ParameterGradientGeometry

class monitorch.lens.ParameterGradientGeometry(inplace: bool = True, normalize_by_size: bool = False, log_scale: bool = False, compute_correlation: bool = True, parameters: str | Iterable[str] = ('weight', 'bias'), line_aggregation: str | Iterable[str] = 'mean', range_aggregation: str | Iterable[str] | None = ('std', 'min-max'))

Bases: AbstractLens

Lens to examine geometry of gradients with respect to parameters.

Computes L2-norm or root-mean-square of gradients on every backward pass through parameter. Optionally computes correlation between gradients from two consecutive backward passes.

Computing correlation requires gradients from both epochs, hence the gradient will be saved after the computation is finished. It drives space consumption linearly by size of studied parameters.

Parameters:

• inplace (bool = True) – Flag indicating if computation should be done in-place or in-memory.

• normalize_by_size (bool = False) – Flag indicating if output norm should be divided by root of number of elements, thus obtaining RMS of output.

• log_scale (bool = False) – Flag indicating if logarithmic scale should be used.

• compute_correlation (bool = True) – Flag indicating if correlation between gradients from consecutive backward passes should be computed.

• parameters (str|Iterable[str] = ('weight', 'bias')) – Parameters which gradient will be studied.

• line_aggregation (str|Iterable[str] = 'mean') – Aggregation method for lines in plots.

• range_aggregation (str|Iterable[str]|None = ('std', 'min-max')) – Aggregation method for bands in plots.

Examples

Default usage is shown below.

>>> inspector = PyTorchInspector(
...     lenses = [
...         ParameterGradientGeometry(),
...     ],
...     module = mynet,
...     visualizer='matplotlib'
... )
>>>
>>> for epoch in range(N_EPOCHS):
...     for data, label in train_dataloader:
...         optimizer.zero_grad()
...         prediction = mynet(data)
...         loss = loss_fn(prediction, label)
...         loss.backward()
...         optimizer.step()
...
...     inspector.tick_epoch()
>>>
>>> inspector.visualizer.show_fig()

detach_from_module()

Detaches lens from module.

Detaches gatherers and resets inner state.

finalize_epoch()

Finaizes computations done through epoch.

Aggregates parameter gradient norms and optionally correlation according to line_aggregation and range_aggregation.

introduce_tags(vizualizer: AbstractVisualizer)

Introduces lens’s plots to visualizer.

For every parameter listed during initialization creates a small numerical plot ‘#PARAMETER_NAME Gradient Norm’ optionally creates a big comparison plot ‘#PARAMETER_NAME Gradient Correlation’.

Parameters:

visualzier (AbstractVisualizer) – A visualizer object to pass tag attributes to.

register_foreign_preprocessor(ext_ppr: AbstractPreprocessor, inspector_state)

Does not interact with foreign preprocessor.

register_leaf_module(module: Module, module_name: str, inspector_state)

Registers (or ignores) module.

Registers any module that has all of the parameters listed during initialization.

Parameters:

• module (torch.nn.Module) – The module object to hook gatherers onto.

• module_name (str) – Name of the module, module’s information will be passed to visaulizer under this name.

register_non_leaf_module(module: Module, module_name: str, inspector_state)

Registers (or ignores) module.

Registers any module that has all of the parameters listed during initialization.

Parameters:

• module (torch.nn.Module) – The module object to hook gatherers onto.

• module_name (str) – Name of the module, module’s information will be passed to visaulizer under this name.

reset_epoch()

Resets inner state.

Resets data computed during last epoch and resets preprocessors.

vizualize(vizualizer: AbstractVisualizer, epoch: int)

Passes computed data to visualizer.

Passes dictionary of per layer data to ‘#PARAMETER_NAME Gradient Norm’, the dictionary may look something like this.

OrderedDict([
    ('lin1',   {'mean' : 0.8}, {'min' : 0.2, 'max' : 0.9}),
    ('lin2',   {'mean' : 0.6}, {'min' : 0.3, 'max' : 0.7}),
])

Gradient correlation dictionary looks the same.

Parameters:

• visualizer (AbstractVisualizer) – The visualizer object responsbile for drawing plots.

• epoch (int) – Computation’s epoch number.