Update chapter6.py

[nesaboz]

model_state will change after training, deepcopy is one way to reload the original one.

[nesaboz]

Note: I do use Python 3.10, so not sure if this is an issue in 3.7.

I ran into a problem by getting following image:

After some digging, seems like model_state gets updated after training. This causes next sbs instance (one for Adam) to start from previous model parameters (which load_state_dict was trying to prevent). Consider following code:

torch.manual_seed(42)
model = nn.Sequential()
model.add_module('linear', nn.Linear(1, 1))
loss_fn = nn.MSELoss(reduction='mean')

model_state = model.state_dict()
optimizer = optim.SGD(model.parameters(), lr= 0.1)
sbs = StepByStep(model, loss_fn, optimizer)
sbs.set_loaders(train_loader, val_loader)
print(model_state)
sbs.train(10)
print(model_state)

OrderedDict([('linear.weight', tensor([[0.7645]])), ('linear.bias', tensor([0.8300]))]) OrderedDict([('linear.weight', tensor([[1.5287]])), ('linear.bias', tensor([1.2382]))])

One solution is to just deepcopy model before any training and reuse it for load_state_dict. After this:

[dvgodoy]

Hi @nesaboz

Thank you for submitting this PR. You're absolutely right, it doesn't work properly in Python 3.10. I always test it on Google Colab (which currently uses Python 3.8) so I didn't notice this. Also, thank you for spotting and removing the duplicate function :-)

Best, Daniel