Pytorch Custom

The pytorch*custom module contains functions that process models created with the pytorch package.

Feature Processing (pytorchcustom.fts)

Contains functions to extract features from a pytorch neural network

fns.pytorchcustom.fts.feature_extractor(model, lay, model_in, norm, device)

Function to extract the features from a given layer of a model

Parameters

  • model (loaded pytorch model) –

  • lay (str) – name of a layer in model

  • model_in (varies) – correctly formatted model input

  • norm (str) – {‘ft01’, ‘all01’, ‘none’} a string to indicate which normalization methodology to use;

  • device (torch.device) – device index to select

Returns

ft_mat (np.ndarray[(any, any, any), float]) – an array of all features extracted from a given layer; the first two dimensions are the size of the feature; the last dimension is the number of features in a layer

See also

fns.mat.normalize_mat() for information about choices for norm

fns.pytorchcustom.fts.parse_features(model, features)

Function to make a list of the features to plot

Prints error message and exits program for features = [‘Grid’]

Parameters

  • model (loaded pytorch model) – features (list[str]): list of features to plot, starting at feature 1

  • features (list[str]) –

Returns

  • n_features (int) – how may features to plot

  • features (list[int]) – list of features to plot, starting at feature 0

Model Prints (pytorchcustom.prints)

Contains functions to print out lists of options for model-related input arguments

fns.pytorchcustom.prints.print_layers(model)

Function that prints a list of layer names in a model

Parameters

model (loaded pytorch model) –

Returns

No Return Objects

fns.pytorchcustom.prints.print_features(model)

Function that prints how many features are extracted from a layer of a model

Parameters

model (loaded pytorch model) –

Returns

No Return Objects

Model Checks (pytorchcustom.checks)

Contains functions to check that the modle-related input arguments passed to are valid

fns.pytorchcustom.checks.check_layer(model, lay)

Function that checks if a layer name is in the model

Parameters

  • model (loaded pytorch model) –

  • lay (str) – name of layer to test

Returns

No Return Objects

fns.pytorchcustom.checks.check_features(model, features)

Function that checks if number of features requested are available from a layer

Parameters

  • model (loaded pytorch model) –

  • features (list[str]) – should be [‘Grid’], [‘All’], or a list of integers; features the script plans on extracting

Returns

No Return Objects

Calico Functions (pytorchcustom.calico)

Contains functions to create a calico network and do prints/checks on the calcio network inputs

fns.pytorchcustom.calico.check_calico_layer(model, lay, branch='None', catlay='None')

Function that checks if the layer is a valid selection for the split layer

Parameters

  • model (loaded pytorch model) –

  • lay (str) – name of layer to test

  • branch (str) – key used to identify which layers are on the secondary branch; use ‘None’ if the model only has one branch

  • catlay (str) – name of layer where the branches of the model are concatenated use ‘None’ if the model only has one branch

Returns

No Return Objects

fns.pytorchcustom.calico.check_calico_features(model, features)

Function that checks if number of features requested are available from a layer

Parameters

  • model (loaded pytorch model) –

  • features (str) – an integer; features the calico model scales

Returns

No Return Objects

fns.pytorchcustom.calico.eval_layer(model, lay, x)

Function to evaluate a pytorch layer given it’s name

Parameters

  • model (loaded pytorch model) –

  • lay (str) – name of layer to evaluate

  • x (Union[torch.FloatTensor, torch.cuda.FloatTensor]) – input to layer

Returns

x (torch.Tensor) – output of layer

Nested Cylinder Model Definition (pytorchcustom.field2PTW_model_definition)

fns.pytorchcustom.field2PTW_model_definition.conv2d_shape(w, h, k, s_w, s_h, p_w, p_h)

Function to calculate the new dimension of an image after a nn.Conv2d

Formula from: https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html
Assumes 2D input and dilation=1
Parameters

  • w (int) – starting width

  • h (int) – starting height

  • k (int) – kernel size

  • s_w (int) – stride size along the width

  • s_h (int) – stride size along the height

  • p_w (int) – padding size along the width

  • p_h (int) – padding size along the height

Returns

  • new_w (int) – number of pixels along the width

  • new_h (int) – number of pixels along the height

  • total (int) – total number of pixels in new image

class fns.pytorchcustom.field2PTW_model_definition.CNN_Interpretability_Module(img_size=(1, 1700, 500), kernel=5, features=12, depth=12, conv_onlyweights=True, batchnorm_onlybias=True, act_layer=<class 'torch.nn.modules.activation.GELU'>)

Convolutional Neural Network Module that creates the “interpretability layers”

Sequence of Conv2D, Batch Normalization, and Activation

Parameters

  • img_size (tuple[int, int, int]) – size of input (channels, height, width)

  • kernel (int) – size of square convolutional kernel

  • features (int) – number of features in the convolutional layers

  • depth (int) – number of interpretability blocks

  • conv_onlyweights (bool) – determines if convolutional layers learn only weights or weights and bias

  • batchnorm_onlybias (bool) – determiens if the batch normalization layers learn only bias or weights and bias

  • act_layer (nn.modules.activation) – torch neural network layer class to use as activation

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class fns.pytorchcustom.field2PTW_model_definition.CNN_Reduction_Module(img_size=(1, 1700, 500), size_threshold=(8, 8), kernel=5, stride=2, features=12, conv_onlyweights=True, batchnorm_onlybias=True, act_layer=<class 'torch.nn.modules.activation.GELU'>)

Convolutional Neural Network Module that creates the “reduction layers”

Sequence of Conv2D, Batch Normalization, and Activation

Parameters

  • img_size (tuple[int, int, int]) – size of input (channels, height, width)

  • size_threshold (tuple[int, int]) – (approximate) size of final, reduced image (height, width)

  • kernel (int) – size of square convolutional kernel

  • stride (int) – size of base stride for convolutional kernel

  • features (int) – number of features in the convolutional layers

  • conv_onlyweights (bool) – determines if convolutional layers learn only weights or weights and bias

  • batchnorm_onlybias (bool) – determiens if the batch normalization layers learn only bias or weights and bias

  • act_layer (nn.modules.activation) – torch neural network layer class to use as activation

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class fns.pytorchcustom.field2PTW_model_definition.field2PTW(img_size=(1, 1700, 500), size_threshold=(8, 8), kernel=5, features=12, interp_depth=12, conv_onlyweights=True, batchnorm_onlybias=True, act_layer=<class 'torch.nn.modules.activation.GELU'>, hidden_features=20)

Convolutional Neural Network Model that uses a single PVI field to predict one scalar value

Parameters

  • img_size (tuple[int, int, int]) – size of input (channels, height, width)

  • size_threshold (tuple[int, int]) – (approximate) size of reduced image (height, width)

  • kernel (int) – size of square convolutional kernel

  • features (int) – number of features in the convolutional layers

  • interp_depth (int) – number of interpretability blocks

  • conv_onlyweights (bool) – determines if convolutional layers learn only weights or weights and bias

  • batchnorm_onlybias (bool) – determiens if the batch normalization layers learn only bias or weights and bias

  • act_layer (nn.modules.activation) – torch neural network layer class to use as activation

  • hidden_features (int) – number of hidden features in the fully connected dense layer

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.