def one_shot(gpt, task):
t = gpt.training
gpt.eval()
-
+ mode='head'
+ dim_in=args.dim_model * (args.nb_blocks * 2 if mode=='deep' else 1)
model = nn.Sequential(
- nn.Linear(args.dim_model, args.dim_model),
+ nn.Linear(dim_in, args.dim_model),
nn.ReLU(),
nn.Linear(args.dim_model, args.dim_model),
nn.ReLU(),
acc_train_loss, nb_train_samples = 0, 0
for input, targets in task.policy_batches(split="train"):
- output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+ output_gpt = gpt(mygpt.BracketedSequence(input), mode=mode).x
output = model(output_gpt)
targets = targets * (input.unsqueeze(-1) == maze.v_empty)
output = output * (input.unsqueeze(-1) == maze.v_empty)
acc_test_loss, nb_test_samples = 0, 0
for input, targets in task.policy_batches(split="test"):
- output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+ output_gpt = gpt(mygpt.BracketedSequence(input), mode=mode).x
output = model(output_gpt)
targets = targets * (input.unsqueeze(-1) == maze.v_empty)
output = output * (input.unsqueeze(-1) == maze.v_empty)
# -------------------
input = task.test_input[:32, : task.height * task.width]
targets = task.test_policies[:32]
- output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+ output_gpt = gpt(mygpt.BracketedSequence(input), mode=mode).x
output = model(output_gpt)
# losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
# losses = losses * (input == maze.v_empty)
m.bias.zero_()
m.weight.fill_(1.0)
- def forward(self, bs, with_readout=True):
+ def forward(self, bs, mode='standard'):
bs.x = F.pad(bs.x, (1, -1))
bs = self.embedding(bs)
- bs = self.trunk(bs)
- if with_readout:
+ if mode=='standard':
+ bs = self.trunk(bs)
bs = self.readout(bs)
+ elif mode=='head':
+ bs = self.trunk(bs)
+ elif mode=='deep':
+ r = []
+ for l in self.trunk:
+ bs = l(bs)
+ r += [ bs.slice() ]
+ bs = BracketedSequence(torch.cat(r, -1))
+ else:
+ raise ValueError
return bs