Skip to content

xai

check_for_nans(data, layer_index)

Checks for NaN values in the tensor data.

Parameters:

Name Type Description Default
data Tensor

The tensor to check for NaN values.

required
layer_index int

The index of the layer for logging purposes.

required

Returns:

Name Type Description
bool bool

True if NaNs are found, False otherwise.

Source code in spotpython/plot/xai.py
21
22
23
24
25
26
27
28
29
30
31
32
33
34
def check_for_nans(data, layer_index) -> bool:
    """Checks for NaN values in the tensor data.

    Args:
        data (torch.Tensor): The tensor to check for NaN values.
        layer_index (int): The index of the layer for logging purposes.

    Returns:
        bool: True if NaNs are found, False otherwise.
    """
    if torch.isnan(data).any():
        print(f"NaN detected after layer {layer_index}")
        return True
    return False

get_activations(net, fun_control, batch_size, device='cpu', normalize=False)

Computes the activations for each layer of the network, the mean activations, and the sizes of the activations for each layer.

Parameters:

Name Type Description Default
net Module

The neural network model.

required
fun_control dict

A dictionary containing the dataset.

required
batch_size int

The batch size for the data loader.

required
device str

The device to run the model on. Defaults to “cpu”.

'cpu'
normalize bool

Whether to normalize the input data. Defaults to False.

False

Returns:

Name Type Description
tuple tuple

A tuple containing the activations, mean activations, and layer sizes for each layer.

Examples:

>>> from spotpython.plot.xai import get_activations
    import torch
    import numpy as np
    import torch.nn as nn
    from spotpython.utils.init import fun_control_init
    from spotpython.data.diabetes import Diabetes
    from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.hyperparameters.values import (
            get_default_hyperparameters_as_array, get_one_config_from_X)
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.data.lightdatamodule import LightDataModule
    from spotpython.plot.xai import get_gradients
    fun_control = fun_control_init(
        _L_in=10, # 10: diabetes
        _L_out=1,
        _torchmetric="mean_squared_error",
        data_set=Diabetes(),
        core_model=NNLinearRegressor,
        hyperdict=LightHyperDict)
    X = get_default_hyperparameters_as_array(fun_control)
    config = get_one_config_from_X(X, fun_control)
    _L_in = fun_control["_L_in"]
    _L_out = fun_control["_L_out"]
    _torchmetric = fun_control["_torchmetric"]
    batch_size = 16
    model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
    activations, mean_activations, layer_sizes = get_activations(net=model, fun_control=fun_control, batch_size=batch_size, device = "cpu")
    plot_nn_values_scatter(nn_values=activations, layer_sizes=layer_sizes, nn_values_names="Activations")
Source code in spotpython/plot/xai.py
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
def get_activations(net, fun_control, batch_size, device="cpu", normalize=False) -> tuple:
    """
    Computes the activations for each layer of the network, the mean activations,
    and the sizes of the activations for each layer.

    Args:
        net (nn.Module): The neural network model.
        fun_control (dict): A dictionary containing the dataset.
        batch_size (int): The batch size for the data loader.
        device (str): The device to run the model on. Defaults to "cpu".
        normalize (bool): Whether to normalize the input data. Defaults to False.

    Returns:
        tuple: A tuple containing the activations, mean activations, and layer sizes for each layer.

    Examples:
        >>> from spotpython.plot.xai import get_activations
            import torch
            import numpy as np
            import torch.nn as nn
            from spotpython.utils.init import fun_control_init
            from spotpython.data.diabetes import Diabetes
            from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.hyperparameters.values import (
                    get_default_hyperparameters_as_array, get_one_config_from_X)
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.data.lightdatamodule import LightDataModule
            from spotpython.plot.xai import get_gradients
            fun_control = fun_control_init(
                _L_in=10, # 10: diabetes
                _L_out=1,
                _torchmetric="mean_squared_error",
                data_set=Diabetes(),
                core_model=NNLinearRegressor,
                hyperdict=LightHyperDict)
            X = get_default_hyperparameters_as_array(fun_control)
            config = get_one_config_from_X(X, fun_control)
            _L_in = fun_control["_L_in"]
            _L_out = fun_control["_L_out"]
            _torchmetric = fun_control["_torchmetric"]
            batch_size = 16
            model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
            activations, mean_activations, layer_sizes = get_activations(net=model, fun_control=fun_control, batch_size=batch_size, device = "cpu")
            plot_nn_values_scatter(nn_values=activations, layer_sizes=layer_sizes, nn_values_names="Activations")
    """
    activations = {}
    mean_activations = {}
    layer_sizes = {}
    net.eval()  # Set the model to evaluation mode

    dataset = fun_control["data_set"]
    data_module = LightDataModule(
        dataset=dataset,
        batch_size=batch_size,
        test_size=fun_control["test_size"],
        scaler=fun_control["scaler"],
        verbosity=10,
    )
    data_module.setup(stage="fit")
    train_loader = data_module.train_dataloader()
    inputs, _ = next(iter(train_loader))
    inputs = inputs.to(device)

    if normalize:
        inputs = (inputs - inputs.mean(dim=0, keepdim=True)) / inputs.std(dim=0, keepdim=True)

    with torch.no_grad():
        inputs = inputs.view(inputs.size(0), -1)
        # Loop through all layers
        for layer_index, layer in enumerate(net.layers[:-1]):
            inputs = layer(inputs)  # Forward pass through the layer

            # Check for NaNs
            if check_for_nans(inputs, layer_index):
                break

            # Collect activations for Linear layers
            if isinstance(layer, nn.Linear):
                activations[layer_index] = inputs.view(-1).cpu().numpy()
                mean_activations[layer_index] = inputs.mean(dim=0).cpu().numpy()
                # Record the size of the activations and set the first dimension to 1
                layer_size = np.array(inputs.size())
                layer_size[0] = 1  # Set the first dimension to 1
                layer_sizes[layer_index] = layer_size

    return activations, mean_activations, layer_sizes

get_all_layers_conductance(spot_tuner, fun_control, device='cpu', remove_spot_attributes=False)

Get the conductance of all layers.

Parameters:

Name Type Description Default
spot_tuner object

The spot tuner object.

required
fun_control dict

A dictionary with the function control.

required
device str

The device to use. Defaults to “cpu”.

'cpu'
remove_spot_attributes bool

Whether to remove the spot attributes. Defaults to False.

False
Source code in spotpython/plot/xai.py
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
def get_all_layers_conductance(spot_tuner, fun_control, device="cpu", remove_spot_attributes=False) -> dict:
    """
    Get the conductance of all layers.

