X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=tasks.py;h=2a1833d831fe2fdfa0c097f7f4b4e8a9327ef88b;hb=694923fcfa606cf8fc9ee6066ef4bbdea27003ce;hp=0f3aaec3ff480ef8209e262baa61e150d23f4be5;hpb=68aa86a6645dfef3f919aad5732a1a09db77bfae;p=culture.git diff --git a/tasks.py b/tasks.py index 0f3aaec..2a1833d 100755 --- a/tasks.py +++ b/tasks.py @@ -1,12 +1,19 @@ #!/usr/bin/env python -import math, os, tqdm +# Any copyright is dedicated to the Public Domain. +# https://creativecommons.org/publicdomain/zero/1.0/ + +# Written by Francois Fleuret + +import math, os, tqdm, warnings import torch, torchvision from torch import nn from torch.nn import functional as F +from mygpt import BracketedSequence + ###################################################################### @@ -15,11 +22,15 @@ def masked_inplace_autoregression( batch_size, input, ar_mask, + temperature, deterministic_synthesis, forbidden_tokens=None, + logit_biases=None, progress_bar_desc="autoregression", device=torch.device("cpu"), ): + assert input.size() == ar_mask.size() + batches = zip(input.split(batch_size), ar_mask.split(batch_size)) if progress_bar_desc is not None: @@ -27,17 +38,35 @@ def masked_inplace_autoregression( batches, dynamic_ncols=True, desc=progress_bar_desc, - total=input.size(0) // batch_size, + total=(input.size(0) + batch_size - 1) // batch_size, ) - for input, ar_mask in batches: - model.masked_inplace_autoregression( - input, ar_mask, forbidden_tokens, deterministic_synthesis - ) + sum_logits = 0 + + with torch.autograd.no_grad(): + t = model.training + model.eval() + + for input, ar_mask in batches: + sum_logits += model.masked_inplace_autoregression( + input=input, + ar_mask=ar_mask, + temperature=temperature, + deterministic_synthesis=deterministic_synthesis, + forbidden_tokens=forbidden_tokens, + forced_biases=logit_biases, + ) + + model.train(t) + + return sum_logits + + +###################################################################### class Task: - def batches(self, split="train"): + def batches(self, split="train", nb_to_use=-1, desc=None): pass def vocabulary_size(self): @@ -51,411 +80,83 @@ class Task: ###################################################################### -import picoclvr - - -class PicoCLVR(Task): - # Make a tensor from a list of strings - def tensorize(self, descr): - token_descr = [s.strip().split(" ") for s in descr] - l = max([len(s) for s in token_descr]) - token_descr = [s + [""] * (l - len(s)) for s in token_descr] - id_descr = [[self.token2id[u] for u in s] for s in token_descr] - return torch.tensor(id_descr, device=self.device) - - # Make a list of strings from a tensor - def detensorize(self, x): - return [" ".join([self.id2token[t.item()] for t in r]) for r in x] - - # trim all the tensors in the tuple z to remove as much token from - # left and right in the first tensor. If z is a tuple, all its - # elements are trimed according to the triming for the first - def trim(self, z, token=""): - n = self.token2id[token] - if type(z) == tuple: - x = z[0] - i = (1 - (F.pad(x, (1, 1), value=n) == n).min(0).values.long()).cumsum(0) - a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min() - return tuple([t[:, a:b] for t in z]) - else: - i = (1 - (F.pad(z, (1, 1), value=n) == n).min(0).values.long()).cumsum(0) - a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min() - return z[:, a:b] - - ###################### - # Not the cleanest part of the code - - # Extract the last image of each sequence, from the last - # included, and set to all the tokens from the beginning of - # that image to the end - def excise_last_image(self, input): - t_img, t_nul = self.token2id[""], self.token2id[""] - nb_img_tokens = self.height * self.width + 1 - - input = input.clone() - t = (input == t_img).long() - tail_masks = (t.cumsum(dim=1) == t.sum(dim=1, keepdim=True)).long() - i = (t * tail_masks).nonzero(as_tuple=True) - j = ( - i[0][:, None], - i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :], - ) - images = self.trim(input[j]) - input[j] = t_nul - loss_masks = 1 - tail_masks - input, loss_masks = self.trim((input, loss_masks)) - return input, loss_masks, images - - def add_true_image(self, input, images, loss_masks): - t_nul = self.token2id[""] - nb_img_tokens = self.height * self.width + 1 - input = F.pad(input, (0, nb_img_tokens), value=t_nul) - loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0) - t = (input == t_nul).long() - i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True) - j = ( - i[0][:, None], - i[1][:, None] + torch.arange(nb_img_tokens, device=input.device)[None, :], - ) - input[j] = images - loss_masks[j] = 1 - input, loss_masks = self.trim((input, loss_masks)) - return input, loss_masks - - def add_generated_image(self, input, loss_masks, model, deterministic_synthesis): - t_img, t_nul = self.token2id[""], self.token2id[""] - nb_img_tokens = self.height * self.width + 1 - - input = F.pad(input, (0, nb_img_tokens), value=t_nul) - loss_masks = F.pad(loss_masks, (0, nb_img_tokens), value=0) - t = (input == t_nul).long() - i = (t.cumsum(dim=1) == 1).nonzero(as_tuple=True) - input[i] = t_img - - j = ( - i[0][:, None], - i[1][:, None] - + 1 - + torch.arange(nb_img_tokens - 1, device=input.device)[None, :], - ) - ar_masks = input.new_zeros(input.size(), dtype=torch.int64) - ar_masks[j] = 1 - forbidden_tokens = ( - torch.arange(self.vocabulary_size(), device=input.device) == t_nul - ) - with torch.autograd.no_grad(): - t = model.training - model.eval() - masked_inplace_autoregression( - model, - self.batch_size, - input, - ar_masks, - deterministic_synthesis, - forbidden_tokens, - progress_bar_desc=None, - device=self.device, - ) - model.train(t) +import world - input, loss_masks = self.trim((input, loss_masks)) - return input, loss_masks +class World(Task): + def save_image(self, input, result_dir, filename, logger): + img = world.seq2img(input.to("cpu"), self.height, self.width) + image_name = os.path.join(result_dir, filename) + torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4) + logger(f"wrote {image_name}") - ###################### + def make_ar_mask(self, input): + b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2 + return b.long()[None, :].expand_as(input) def __init__( self, nb_train_samples, nb_test_samples, batch_size, - height, - width, - nb_colors=5, + result_dir=None, logger=None, device=torch.device("cpu"), - pruner_train=None, - pruner_eval=None, ): - def generate_descr(nb, cache_suffix, pruner): - return picoclvr.generate( - nb, - height=self.height, - width=self.width, - nb_colors=nb_colors, - pruner=pruner, - ) + super().__init__() - self.height = height - self.width = width self.batch_size = batch_size self.device = device - self.pruner_train = pruner_train - self.pruner_eval = pruner_eval - - param = { - "nb_train_samples": nb_train_samples, - "nb_test_samples": nb_test_samples, - "height": height, - "width": width, - "nb_colors": nb_colors, - "batch_size": batch_size, - "rng_state": list(torch.get_rng_state()), - } - - if logger is not None: - logger( - f"generating {nb_train_samples+nb_test_samples} samples (can take some time)" - ) - - self.train_descr = generate_descr( - nb_train_samples, "train", pruner=self.pruner_train - ) - self.test_descr = generate_descr(nb_test_samples, "test", pruner=None) - - # Build the tokenizer - tokens = {"", ""} - for d in [self.train_descr, self.test_descr]: - for s in d: - for t in s.strip().split(" "): - tokens.add(t) - # make this set a sorted list to get the same tensors given - # the same descr - tokens = list(tokens) - tokens.sort() - self.token2id = dict([(t, n) for n, t in enumerate(tokens)]) - self.id2token = dict([(n, t) for n, t in enumerate(tokens)]) - - # Tokenize the train and test sets - self.train_input = self.tensorize(self.train_descr) - self.test_input = self.tensorize(self.test_descr) - - def batches(self, split="train"): - assert split in {"train", "test"} - input = self.train_input if split == "train" else self.test_input - for batch in tqdm.tqdm( - input.split(self.batch_size), dynamic_ncols=True, desc=f"epoch-{split}" - ): - yield self.trim(batch) - - def vocabulary_size(self): - return len(self.token2id) - - def compute_missing_properties( - self, n_epoch, model, logger, deterministic_synthesis, pruner=None - ): - acc_nb_requested_properties = [] - acc_nb_missing_properties = [] - acc_nb_results = 0 - - for input in tqdm.tqdm( - self.test_input.split(self.batch_size), - dynamic_ncols=True, - desc=f"test-properties", - ): - tape, loss_masks, _ = self.excise_last_image(input) - tape, loss_masks = self.add_generated_image( - tape, loss_masks, model, deterministic_synthesis - ) - result_descr = self.detensorize(tape) - np = picoclvr.nb_properties( - result_descr, - height=self.height, - width=self.width, - pruner=pruner, - ) - nb_requested_properties, _, nb_missing_properties = zip(*np) - acc_nb_requested_properties += nb_requested_properties - acc_nb_missing_properties += nb_missing_properties - acc_nb_results += len(result_descr) - - nb_requested_properties = sum(acc_nb_requested_properties) - nb_missing_properties = sum(acc_nb_missing_properties) - - prefix = "" if pruner is None else "pruned_" - logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}") - logger( - f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}" - ) - logger( - f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%" - ) - - ###################################################################### - - def produce_results( - self, n_epoch, model, result_dir, logger, deterministic_synthesis - ): - self.compute_missing_properties(n_epoch, model, logger, deterministic_synthesis) - - if