org.sonar.l10n.py.rules.python.S6982.html Maven / Gradle / Ivy
This rule raises an issue when a PyTorch model state is loaded and torch.nn.Module.eval() or torch.nn.Module.train() is
not called.
Why is this an issue?
When using PyTorch it is common practice to load and save a model’s state from/to a .pth file. Doing so allows, for example, to
instantiate an untrained model and load learned parameters coming from another pre-trained model. Once the learned parameters are loaded to the model
it is important, before inferencing, to clearly state the intention by calling torch.nn.Module.eval() method to set the model in
evaluation mode or calling torch.nn.Module.train() to indicate the training will resume. Failing to call
torch.nn.Module.eval() would leave the model in training mode which may not be the intention.
How to fix it
Call the torch.nn.Module.eval() or torch.nn.Module.train() method on the model.
Code examples
Noncompliant code example
import torch
import torchvision.models as models
model = models.vgg16()
model.load_state_dict(torch.load('model_weights.pth')) # Noncompliant: model.train() or model.eval() was not called.
Compliant solution
import torch
import torchvision.models as models
model = models.vgg16()
model.load_state_dict(torch.load('model_weights.pth'))
model.eval()
Resources
Documentation
- PyTorch Documentation - eval - reference
- PyTorch Documentation - train - reference
- PyTorch Documentation - Autograd - Evaluation
Mode