928,
 0.9630153179168701]

model2.evaluate(w.test)

4/4 [==============================] - 0s 20ms/step - loss: 10.9378 - output_1_loss: 5.5514 - output_2_loss: 5.3864 - output_1_mae: 1.7609 - output_2_mae: 1.6672





[10.937788963317871,
 5.551424503326416,
 5.386363506317139,
 1.7609360218048096,
 1.66715669631958]

Plot

Let plot some previsions

import matplotlib.pyplot as plt

iterator = iter(w.train)
x, y = iterator.get_next()

input_width = 80

plt.plot(range(input_width, input_width + 40), model1.predict(x)[1].values[0], label='forecast')
# Usefull only if stack = True
plt.plot(range(input_width), model1.predict(x)[0].values[0], label='backcast')
plt.plot(range(input_width, input_width + 40), y[1][0], label='labels')
plt.plot(range(input_width), x[0], label='inputs')
plt.legend()

<matplotlib.legend.Legend at 0x25a5293e4f0>

Production

prod = w.production(data)
plt.plot(model1.predict(prod)[0].values[1])

[<matplotlib.lines.Line2D at 0x25a5363f2b0>]

A pool is better than a single model

Generally, a pool is better than a single model : It is a regularization method. In this toy example, it won't because we are not optimizing the training phase.

from autopycoin.models import PoolNBEATS

# Create a callable which define label_width and create a Tensorflow Model.
model = lambda label_width: nbeats.create_interpretable_nbeats(
            label_width=label_width,
            forecast_periods=[10],
            backcast_periods=[10],
            forecast_fourier_order=[10],
            backcast_fourier_order=[10],
            p_degree=1,
            trend_n_neurons=200,
            seasonality_n_neurons=200,
            drop_rate=0.,
            share=True)

# Define an aggregation method with fn_agg, the number of models with n_models and the 
# label width with label_width. label_width is mandatory if you use callables in nbeats_models !
model = PoolNBEATS(n_models=3, label_width=40, nbeats_models=model,
            fn_agg=tf.reduce_mean)

model.compile(tf.keras.optimizers.Adam(
    learning_rate=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-07, amsgrad=True,
    name='Adam'), loss=['mse', 'mse'], metrics=['mae'])

# The fit is printing 6 different maes results: 2 for each model (backcast, forecast).
# The loss is aggregated then it is printing 4 loss results: 1 for the entire model, 
model.fit(w.train, validation_data=w.valid, epochs=20)

