Open TuKo opened 10 months ago
2(1d1ts1)/1^-b^^-2/9 + ddx(Ass.V).dmb R2 = rm^2 + 2/1 (oh^-eipi^n2-2mf.rsc n^3x.u'/ r^2, 64pi rnt?d3/4.int -L^2 = 0rc . g3/2pi / Am^-pu.u.lk z = r^2 v^2 - 2.zip 3+1/2Nmf = -Aq' . 1/2hco.dnc Trc . dt^2 = oH OH = 5Tu' . 3/5q^2 dryHe = Frv^vu^2 10KT = / . 2/1 Api/-pu.ucc DM = ds^5/4 - pu^-2mû'. -hu! arcsen^2 = Apu'-mf'/1 5/0- . Mo - ddt^2E = GEo . Q^nk . pEc - Ec2 dt = 3/2j - dc2 . t^h + KPR+-4/27 mu^2 = -t^2 + arcsen r^216pirsc^-2/1 s = -2/1 rsc^2 - TKqu' +DEc o = -pmiEc't^3/2 a = d-pmu . QPRRct q = +- -2/1dp (-R) cos159 = ^8pi^KPR 24/7 + 2/3 pidrt^2ms / r^2arsen3/0 FRmo = Amu + KP445 -r^2h gcr =dt^3 . -1/2dt
@lrzpellegrini do you know if there are any workaround for this issue?
When trying to save a checkpoint the,
torch.save(checkpoint_data, fname, pickle_module=dill)
call insave_checkpoint()
function fails. It complains about a recursive definition when callingtorch.save()
inside the respective function.Additional context: I've followed the checkpoint example with the few changes below. The model is a CNN model (inheriting from
DynamicModule
) with an IncrementalClassifier as the output layer. This is in order to do class-incremental learning. Starts with 2 output classes and adapts in the train-eval loop. The dataset is MNIST loaded from torchvision (not avalanche) and the benchmark is created withnc_benchmark
with 10 experiences andtask_labels=False
. OnlyToTensor()
transform is used for the data. The strategy isNaive
. I'm using aReplayBuffer
, which I try removing it but still fails. The evaluation plugin has theaccuracy_metrics
, theforgetting_metrics
, and an interactive logger. The main loop calls adaptation(), train(), eval() in that order.System: Mac 2023 with M2
Versions: