MLPRegressor¶
Multi-layer Perceptron for regression.
This model is still work in progress. Here are some features that still need implementing:
learn_one
andpredict_one
just cast the inputdict
to a single row dataframe and then calllearn_many
andpredict_many
respectively. This is very inefficient. - Not all of the optimizers in theoptim
module can be used as they are not all vectorised. - Emerging and disappearing features are not supported. Each instance/batch has to have the same features. - The gradient haven't been numerically checked.
Parameters¶
-
hidden_dims
The dimensions of the hidden layers. For example, specifying
(10, 20)
means that there are two hidden layers with 10 and 20 neurons, respectively. Note that the number of layers the network contains is equal to the number of hidden layers plus two (to account for the input and output layers). -
activations
The activation functions to use at each layer, including the input and output layers. Therefore you need to specify three activation if you specify one hidden layer.
-
loss (optim.losses.Loss) – defaults to
None
Loss function. Defaults to
optim.losses.Squared
. -
optimizer (optim.Optimizer) – defaults to
None
Optimizer. Defaults to
optim.SGD(.01)
. -
seed (int) – defaults to
None
Random number generation seed. Set this for reproducibility.
Attributes¶
-
n_layers
Return the number of layers in the network. The number of layers is equal to the number of hidden layers plus 2. The 2 accounts for the input layer and the output layer.
Examples¶
>>> from river import datasets
>>> from river import evaluate
>>> from river import neural_net as nn
>>> from river import optim
>>> from river import preprocessing as pp
>>> from river import metrics
>>> model = (
... pp.StandardScaler() |
... nn.MLPRegressor(
... hidden_dims=(5,),
... activations=(
... nn.activations.ReLU,
... nn.activations.ReLU,
... nn.activations.Identity
... ),
... optimizer=optim.SGD(1e-3),
... seed=42
... )
... )
>>> dataset = datasets.TrumpApproval()
>>> metric = metrics.MAE()
>>> evaluate.progressive_val_score(dataset, model, metric)
MAE: 1.589827
You can also use this to process mini-batches of data.
>>> model = (
... pp.StandardScaler() |
... nn.MLPRegressor(
... hidden_dims=(10,),
... activations=(
... nn.activations.ReLU,
... nn.activations.ReLU,
... nn.activations.ReLU
... ),
... optimizer=optim.SGD(1e-4),
... seed=42
... )
... )
>>> dataset = datasets.TrumpApproval()
>>> batch_size = 32
>>> for epoch in range(10):
... for xb in pd.read_csv(dataset.path, chunksize=batch_size):
... yb = xb.pop('five_thirty_eight')
... y_pred = model.predict_many(xb)
... model = model.learn_many(xb, yb)
>>> model.predict_many(xb)
five_thirty_eight
992 39.361609
993 46.398536
994 42.094086
995 40.195802
996 40.782954
997 40.839678
998 40.896403
999 48.362659
1000 42.021849
Methods¶
call
Make predictions.
Parameters
- X (pandas.core.frame.DataFrame)
clone
Return a fresh estimator with the same parameters.
The clone has the same parameters but has not been updated with any data. This works by looking at the parameters from the class signature. Each parameter is either - recursively cloned if it's a River classes. - deep-copied via copy.deepcopy
if not. If the calling object is stochastic (i.e. it accepts a seed parameter) and has not been seeded, then the clone will not be idempotent. Indeed, this method's purpose if simply to return a new instance with the same input parameters.
learn_many
Train the network.
Parameters
- X (pandas.core.frame.DataFrame)
- y (pandas.core.frame.DataFrame)
learn_one
Fits to a set of features x
and a real-valued target y
.
Parameters
- x (dict)
- y (numbers.Number)
Returns
Regressor: self
predict_many
predict_one
Predicts the target value of a set of features x
.
Parameters
- x (dict)
Returns
Number: The prediction.