Loading a pre-trained model Customized

1

Hi guys, I have a problem when I load my model:

This is the code when I trained my model:

model = models.densenet121(pretrained = True)

train_on_gpu = torch.cuda.is_available()

for param in model.parameters():
param.requires_grad = False

from collections import OrderedDict
classifier = nn.Sequential(OrderedDict([
(‘c1’,nn.Linear(1024,740)),
(‘relu’, nn.ReLU()),
(‘c2’,nn.Linear(740,400)),
(‘relu1’,nn.ReLU()),
(‘c3’, nn.Linear(400,2)),
(‘output’, nn.LogSoftmax(dim=1))
]))

model.classifier = classifier

…
The train code
…

torch.save( model.state_dict(),‘classificador-1.pth’)

When I load the model:

state_dict = (checkpoint[‘state_dict’])
model = models.densenet121(pretrained = False)
model.load_state_dict(new_state_dict,strict=False)
model.fc = classifier
for parameter in model.parameters():
parameter.requires_grad = False
model.eval()

But I have wrong predicted class when I test the model,
because I just wanted two exits but I’m getting more.

2

It seems you can load the saved model, but what is not clear where is the variable classifier initialized before you use when loading:

state_dict = (checkpoint[‘state_dict’])
model = models.densenet121(pretrained = False)
model.load_state_dict(new_state_dict,strict=False)
model.fc = **classifier**
for parameter in model.parameters():
parameter.requires_grad = False
model.eval()

If the linear layer is not initialized/loaded with trained weights then the predictions will not be accurate.