X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=beaver.py;h=33d174d89aa204e8129ba56046ff3abdeea31aca;hb=d63c681fdb2d6b5590991eaa4a2d9a5376678c67;hp=afec61d4a506161e0da2e449d2dfa3445e386110;hpb=71a5d04a1decec9d71be93cb816a15a8c0de83a2;p=beaver.git diff --git a/beaver.py b/beaver.py index afec61d..33d174d 100755 --- a/beaver.py +++ b/beaver.py @@ -81,6 +81,8 @@ parser.add_argument("--maze_width", type=int, default=21) parser.add_argument("--maze_nb_walls", type=int, default=15) +parser.add_argument("--oneshot_mode", type=str, default="head") + ###################################################################### args = parser.parse_args() @@ -172,9 +174,9 @@ def compute_perplexity(model, split="train"): def one_shot(gpt, task): t = gpt.training gpt.eval() - + dim_in = args.dim_model * (args.nb_blocks * 2 if args.oneshot_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(), @@ -186,16 +188,19 @@ 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.unsqueeze(-1) == maze.v_empty + output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x output = model(output_gpt) - targets = targets * (input.unsqueeze(-1) == maze.v_empty) - output = output * (input.unsqueeze(-1) == maze.v_empty) + targets = policies.permute(0, 2, 1) * mask + output = output * mask # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean() - loss = ( - -(output.log_softmax(-1) * targets).sum() - / (input == maze.v_empty).sum() - ) + loss = -(output.log_softmax(-1) * targets).sum() / mask.sum() acc_train_loss += loss.item() * input.size(0) nb_train_samples += input.size(0) @@ -204,16 +209,14 @@ 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.unsqueeze(-1) == maze.v_empty + output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x output = model(output_gpt) - targets = targets * (input.unsqueeze(-1) == maze.v_empty) - output = output * (input.unsqueeze(-1) == maze.v_empty) + targets = policies.permute(0, 2, 1) * mask + output = output * mask # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean() - loss = ( - -(output.log_softmax(-1) * targets).sum() - / (input == maze.v_empty).sum() - ) + loss = -(output.log_softmax(-1) * targets).sum() / mask.sum() acc_test_loss += loss.item() * input.size(0) nb_test_samples += input.size(0) @@ -223,11 +226,11 @@ 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_mode).x output = model(output_gpt) # losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1) - # losses = losses * (input == maze.v_empty) + # losses = losses * mask # losses = losses / losses.max() # losses = (output.softmax(-1) - targets).abs().max(-1).values # losses = (losses >= 0.05).float() @@ -237,7 +240,9 @@ def one_shot(gpt, task): losses = losses.reshape(-1, args.maze_height, args.maze_width) input = input.reshape(-1, args.maze_height, args.maze_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_mode}_{n_epoch:04d}.png" + ), mazes=input, score_paths=losses, ) @@ -296,7 +301,7 @@ 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, @@ -306,7 +311,7 @@ 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 @@ -323,16 +328,16 @@ 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}", ):