    Args:
        spot_tuner (object):
            The spot tuner object.
        fun_control (dict):
            A dictionary with the function control.
        device (str, optional):
            The device to use. Defaults to "cpu".
        remove_spot_attributes (bool, optional):
            Whether to remove the spot attributes. Defaults to False.
    """
    config = get_tuned_architecture(spot_tuner, fun_control)
    train_model(config, fun_control, timestamp=False)
    model_loaded = load_light_from_checkpoint(config, fun_control, postfix="_TRAIN")
    if remove_spot_attributes:
        removed_attributes, model = get_removed_attributes_and_base_net(net=model_loaded)
    else:
        model = model_loaded
    model = model.to(device)
    model.eval()
    _, index, _ = get_weights(model, return_index=True)
    layer_conductance = {}
    for i in index:
        layer_conductance[i] = get_layer_conductance(spot_tuner, fun_control, layer_idx=i)
    return layer_conductance

get_attributions(spot_tuner, fun_control, attr_method='IntegratedGradients', baseline=None, abs_attr=True, n_rel=5, device='cpu', normalize=True, remove_spot_attributes=False)

Get the attributions of a neural network.

Parameters:

Name Type Description Default
spot_tuner object

The spot tuner object.

required
fun_control dict

A dictionary with the function control.

required
attr_method str

The attribution method. Defaults to “IntegratedGradients”.

'IntegratedGradients'
baseline Tensor

The baseline for the attribution methods. Defaults to None.

None
abs_attr bool

Whether the method should sort by the absolute attribution values. Defaults to True.

True
n_rel int

The number of relevant features. Defaults to 5.

5
device str

The device to use. Defaults to “cpu”.

'cpu'
normalize bool

Whether to normalize the input data. Defaults to True.

True
remove_spot_attributes bool

Whether to remove the spot attributes. If True, a torch model is created via get_removed_attributes. Defaults to False.

False

Returns:

Type Description
DataFrame

pd.DataFrame (object): A DataFrame with the attributions.

Source code in spotpython/plot/xai.py
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
def get_attributions(
    spot_tuner,
    fun_control,
    attr_method="IntegratedGradients",
    baseline=None,
    abs_attr=True,
    n_rel=5,
    device="cpu",
    normalize=True,
    remove_spot_attributes=False,
) -> pd.DataFrame:
    """Get the attributions of a neural network.

    Args:
        spot_tuner (object):
            The spot tuner object.
        fun_control (dict):
            A dictionary with the function control.
        attr_method (str, optional):
            The attribution method. Defaults to "IntegratedGradients".
        baseline (torch.Tensor, optional):
            The baseline for the attribution methods. Defaults to None.
        abs_attr (bool, optional):
            Whether the method should sort by the absolute attribution values. Defaults to True.
        n_rel (int, optional):
            The number of relevant features. Defaults to 5.
        device (str, optional):
            The device to use. Defaults to "cpu".
        normalize (bool, optional):
            Whether to normalize the input data. Defaults to True.
        remove_spot_attributes (bool, optional):
            Whether to remove the spot attributes.
            If True, a torch model is created via `get_removed_attributes`. Defaults to False.

    Returns:
        pd.DataFrame (object): A DataFrame with the attributions.
    """
    try:
        fun_control["data_set"].names
    except AttributeError:
        fun_control["data_set"].names = None
    feature_names = fun_control["data_set"].names
    total_attributions = None
    config = get_tuned_architecture(spot_tuner, fun_control)
    train_model(config, fun_control, timestamp=False)
    model_loaded = load_light_from_checkpoint(config, fun_control, postfix="_TRAIN")
    if remove_spot_attributes:
        removed_attributes, model = get_removed_attributes_and_base_net(net=model_loaded)
    else:
        model = model_loaded
    model = model.to(device)
    model.eval()
    # get feature names
    dataset = fun_control["data_set"]
    try:
        n_features = dataset.data.shape[1]
    except AttributeError:
        n_features = dataset.tensors[0].shape[1]
    if feature_names is None:
        feature_names = [f"x{i}" for i in range(n_features)]
    # get batch size
    batch_size = config["batch_size"]
    # test_loader = DataLoader(dataset, batch_size=batch_size, shuffle=False)

    data_module = LightDataModule(
        dataset=dataset,
        batch_size=batch_size,
        test_size=fun_control["test_size"],
        scaler=fun_control["scaler"],
        verbosity=10,
    )
    data_module.setup(stage="test")
    test_loader = data_module.test_dataloader()

    if attr_method == "IntegratedGradients":
        attr = IntegratedGradients(model)
    elif attr_method == "DeepLift":
        attr = DeepLift(model)
    elif attr_method == "GradientShap":  # Todo: would need a baseline
        if baseline is None:
            raise ValueError("baseline cannot be 'None' for GradientShap")
        attr = GradientShap(model)
    elif attr_method == "FeatureAblation":
        attr = FeatureAblation(model)
    else:
        raise ValueError(
            """
            Unsupported attribution method.
            Please choose from 'IntegratedGradients', 'DeepLift', 'GradientShap', or 'FeatureAblation'.
            """
        )
    for inputs, _ in test_loader:
        if normalize:
            inputs = (inputs - inputs.mean()) / inputs.std()
        inputs.requires_grad_()
        attributions = attr.attribute(inputs, return_convergence_delta=False, baselines=baseline)
        if total_attributions is None:
            total_attributions = attributions
        else:
            if len(attributions) == len(total_attributions):
                total_attributions += attributions

    # Calculation of average attribution across all batches
    avg_attributions = total_attributions.mean(dim=0).detach().numpy()

    # Transformation to the absolute attribution values if abs_attr is True
    # Get indices of the n most important features
    if abs_attr is True:
        abs_avg_attributions = abs(avg_attributions)
        top_n_indices = abs_avg_attributions.argsort()[-n_rel:][::-1]
    else:
        top_n_indices = avg_attributions.argsort()[-n_rel:][::-1]

    # Get the importance values for the top n features
    top_n_importances = avg_attributions[top_n_indices]

    df = pd.DataFrame(
        {
            "Feature Index": top_n_indices,
            "Feature": [feature_names[i] for i in top_n_indices],
            attr_method + "Attribution": top_n_importances,
        }
    )
    return df

get_gradients(net, fun_control, batch_size, device='cpu', normalize=False)

Get the gradients of a neural network and the size of each layer.

