spacekit.builder.architect

class spacekit.builder.architect.Builder(train_data=None, test_data=None, blueprint=None, algorithm=None, model_path=None, name=None, logname='Builder', **log_kws)

Bases: object

Class for building and training a neural network.

batch_fit(**kwargs)

decay_learning_rate()

Set learning schedule with exponential decay

Returns:

exponential decay learning rate schedule

Return type:

keras.optimizers.schedules.ExponentialDecay

find_tx_file(name='tx_data.json')

fit(**kwargs)

fit_params(batch_size=32, epochs=60, lr=0.0001, decay=[100000, 0.96], early_stopping=None, verbose=2, ensemble=False)

Set custom model fitting parameters as Builder object attributes.

Parameters:

• batch_size (int, optional) – size of each training batch, by default 32

• epochs (int, optional) – number of epochs, by default 60

• lr (float, optional) – initial learning rate, by default 1e-4

• decay (list, optional) – decay_steps, decay_rate, by default [100000, 0.96]

• early_stopping (str, optional) – use an early stopping callback, by default None

• verbose (int, optional) – set the verbosity level, by default 2

• ensemble (bool, optional) – ensemble type network, by default False

Returns:

spacekit.builder.architect.Builder class object with updated fitting parameters.

Return type:

self

get_blueprint(architecture, fitting=True)

get_build_params()

Build params for initialized model

Returns:

dictionary of key-val pairs for initialized model’s build parameters

Return type:

dict

get_fit_params()

Fit parameters of built model

Returns:

key-val pairs of model fitting hyperparameters

Return type:

dict

load_pretrained_network(arch=None)

load_saved_model(arch=None, compile_params=None, custom_obj={}, extract_to='models', keras_archive=True)

Load saved keras model from local disk (located at the model_path attribute) or a pre-trained model from spacekit.skopes.trained_networks (if model_path attribute is None). Example for compile_params: dict(loss="binary_crossentropy",metrics=["accuracy"],        optimizer=Adam(learning_rate=optimizers.schedules.ExponentialDecay(1e-4,         decay_steps=100000, decay_rate=0.96, staircase=True)))

Parameters:

• arch (str, optional) – select a pre-trained model from the spacekit library (“svm_align”, “jwst_cal”, or “hst_cal”), by default None

• compile_params (dict, optional) – Compile the model using kwarg parameters, by default None

• custom_obj (dict, optional) – custom objects keyword arguments to be passed into Keras load_model, by default {}

• extract_to (str or path, optional) – directory location to extract into, by default “models”

• keras_archive (bool, optional) – for models saved in the newer high-level .keras compressed archive format (recommended), by default True

Returns:

pre-trained (and/or compiled) functional Keras model.

Return type:

Keras functional Model object

model_diagram(model=None, output_path=None, show_shapes=True, show_dtype=False, LR=False, expand_nested=True, show_layer_names=False)

save_keras_model(model_path)

save_model(output_path='.', keras_archive=True, parent_dir='', weights=True)

The model architecture, and training configuration (including the optimizer, losses, and metrics) are stored in saved_model.pb. The weights are saved in the variables/ directory.

Parameters:

• output_path (str, optional) – where to save the model files, by default “.”

• keras_archive (bool, optional) – save model using new (preferred) keras archive format, by default True

• parent_dir (str, optional) – store model in a subdirectory of output_path using this name, by default “”

• weights (bool, optional) – save weights learned by the model separately also, by default True

set_build_params(input_shape=None, output_shape=None, layers=None, kernel_size=None, activation=None, cost_function=None, strides=None, optimizer=None, lr_sched=None, loss=None, metrics=None, input_name=None, output_name=None, name=None, algorithm=None)

Set custom build parameters for a Builder object.

Parameters:

• input_shape (tuple, optional) – shape of the inputs, by default None

• output_shape (tuple, optional) – shape of the output, by default None

• layers (list, optional) – sizes of hidden (dense) layers, by default None

• kernel_size (int, optional) – size of the kernel, by default None

• activation (string, optional) – dense layer activation, by default None

• cost_function (str, optional) – function to update weights (calculate cost), by default None

• strides (int, optional) – number of strides, by default None

• optimizer (object, optional) – type of optimizer to use, by default None

• lr_sched (bool, optional) – use a learning_rate schedule such as ExponentialDecay, by default None

• loss (string, optional) – loss metric to monitor, by default None

• metrics (list, optional) – metrics for model to train on, by default None

• algorithm (str, optional) – analysis type, used for determining spacekit.analyzer.Compute class e.g. “linreg” for linear regression or “multiclass” for multi-label classification, by default None

Returns:

spacekit.builder.architect.Builder class object with updated attributes

Return type:

self

set_callbacks()

Set an early stopping callback by monitoring the model training for either accuracy or loss. For classifiers, use ‘val_accuracy’ or ‘val_loss’. For regression use ‘val_loss’ or ‘val_rmse’. Default patience is 15 and can be modified by setting the self.patience attribute to a custom value (int).

