def initialize(cls, model, optimizer):
train_steps = 0
max_grad_norm = 0
random_state = RandomState()
return TrainState(model=model,
optimizer=optimizer,
train_steps=train_steps,
random_state=random_state,
max_grad_norm=max_grad_norm)
def save(self, path):
gtd.io.makedirs(path)
# Store the latest random state
self.random_state = RandomState()
# save model
torch.save(self.model.state_dict(), join(path, 'model'))
torch.save(self.optimizer.state_dict(), join(path, 'optimizer'))
# pickle remaining attributes
d = {attr: getattr(self, attr) for attr in ['train_steps', 'random_state', 'max_grad_norm']}
with open(join(path, 'metadata.p'), 'wb') as f:
pickle.dump(d, f)
def save(self, path):
gtd.io.makedirs(path)
# Store the latest random state
self.random_state = RandomState()
# save model
chkpt_dct = {'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict(),
}
torch.save(chkpt_dct, join(path, 'checkpoint.pth.tar'))
# torch.save(self.model.state_dict(), join(path, 'model'))
# torch.save(self.optimizer.state_dict(), join(path, 'optimizer'))
# pickle remaining attributes
d = {attr: getattr(self, attr) for attr in ['train_steps', 'random_state', 'max_grad_norm']}
import json