X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=c1f4dc7f0540e2dcbbdf8c71b9a3c1ca29db457b;hb=f91736e6e56152746b3c44342748b70ad1c89888;hp=ae4254430653eb236c87b3dfaa31d295f654e05e;hpb=c921b95d0ea5b94a893447fbd4792e5047ba6e99;p=picoclvr.git diff --git a/main.py b/main.py index ae42544..c1f4dc7 100755 --- a/main.py +++ b/main.py @@ -8,7 +8,7 @@ # torch.backends.cuda.matmul.allow_tf23 # torch.autocast(torch.bfloat16) -import math, sys, argparse, time, tqdm, itertools, os +import math, sys, argparse, time, tqdm, os import torch, torchvision from torch import nn @@ -27,24 +27,27 @@ else: ###################################################################### parser = argparse.ArgumentParser( - description="An implementation of GPT with cache to solve a toy geometric reasoning task." + description="An implementation of GPT with cache.", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, ) -parser.add_argument("--task", type=str, default="picoclvr") +parser.add_argument( + "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake, stack" +) -parser.add_argument("--log_filename", type=str, default="train.log") +parser.add_argument("--log_filename", type=str, default="train.log", help=" ") -parser.add_argument("--result_dir", type=str, default="results_default") +parser.add_argument("--result_dir", type=str, default=None) parser.add_argument("--seed", type=int, default=0) -parser.add_argument("--nb_epochs", type=int, default=25) +parser.add_argument("--nb_epochs", type=int, default=None) -parser.add_argument("--batch_size", type=int, default=25) +parser.add_argument("--batch_size", type=int, default=None) -parser.add_argument("--nb_train_samples", type=int, default=250000) +parser.add_argument("--nb_train_samples", type=int, default=None) -parser.add_argument("--nb_test_samples", type=int, default=10000) +parser.add_argument("--nb_test_samples", type=int, default=None) parser.add_argument("--optim", type=str, default="adam") @@ -92,12 +95,79 @@ parser.add_argument("--maze_width", type=int, default=21) parser.add_argument("--maze_nb_walls", type=int, default=15) +############################## +# Snake options + +parser.add_argument("--snake_height", type=int, default=6) + +parser.add_argument("--snake_width", type=int, default=8) + +parser.add_argument("--snake_nb_colors", type=int, default=5) + +parser.add_argument("--snake_length", type=int, default=200) + +############################## +# Snake options + +parser.add_argument("--stack_nb_steps", type=int, default=100) + +parser.add_argument("--stack_nb_stacks", type=int, default=1) + +parser.add_argument("--stack_nb_digits", type=int, default=3) + +parser.add_argument("--stack_fraction_values_for_train", type=float, default=None) + ###################################################################### args = parser.parse_args() assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"} +if args.result_dir is None: + args.result_dir = f"results_{args.task}" + +###################################################################### + +default_args = { + "picoclvr": { + "nb_epochs": 25, + "batch_size": 25, + "nb_train_samples": 250000, + "nb_test_samples": 10000, + }, + "mnist": { + "nb_epochs": 25, + "batch_size": 10, + "nb_train_samples": 250000, + "nb_test_samples": 10000, + }, + "maze": { + "nb_epochs": 25, + "batch_size": 25, + "nb_train_samples": 250000, + "nb_test_samples": 10000, + }, + "snake": { + "nb_epochs": 5, + "batch_size": 25, + "nb_train_samples": 250000, + "nb_test_samples": 10000, + }, + "stack": { + "nb_epochs": 5, + "batch_size": 25, + "nb_train_samples": 100000, + "nb_test_samples": 1000, + }, +} + +if args.task in default_args: + for k, v in default_args[args.task].items(): + if getattr(args, k) is None: + setattr(args, k, v) + +###################################################################### + try: os.mkdir(args.result_dir) except FileExistsError: @@ -135,10 +205,29 @@ for n in vars(args): ###################################################################### +# ra_mask is boolean, with 1s on the values to generate + + def masked_inplace_autoregression( - model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu") + model, + batch_size, + input, + ar_mask, + forbidden_tokens=None, + progress_bar_desc="autoregression", + device=torch.device("cpu"), ): - for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)): + # p = logits.softmax(1) + # entropy[:,s]= p.xlogy(p).sum(1) / math.log(2) + batches = zip(input.split(batch_size), ar_mask.split(batch_size)) + if progress_bar_desc is not None: + batches = tqdm.tqdm( + batches, + dynamic_ncols=True, + desc=progress_bar_desc, + total=input.size(0) // batch_size, + ) + for input, ar_mask in batches: i = (ar_mask.sum(0) > 0).nonzero() if i.min() > 0: model( @@ -274,6 +363,7 @@ class TaskPicoCLVR(Task): input, ar_masks, forbidden_tokens, + progress_bar_desc=None, device=self.device, ) model.train(t) @@ -451,9 +541,49 @@ class TaskPicoCLVR(Task): 0, ) - image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png") + image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png") torchvision.utils.save_image( - img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0 + img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=0.0 + ) + log_string(f"wrote {image_name}") + + +###################################################################### + + +class TaskMNIST(Task): + def __init__(self, batch_size, device=torch.device("cpu")): + self.device = device + self.batch_size = batch_size + + def batches(self, split="train"): + assert split in {"train", "test"} + data_set = torchvision.datasets.MNIST( + root="./data", train=(split == "train"), download=True + ) + data_input = data_set.data.view(-1, 28 * 28).long() + if args.nb_train_samples is not None: + data_input = data_input[: args.nb_train_samples] + for batch in tqdm.tqdm( + data_input.split(self.batch_size), desc=f"epoch-{split}" + ): + yield batch + + def vocabulary_size(self): + return 256 + + def produce_results(self, n_epoch, model): + results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64) + ar_mask = torch.full_like(results, 1) + masked_inplace_autoregression( + model, self.batch_size, results, ar_mask, device=self.device + ) + image_name = os.path.join(args.result_dir, f"mnist_result_{n_epoch:04d}.png") + torchvision.utils.save_image( + 1 - results.reshape(-1, 1, 28, 28) / 255.0, + image_name, + nrow=16, + pad_value=0.8, ) log_string(f"wrote {image_name}") @@ -486,7 +616,7 @@ class TaskMaze(Task): self.width = width self.device = device - train_mazes, train_paths, train_policies = maze.create_maze_data( + train_mazes, train_paths, _ = maze.create_maze_data( nb_train_samples, height=height, width=width, @@ -494,9 +624,8 @@ 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).to(device) - test_mazes, test_paths, test_policies = maze.create_maze_data( + test_mazes, test_paths, _ = maze.create_maze_data( nb_test_samples, height=height, width=width, @@ -504,9 +633,8 @@ 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).to(device) - self.nb_codes = self.train_input.max() + 1 + self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1 def batches(self, split="train", nb_to_use=-1, desc=None): assert split in {"train", "test"} @@ -520,64 +648,88 @@ class TaskMaze(Task): ): yield batch - def policy_batches(self, split="train", nb_to_use=-1, desc=None): - 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] - - if desc is None: - desc = f"epoch-{split}" - for batch in tqdm.tqdm( - zip(input.split(self.batch_size), policies.split(self.batch_size)), - dynamic_ncols=True, - desc=desc, - ): - yield batch - def vocabulary_size(self): return self.nb_codes def compute_error(self, model, split="train", nb_to_use=-1): nb_total, nb_correct = 0, 0 - for input in task.batches(split, nb_to_use): + count = torch.zeros( + self.width * self.height, + self.width * self.height, + device=self.device, + dtype=torch.int64, + ) + for input in tqdm.tqdm( + task.batches(split, nb_to_use), + dynamic_ncols=True, + desc=f"test-mazes", + ): result = input.clone() ar_mask = result.new_zeros(result.size()) ar_mask[:, self.height * self.width :] = 1 result *= 1 - ar_mask masked_inplace_autoregression( - model, self.batch_size, result, ar_mask, device=self.device + model, + self.batch_size, + result, + ar_mask, + progress_bar_desc=None, + device=self.device, ) mazes, paths = self.seq2map(result) - nb_correct += maze.path_correctness(mazes, paths).long().sum() + path_correctness = maze.path_correctness(mazes, paths) + nb_correct += path_correctness.long().sum() nb_total += mazes.size(0) - return nb_total, nb_correct + optimal_path_lengths = ( + (input[:, self.height * self.width :] == maze.v_path).long().sum(1) + ) + predicted_path_lengths = ( + (result[:, self.height * self.width :] == maze.v_path).long().sum(1) + ) + optimal_path_lengths = optimal_path_lengths[path_correctness] + predicted_path_lengths = predicted_path_lengths[path_correctness] + count[optimal_path_lengths, predicted_path_lengths] += 1 + + if count.max() == 0: + count = None + else: + count = count[ + : count.sum(1).nonzero().max() + 1, : count.sum(0).nonzero().max() + 1 + ] + + return