Model Visualization

class ai4water.postprocessing.visualize.Visualize(model, save=True, show=True, verbosity=None)

Bases: Plots

Hepler class to peek inside the machine learning mdoel.

If the machine learning model consists of layers of neural networks, then this class can be used to plot following 4 items

• outputs of individual layers

• gradients of outputs of individual layers

• weights and biases of individual layers

• gradients of weights of individual layers

If the machine learning model consists of tree, then this class can be used to plot the learned tree of the model.

- get_activations

- activations

- get_activation_gradients

- activation_gradients

- get_weights

- weights

- get_weight_gradients

- weight_gradients

- decision_tree

__init__(model, save=True, show=True, verbosity=None)

Parameters:

• model – the learned machine learning model.

• save – bool

• show – bool

• verbosity – bool/int

activation_gradients(layer_names: Union[str, list], data=None, data_type='training', x=None, y=None, examples_to_use=None, plot_type='2D')

Plots the gradients o activations/outputs of layers

Parameters:

• layer_names – the layer name for which the gradients of its outputs are to be plotted.

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract x and y which will be given to NN to get gradients of ativations

• data_type – str either training, validation or test. Only relevant if data argument is active

• x – alternative to data

• y – alternative to data

• examples_to_use – the examples from the data to use. If None, then all examples will be used, which is equal to the length of data.

• plot_type –

activation_gradients_1D(layer_names, data=None, data_type: str = 'training', x=None, y=None, examples_to_use=None)

Plots gradients of layer outputs as 1D

Parameters:

• layer_names –

• examples_to_use –

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract input

• data_type – str either training, validation or test. Only relevant if data argument is active

• x –

• y –

activation_gradients_2D(layer_names=None, data=None, data_type: str = 'training', x=None, y=None, examples_to_use=None)

Plots activations of intermediate layers except input and output

Parameters:

• layer_names –

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract input

• data_type – str either training, validation or test. Only relevant if data argument is active

• x –

• y –

• examples_to_use – if integer, it will be the number of examples to use. If array like, it will be index of examples to use

activations(layer_names=None, x=None, data=None, data_type: str = 'training', examples_to_use: Optional[Union[int, list, ndarray, range]] = None, **kwargs)

Plots outputs of any layer of neural network.

Parameters:

• x – if given, will override, ‘data’.

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract input

• data_type – str either training, validation or test. Only relevant if data argument is active

• layer_names – name of layer whose output is to be plotted. If None, it will plot outputs of all layers

• examples_to_use – If integer, it will be the number of examples to use. If array like, it will be the indices of examples to use.

decision_tree(show=False, **kwargs)

Plots the decision tree

decision_tree_leaves(data=None, data_type: str = 'training')

Plots dtreeviz related plots if dtreeviz is installed

data :

raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract input

data_typestr

either training, validation or test. Only relevant if data argument is active

features_1d(data, save=True, name='', **kwargs)

features_2d(data, name, save=True, slices=24, slice_dim=0, **kwargs)

Calls the features_2d from see-rnn

find_num_lstms(layer_names=None) → list

Finds names of lstm layers in model

get_activation_gradients(layer_names: Optional[Union[str, list]] = None, x=None, y=None, data=None, data_type: str = 'training') → dict

Finds gradients of outputs of a layer.

either x,y or data is required

Parameters:

• layer_names – The layer for which, the gradients of its outputs are to be calculated.

• x – input data. Will overwrite data

• y – corresponding label of x. Will overwrite data.

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract x and y

• data_type – str either training, validation or test. Only relevant if data argument is active

get_activations(layer_names: Optional[Union[str, list]] = None, x=None, data=None, data_type: str = 'training', batch_size: Optional[int] = None) → dict

gets the activations/outputs of any layer of the Keras Model.

Parameters:

• layer_names – name of list of names of layers whose activations are to be returned.

• x – The input that will be fed to NN to extract activations. If provided, it will override data.

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract input

• data_type – str either training, validation or test. Only relevant if data argument is active

• batch_size – int

Returns:

a dictionary whose keys are names of layers and values are weights of those layers as numpy arrays

get_rnn_weights(weights: dict, layer_names=None) → dict

Finds RNN related weights.

It combines kernel recurrent curnel and bias of each layer into a list.

get_weight_gradients(x=None, y=None, data=None, data_type: str = 'training') → dict

Returns the gradients of weights.

Parameters:

• x – inputs, if not given, then data must be given

• y – target

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract x and y

• data_type – str either training, validation or test. Only relevant if data argument is active

Returns:

dictionary whose keys are names of layers and values are gradients of weights as numpy arrays.

get_weights() → dict

returns all trainable weights as arrays in a dictionary

rnn_histogram(data, save=True, name='', **kwargs)

rnn_weight_grads_as_hist(layer_name=None, data=None, data_type='training', x=None, y=None)

rnn_weights_histograms(layer_name)

weight_gradients(layer_names: Optional[Union[str, list]] = None, data=None, data_type='training', x=None, y=None)

Plots gradient of all trainable weights

Parameters:

• layer_names – the layer whose weeights are to be considered.

• data – raw unprepared data from which will be forwarded to ai4water.preprocessing.DataSet() to extract x and y

• data_type – str either training, validation or test. Only relevant if data argument is active

• x – alternative to data

• y – alternative to data

weights(layer_names: Optional[Union[str, list]] = None, **kwargs)

Plots the weights of a specific layer or all layers.

Parameters:

layer_names – The layer whose weights are to be viewed.