def one_shot(gpt, task):
t = gpt.training
gpt.eval()
+
model = nn.Sequential(
+ nn.Linear(args.dim_model, args.dim_model),
+ nn.ReLU(),
nn.Linear(args.dim_model, args.dim_model),
nn.ReLU(),
nn.Linear(args.dim_model, 4),
).to(device)
- print(f"{args.nb_epochs=}")
-
for n_epoch in range(args.nb_epochs):
- print(f"{n_epoch=}")
learning_rate = learning_rate_schedule[n_epoch]
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
for input, targets in task.policy_batches(split="train"):
output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
output = model(output_gpt)
+ targets = targets * (input.unsqueeze(-1) == maze.v_empty)
+ output = output * (input.unsqueeze(-1) == maze.v_empty)
+ # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
loss = (
- -(output.log_softmax(-1) * targets).sum(-1).mean()
- + targets.xlogy(targets).sum(-1).mean()
+ -(output.log_softmax(-1) * targets).sum()
+ / (input == maze.v_empty).sum()
)
acc_train_loss += loss.item() * input.size(0)
nb_train_samples += input.size(0)
for input, targets in task.policy_batches(split="test"):
output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
output = model(output_gpt)
+ targets = targets * (input.unsqueeze(-1) == maze.v_empty)
+ output = output * (input.unsqueeze(-1) == maze.v_empty)
+ # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
loss = (
- -(output.log_softmax(-1) * targets).sum(-1).mean()
- + targets.xlogy(targets).sum(-1).mean()
+ -(output.log_softmax(-1) * targets).sum()
+ / (input == maze.v_empty).sum()
)
acc_test_loss += loss.item() * input.size(0)
nb_test_samples += input.size(0)
)
# -------------------
- input, targets = next(task.policy_batches(split="test"))
+ 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 = model(output_gpt)
- losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
- losses = losses / losses.max()
- print(f"{input.size()=} {losses.size()=} {losses.min()=} {losses.max()=}")
- losses = losses * (input == 0)
+ # losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
+ # losses = losses * (input == maze.v_empty)
+ # losses = losses / losses.max()
+ # losses = (output.softmax(-1) - targets).abs().max(-1).values
+ # losses = (losses >= 0.05).float()
+ losses = (
+ (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
+ ).float()
losses = losses.reshape(-1, args.maze_height, args.maze_width)
input = input.reshape(-1, args.maze_height, args.maze_width)
maze.save_image(