Epoch 1/20
9/9 [==============================] - 43s 829ms/step - loss: 129.9779 - output_1_1_loss: 54.8111 - output_1_2_loss: 58.6071 - output_2_1_loss: 16.5597 - output_1_1_mae: 5.1632 - output_1_2_mae: 5.4097 - output_2_1_mae: 3.1162 - output_2_2_mae: 3.6870 - output_3_1_mae: 4.8781 - output_3_2_mae: 4.5370 - val_loss: 46.9568 - val_output_1_1_loss: 14.0975 - val_output_1_2_loss: 24.6893 - val_output_2_1_loss: 8.1700 - val_output_1_1_mae: 2.9376 - val_output_1_2_mae: 3.7093 - val_output_2_1_mae: 2.3219 - val_output_2_2_mae: 2.5332 - val_output_3_1_mae: 3.8348 - val_output_3_2_mae: 3.3322
Epoch 2/20
9/9 [==============================] - 1s 74ms/step - loss: 25.4214 - output_1_1_loss: 7.0674 - output_1_2_loss: 14.4327 - output_2_1_loss: 3.9213 - output_1_1_mae: 2.0938 - output_1_2_mae: 2.8991 - output_2_1_mae: 1.5464 - output_2_2_mae: 1.9865 - output_3_1_mae: 2.3966 - output_3_2_mae: 2.8036 - val_loss: 17.6809 - val_output_1_1_loss: 4.1897 - val_output_1_2_loss: 11.3001 - val_output_2_1_loss: 2.1912 - val_output_1_1_mae: 1.6525 - val_output_1_2_mae: 2.6448 - val_output_2_1_mae: 1.1933 - val_output_2_2_mae: 1.8943 - val_output_3_1_mae: 1.8263 - val_output_3_2_mae: 2.2586
Epoch 3/20
9/9 [==============================] - 1s 86ms/step - loss: 11.8408 - output_1_1_loss: 3.4665 - output_1_2_loss: 6.3762 - output_2_1_loss: 1.9981 - output_1_1_mae: 1.4820 - output_1_2_mae: 1.9251 - output_2_1_mae: 1.1135 - output_2_2_mae: 1.4251 - output_3_1_mae: 1.5350 - output_3_2_mae: 1.9391 - val_loss: 8.2325 - val_output_1_1_loss: 2.1980 - val_output_1_2_loss: 4.5575 - val_output_2_1_loss: 1.4771 - val_output_1_1_mae: 1.2003 - val_output_1_2_mae: 1.7161 - val_output_2_1_mae: 0.9559 - val_output_2_2_mae: 1.8211 - val_output_3_1_mae: 1.4079 - val_output_3_2_mae: 1.7984
Epoch 4/20
9/9 [==============================] - 1s 90ms/step - loss: 6.9427 - output_1_1_loss: 2.2510 - output_1_2_loss: 3.3005 - output_2_1_loss: 1.3912 - output_1_1_mae: 1.2012 - output_1_2_mae: 1.4370 - output_2_1_mae: 0.9312 - output_2_2_mae: 1.2682 - output_3_1_mae: 1.2522 - output_3_2_mae: 1.3777 - val_loss: 5.8157 - val_output_1_1_loss: 1.7247 - val_output_1_2_loss: 3.0949 - val_output_2_1_loss: 0.9962 - val_output_1_1_mae: 1.0654 - val_output_1_2_mae: 1.3998 - val_output_2_1_mae: 0.7977 - val_output_2_2_mae: 1.6829 - val_output_3_1_mae: 1.1519 - val_output_3_2_mae: 1.4012
Epoch 5/20
9/9 [==============================] - 1s 86ms/step - loss: 5.0599 - output_1_1_loss: 1.7671 - output_1_2_loss: 2.1756 - output_2_1_loss: 1.1172 - output_1_1_mae: 1.0628 - output_1_2_mae: 1.1825 - output_2_1_mae: 0.8378 - output_2_2_mae: 1.4185 - output_3_1_mae: 1.0667 - output_3_2_mae: 1.1119 - val_loss: 4.7978 - val_output_1_1_loss: 1.3100 - val_output_1_2_loss: 2.6150 - val_output_2_1_loss: 0.8728 - val_output_1_1_mae: 0.9288 - val_output_1_2_mae: 1.3082 - val_output_2_1_mae: 0.7442 - val_output_2_2_mae: 1.0579 - val_output_3_1_mae: 0.9992 - val_output_3_2_mae: 1.2378
Epoch 6/20