Parameters:

Name Type Description Default
net object

A neural network.

required
fun_control dict

A dictionary with the function control.

required
batch_size int

The batch size.

required
device str

The device to use. Defaults to “cpu”.

'cpu'
normalize bool

Whether to normalize the input data. Defaults to False.

False

Returns:

Name Type Description
tuple tuple

A tuple containing: - grads: A dictionary with the gradients of the neural network. - layer_sizes: A dictionary with layer names as keys and their sizes as entries in NumPy array format.

Examples:

>>> from spotpython.plot.xai import get_gradients
    import torch
    import numpy as np
    import torch.nn as nn
    from spotpython.utils.init import fun_control_init
    from spotpython.data.diabetes import Diabetes
    from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.hyperparameters.values import (
            get_default_hyperparameters_as_array, get_one_config_from_X)
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.data.lightdatamodule import LightDataModule
    # from spotpython.plot.xai import get_gradients
    fun_control = fun_control_init(
        _L_in=10, # 10: diabetes
        _L_out=1,
        _torchmetric="mean_squared_error",
        data_set=Diabetes(),
        core_model=NNLinearRegressor,
        hyperdict=LightHyperDict)
    X = get_default_hyperparameters_as_array(fun_control)
    config = get_one_config_from_X(X, fun_control)
    _L_in = fun_control["_L_in"]
    _L_out = fun_control["_L_out"]
    _torchmetric = fun_control["_torchmetric"]
    batch_size = 16
    model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
    gradients, layer_sizes = get_gradients(net=model)
    gradients, layer_sizes
Source code in spotpython/plot/xai.py
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
def get_gradients(net, fun_control, batch_size, device="cpu", normalize=False) -> tuple:
    """
    Get the gradients of a neural network and the size of each layer.

    Args:
        net (object):
            A neural network.
        fun_control (dict):
            A dictionary with the function control.
        batch_size (int, optional):
            The batch size.
        device (str, optional):
            The device to use. Defaults to "cpu".
        normalize (bool, optional):
            Whether to normalize the input data. Defaults to False.

    Returns:
        tuple: A tuple containing:
            - grads: A dictionary with the gradients of the neural network.
            - layer_sizes: A dictionary with layer names as keys and their sizes as entries in NumPy array format.

    Examples:
        >>> from spotpython.plot.xai import get_gradients
            import torch
            import numpy as np
            import torch.nn as nn
            from spotpython.utils.init import fun_control_init
            from spotpython.data.diabetes import Diabetes
            from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.hyperparameters.values import (
                    get_default_hyperparameters_as_array, get_one_config_from_X)
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.data.lightdatamodule import LightDataModule
            # from spotpython.plot.xai import get_gradients
            fun_control = fun_control_init(
                _L_in=10, # 10: diabetes
                _L_out=1,
                _torchmetric="mean_squared_error",
                data_set=Diabetes(),
                core_model=NNLinearRegressor,
                hyperdict=LightHyperDict)
            X = get_default_hyperparameters_as_array(fun_control)
            config = get_one_config_from_X(X, fun_control)
            _L_in = fun_control["_L_in"]
            _L_out = fun_control["_L_out"]
            _torchmetric = fun_control["_torchmetric"]
            batch_size = 16
            model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
            gradients, layer_sizes = get_gradients(net=model)
            gradients, layer_sizes
    """
    net.eval()
    dataset = fun_control["data_set"]
    data_module = LightDataModule(
        dataset=dataset,
        batch_size=batch_size,
        test_size=fun_control["test_size"],
        scaler=fun_control["scaler"],
        verbosity=10,
    )
    data_module.setup(stage="fit")
    train_loader = data_module.train_dataloader()
    inputs, targets = next(iter(train_loader))
    if normalize:
        inputs = (inputs - inputs.mean(dim=0, keepdim=True)) / inputs.std(dim=0, keepdim=True)
    inputs, targets = inputs.to(device), targets.to(device)

    # Pass one batch through the network, and calculate the gradients for the weights
    net.zero_grad()
    preds = net(inputs)
    preds = preds.squeeze(-1)  # Remove the last dimension if it's 1
    loss = F.mse_loss(preds, targets)
    loss.backward()

    grads = {}
    layer_sizes = {}
    for name, params in net.named_parameters():
        if "weight" in name:
            # Collect gradient information
            grads[name] = params.grad.view(-1).cpu().clone().numpy()
            # Collect size information
            layer_sizes[name] = np.array(params.size())

    net.zero_grad()
    return grads, layer_sizes

get_layer_conductance(spot_tuner, fun_control, layer_idx, device='cpu', normalize=True, remove_spot_attributes=False)

Compute the average layer conductance attributions for a specified layer in the model.

Parameters:

Name Type Description Default
spot_tuner Spot

The spot tuner object containing the trained model.

required
fun_control dict

The fun_control dictionary containing the hyperparameters used to train the model.

required
layer_idx int

Index of the layer for which to compute layer conductance attributions.

required
device str

The device to use. Defaults to “cpu”.

'cpu'
normalize bool

Whether to normalize the input data. Defaults to True.

True
remove_spot_attributes bool

Whether to remove the spot attributes. Defaults to False.

False

Returns:

Type Description
ndarray

numpy.ndarray: An array containing the average layer conductance attributions for the specified layer. The shape of the array corresponds to the shape of the attributions.

Source code in spotpython/plot/xai.py
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
def get_layer_conductance(
    spot_tuner, fun_control, layer_idx, device="cpu", normalize=True, remove_spot_attributes=False
) -> np.ndarray:
    """
    Compute the average layer conductance attributions for a specified layer in the model.

    Args:
        spot_tuner (spot.Spot):
            The spot tuner object containing the trained model.
        fun_control (dict):
            The fun_control dictionary containing the hyperparameters used to train the model.
        layer_idx (int):
            Index of the layer for which to compute layer conductance attributions.
        device (str, optional):
            The device to use. Defaults to "cpu".
        normalize (bool, optional):
            Whether to normalize the input data. Defaults to True.
        remove_spot_attributes (bool, optional):
            Whether to remove the spot attributes. Defaults to False.

