AttributeError: '_IncompatibleKeys' object has no attribute 'eval' CNN

vision
1

Hi, after the last epoch, i have an array of local model weights, which i use to load them onto an empty model to test it later on.

global_model = CNNMnist(args=args)
local_model1 = CNNMnist(args=args)
.........
#training
local_weights, local_losses = [], []
local_model = LocalUpdate(args=args, dataset=train_dataset,
                                      idxs=user_groups[idx], logger=logger)
            w, loss = local_model.update_weights(
                model=copy.deepcopy(global_model), global_round=epoch)
            local_weights.append(copy.deepcopy(w))

            local_losses.append(copy.deepcopy(loss))
# update global weights
        global_weights = average_weights(local_weights)
 # update global weights
        global_model.load_state_dict(global_weights)

The following line is where i start tinkering:

local_model1 = local_model1.load_state_dict(local_weights[0])

then i send the model to a function to get the metrics:
test_acc, test_loss, precision, recall, f1, confusion = sam_eval(args,local_model1,test_dataset)
the error is thrown right towards the beginning of the function:

def sam_eval(args, model, test_dataset):
    device = 'cuda' if args.gpu else 'cpu'
    criterion = nn.NLLLoss().to(device)
    model.eval() #this is the line throwing the error
    loss, total, correct = 0.0, 0.0, 0.0
    ..........

p.s. when i do the same thing with global_model, it works just fine. And i also tried adding local_model.to(device) line to my code.
Any help would be really appreciated.

2

I don’t think load_state_dict returns a reference to the model so you likely want to just change
local_model1 = local_model1.load_state_dict(local_weights[0]) to local_model1.load_state_dict(local_weights[0]).

3

thank you so much. this worked.
p.s. the code does not work if i don’t do:

local_model1.to(device)