9/9 [==============================] - 1s 87ms/step - loss: 4.2656 - output_1_1_loss: 1.4522 - output_1_2_loss: 1.7817 - output_2_1_loss: 1.0316 - output_1_1_mae: 0.9619 - output_1_2_mae: 1.0819 - output_2_1_mae: 0.8059 - output_2_2_mae: 1.3582 - output_3_1_mae: 0.9386 - output_3_2_mae: 1.0177 - val_loss: 3.9572 - val_output_1_1_loss: 1.1128 - val_output_1_2_loss: 2.0426 - val_output_2_1_loss: 0.8018 - val_output_1_1_mae: 0.8523 - val_output_1_2_mae: 1.1461 - val_output_2_1_mae: 0.7204 - val_output_2_2_mae: 1.6967 - val_output_3_1_mae: 0.8754 - val_output_3_2_mae: 1.1660
Epoch 7/20
9/9 [==============================] - 1s 100ms/step - loss: 3.8616 - output_1_1_loss: 1.3152 - output_1_2_loss: 1.5499 - output_2_1_loss: 0.9965 - output_1_1_mae: 0.9157 - output_1_2_mae: 1.0036 - output_2_1_mae: 0.7941 - output_2_2_mae: 1.1896 - output_3_1_mae: 0.8731 - output_3_2_mae: 0.9514 - val_loss: 3.8208 - val_output_1_1_loss: 1.0152 - val_output_1_2_loss: 2.0275 - val_output_2_1_loss: 0.7781 - val_output_1_1_mae: 0.8203 - val_output_1_2_mae: 1.1385 - val_output_2_1_mae: 0.7071 - val_output_2_2_mae: 1.2886 - val_output_3_1_mae: 0.8086 - val_output_3_2_mae: 1.1600
Epoch 8/20
9/9 [==============================] - 1s 102ms/step - loss: 3.5669 - output_1_1_loss: 1.2355 - output_1_2_loss: 1.3619 - output_2_1_loss: 0.9696 - output_1_1_mae: 0.8884 - output_1_2_mae: 0.9339 - output_2_1_mae: 0.7838 - output_2_2_mae: 0.9751 - output_3_1_mae: 0.8383 - output_3_2_mae: 0.9076 - val_loss: 3.5757 - val_output_1_1_loss: 0.9730 - val_output_1_2_loss: 1.8360 - val_output_2_1_loss: 0.7667 - val_output_1_1_mae: 0.8063 - val_output_1_2_mae: 1.0796 - val_output_2_1_mae: 0.7042 - val_output_2_2_mae: 1.0481 - val_output_3_1_mae: 0.7874 - val_output_3_2_mae: 1.1607
Epoch 9/20
9/9 [==============================] - 1s 90ms/step - loss: 3.3985 - output_1_1_loss: 1.1943 - output_1_2_loss: 1.2635 - output_2_1_loss: 0.9407 - output_1_1_mae: 0.8756 - output_1_2_mae: 0.8978 - output_2_1_mae: 0.7717 - output_2_2_mae: 0.8747 - output_3_1_mae: 0.8182 - output_3_2_mae: 0.8798 - val_loss: 3.6140 - val_output_1_1_loss: 0.9473 - val_output_1_2_loss: 1.9000 - val_output_2_1_loss: 0.7666 - val_output_1_1_mae: 0.7956 - val_output_1_2_mae: 1.0952 - val_output_2_1_mae: 0.7048 - val_output_2_2_mae: 1.0705 - val_output_3_1_mae: 0.7708 - val_output_3_2_mae: 1.1711
Epoch 10/20
9/9 [==============================] - 1s 88ms/step - loss: 3.2628 - output_1_1_loss: 1.1613 - output_1_2_loss: 1.1828 - output_2_1_loss: 0.9187 - output_1_1_mae: 0.8640 - output_1_2_mae: 0.8752 - output_2_1_mae: 0.7643 - output_2_2_mae: 0.8340 - output_3_1_mae: 0.8036 - output_3_2_mae: 0.8597 - val_loss: 3.6254 - val_output_1_1_loss: 0.9419 - val_output_1_2_loss: 1.9278 - val_output_2_1_loss: 0.7556 - val_output_1_1_mae: 0.7926 - val_output_1_2_mae: 1.1058 - val_output_2_1_mae: 0.7003 - val_output_2_2_mae: 1.0737 - val_output_3_1_mae: 0.7666 - val_output_3_2_mae: 1.1687
Epoch 11/20