    Returns:
        numpy.ndarray:
            An array containing the average layer conductance attributions for the specified layer.
            The shape of the array corresponds to the shape of the attributions.
    """
    try:
        fun_control["data_set"].names
    except AttributeError:
        fun_control["data_set"].names = None
    feature_names = fun_control["data_set"].names

    config = get_tuned_architecture(spot_tuner, fun_control)
    train_model(config, fun_control, timestamp=False)
    model_loaded = load_light_from_checkpoint(config, fun_control, postfix="_TRAIN")
    if remove_spot_attributes:
        removed_attributes, model = get_removed_attributes_and_base_net(net=model_loaded)
    else:
        model = model_loaded
    model = model.to(device)
    model.eval()

    dataset = fun_control["data_set"]
    try:
        n_features = dataset.data.shape[1]
    except AttributeError:
        n_features = dataset.tensors[0].shape[1]
    if feature_names is None:
        feature_names = [f"x{i}" for i in range(n_features)]
    batch_size = config["batch_size"]
    test_loader = DataLoader(dataset, batch_size=batch_size)
    total_layer_attributions = None
    layers = model.layers
    print("Conductance analysis for layer: ", layers[layer_idx])
    lc = LayerConductance(model, layers[layer_idx])

    for inputs, labels in test_loader:
        if normalize:
            inputs = (inputs - inputs.mean()) / inputs.std()
        lc_attr_test = lc.attribute(inputs, n_steps=10, attribute_to_layer_input=True)
        if total_layer_attributions is None:
            total_layer_attributions = lc_attr_test
        else:
            if len(lc_attr_test) == len(total_layer_attributions):
                total_layer_attributions += lc_attr_test

    avg_layer_attributions = total_layer_attributions.mean(dim=0).detach().numpy()

    return avg_layer_attributions

get_weights(net, return_index=False)

Get the weights of a neural network and the size of each layer.

Parameters:

Name Type Description Default
net object

A neural network.

required
return_index bool

Whether to return the index. Defaults to False.

False

Returns:

Name Type Description
tuple tuple

A tuple containing: - weights: A dictionary with the weights of the neural network. - index: The layer index list (only if return_index is True). - layer_sizes: A dictionary with layer names as keys and their sizes as entries in NumPy array format.

Examples:

>>> from spotpython.plot.xai import get_weights
    import torch
    import numpy as np
    import torch.nn as nn
    from spotpython.utils.init import fun_control_init
    from spotpython.data.diabetes import Diabetes
    from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.hyperparameters.values import (
            get_default_hyperparameters_as_array, get_one_config_from_X)
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.data.lightdatamodule import LightDataModule
    from spotpython.plot.xai import get_gradients
    fun_control = fun_control_init(
        _L_in=10, # 10: diabetes
        _L_out=1,
        _torchmetric="mean_squared_error",
        data_set=Diabetes(),
        core_model=NNLinearRegressor,
        hyperdict=LightHyperDict)
    X = get_default_hyperparameters_as_array(fun_control)
    config = get_one_config_from_X(X, fun_control)
    _L_in = fun_control["_L_in"]
    _L_out = fun_control["_L_out"]
    _torchmetric = fun_control["_torchmetric"]
    batch_size = 16
    model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
    weights, layer_sizes = get_weights(net=model)
    weights, layer_sizes
Source code in spotpython/plot/xai.py
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
def get_weights(net, return_index=False) -> tuple:
    """
    Get the weights of a neural network and the size of each layer.

    Args:
        net (object):
            A neural network.
        return_index (bool, optional):
            Whether to return the index. Defaults to False.

    Returns:
        tuple:
            A tuple containing:
            - weights: A dictionary with the weights of the neural network.
            - index: The layer index list (only if return_index is True).
            - layer_sizes: A dictionary with layer names as keys and their sizes as entries in NumPy array format.

    Examples:
        >>> from spotpython.plot.xai import get_weights
            import torch
            import numpy as np
            import torch.nn as nn
            from spotpython.utils.init import fun_control_init
            from spotpython.data.diabetes import Diabetes
            from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.hyperparameters.values import (
                    get_default_hyperparameters_as_array, get_one_config_from_X)
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.data.lightdatamodule import LightDataModule
            from spotpython.plot.xai import get_gradients
            fun_control = fun_control_init(
                _L_in=10, # 10: diabetes
                _L_out=1,
                _torchmetric="mean_squared_error",
                data_set=Diabetes(),
                core_model=NNLinearRegressor,
                hyperdict=LightHyperDict)
            X = get_default_hyperparameters_as_array(fun_control)
            config = get_one_config_from_X(X, fun_control)
            _L_in = fun_control["_L_in"]
            _L_out = fun_control["_L_out"]
            _torchmetric = fun_control["_torchmetric"]
            batch_size = 16
            model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
            weights, layer_sizes = get_weights(net=model)
            weights, layer_sizes
    """
    weights = {}
    index = []
    layer_sizes = {}

    for name, param in net.named_parameters():
        if name.endswith(".bias"):
            continue

        # Extract layer number
        layer_number = int(name.split(".")[1])
        index.append(layer_number)

        # Create dictionary key for this layer
        key_name = f"Layer {layer_number}"

        # Store weight information
        weights[key_name] = param.detach().view(-1).cpu().numpy()

        # Store layer size as a NumPy array
        layer_sizes[key_name] = np.array(param.size())

    if return_index:
        return weights, index, layer_sizes
    else:
        return weights, layer_sizes

get_weights_conductance_last_layer(spot_tuner, fun_control, device='cpu', remove_spot_attributes=False)

Get the weights and the conductance of the last layer.

Parameters:

Name Type Description Default
spot_tuner object

The spot tuner object.

required
fun_control dict

A dictionary with the function control.

required
device str

The device to use. Defaults to “cpu”.

'cpu'
remove_spot_attributes bool

Whether to remove the spot attributes. Defaults to False.

False
Source code in spotpython/plot/xai.py
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
def get_weights_conductance_last_layer(spot_tuner, fun_control, device="cpu", remove_spot_attributes=False) -> tuple:
    """
    Get the weights and the conductance of the last layer.

    Args:
        spot_tuner (object):
            The spot tuner object.
        fun_control (dict):
            A dictionary with the function control.
        device (str, optional):
            The device to use. Defaults to "cpu".
        remove_spot_attributes (bool, optional):
            Whether to remove the spot attributes. Defaults to False.
    """
    config = get_tuned_architecture(spot_tuner, fun_control)
    train_model(config, fun_control, timestamp=False)
    model_loaded = load_light_from_checkpoint(config, fun_control, postfix="_TRAIN")
    if remove_spot_attributes:
        removed_attributes, model = get_removed_attributes_and_base_net(net=model_loaded)
    else:
        model = model_loaded
    model = model.to(device)
    model.eval()

    weights, index, _ = get_weights(model, return_index=True)
    layer_idx = index[-1]
    weights_last = weights[f"Layer {layer_idx}"]
    weights_last
    layer_conductance_last = get_layer_conductance(spot_tuner, fun_control, layer_idx=layer_idx)

    return weights_last, layer_conductance_last

is_square(n)

Check if a number is a square number.

