Loading a model with pytorch

Question

I'm having a problem loading my model in the Image Classifier project. First, I saved it:

model.class_to_idx = train_data.class_to_idx

checkpoint = {'arch': 'vgg19',
              'learn_rate': learn_rate,
              'epochs': epochs,
              'state_dict': model.state_dict(),
              'class_to_idx': model.class_to_idx,
              'optimizer': optimizer.state_dict(),
              'input_size': 25088,
              'output_size': 102,
              'momentum': momentum,
              'batch_size':64,
              'classifier' : classifier
             }

torch.save(checkpoint, 'checkpoint.pth')

Then I tried to load the project I had saved:

def load_checkpoint(filepath):
    checkpoint = torch.load(filepath)

    learn_rate = checkpoint['learn_rate']

    optimizer.load_state_dict(checkpoint['optimizer'])

    model = models.vgg16(pretrained=True)
    model.epochs = checkpoint['epochs']
    model.load_state_dict(checkpoint['state_dict'])
    model.class_to_idx = checkpoint['class_to_idx']
    model.classifier = checkpoint['classifier']

    return learn_rate, optimizer, model

learn_rate, optimizer, model = load_checkpoint('checkpoint.pth')

And I get an error when I try to load:

<ipython-input-75-5bd1aa042c7f> in load_checkpoint(filepath)
      9     model = models.vgg16(pretrained=True)
     10     model.epochs = checkpoint['epochs']
---> 11     model.load_state_dict(checkpoint['state_dict'])
     12     model.class_to_idx = checkpoint['class_to_idx']
     13     model.classifier = checkpoint['classifier']

RuntimeError: Error(s) in loading state_dict for VGG:
    Missing key(s) in state_dict: "classifier.0.weight", "classifier.0.bias", "classifier.3.weight", "classifier.3.bias", "classifier.6.weight", "classifier.6.bias". 
    Unexpected key(s) in state_dict: "classifier.fc1.weight", "classifier.fc1.bias", "classifier.fc2.weight", "classifier.fc2.bias".

This seems to be classifier issue. Does anyone know what's going on?

It looks like model.classifier has changed since you saved. Specifically it looks like while you're loading, model.classifier is an nn.Sequential but when you saved your model model.classifier was a custom module containing fc1 and fc2 layers. — jodag
– jodag, Commented Feb 8, 2020 at 23:35

ychnh · Accepted Answer · 2020-02-11 01:06:26Z

2

jodag's comment points at the heart of the issue. If fc1 fc2 correspond to classifier.0 classifier.3, classifier.6 you can adjust the dictionary to link them. When loading the weights to the model make sure to add the option strict=False.

You will need to retrain your model for the classifier - because your state dict misses weights for 3 layers but have 2 unused layer weights - but it should converge really quickly (from personal experience).

answered Feb 11, 2020 at 1:06

ychnh

1971 silver badge12 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Loading a model with pytorch

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related