9/9 [==============================] - 1s 86ms/step - loss: 3.1657 - output_1_1_loss: 1.1376 - output_1_2_loss: 1.1247 - output_2_1_loss: 0.9034 - output_1_1_mae: 0.8553 - output_1_2_mae: 0.8621 - output_2_1_mae: 0.7592 - output_2_2_mae: 0.8126 - output_3_1_mae: 0.7948 - output_3_2_mae: 0.8450 - val_loss: 3.5362 - val_output_1_1_loss: 0.9206 - val_output_1_2_loss: 1.8659 - val_output_2_1_loss: 0.7497 - val_output_1_1_mae: 0.7823 - val_output_1_2_mae: 1.0899 - val_output_2_1_mae: 0.6958 - val_output_2_2_mae: 1.0339 - val_output_3_1_mae: 0.7653 - val_output_3_2_mae: 1.1516
Epoch 12/20
9/9 [==============================] - 1s 87ms/step - loss: 3.1066 - output_1_1_loss: 1.1205 - output_1_2_loss: 1.0941 - output_2_1_loss: 0.8920 - output_1_1_mae: 0.8487 - output_1_2_mae: 0.8511 - output_2_1_mae: 0.7547 - output_2_2_mae: 0.7885 - output_3_1_mae: 0.7868 - output_3_2_mae: 0.8316 - val_loss: 3.4825 - val_output_1_1_loss: 0.9152 - val_output_1_2_loss: 1.8193 - val_output_2_1_loss: 0.7479 - val_output_1_1_mae: 0.7806 - val_output_1_2_mae: 1.0748 - val_output_2_1_mae: 0.6967 - val_output_2_2_mae: 1.0604 - val_output_3_1_mae: 0.7619 - val_output_3_2_mae: 1.1315
Epoch 13/20
9/9 [==============================] - 1s 89ms/step - loss: 3.0458 - output_1_1_loss: 1.1044 - output_1_2_loss: 1.0592 - output_2_1_loss: 0.8822 - output_1_1_mae: 0.8432 - output_1_2_mae: 0.8309 - output_2_1_mae: 0.7507 - output_2_2_mae: 0.7824 - output_3_1_mae: 0.7802 - output_3_2_mae: 0.8183 - val_loss: 3.5015 - val_output_1_1_loss: 0.9050 - val_output_1_2_loss: 1.8511 - val_output_2_1_loss: 0.7454 - val_output_1_1_mae: 0.7771 - val_output_1_2_mae: 1.0792 - val_output_2_1_mae: 0.6925 - val_output_2_2_mae: 1.0433 - val_output_3_1_mae: 0.7592 - val_output_3_2_mae: 1.1145
Epoch 14/20
9/9 [==============================] - 1s 88ms/step - loss: 2.9937 - output_1_1_loss: 1.0914 - output_1_2_loss: 1.0284 - output_2_1_loss: 0.8740 - output_1_1_mae: 0.8393 - output_1_2_mae: 0.8118 - output_2_1_mae: 0.7473 - output_2_2_mae: 0.7690 - output_3_1_mae: 0.7750 - output_3_2_mae: 0.8043 - val_loss: 3.4499 - val_output_1_1_loss: 0.9009 - val_output_1_2_loss: 1.8061 - val_output_2_1_loss: 0.7429 - val_output_1_1_mae: 0.7749 - val_output_1_2_mae: 1.0674 - val_output_2_1_mae: 0.6918 - val_output_2_2_mae: 1.0505 - val_output_3_1_mae: 0.7601 - val_output_3_2_mae: 1.1067
Epoch 15/20
9/9 [==============================] - 1s 89ms/step - loss: 2.9859 - output_1_1_loss: 1.0803 - output_1_2_loss: 1.0397 - output_2_1_loss: 0.8658 - output_1_1_mae: 0.8357 - output_1_2_mae: 0.8106 - output_2_1_mae: 0.7445 - output_2_2_mae: 0.7634 - output_3_1_mae: 0.7699 - output_3_2_mae: 0.7922 - val_loss: 3.4064 - val_output_1_1_loss: 0.8995 - val_output_1_2_loss: 1.7671 - val_output_2_1_loss: 0.7399 - val_output_1_1_mae: 0.7744 - val_output_1_2_mae: 1.0629 - val_output_2_1_mae: 0.6911 - val_output_2_2_mae: 1.0509 - val_output_3_1_mae: 0.7584 - val_output_3_2_mae: 1.1074
Epoch 16/20
9/9 [==============================] - 1s 86ms/step - loss: 2.9870 - output_1_1_loss: 1.0734 - output_1_2_loss: 1.0549 - output_2_1_loss: 0.8587 - output_1_1_mae: 0.8338 - output_1_2_mae: 0.8185 - output_2_1_mae: 0.7419 - output_2_2_mae: 0.7561 - output_3_1_mae: 0.7658 - output_3_2_mae: 0.7829 - val_loss: 3.6377 - val_output_1_1_loss: 0.9016 - val_output_1_2_loss: 2.0002 - val_output_2_1_loss: 0.7359 - val_output_1_1_mae: 0.7739 - val_output_1_2_mae: 1.1345 - val_output_2_1_mae: 0.6889 - val_output_2_2_mae: 1.0411 - val_output_3_1_mae: 0.7554 - val_output_3_2_mae: 1.1253