Parameters:

Name Type Description Default
n int

The number to check.

required

Returns:

Name Type Description
bool bool

True if the number is a square number, False otherwise.

Examples:

>>> is_square(4)
True
>>> is_square(5)
False
Source code in spotpython/plot/xai.py
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
def is_square(n) -> bool:
    """Check if a number is a square number.

    Args:
        n (int): The number to check.

    Returns:
        bool: True if the number is a square number, False otherwise.

    Examples:
        >>> is_square(4)
        True
        >>> is_square(5)
        False
    """
    return n == int(math.sqrt(n)) ** 2

plot_attributions(df, attr_method='IntegratedGradients')

Plot the attributions of a neural network.

Parameters:

Name Type Description Default
df DataFrame

A DataFrame with the attributions.

required
attr_method str

The attribution method. Defaults to “IntegratedGradients”.

'IntegratedGradients'

Returns:

Type Description
None

None

Source code in spotpython/plot/xai.py
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
def plot_attributions(df, attr_method="IntegratedGradients") -> None:
    """
    Plot the attributions of a neural network.

    Args:
        df (pd.DataFrame):
            A DataFrame with the attributions.
        attr_method (str, optional):
            The attribution method. Defaults to "IntegratedGradients".

    Returns:
        None

    """
    sns.set_theme(style="whitegrid")
    plt.figure(figsize=(10, 6))
    sns.barplot(x=attr_method + "Attribution", y="Feature", data=df, palette="viridis", hue="Feature")
    plt.title(f"Top {df.shape[0]} Features by {attr_method} Attribution")
    plt.xlabel(f"{attr_method} Attribution Value")
    plt.ylabel("Feature")
    plt.show()

plot_conductance_last_layer(weights_last, layer_conductance_last, figsize=(12, 6), show=True)

Plot the conductance of the last layer.

Parameters:

Name Type Description Default
weights_last ndarray

The weights of the last layer.

required
layer_conductance_last ndarray

The conductance of the last layer.

required
figsize tuple

The figure size. Defaults to (12, 6).

(12, 6)
show bool

Whether to show the plot. Defaults

True

Examples:

>>> import numpy as np
    from spotpython.plot.xai import plot_conductance_last_layer
    weights_last = np.random.rand(10)
    layer_conductance_last = np.random.rand(10)
    plot_conductance_last_layer(weights_last, layer_conductance_last, show=True)
Source code in spotpython/plot/xai.py
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
def plot_conductance_last_layer(weights_last, layer_conductance_last, figsize=(12, 6), show=True) -> None:
    """
    Plot the conductance of the last layer.

    Args:
        weights_last (np.ndarray):
            The weights of the last layer.
        layer_conductance_last (np.ndarray):
            The conductance of the last layer.
        figsize (tuple, optional):
            The figure size. Defaults to (12, 6).
        show (bool, optional):
            Whether to show the plot. Defaults

    Examples:
        >>> import numpy as np
            from spotpython.plot.xai import plot_conductance_last_layer
            weights_last = np.random.rand(10)
            layer_conductance_last = np.random.rand(10)
            plot_conductance_last_layer(weights_last, layer_conductance_last, show=True)
    """
    fig, ax = plt.subplots(figsize=figsize)
    ax.bar(range(len(weights_last)), weights_last / weights_last.max(), label="Weights", alpha=0.5)
    ax.bar(
        range(len(layer_conductance_last)),
        layer_conductance_last / layer_conductance_last.max(),
        label="Layer Conductance",
        alpha=0.5,
    )
    ax.set_xlabel("Weight Index")
    ax.set_ylabel("Normalized Value")
    ax.set_title("Layer Conductance vs. Weights")
    ax.legend()
    ax.xaxis.set_major_locator(MaxNLocator(integer=True))
    if show:
        plt.show()

plot_nn_values_hist(nn_values, net, nn_values_names='', color='C0', columns=2)

Plot the values of a neural network. Can be used to plot the weights, gradients, or activations of a neural network.

Parameters:

Name Type Description Default
nn_values dict

A dictionary with the values of the neural network. For example, the weights, gradients, or activations.

required
net object

A neural network.

required
color str

The color to use. Defaults to “C0”.

'C0'
columns int

The number of columns. Defaults to 2.

2
Source code in spotpython/plot/xai.py
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
def plot_nn_values_hist(nn_values, net, nn_values_names="", color="C0", columns=2) -> None:
    """
    Plot the values of a neural network.
    Can be used to plot the weights, gradients, or activations of a neural network.

    Args:
        nn_values (dict):
            A dictionary with the values of the neural network. For example,
            the weights, gradients, or activations.
        net (object):
            A neural network.
        color (str, optional):
            The color to use. Defaults to "C0".
        columns (int, optional):
            The number of columns. Defaults to 2.

    """
    n = len(nn_values)
    print(f"n:{n}")
    rows = n // columns + int(n % columns > 0)
    fig, ax = plt.subplots(rows, columns, figsize=(columns * 2.7, rows * 2.5))
    fig_index = 0
    for key in nn_values:
        key_ax = ax[fig_index // columns][fig_index % columns]
        sns.histplot(data=nn_values[key], bins=50, ax=key_ax, color=color, kde=True, stat="density")
        hidden_dim_str = (
            r"(%i $\to$ %i)" % (nn_values[key].shape[1], nn_values[key].shape[0])
            if len(nn_values[key].shape) > 1
            else ""
        )
        key_ax.set_title(f"{key} {hidden_dim_str}")
        # key_ax.set_title(f"Layer {key} - {net.layers[key].__class__.__name__}")
        fig_index += 1
    fig.suptitle(f"{nn_values_names} distribution for activation function {net.hparams.act_fn}", fontsize=14)
    fig.subplots_adjust(hspace=0.4, wspace=0.4)
    plt.show()

plot_nn_values_scatter(nn_values, layer_sizes, nn_values_names='', absolute=True, cmap='gray', figsize=(6, 6), return_reshaped=False, show=True)

Plot the values of a neural network including a marker for padding values. For simplicity, this example will annotate ‘P’ directly on the plot for padding values using a unique marker value approach.

Parameters:

Name Type Description Default
nn_values dict

A dictionary with the values of the neural network. For example, the weights, gradients, or activations.

required
layer_sizes dict

A dictionary with layer names as keys and their sizes as entries in NumPy array format.

required
nn_values_names str

The name of the values. Defaults to “”.

