X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=38dccb9f8eea57437ce7574d4f87208ab0077b38;hb=87214829798bca7e3eb853df4a27bcb918bb9f67;hp=45bddb762721c468d8990de9cc88044c5cbbd635;hpb=f23843d33a4fa5a38f5034deab8f473793732ee3;p=picoclvr.git diff --git a/main.py b/main.py index 45bddb7..38dccb9 100755 --- a/main.py +++ b/main.py @@ -32,7 +32,7 @@ parser = argparse.ArgumentParser( ) parser.add_argument( - "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake" + "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake, stack" ) parser.add_argument("--log_filename", type=str, default="train.log", help=" ") @@ -45,9 +45,9 @@ parser.add_argument("--nb_epochs", type=int, default=None) 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") @@ -106,6 +106,17 @@ 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() @@ -135,18 +146,32 @@ 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, }, } @@ -191,7 +216,7 @@ def masked_inplace_autoregression( # 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: - tqdm.tqdm( + batches = tqdm.tqdm( batches, dynamic_ncols=True, desc=progress_bar_desc, @@ -841,6 +866,133 @@ class TaskSnake(Task): ###################################################################### +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, + ) + + mask = self.test_input.clone() + stack.remove_popped_values(mask, self.nb_stacks, self.nb_digits) + mask = mask != self.test_input + counts = self.test_stack_counts.flatten()[mask.flatten()] + counts = F.one_hot(counts).sum(0) + log_string(f"stack_count {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 nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%" + ) + + #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! + l = 50 + l = l - l % (1 + self.nb_digits) + input = self.test_input[:10, :l] + 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)) @@ -902,6 +1054,18 @@ elif args.task == "snake": 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}")