Hi, I got a strange error when I load my saved model

#1

After training, I saved my model with following code:

state = {'value_net1': agent.value_net1.state_dict(),
         'value_optimizer1': agent.value_optimizer1.state_dict(),
         'value_net2': agent.value_net2.state_dict(),
         'value_optimizer2': agent.value_optimizer2.state_dict(),
         'q_approximation': agent.q_approximation.state_dict(),
         'q_optimizer': agent.q_optimizer.state_dict(),
         }
torch.save(state, self.output_path + 'epi{}_{}.pkl'.format(self.episode_number,name))

then I need to load the data with sentence:

checkpoint = torch.load(path_of_data, map_location=torch.device('cpu'))

I got an error:

微信截图_20220818222333
微信截图_202208182223331136×264 32.6 KB

It is really strange. I alway save and load my model in this way, but I never got an error. I don’t know what wrong with my code this time.

#2

I’m unsure what’s causing the error as I cannot reproduce it locally by trying to pass invalid inputs to torch.load. Could you check what path_of_data contains?

#3

Thank you for your reply! I stored the variable “state” in the data_path. the “state” is a dictionary variable, which includes three model : value_net1,value_net2 and q_approximation, and their optimizers.

#4

I cannot reproduce the issue using some random modules and optimizers for the mentioned classes:

value_net1 = nn.Linear(1, 1)
value_optimizer1 = torch.optim.SGD(value_net1.parameters(), lr=1e-3)

value_net2 = nn.Linear(2, 2)
value_optimizer2 = torch.optim.SGD(value_net2.parameters(), lr=1e-3)

q_approximation = nn.Linear(3, 3)
q_optimizer = torch.optim.SGD(q_approximation.parameters(), lr=1e-3)


state = {'value_net1': value_net1.state_dict(),
         'value_optimizer1': value_optimizer1.state_dict(),
         'value_net2': value_net2.state_dict(),
         'value_optimizer2': value_optimizer2.state_dict(),
         'q_approximation': q_approximation.state_dict(),
         'q_optimizer': q_optimizer.state_dict(),
         }
torch.save(state, "tmp.pt")

checkpoint = torch.load("tmp.pt")

and can properly load the checkpoint.

#5

Thank you! Maybe there is something wrong with my model. But what strange is that these model perform welI during training, and can be saved. I seems need to check my models again. thank you again!