''
absolute bool

Whether to use the absolute values. Defaults to True.

True
cmap str

The colormap to use. Defaults to “gray”.

'gray'
figsize tuple

The figure size. Defaults to (6, 6).

(6, 6)
return_reshaped bool

Whether to return the reshaped values. Defaults to False.

False
show bool

Whether to show the plot. Defaults to True.

True

Returns:

Name Type Description
dict dict

A dictionary with the reshaped values.

Examples:

>>> from spotpython.utils.init import fun_control_init
    from spotpython.data.diabetes import Diabetes
    from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.hyperparameters.values import (
            get_default_hyperparameters_as_array, get_one_config_from_X)
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    # from spotpython.plot.xai import get_gradients
    fun_control = fun_control_init(
        _L_in=10, # 10: diabetes
        _L_out=1,
        _torchmetric="mean_squared_error",
        data_set=Diabetes(),
        core_model=NNLinearRegressor,
        hyperdict=LightHyperDict)
    X = get_default_hyperparameters_as_array(fun_control)
    config = get_one_config_from_X(X, fun_control)
    _L_in = fun_control["_L_in"]
    _L_out = fun_control["_L_out"]
    _torchmetric = fun_control["_torchmetric"]
    batch_size = 16
    model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
    gradients, layer_sizes = get_gradients(model, fun_control=fun_control, batch_size=batch_size, device = "cpu")
    plot_nn_values_scatter(nn_values=gradients, layer_sizes=layer_sizes, nn_values_names="Weights")
Source code in spotpython/plot/xai.py
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
def plot_nn_values_scatter(
    nn_values,
    layer_sizes,
    nn_values_names="",
    absolute=True,
    cmap="gray",
    figsize=(6, 6),
    return_reshaped=False,
    show=True,
) -> dict:
    """
    Plot the values of a neural network including a marker for padding values.
    For simplicity, this example will annotate 'P' directly on the plot for padding values
    using a unique marker value approach.

    Args:
        nn_values (dict):
            A dictionary with the values of the neural network. For example,
            the weights, gradients, or activations.
        layer_sizes (dict):
            A dictionary with layer names as keys and their sizes as entries in NumPy array format.
        nn_values_names (str, optional):
            The name of the values. Defaults to "".
        absolute (bool, optional):
            Whether to use the absolute values. Defaults to True.
        cmap (str, optional):
            The colormap to use. Defaults to "gray".
        figsize (tuple, optional):
            The figure size. Defaults to (6, 6).
        return_reshaped (bool, optional):
            Whether to return the reshaped values. Defaults to False.
        show (bool, optional):
            Whether to show the plot. Defaults to True.

    Returns:
        dict: A dictionary with the reshaped values.

    Examples:
        >>> from spotpython.utils.init import fun_control_init
            from spotpython.data.diabetes import Diabetes
            from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.hyperparameters.values import (
                    get_default_hyperparameters_as_array, get_one_config_from_X)
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            # from spotpython.plot.xai import get_gradients
            fun_control = fun_control_init(
                _L_in=10, # 10: diabetes
                _L_out=1,
                _torchmetric="mean_squared_error",
                data_set=Diabetes(),
                core_model=NNLinearRegressor,
                hyperdict=LightHyperDict)
            X = get_default_hyperparameters_as_array(fun_control)
            config = get_one_config_from_X(X, fun_control)
            _L_in = fun_control["_L_in"]
            _L_out = fun_control["_L_out"]
            _torchmetric = fun_control["_torchmetric"]
            batch_size = 16
            model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
            gradients, layer_sizes = get_gradients(model, fun_control=fun_control, batch_size=batch_size, device = "cpu")
            plot_nn_values_scatter(nn_values=gradients, layer_sizes=layer_sizes, nn_values_names="Weights")
    """
    if cmap == "gray":
        cmap = "gray"
    elif cmap == "BlueWhiteRed":
        cmap = colors.LinearSegmentedColormap.from_list("", ["blue", "white", "red"])
    elif cmap == "GreenYellowRed":
        cmap = colors.LinearSegmentedColormap.from_list("", ["green", "yellow", "red"])
    else:
        cmap = "viridis"

    res = {}
    padding_marker = np.nan  # Use NaN as a special marker for padding
    for layer, values in nn_values.items():
        if layer not in layer_sizes:
            print(f"Layer {layer} size not defined, skipping.")
            continue

        layer_shape = layer_sizes[layer]
        height, width = layer_shape if len(layer_shape) == 2 else (layer_shape[0], 1)  # Support linear layers

        print(f"{len(values)} values in Layer {layer}. Geometry: ({height}, {width})")

        total_size = height * width
        if len(values) < total_size:
            padding_needed = total_size - len(values)
            print(f"{padding_needed} padding values added to Layer {layer}.")
            values = np.append(values, [padding_marker] * padding_needed)  # Append padding values

        if absolute:
            reshaped_values = np.abs(values).reshape((height, width))
            # Mark padding values distinctly by setting them back to NaN
            reshaped_values[reshaped_values == np.abs(padding_marker)] = np.nan
        else:
            reshaped_values = values.reshape((height, width))

        _, ax = plt.subplots(figsize=figsize)
        cax = ax.imshow(reshaped_values, cmap=cmap, interpolation="nearest")

        for i in range(height):
            for j in range(width):
                if np.isnan(reshaped_values[i, j]):
                    ax.text(j, i, "P", ha="center", va="center", color="red")

        plt.colorbar(cax, label="Value")
        plt.title(f"{nn_values_names} Plot for {layer}")
        if show:
            plt.show()

        # Add reshaped_values to the dictionary res
        res[layer] = reshaped_values
    if return_reshaped:
        return res

sort_layers(data_dict)

Sorts a dictionary with keys in the format “Layer X” based on the numerical value X.

Parameters:

Name Type Description Default
data_dict dict

A dictionary with keys in the format “Layer X”.

required

Returns:

Name Type Description
dict dict

A dictionary with the keys sorted based on the numerical value X.

Examples:

>>> data_dict = {
...     "Layer 1": [1, 2, 3],
...     "Layer 3": [4, 5, 6],
...     "Layer 2": [7, 8, 9]
... }
>>> sort_layers(data_dict)
{'Layer 1': [1, 2, 3], 'Layer 2': [7, 8, 9], 'Layer 3': [4,
Source code in spotpython/plot/xai.py
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
def sort_layers(data_dict) -> dict:
    """
    Sorts a dictionary with keys in the format "Layer X" based on the numerical value X.

