X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=b505156b49abb389f90f0640fae3d4d59dc9a4e4;hb=126857a5ef0f205a1d77f62aaf1ee283061396d8;hp=4f694dab0f4a42a34ddc9c9375c793180456dc66;hpb=731aa9de1343e3e7bd5102b5e553d14893c0680a;p=beaver.git diff --git a/beaver.py b/beaver.py index 4f694da..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() @@ -169,8 +173,99 @@ def compute_perplexity(model, split="train"): ###################################################################### -def one_shot(gpt, task): - pass +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(dim_in, args.dim_model), + nn.ReLU(), + nn.Linear(args.dim_model, args.dim_model), + nn.ReLU(), + nn.Linear(args.dim_model, 4), + ).to(device) + + for n_epoch in range(args.nb_epochs): + learning_rate = learning_rate_schedule[n_epoch] + optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) + + acc_train_loss, nb_train_samples = 0, 0 + 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) + + loss = compute_loss(output, policies, mask) + acc_train_loss += loss.item() * input.size(0) + nb_train_samples += input.size(0) + + optimizer.zero_grad() + loss.backward() + optimizer.step() + + acc_test_loss, nb_test_samples = 0, 0 + 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) + loss = compute_loss(output, policies, mask) + acc_test_loss += loss.item() * input.size(0) + nb_test_samples += input.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].permute(0, 2, 1) + output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x + output = model(output_gpt) + 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_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png", + ), + mazes=input, + score_paths=scores, + ) + # ------------------- + + gpt.train(t) ###################################################################### @@ -215,25 +310,25 @@ class TaskMaze(Task): self.width = width self.device = device - mazes_train, paths_train = maze.create_maze_data( + train_mazes, train_paths, train_policies = maze.create_maze_data( nb_train_samples, height=height, width=width, nb_walls=nb_walls, progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"), ) - mazes_train, paths_train = mazes_train.to(device), paths_train.to(device) - self.train_input = self.map2seq(mazes_train, paths_train) + self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device)) + self.train_policies = train_policies.flatten(-2).to(device) - mazes_test, paths_test = maze.create_maze_data( + test_mazes, test_paths, test_policies = maze.create_maze_data( nb_test_samples, height=height, width=width, nb_walls=nb_walls, progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"), ) - mazes_test, paths_test = mazes_test.to(device), paths_test.to(device) - self.test_input = self.map2seq(mazes_test, paths_test) + self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device)) + self.test_policies = test_policies.flatten(-2).to(device) self.nb_codes = self.train_input.max() + 1 @@ -247,6 +342,24 @@ class TaskMaze(Task): ): yield batch + 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 + policies = self.train_policies if split == "train" else self.test_policies + input = input[:, : self.height * self.width] + policies = policies * (input != maze.v_wall)[:, None] + + if nb_to_use > 0: + input = input[:nb_to_use] + policies = policies[:nb_to_use] + + for batch in tqdm.tqdm( + zip(input.split(self.batch_size), policies.split(self.batch_size)), + dynamic_ncols=True, + desc=f"epoch-{split}", + ): + yield batch + def vocabulary_size(self): return self.nb_codes @@ -294,10 +407,10 @@ class TaskMaze(Task): _, predicted_paths = self.seq2map(result) maze.save_image( os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"), - mazes, - paths, - predicted_paths, - maze.path_correctness(mazes, predicted_paths), + mazes=mazes, + target_paths=paths, + predicted_paths=predicted_paths, + path_correct=maze.path_correctness(mazes, predicted_paths), ) model.train(t) @@ -369,8 +482,6 @@ else: ###################################################################### -nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default - token_count = 0 for input in task.batches(split="train"): token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1)) @@ -404,13 +515,13 @@ 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) ############################## -if nb_epochs_finished >= nb_epochs: +if nb_epochs_finished >= args.nb_epochs: n_epoch = nb_epochs_finished train_perplexity = compute_perplexity(model, split="train") test_perplexity = compute_perplexity(model, split="test") @@ -425,7 +536,7 @@ if nb_epochs_finished >= nb_epochs: ############################## -for n_epoch in range(nb_epochs_finished, nb_epochs): +for n_epoch in range(nb_epochs_finished, args.nb_epochs): learning_rate = learning_rate_schedule[n_epoch] log_string(f"learning_rate {learning_rate}") @@ -437,7 +548,7 @@ for n_epoch in range(nb_epochs_finished, 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()