| 13 | |
| 14 | |
| 15 | class Net(torch.nn.Module): |
| 16 | |
| 17 | def __init__(self, |
| 18 | n_inputs, |
| 19 | n_outputs, |
| 20 | n_tasks, |
| 21 | args): |
| 22 | super(Net, self).__init__() |
| 23 | self.net = args.net |
| 24 | |
| 25 | # setup optimizer |
| 26 | self.opt = torch.optim.SGD(self.parameters(), lr=args.lr) |
| 27 | |
| 28 | # setup losses |
| 29 | self.bce = torch.nn.CrossEntropyLoss() |
| 30 | |
| 31 | self.nc_per_task = n_outputs |
| 32 | self.n_outputs = n_outputs |
| 33 | self.n_iter = args.n_iter |
| 34 | self.batch_size = args.batch_size # How many to process per update |
| 35 | |
| 36 | def forward(self, x, t): |
| 37 | output = self.net(x) |
| 38 | return output |
| 39 | |
| 40 | def observe(self, x, t, y): |
| 41 | self.train() |
| 42 | for batch_iter in range(self.n_iter): # How many times reprocess same batch |
| 43 | permutation = torch.randperm(x.size()[0]) # Shuffle |
| 44 | |
| 45 | for i in range(0, x.size()[0], self.batch_size): # Iterate mini-batches |
| 46 | self.zero_grad() |
| 47 | indices = permutation[i:i + self.batch_size] |
| 48 | batch_x, batch_y = x[indices], y[indices] |
| 49 | ptloss = self.bce(self.forward( |
| 50 | batch_x, t), |
| 51 | batch_y) |
| 52 | ptloss.backward() |
| 53 | self.opt.step() |
| 54 | |
| 55 | def get_hyperparam_list(self, args): |
| 56 | return [] |
nothing calls this directly
no outgoing calls
no test coverage detected