    Args:
        data_dict (dict): A dictionary with keys in the format "Layer X".

    Returns:
        dict: A dictionary with the keys sorted based on the numerical value X.

    Examples:
        >>> data_dict = {
        ...     "Layer 1": [1, 2, 3],
        ...     "Layer 3": [4, 5, 6],
        ...     "Layer 2": [7, 8, 9]
        ... }
        >>> sort_layers(data_dict)
        {'Layer 1': [1, 2, 3], 'Layer 2': [7, 8, 9], 'Layer 3': [4,

    """
    # Use a lambda function to extract the number X from "Layer X" and sort based on that number
    sorted_items = sorted(data_dict.items(), key=lambda item: int(item[0].split()[1]))
    # Create a new dictionary from the sorted items
    sorted_dict = dict(sorted_items)
    return sorted_dict

visualize_activations_distributions(activations, net, color='C0', columns=4, bins=50, show=True)

Plots the distribution of activations for each layer that were determined via the get_activations function.

Parameters:

Name Type Description Default
activations dict

A dictionary containing activations for each layer.

required
net Module

The neural network model.

required
color str

The color for the plot histogram. Defaults to “C0”.

'C0'
columns int

The number of columns for the subplots. Defaults to 4.

4
bins int

The number of bins for the histogram. Defaults to 50.

50
show bool

Whether to show the plot. Defaults to True.

True

Returns:

Type Description
None

None

Source code in spotpython/plot/xai.py
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
def visualize_activations_distributions(activations, net, color="C0", columns=4, bins=50, show=True) -> None:
    """Plots the distribution of activations for each layer
        that were determined via the get_activations function.

    Args:
        activations (dict): A dictionary containing activations for each layer.
        net (nn.Module): The neural network model.
        color (str): The color for the plot histogram. Defaults to "C0".
        columns (int): The number of columns for the subplots. Defaults to 4.
        bins (int): The number of bins for the histogram. Defaults to 50.
        show (bool): Whether to show the plot. Defaults to True.

    Returns:
        None
    """
    rows = math.ceil(len(activations) / columns)
    fig, ax = plt.subplots(rows, columns, figsize=(columns * 2.7, rows * 2.5))
    fig_index = 0
    for key in activations:
        key_ax = ax[fig_index // columns][fig_index % columns]
        sns.histplot(data=activations[key], bins=bins, ax=key_ax, color=color, kde=True, stat="density")
        key_ax.set_title(f"Layer {key} - {net.layers[key].__class__.__name__}")
        fig_index += 1
    fig.suptitle("Activation distribution", fontsize=14)
    fig.subplots_adjust(hspace=0.4, wspace=0.4)
    if show:
        plt.show()
    plt.close()

visualize_gradient_distributions(net, fun_control, batch_size, device='cpu', color='C0', xlabel=None, stat='count', use_kde=True, columns=2, normalize=True)

Plot the gradients distributions of a neural network.

Parameters:

Name Type Description Default
net object

A neural network.

required
fun_control dict

A dictionary with the function control.

required
batch_size int

The batch size.

required
device str

The device to use. Defaults to “cpu”.

'cpu'
color str

The color to use. Defaults to “C0”.

'C0'
xlabel str

The x label. Defaults to None.

None
stat str

The stat. Defaults to “count”.

'count'
use_kde bool

Whether to use kde. Defaults to True.

True
columns int

The number of columns. Defaults to 2.

2
normalize bool

Whether to normalize the input data. Defaults to True.

True

Returns:

Type Description
None

None

Source code in spotpython/plot/xai.py
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
def visualize_gradient_distributions(
    net,
    fun_control,
    batch_size,
    device="cpu",
    color="C0",
    xlabel=None,
    stat="count",
    use_kde=True,
    columns=2,
    normalize=True,
) -> None:
    """
    Plot the gradients distributions of a neural network.

    Args:
        net (object):
            A neural network.
        fun_control (dict):
            A dictionary with the function control.
        batch_size (int, optional):
            The batch size.
        device (str, optional):
            The device to use. Defaults to "cpu".
        color (str, optional):
            The color to use. Defaults to "C0".
        xlabel (str, optional):
            The x label. Defaults to None.
        stat (str, optional):
            The stat. Defaults to "count".
        use_kde (bool, optional):
            Whether to use kde. Defaults to True.
        columns (int, optional):
            The number of columns. Defaults to 2.
        normalize (bool, optional):
            Whether to normalize the input data. Defaults to True.

    Returns:
        None

    """
    grads, _ = get_gradients(net, fun_control, batch_size, device, normalize=normalize)
    plot_nn_values_hist(grads, net, nn_values_names="Gradients", color=color, columns=columns)

visualize_gradients(net, fun_control, batch_size, absolute=True, cmap='gray', figsize=(6, 6), device='cpu', normalize=True)

Scatter plots the gradients of a neural network.

Parameters:

Name Type Description Default
net object

A neural network.

required
fun_control dict

A dictionary with the function control.

required
batch_size int

The batch size.

required
absolute bool

Whether to use the absolute values. Defaults to True.

True
cmap str

The colormap to use. Defaults to “gray”.

'gray'
figsize tuple

The figure size. Defaults to (6, 6).

(6, 6)
device str

The device to use. Defaults to “cpu”.

'cpu'
normalize bool

Whether to normalize the input data. Defaults to True.

True

Returns:

Type Description
None

None

Source code in spotpython/plot/xai.py
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
def visualize_gradients(
    net, fun_control, batch_size, absolute=True, cmap="gray", figsize=(6, 6), device="cpu", normalize=True
) -> None:
    """
    Scatter plots the gradients of a neural network.

    Args:
        net (object):
            A neural network.
        fun_control (dict):
            A dictionary with the function control.
        batch_size (int, optional):
            The batch size.
        absolute (bool, optional):
            Whether to use the absolute values. Defaults to True.
        cmap (str, optional):
            The colormap to use. Defaults to "gray".
        figsize (tuple, optional):
            The figure size. Defaults to (6, 6).
        device (str, optional):
            The device to use. Defaults to "cpu".
        normalize (bool, optional):
            Whether to normalize the input data. Defaults to True.

