torchfunc.performance

Modules

• torchfunc.performance.layers
• torchfunc.performance.technology

This package allows you to get info and tips about performance of your neural networks.

Following functions should be considered as general recommendations. For specific/customized tips, use specific submodules.

torchfunc.performance.report(module: torch.nn.modules.module.Module) → Dict[str, Any]

Run essential module’s performance analysis with default settings.

Following tests will be performed:

• Module being an instance of torch.nn.ScriptModule

• Apex (mixed precision training) availability

• Any inplace ops used

• Analysis of compliance with TensorCores technology

• Any depthwise convolution used

Report returns data in machine-ready type; if you wish to have easy to follow guidelines use function tips.

Example:

model = torch.nn.Sequential(
    torch.nn.Linear(784, 100),
    torch.nn.ReLU(),
    torch.nn.Linear(100, 50),
    torch.nn.ReLU(),
    torch.nn.Linear(50, 10),
)
report = torchfunc.performance.report(model)

Parameters

module (torch.nn.Module) – Module to be tested against test suite.

Returns

Dictionary with keys:

• torchscript: True if module is an instance of torch.jit.ScriptModule

• apex: True if apex installed (recommended mixed precision training library from NVidia for PyTorch)

• tensorcores: same as torchscript.performance.technology.TensorCores

• inplace: same as torchscript.performance.layers.Inplace

• depthwise: same as torchscript.performance.layers.Depthwise

Return type

Dict[str, Any]

torchfunc.performance.tips(module: torch.nn.modules.module.Module, general: bool = True, specific: bool = True)

Return string describing possible performance improvements one can undertake.

Internally report will be called and it’s output parsed and described. It is the easiest way to get information about your module/network and to quickly check possible performance improvements you could use.

Example:

model = torch.nn.Sequential(
    torch.nn.Linear(784, 100),
    torch.nn.ReLU(),
    torch.nn.Linear(100, 50),
    torch.nn.ReLU(),
    torch.nn.Linear(50, 10),
)
print(torchfunc.performance.tips(model)) # Display found vulnerabilities

Parameters

• module (torch.nn.Module) – Module to be tested against test suite.

• general (bool, optional) – Return general (not specific to your module) tips. Default: True

• specific (bool, optional) – Return specific tips for your module. Default: True

Returns

Human readable version of report, highlighting steps one can take to improve module’s performance.

Return type

str