parser.add_argument("--overwrite_results", action="store_true", default=False)
-parser.add_argument("--one_shot", action="store_true", default=False)
-
parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
##############################
parser.add_argument("--maze_nb_walls", type=int, default=15)
+parser.add_argument("--oneshot", action="store_true", default=False)
+
+parser.add_argument("--oneshot_input", type=str, default="head")
+
+parser.add_argument("--oneshot_output", type=str, default="policy")
+
######################################################################
args = parser.parse_args()
######################################################################
-def one_shot(gpt, task):
+def oneshot_policy_loss(mazes, output, policies, height, width):
+ masks = (mazes == maze.v_empty).unsqueeze(-1)
+ targets = policies.permute(0, 2, 1) * masks
+ output = output * masks
+ return -(output.log_softmax(-1) * targets).sum() / masks.sum()
+
+
+def oneshot_trace_loss(mazes, output, policies, height, width):
+ masks = mazes == maze.v_empty
+ targets = maze.stationary_densities(
+ mazes.view(-1, height, width), policies.view(-1, 4, height, width)
+ ).flatten(-2)
+ targets = targets * masks
+ output = output.squeeze(-1) * masks
+ return (output - targets).abs().sum() / masks.sum()
+
+
+def oneshot(gpt, task):
t = gpt.training
gpt.eval()
+
+ if args.oneshot_input == "head":
+ dim_in = args.dim_model
+ elif args.oneshot_input == "deep":
+ dim_in = args.dim_model * args.nb_blocks * 2
+ else:
+ raise ValueError(f"{args.oneshot_input=}")
+
+ if args.oneshot_output == "policy":
+ dim_out = 4
+ compute_loss = oneshot_policy_loss
+ elif args.oneshot_output == "trace":
+ dim_out = 1
+ compute_loss = oneshot_trace_loss
+ else:
+ raise ValueError(f"{args.oneshot_output=}")
+
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(),
- nn.Linear(args.dim_model, 4),
+ nn.Linear(args.dim_model, dim_out),
).to(device)
for n_epoch in range(args.nb_epochs):
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
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
+ for mazes, policies in task.policy_batches(split="train"):
+ ####
+ # print(f'{mazes.size()=} {policies.size()=}')
+ # s = maze.stationary_densities(
+ # exit(0)
+ ####
+ masks = mazes == maze.v_empty
+ output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
output = model(output_gpt)
- targets = targets * (input.unsqueeze(-1) == maze.v_empty)
- output = output * (input.unsqueeze(-1) == maze.v_empty)
- loss = (
- -(output.log_softmax(-1) * targets).sum()
- / (input == maze.v_empty).sum()
- + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
- )
- acc_train_loss += loss.item() * input.size(0)
- nb_train_samples += input.size(0)
+
+ loss = compute_loss(mazes, output, policies, task.height, task.width)
+ acc_train_loss += loss.item() * mazes.size(0)
+ nb_train_samples += mazes.size(0)
optimizer.zero_grad()
loss.backward()
optimizer.step()
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
+ for mazes, policies in task.policy_batches(split="test"):
+ output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
output = model(output_gpt)
- targets = targets * (input.unsqueeze(-1) == maze.v_empty)
- output = output * (input.unsqueeze(-1) == maze.v_empty)
- loss = (
- -(output.log_softmax(-1) * targets).sum()
- / (input == maze.v_empty).sum()
- + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
- )
- acc_test_loss += loss.item() * input.size(0)
- nb_test_samples += input.size(0)
+ loss = compute_loss(mazes, output, policies, task.height, task.width)
+ acc_test_loss += loss.item() * mazes.size(0)
+ nb_test_samples += mazes.size(0)
log_string(
f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
)
# -------------------
- input = task.test_input[:32, : task.height * task.width]
- targets = task.test_policies[:32]
- output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+ mazes = task.test_input[:32, : task.height * task.width]
+ policies = task.test_policies[:32]
+ output_gpt = gpt(mygpt.BracketedSequence(mazes), mode=args.oneshot_input).x
output = model(output_gpt)
- losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
- losses = losses * (input == maze.v_empty)
- losses = losses / losses.max()
- losses = losses.reshape(-1, args.maze_height, args.maze_width)
- input = input.reshape(-1, args.maze_height, args.maze_width)
+ if args.oneshot_output == "policy":
+ targets = policies.permute(0, 2, 1)
+ scores = (
+ (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
+ ).float()
+ elif args.oneshot_output == "trace":
+ targets = maze.stationary_densities(
+ mazes.view(-1, task.height, task.width),
+ policies.view(-1, 4, task.height, task.width),
+ ).flatten(-2)
+ scores = output.flatten(-2)
+ else:
+ raise ValueError(f"{args.oneshot_output=}")
+
+ scores = scores.reshape(-1, task.height, task.width)
+ mazes = mazes.reshape(-1, task.height, task.width)
+ targets = targets.reshape(-1, task.height, task.width)
maze.save_image(
- os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"),
- mazes=input,
- score_paths=losses,
+ os.path.join(
+ args.result_dir,
+ f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
+ ),
+ mazes=mazes,
+ score_paths=scores,
+ score_truth=targets,
)
# -------------------
progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
)
self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
- self.train_policies = train_policies.flatten(-2).permute(0, 2, 1).to(device)
+ self.train_policies = train_policies.flatten(-2).to(device)
test_mazes, test_paths, test_policies = maze.create_maze_data(
nb_test_samples,
progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
)
self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
- self.test_policies = test_policies.flatten(-2).permute(0, 2, 1).to(device)
+ self.test_policies = test_policies.flatten(-2).to(device)
self.nb_codes = self.train_input.max() + 1
def policy_batches(self, split="train", nb_to_use=-1):
assert split in {"train", "test"}
input = self.train_input if split == "train" else self.test_input
- targets = self.train_policies if split == "train" else self.test_policies
+ policies = self.train_policies if split == "train" else self.test_policies
input = input[:, : self.height * self.width]
- targets = targets * (input != maze.v_wall)[:, :, None]
+ policies = policies * (input != maze.v_wall)[:, None]
if nb_to_use > 0:
input = input[:nb_to_use]
- targets = targets[:nb_to_use]
+ policies = policies[:nb_to_use]
for batch in tqdm.tqdm(
- zip(input.split(self.batch_size), targets.split(self.batch_size)),
+ zip(input.split(self.batch_size), policies.split(self.batch_size)),
dynamic_ncols=True,
desc=f"epoch-{split}",
):
##############################
-if args.one_shot:
- one_shot(model, task)
+if args.oneshot:
+ oneshot(model, task)
exit(0)
##############################
elif args.optim == "adamw":
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
else:
- raise ValueError(f"Unknown optimizer {args.optim}.")
+ raise ValueError(f"{args.optim=}")
model.train()