Epoch 17/20
9/9 [==============================] - 1s 90ms/step - loss: 2.9628 - output_1_1_loss: 1.0636 - output_1_2_loss: 1.0463 - output_2_1_loss: 0.8529 - output_1_1_mae: 0.8296 - output_1_2_mae: 0.8272 - output_2_1_mae: 0.7396 - output_2_2_mae: 0.7494 - output_3_1_mae: 0.7621 - output_3_2_mae: 0.7780 - val_loss: 3.8666 - val_output_1_1_loss: 0.9005 - val_output_1_2_loss: 2.2291 - val_output_2_1_loss: 0.7369 - val_output_1_1_mae: 0.7731 - val_output_1_2_mae: 1.2023 - val_output_2_1_mae: 0.6899 - val_output_2_2_mae: 1.0530 - val_output_3_1_mae: 0.7531 - val_output_3_2_mae: 1.1594
Epoch 18/20
9/9 [==============================] - 1s 93ms/step - loss: 3.0166 - output_1_1_loss: 1.0595 - output_1_2_loss: 1.1103 - output_2_1_loss: 0.8468 - output_1_1_mae: 0.8277 - output_1_2_mae: 0.8641 - output_2_1_mae: 0.7373 - output_2_2_mae: 0.7442 - output_3_1_mae: 0.7590 - output_3_2_mae: 0.7782 - val_loss: 3.4706 - val_output_1_1_loss: 0.8868 - val_output_1_2_loss: 1.8534 - val_output_2_1_loss: 0.7304 - val_output_1_1_mae: 0.7681 - val_output_1_2_mae: 1.0926 - val_output_2_1_mae: 0.6866 - val_output_2_2_mae: 1.0440 - val_output_3_1_mae: 0.7500 - val_output_3_2_mae: 1.2009
Epoch 19/20
9/9 [==============================] - 1s 92ms/step - loss: 3.1871 - output_1_1_loss: 1.0524 - output_1_2_loss: 1.2934 - output_2_1_loss: 0.8412 - output_1_1_mae: 0.8256 - output_1_2_mae: 0.9245 - output_2_1_mae: 0.7353 - output_2_2_mae: 0.7387 - output_3_1_mae: 0.7563 - output_3_2_mae: 0.7871 - val_loss: 3.6106 - val_output_1_1_loss: 0.8820 - val_output_1_2_loss: 1.9958 - val_output_2_1_loss: 0.7328 - val_output_1_1_mae: 0.7685 - val_output_1_2_mae: 1.1143 - val_output_2_1_mae: 0.6885 - val_output_2_2_mae: 1.0465 - val_output_3_1_mae: 0.7469 - val_output_3_2_mae: 1.1983
Epoch 20/20
9/9 [==============================] - 1s 87ms/step - loss: 3.1260 - output_1_1_loss: 1.0483 - output_1_2_loss: 1.2413 - output_2_1_loss: 0.8363 - output_1_1_mae: 0.8245 - output_1_2_mae: 0.8869 - output_2_1_mae: 0.7333 - output_2_2_mae: 0.7342 - output_3_1_mae: 0.7541 - output_3_2_mae: 0.8061 - val_loss: 3.9547 - val_output_1_1_loss: 0.8803 - val_output_1_2_loss: 2.3441 - val_output_2_1_loss: 0.7303 - val_output_1_1_mae: 0.7685 - val_output_1_2_mae: 1.2228 - val_output_2_1_mae: 0.6864 - val_output_2_2_mae: 1.0471 - val_output_3_1_mae: 0.7449 - val_output_3_2_mae: 1.1286





<keras.callbacks.History at 0x25a5f4cb4f0>

import matplotlib.pyplot as plt

iterator = iter(w.train)
x, y = iterator.get_next()

input_width = 80

plt.plot(range(input_width, input_width + 40), model.predict(x)[1][0], label='forecast')
plt.plot(range(input_width, input_width + 40), y[1][0], label='labels')
plt.plot(range(input_width), x[0], label='inputs')
plt.legend()

<matplotlib.legend.Legend at 0x25a852bb2e0>