    Returns:
        None
    """
    grads, layer_sizes = get_gradients(
        net=net,
        fun_control=fun_control,
        batch_size=batch_size,
        device=device,
        normalize=normalize,
    )
    plot_nn_values_scatter(
        nn_values=grads,
        layer_sizes=layer_sizes,
        nn_values_names="Gradients",
        absolute=absolute,
        cmap=cmap,
        figsize=figsize,
    )

visualize_mean_activations(mean_activations, layer_sizes, absolute=True, cmap='gray', figsize=(6, 6))

Scatter plots the mean activations of a neural network for each layer. means_activations is a dictionary with the mean activations of the neural network computed via the get_activations function.

Parameters:

Name Type Description Default
mean_activations dict

A dictionary with the mean activations of the neural network.

required
layer_sizes dict

A dictionary with layer names as keys and their sizes as entries in NumPy array format.

required
absolute bool

Whether to use the absolute values. Defaults to True.

True
cmap str

The colormap to use. Defaults to “gray”.

'gray'
figsize tuple

The figure size. Defaults to (6, 6).

(6, 6)

Returns:

Type Description
None

None

Examples:

>>> from spotpython.plot.xai import get_activations
    activations, mean_activations, layer_sizes = get_activations(net, fun_control)
    visualize_mean_activations(mean_activations, layer_sizes)
Source code in spotpython/plot/xai.py
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
def visualize_mean_activations(mean_activations, layer_sizes, absolute=True, cmap="gray", figsize=(6, 6)) -> None:
    """
    Scatter plots the mean activations of a neural network for each layer.
    means_activations is a dictionary with the mean activations of the neural network computed via
    the get_activations function.

    Args:
        mean_activations (dict):
            A dictionary with the mean activations of the neural network.
        layer_sizes (dict):
            A dictionary with layer names as keys and their sizes as entries in NumPy array format.
        absolute (bool, optional):
            Whether to use the absolute values. Defaults to True.
        cmap (str, optional):
            The colormap to use. Defaults to "gray".
        figsize (tuple, optional):
            The figure size. Defaults to (6, 6).

    Returns:
        None

    Examples:
        >>> from spotpython.plot.xai import get_activations
            activations, mean_activations, layer_sizes = get_activations(net, fun_control)
            visualize_mean_activations(mean_activations, layer_sizes)

    """
    plot_nn_values_scatter(
        nn_values=mean_activations,
        layer_sizes=layer_sizes,
        nn_values_names="Average Activations",
        absolute=absolute,
        cmap=cmap,
        figsize=figsize,
    )

visualize_weights(net, absolute=True, cmap='gray', figsize=(6, 6))

Scatter plots the weights of a neural network.

Parameters:

Name Type Description Default
net object

A neural network.

required
absolute bool

Whether to use the absolute values. Defaults to True.

True
cmap str

The colormap to use. Defaults to “gray”.

'gray'
figsize tuple

The figure size. Defaults to (6, 6).

(6, 6)

Returns:

Type Description
None

None

Examples:

>>> from spotpython.utils.init import fun_control_init
    from spotpython.data.diabetes import Diabetes
    from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
    from spotpython.hyperdict.light_hyper_dict import LightHyperDict
    from spotpython.hyperparameters.values import (
            get_default_hyperparameters_as_array, get_one_config_from_X)
    from spotpython.plot.xai import visualize_weights
    fun_control = fun_control_init(
        _L_in=10, # 10: diabetes
        _L_out=1,
        _torchmetric="mean_squared_error",
        data_set=Diabetes(),
        core_model=NNLinearRegressor,
        hyperdict=LightHyperDict)
    X = get_default_hyperparameters_as_array(fun_control)
    config = get_one_config_from_X(X, fun_control)
    _L_in = fun_control["_L_in"]
    _L_out = fun_control["_L_out"]
    _torchmetric = fun_control["_torchmetric"]
    batch_size = 16
    model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
    visualize_weights(net=model, absolute=True, cmap="gray", figsize=(6, 6))
Source code in spotpython/plot/xai.py
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
def visualize_weights(net, absolute=True, cmap="gray", figsize=(6, 6)) -> None:
    """
    Scatter plots the weights of a neural network.

    Args:
        net (object):
            A neural network.
        absolute (bool, optional):
            Whether to use the absolute values. Defaults to True.
        cmap (str, optional):
            The colormap to use. Defaults to "gray".
        figsize (tuple, optional):
            The figure size. Defaults to (6, 6).

    Returns:
        None

    Examples:
        >>> from spotpython.utils.init import fun_control_init
            from spotpython.data.diabetes import Diabetes
            from spotpython.light.regression.nn_linear_regressor import NNLinearRegressor
            from spotpython.hyperdict.light_hyper_dict import LightHyperDict
            from spotpython.hyperparameters.values import (
                    get_default_hyperparameters_as_array, get_one_config_from_X)
            from spotpython.plot.xai import visualize_weights
            fun_control = fun_control_init(
                _L_in=10, # 10: diabetes
                _L_out=1,
                _torchmetric="mean_squared_error",
                data_set=Diabetes(),
                core_model=NNLinearRegressor,
                hyperdict=LightHyperDict)
            X = get_default_hyperparameters_as_array(fun_control)
            config = get_one_config_from_X(X, fun_control)
            _L_in = fun_control["_L_in"]
            _L_out = fun_control["_L_out"]
            _torchmetric = fun_control["_torchmetric"]
            batch_size = 16
            model = fun_control["core_model"](**config, _L_in=_L_in, _L_out=_L_out, _torchmetric=_torchmetric)
            visualize_weights(net=model, absolute=True, cmap="gray", figsize=(6, 6))
    """
    weights, layer_sizes = get_weights(net)
    plot_nn_values_scatter(
        nn_values=weights,
        layer_sizes=layer_sizes,
        nn_values_names="Weights",
        absolute=absolute,
        cmap=cmap,
        figsize=figsize,
    )

visualize_weights_distributions(net, color='C0', columns=2)

Plot the weights distributions of a neural network.

Parameters:

Name Type Description Default
net object

A neural network.

required
color str

The color to use. Defaults to “C0”.

'C0'
columns int

The number of columns. Defaults to 2.

2

Returns:

Type Description
None

None

Source code in spotpython/plot/xai.py
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
def visualize_weights_distributions(net, color="C0", columns=2) -> None:
    """
    Plot the weights distributions of a neural network.

    Args:
        net (object):
            A neural network.
        color (str, optional):
            The color to use. Defaults to "C0".
        columns (int, optional):
            The number of columns. Defaults to 2.

    Returns:
        None

    """
    weights, _ = get_weights(net)
    plot_nn_values_hist(weights, net, nn_values_names="Weights", color=color, columns=columns)