X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=b505156b49abb389f90f0640fae3d4d59dc9a4e4;hb=126857a5ef0f205a1d77f62aaf1ee283061396d8;hp=c3b7e09c1714199729a737a70725b3944ab787b7;hpb=a0e547917131af0b353e3bf31a062c9b35c8dd18;p=beaver.git diff --git a/beaver.py b/beaver.py index c3b7e09..b505156 100755 --- a/beaver.py +++ b/beaver.py @@ -68,8 +68,6 @@ parser.add_argument("--no_checkpoint", action="store_true", default=False) 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") ############################## @@ -81,6 +79,12 @@ parser.add_argument("--maze_width", type=int, default=21) 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() @@ -169,11 +173,36 @@ def compute_perplexity(model, split="train"): ###################################################################### -def one_shot(gpt, task): +def oneshot_policy_loss(output, policies, mask): + targets = policies.permute(0, 2, 1) * mask.unsqueeze(-1) + output = output * mask.unsqueeze(-1) + return -(output.log_softmax(-1) * targets).sum() / mask.sum() + + +# loss = (output.softmax(-1) - targets).abs().max(-1).values.mean() + + +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 + 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(), @@ -185,16 +214,17 @@ def one_shot(gpt, task): 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 input, policies in task.policy_batches(split="train"): + #### + # print(f'{input.size()=} {policies.size()=}') + # s = maze.stationary_densities( + # exit(0) + #### + mask = input == maze.v_empty + output_gpt = gpt(mygpt.BracketedSequence(input), 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() - ) + + loss = compute_loss(output, policies, mask) acc_train_loss += loss.item() * input.size(0) nb_train_samples += input.size(0) @@ -203,16 +233,11 @@ def one_shot(gpt, task): 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 input, policies in task.policy_batches(split="test"): + mask = input == maze.v_empty + output_gpt = gpt(mygpt.BracketedSequence(input), 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() - ) + loss = compute_loss(output, policies, mask) acc_test_loss += loss.item() * input.size(0) nb_test_samples += input.size(0) @@ -222,18 +247,21 @@ def one_shot(gpt, task): # ------------------- input = task.test_input[:32, : task.height * task.width] - targets = task.test_policies[:32] - output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x + targets = task.test_policies[:32].permute(0, 2, 1) + output_gpt = gpt(mygpt.BracketedSequence(input), 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) + scores = ( + (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0 + ).float() + scores = scores.reshape(-1, task.height, task.width) + input = input.reshape(-1, task.height, task.width) maze.save_image( - os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"), + os.path.join( + args.result_dir, + f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png", + ), mazes=input, - score_paths=losses, + score_paths=scores, ) # ------------------- @@ -290,7 +318,7 @@ class TaskMaze(Task): 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, @@ -300,7 +328,7 @@ class TaskMaze(Task): 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 @@ -317,16 +345,16 @@ class TaskMaze(Task): 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}", ): @@ -487,8 +515,8 @@ log_string(f"learning_rate_schedule {learning_rate_schedule}") ############################## -if args.one_shot: - one_shot(model, task) +if args.oneshot: + oneshot(model, task) exit(0) ############################## @@ -520,7 +548,7 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs): 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()