nb_total, nb_correct, count def produce_results(self, n_epoch, model): with torch.autograd.no_grad(): t = model.training model.eval() - train_nb_total, train_nb_correct = self.compute_error( + train_nb_total, train_nb_correct, count = self.compute_error( model, "train", nb_to_use=1000 ) log_string( - f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%" + f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%" ) - test_nb_total, test_nb_correct = self.compute_error( + test_nb_total, test_nb_correct, count = self.compute_error( model, "test", nb_to_use=1000 ) log_string( - f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%" + f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%" ) + if count is not None: + proportion_optimal = count.diagonal().sum().float() / count.sum() + log_string(f"proportion_optimal_test {proportion_optimal*100:.02f}%") + with open( + os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.txt"), "w" + ) as f: + for i in range(count.size(0)): + for j in range(count.size(1)): + eol = " " if j < count.size(1) - 1 else "\n" + f.write(f"{count[i,j]}{eol}") + input = self.test_input[:48] result = input.clone() ar_mask = result.new_zeros(result.size()) @@ -589,13 +741,15 @@ class TaskMaze(Task): mazes, paths = self.seq2map(input) _, predicted_paths = self.seq2map(result) - filename = f"result_{n_epoch:04d}.png" + + filename = os.path.join(args.result_dir, f"maze_result_{n_epoch:04d}.png") maze.save_image( - os.path.join(args.result_dir, filename), + filename, mazes=mazes, target_paths=paths, predicted_paths=predicted_paths, path_correct=maze.path_correctness(mazes, predicted_paths), + path_optimal=maze.path_optimality(paths, predicted_paths), ) log_string(f"wrote {filename}") @@ -605,6 +759,242 @@ class TaskMaze(Task): ###################################################################### +import snake + + +class TaskSnake(Task): + def __init__( + self, + nb_train_samples, + nb_test_samples, + batch_size, + height, + width, + nb_colors, + length, + prompt_length, + device=torch.device("cpu"), + ): + self.batch_size = batch_size + self.height = height + self.width = width + self.device = device + self.prompt_length = prompt_length + + self.train_input, self.train_prior_visits, _, _ = snake.generate_sequences( + nb_train_samples, + height, + width, + nb_colors, + length, + prompt_length, + self.device, + ) + self.test_input, self.test_prior_visits, _, _ = snake.generate_sequences( + nb_test_samples, + height, + width, + nb_colors, + length, + prompt_length, + self.device, + ) + + self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1 + + def batches(self, split="train", nb_to_use=-1, desc=None): + assert split in {"train", "test"} + input = self.train_input if split == "train" else self.test_input + if nb_to_use > 0: + input = input[:nb_to_use] + if desc is None: + desc = f"epoch-{split}" + for batch in tqdm.tqdm( + input.split(self.batch_size), dynamic_ncols=True, desc=desc + ): + yield batch + + def vocabulary_size(self): + return self.nb_codes + + def produce_results(self, n_epoch, model): + with torch.autograd.no_grad(): + t = model.training + model.eval() + + def compute_nb_correct(input, prior_visits): + result = input.clone() + i = torch.arange(result.size(1), device=result.device)[None, :] + ar_mask = ( + torch.logical_and(i >= self.prompt_length * 2, i % 2 == 0) + .long() + .expand_as(result) + ) + result *= 1 - ar_mask + + # snake.solver(result,ar_mask) + + masked_inplace_autoregression( + model, self.batch_size, result, ar_mask, device=self.device + ) + + nb_total = ((prior_visits > 0) * ar_mask).sum() + + nb_correct = ( + (result == input).long() * (prior_visits > 0) * ar_mask + ).sum() + + # nb_total = result.size(0) + # nb_correct = ((result - input).abs().sum(1) == 0).sum() + + return nb_total, nb_correct + + # train_nb_total, train_nb_correct = compute_nb_correct( + # self.train_input, self.train_prior_visits + # ) + + # log_string( + # f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%" + # ) + + test_nb_total, test_nb_correct = compute_nb_correct( + self.test_input[:1000], self.test_prior_visits[:1000] + ) + + log_string( + f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%" + ) + + model.train(t) + + +###################################################################### + + +import stack + + +class TaskStack(Task): + def __init__( + self, + nb_train_samples, + nb_test_samples, + batch_size, + nb_steps, + nb_stacks, + nb_digits, + fraction_values_for_train=None, + device=torch.device("cpu"), + ): + self.batch_size = batch_size + self.nb_steps = nb_steps + self.nb_stacks = nb_stacks + self.nb_digits = nb_digits + self.device = device + + if fraction_values_for_train is None: + values_for_train = None + values_for_test = None + else: + all = torch.randperm(10**nb_digits) + nb_for_train = int(all.size(0) * fraction_values_for_train) + values_for_train = all[:nb_for_train] + values_for_test = all[nb_for_train:] + + self.train_input, self.train_stack_counts = stack.generate_sequences( + nb_train_samples, + nb_steps, + nb_stacks, + nb_digits, + values_for_train, + self.device, + ) + + self.test_input, self.test_stack_counts = stack.generate_sequences( + nb_test_samples, + nb_steps, + nb_stacks, + nb_digits, + values_for_test, + self.device, + ) + + i = torch.logical_and(self.test_input % 2 == 1, self.test_input < 2 * nb_stacks) + counts = self.test_stack_counts.flatten()[i.flatten()] + counts = F.one_hot(counts).sum(0) + log_string(f"pop_stack_counts {counts}") + + self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1 + + def batches(self, split="train", nb_to_use=-1, desc=None): + assert split in {"train", "test"} + input = self.train_input if split == "train" else self.test_input + if nb_to_use > 0: + input = input[:nb_to_use] + if desc is None: + desc = f"epoch-{split}" + for batch in tqdm.tqdm( + input.split(self.batch_size), dynamic_ncols=True, desc=desc + ): + yield batch + + def vocabulary_size(self): + return self.nb_codes + + def produce_results(self, n_epoch, model): + with torch.autograd.no_grad(): + t = model.training + model.eval() + + def compute_nb_correct(input): + result = input.clone() + stack.remove_popped_values(result, self.nb_stacks, self.nb_digits) + ar_mask = (result != input).long() + masked_inplace_autoregression( + model, self.batch_size, result, ar_mask, device=self.device + ) + + errors = ((result != input).long() * ar_mask).reshape( + -1, 1 + self.nb_digits + ) + ar_mask = ar_mask.reshape(-1, 1 + self.nb_digits) + + nb_total = ar_mask.max(1).values.sum() + nb_correct = nb_total - errors.max(1).values.sum() + + return nb_total, nb_correct + + test_nb_total, test_nb_correct = compute_nb_correct(self.test_input[:1000]) + + log_string( + f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%" + ) + + ############################################################## + # Log a few generated sequences + input = self.test_input[:10, : 12 * (1 + self.nb_digits)] + result = input.clone() + stack.remove_popped_values(result, self.nb_stacks, self.nb_digits) + ar_mask = (result != input).long() + for n in range(result.size(0)): + log_string( + f"test_before {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}" + ) + masked_inplace_autoregression( + model, self.batch_size, result, ar_mask, device=self.device + ) + for n in range(result.size(0)): + log_string( + f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}" + ) + ############################################################## + + model.train(t) + + +###################################################################### + + def picoclvr_pruner_horizontal_green(p): return not ("green" in p and ("left" in p or "right" in p)) @@ -636,6 +1026,12 @@ if args.task == "picoclvr": pruner_eval=picoclvr_pruner_eval, ) +elif args.task == "mnist": + task = TaskMNIST( + batch_size=args.batch_size, + device=device, + ) + elif args.task == "maze": task = TaskMaze( nb_train_samples=args.nb_train_samples, @@ -647,6 +1043,31 @@ elif args.task == "maze": device=device, ) +elif args.task == "snake": + task = TaskSnake( + nb_train_samples=args.nb_train_samples, + nb_test_samples=args.nb_test_samples, + batch_size=args.batch_size, + height=args.snake_height, + width=args.snake_width, + nb_colors=args.snake_nb_colors, + length=args.snake_length, + prompt_length=args.snake_length // 2, + device=device, + ) + +elif args.task == "stack": + task = TaskStack( + nb_train_samples=args.nb_train_samples, + nb_test_samples=args.nb_test_samples, + batch_size=args.batch_size, + nb_steps=args.stack_nb_steps, + nb_stacks=args.stack_nb_stacks, + nb_digits=args.stack_nb_digits, + fraction_values_for_train=args.stack_fraction_values_for_train, + device=device, + ) + else: raise ValueError(f"Unknown task {args.task}") @@ -782,9 +1203,6 @@ for n_epoch in range(nb_epochs_finished, nb_epochs): for input in task.batches(split="test"): input = input.to(device) - # input, loss_masks, true_images = task.excise_last_image(input) - # input, loss_masks = task.add_true_image(input, true_images, loss_masks) - output = model(mygpt.BracketedSequence(input)).x loss = F.cross_entropy(output.transpose(1, 2), input) acc_test_loss += loss.item() * input.size(0)