Returns:

[callbacks.ModelCheckpoint, callbacks.EarlyStopping]

Return type:

list

unzip_model_files(extract_to='models')

Extracts a keras model object from a zip archive

Parameters:

extract_to (str, optional) – directory location to extract into, by default “models”

Returns:

path to where the model archive has been extracted

Return type:

string

class spacekit.builder.architect.BuilderEnsemble(X_train=None, y_train=None, X_test=None, y_test=None, params=None, input_name='svm_mixed_inputs', output_name='ensemble_output', **builder_kwargs)

Bases: Builder

Subclass for building and training an ensemble model (stacked MLP and 3D CNN)

Parameters:

Builder (class) – spacekit.builder.architect.Builder class object

batch()

Gives equal number of positive and negative samples rotating randomly The output of the generator must be either - a tuple (inputs, targets) - a tuple (inputs, targets, sample_weights).

This tuple (a single output of the generator) makes a single batch. The last batch of the epoch is commonly smaller than the others, if the size of the dataset is not divisible by the batch size. The generator loops over its data indefinitely. An epoch finishes when steps_per_epoch batches have been seen by the model.

build()

Builds and compiles the ensemble model

Returns:

spacekit.builder.architect.Builder.BuilderEnsemble object with model attribute initialized

Return type:

self

ensemble_cnn()

Compiles the CNN branch of the ensemble model

Returns:

spacekit.builder.architect.Builder.BuilderEnsemble object with cnn attribute initialized

Return type:

self

ensemble_mlp()

Compiles the MLP branch of the ensemble model

Returns:

spacekit.builder.architect.Builder.BuilderEnsemble object with mlp attribute initialized

Return type:

self

class spacekit.builder.architect.BuilderMLP(X_train=None, y_train=None, X_test=None, y_test=None, blueprint='mlp', **builder_kwargs)

Bases: Builder

Subclass for building and training MLP neural networks

Parameters:

Builder (class) – spacekit.builder.architect.Builder class object

batch(augment=True)

Randomly rotates through positive and negative samples indefinitely, generating a single batch tuple of (inputs, targets) or (inputs, targets, sample_weights). If the size of the dataset is not divisible by the batch size, the last batch will be smaller than the others. An epoch finishes once steps_per_epoch have been seen by the model.

Parameters:

augment (bool, optional) – whether or not to apply stochastic data augmentation on input features, by default True

Yields:

tuple – a single batch tuple of (inputs, targets) or (inputs, targets, sample_weights).

build(layer_kwargs={})

Build and compile an MLP network

Returns:

compiled model object

Return type:

tf.keras.model

class spacekit.builder.architect.BuilderCNN3D(X_train=None, y_train=None, X_test=None, y_test=None, blueprint='cnn3d', **builder_kwargs)

Bases: Builder

Subclass for building and training 3D convolutional neural networks

Parameters:

Builder (class) – spacekit.builder.architect.Builder class object

batch()

Gives equal number of positive and negative samples rotating randomly The output of the generator must be either - a tuple (inputs, targets) - a tuple (inputs, targets, sample_weights).

This tuple (a single output of the generator) makes a single batch. The last batch of the epoch is commonly smaller than the others, if the size of the dataset is not divisible by the batch size. The generator loops over its data indefinitely. An epoch finishes when steps_per_epoch batches have been seen by the model.

build()

Builds a 3D convolutional neural network for RGB image triplets

class spacekit.builder.architect.BuilderCNN2D(X_train=None, y_train=None, X_test=None, y_test=None, blueprint='cnn2d', **builder_kwargs)

Bases: Builder

Subclass Builder object for 2D Convolutional Neural Networks

Parameters:

Builder (class) – spacekit.builder.architect.Builder.Builder object

batch()

Gives equal number of positive and negative samples rotating randomly The output of the generator must be either - a tuple (inputs, targets) - a tuple (inputs, targets, sample_weights).

This tuple (a single output of the generator) makes a single batch. Therefore, all arrays in this tuple must have the same length (equal to the size of this batch). Different batches may have different sizes.

For example, the last batch of the epoch is commonly smaller than the others, if the size of the dataset is not divisible by the batch size. The generator is expected to loop over its data indefinitely. An epoch finishes when steps_per_epoch batches have been seen by the model.

build()

Builds and compiles a 2-dimensional Keras Functional CNN

class spacekit.builder.architect.MemoryClassifier(X_train=None, y_train=None, X_test=None, y_test=None, blueprint='hst_mem_clf', test_idx=None, **builder_kwargs)

Bases: BuilderMLP

Builds an MLP neural network classifier object with pre-tuned params for Calcloud’s Memory Bin classifier

Parameters:

MultiLayerPerceptron (object) – mlp multi-classification builder object

class spacekit.builder.architect.MemoryRegressor(X_train=None, y_train=None, X_test=None, y_test=None, blueprint='hst_mem_reg', test_idx=None, **builder_kwargs)

Bases: BuilderMLP

Builds an MLP neural network regressor object with pre-tuned params for estimating memory allocation (GB) in Calcloud

Parameters:

MultiLayerPerceptron (object) – mlp linear regression builder object

class spacekit.builder.architect.WallclockRegressor(X_train=None, y_train=None, X_test=None, y_test=None, blueprint='hst_wall_reg', test_idx=None, **builder_kwargs)

Bases: BuilderMLP

Builds an MLP neural network regressor object with pre-tuned params for estimating wallclock allocation (minimum execution time in seconds) in Calcloud.

Parameters:

MultiLayerPerceptron (object) – mlp linear regression builder object