Model

class Model(layers: Sequence[bluebird.layers.layer.Layer])

Bases: object

Base model class.

All other models inherit from it.

Example:

class CustomModel(Model):
    def __init__(self, layers: Sequence[Layer]):
        super.__init__()

    def build(self,
            iterator: DataIterator = BatchIterator(),
            loss: Loss = MSE(),
            optimizer: Optimizer = SGD()) -> None:
        super.build()

    def step(self, batch) -> float:
        ...

        return self.loss.loss(predicted, batch.targets)

__init__(layers: Sequence[bluebird.layers.layer.Layer]) → None

Initalizes the object.

Parameters

layers (Sequence[Layer]) – sequence of layers of the newtwork

Methods

__init__(layers)	Initalizes the object.
backward(grad)	Backward propagates through the network.
build([loss, optimizer])	Used to build the model.
fit(loader, num_epochs)	Used to train the model.
forward(inputs)	Forward propagates through the network.
get_layers()	Returns reversed layers.
get_params_and_grads()	Returns parameters and gradients for each layer.
load(path)	Load the model from a file.
predict(inputs)	Used to predict values after you finished the training.
save([path, filename])	Save the model to a file.
step(batch)	Step function is called during each training step.
summary(shape)	Prints the information about the model (Layes, shapes, params).

backward(grad: numpy.ndarray) → numpy.ndarray

Backward propagates through the network.

Parameters

inputs (Tensor) – Gradient that you get from the loss function

Returns

Gradient from the first layer

Return type

Tensor

build(loss: bluebird.loss.loss.Loss = <bluebird.loss.mse.MSE object>, optimizer: bluebird.optimizers.optimizer.Optimizer = <bluebird.optimizers.sgd.SGD object>) → None

Used to build the model.

Defines aditional parameters and initializes the weights.

Parameters

• loss (Loss, otpional) – defines loss function Defaults to MSE()

• optimizer (Optimizer, optional) – defines how the weights should be updated Defaults to SGD()

fit(loader: bluebird.dataloader.dataloaderbase.DataLoaderBase, num_epochs: int) → None

Used to train the model.

Parameters

• loader (DataLoaderBase) – data loader type object used to load data for training

• num_epochs (int) – number of epochs you want to train

forward(inputs: numpy.ndarray) → numpy.ndarray

Forward propagates through the network.

Parameters

inputs (Tensor) – Network input

Returns

Network output

Return type

Tensor

get_layers() → Iterator[bluebird.layers.layer.Layer]

Returns reversed layers.

Returns

List of layer objects

Return type

Iterator[Tensor]

get_params_and_grads() → Iterator[numpy.ndarray]

Returns parameters and gradients for each layer.

Returns

List of parameter, gradient pairs

Return type

Iterator[Tensor]

load(path: str) → None

Load the model from a file.

Parameters

path (str) – path to the file where the model is saved

predict(inputs: numpy.ndarray) → numpy.ndarray

Used to predict values after you finished the training.

Parameters

inputs (Tensor) – values you wish to predict

Returns

Predicted values

Return type

Tensor

save(path: str = '.', filename: str = 'model') → None

Save the model to a file.

It saves every layer, and everything within params object.

If you want your custom layer to be saved make sure, that everything you want to save is in params.

Everything is saved as a json object as follows:

{
    'layerName': [
        {
            name: 'paramName',
            value: '0.5888 2.59874.......'
        },
        ....
    ]
}

Parameters

• path (str, optional) – location where you want your model to be saved to, defaults to ‘.’ which means this directory

• name (str, optional) – name of the file, defaults to ‘model’, note: extension is added manually

step(batch: bluebird.data.Batch) → float

Step function is called during each training step.

It calculates the loss and updates weights and biases.

You must return loss.

Parameters

batch (Batch) – recives batch object, each batch object has inputs and targets

Raises

NotImplementedError –

summary(shape: tuple) → None

Prints the information about the model (Layes, shapes, params).

Parameters

shape (tuple) – shape of the input vector