self.pruner_eval is not None: - self.compute_missing_properties(n_epoch, model, self.pruner_eval) - - nb_tokens_to_generate = self.height * self.width + 3 - result_descr = [] - nb_per_primer = 8 - primer = [] - - for primer_descr in [ - "red above green green top blue right of red", - "there is red there is yellow there is blue", - "red below yellow yellow below green green below blue red right yellow left green right blue left", - "green bottom yellow bottom green left of blue yellow right of blue blue top", - ]: - primer += [primer_descr] * nb_per_primer - - tape = self.tensorize(primer) - loss_masks = 1 - (tape == self.token2id[""]).long() - tape, loss_masks = self.add_generated_image( - tape, loss_masks, model, deterministic_synthesis - ) - result_descr = self.detensorize(tape) - - np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width) - - acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np) - acc_nb_results = len(result_descr) - - nb_requested_properties = sum(acc_nb_requested_properties) - nb_missing_properties = sum(acc_nb_missing_properties) - - prefix = "demo_" - logger(f"nb_{prefix}samples {n_epoch} {acc_nb_results}") - logger( - f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}" - ) - logger( - f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%" - ) - - img = picoclvr.descr2img(result_descr, height=self.height, width=self.width) + self.height = 6 + self.width = 8 - if img.dim() == 5: - if img.size(1) == 1: - img = F.pad(img.squeeze(1), pad=(1, 1, 1, 1), value=64) - else: - img = torch.cat( - [ - torchvision.utils.make_grid(x, padding=1, pad_value=64)[None] - for x in img - ], - 0, - ) - - image_name = os.path.join(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=0.0 - ) - logger(f"wrote {image_name}") + self.train_input = world.generate_seq( + nb_train_samples, height=self.height, width=self.width + ).to(device) + self.test_input = world.generate_seq( + nb_test_samples, height=self.height, width=self.width + ).to(device) -###################################################################### + self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1 + self.train_quizzes = [] + self.test_quizzes = [] -class MNIST(Task): - def __init__( - self, nb_train_samples, nb_test_samples, batch_size, device=torch.device("cpu") - ): - self.nb_train_samples = (nb_train_samples,) - self.nb_test_samples = (nb_test_samples,) - self.batch_size = batch_size - self.device = device - data_set = torchvision.datasets.MNIST(root="./data", train=True, download=True) - self.train_input = data_set.data[:nb_train_samples].view(-1, 28 * 28).long() - data_set = torchvision.datasets.MNIST(root="./data", train=False, download=True) - self.test_input = data_set.data[:nb_test_samples].view(-1, 28 * 28).long() + if result_dir is not None: + self.save_image( + self.train_input[:72], result_dir, f"world_train.png", logger + ) - def batches(self, split="train", nb_to_use=-1, desc=None): + def batches(self, split="train", 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 256 - - def produce_results( - self, n_epoch, model, result_dir, logger, deterministic_synthesis - ): - 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, - deterministic_synthesis, - device=self.device, - ) - image_name = os.path.join(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, - ) - logger(f"wrote {image_name}") - - -###################################################################### - -import maze - + if split == "train": + input = self.train_input + quizzes = self.train_quizzes + else: + input = self.test_input + quizzes = self.test_quizzes -class Maze(Task): - def map2seq(self, *m): - return torch.cat([x.flatten(1) for x in m], 1) + if len(quizzes) > 0: + quizzes = torch.cat(quizzes, dim=0) + if quizzes.size(0) > input.size(0) // 2: + i = torch.randperm(input.size(0))[: input.size(0) // 2] + quizzes = quizzes[i] - def seq2map(self, s): - s = s.reshape(s.size(0), -1, self.height, self.width) - return (s[:, k] for k in range(s.size(1))) + i = torch.randperm(input.size(0))[: input.size(0) - quizzes.size(0)] + input = input[i] - def __init__( - self, - nb_train_samples, - nb_test_samples, - batch_size, - height, - width, - nb_walls, - device=torch.device("cpu"), - ): - self.batch_size = batch_size - self.height = height - self.width = width - self.device = device + self.nb_batch_samples_world = input.size(0) + self.nb_batch_samples_quizzes = quizzes.size(0) - train_mazes, train_paths, _ = 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"), - ) - self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device)) - - test_mazes, test_paths, _ = 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"), - ) - self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device)) + input = torch.cat([input, quizzes], dim=0) + else: + self.nb_batch_samples_world = input.size(0) + self.nb_batch_samples_quizzes = 0 - self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1 + # Shuffle + input = input[torch.randperm(input.size(0))] - 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( @@ -466,564 +167,200 @@ class Maze(Task): def vocabulary_size(self): return self.nb_codes - def compute_error( - self, model, split="train", nb_to_use=-1, deterministic_synthesis=False - ): - nb_total, nb_correct = 0, 0 - count = torch.zeros( - self.width * self.height, - self.width * self.height, - device=self.device, - dtype=torch.int64, - ) - - for input in self.batches(split, nb_to_use): - 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, - deterministic_synthesis, - progress_bar_desc=None, - device=self.device, - ) - mazes, paths = self.seq2map(result) - path_correctness = maze.path_correctness(mazes, paths) - nb_correct += path_correctness.long().sum() - nb_total += mazes.size(0) - - 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, result_dir, logger, deterministic_synthesis + self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000 ): - with torch.autograd.no_grad(): - t = model.training - model.eval() - - train_nb_total, train_nb_correct, count = self.compute_error( - model, - "train", - nb_to_use=1000, - deterministic_synthesis=deterministic_synthesis, - ) - logger( - 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}%" - ) + def compute_accuracy(input, logger=None): + input = input[:nmax] + ar_mask = self.make_ar_mask(input) + result = input.clone() * (1 - ar_mask) - test_nb_total, test_nb_correct, count = self.compute_error( - model, - "test", - nb_to_use=1000, - deterministic_synthesis=deterministic_synthesis, - ) - logger( - 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() - logger(f"proportion_optimal_test {proportion_optimal*100:.02f}%") - with open( - os.path.join(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()) - ar_mask[:, self.height * self.width :] = 1 - result *= 1 - ar_mask masked_inplace_autoregression( - model, - self.batch_size, - result, - ar_mask, - deterministic_synthesis, + model=model, + batch_size=self.batch_size, + input=result, + ar_mask=ar_mask, + temperature=1.0, + deterministic_synthesis=deterministic_synthesis, + progress_bar_desc=None, device=self.device, ) - mazes, paths = self.seq2map(input) - _, predicted_paths = self.seq2map(result) - - filename = os.path.join(result_dir, f"maze_result_{n_epoch:04d}.png") - maze.save_image( - 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), + nb_total, nb_correct = ( + input.size(0), + (input == result).long().min(dim=1).values.sum(), ) - logger(f"wrote {filename}") - model.train(t) + return nb_total, nb_correct + train_nb_total, train_nb_correct = compute_accuracy(self.train_input) -###################################################################### - - -import snake - - -class Snake(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, + logger( + 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}%" ) - 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 + test_nb_total, test_nb_correct = compute_accuracy(self.test_input, logger) - def vocabulary_size(self): - return self.nb_codes + logger( + 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}%" + ) - def produce_results( - self, n_epoch, model, result_dir, logger, deterministic_synthesis - ): - 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, - deterministic_synthesis, - 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 - # ) - - # logger( - # 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] - ) + main_test_accuracy = test_nb_correct / test_nb_total + logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}") - logger( - 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) + input = self.test_input[:96] + ar_mask = self.make_ar_mask(input) + result = input.clone() * (1 - ar_mask) + masked_inplace_autoregression( + model=model, + batch_size=self.batch_size, + input=result, + ar_mask=ar_mask, + temperature=1.0, + deterministic_synthesis=deterministic_synthesis, + progress_bar_desc=None, + device=self.device, + ) -###################################################################### + self.save_image( + result[:72], + result_dir, + f"world_prediction_{n_epoch:04d}_{model.id:02d}.png", + logger, + ) + return main_test_accuracy -import stack + def renew_samples(self, nb, for_train=True): + input = self.train_input if for_train else self.test_input + nb = min(nb, input.size(0)) + input[:-nb] = input[nb:].clone() + input[-nb:] = world.generate_seq(nb, height=self.height, width=self.width).to( + self.device + ) + def store_new_quizzes(self, new_quizzes, for_train=True): + if for_train: + self.train_quizzes.append(new_quizzes) + else: + self.test_quizzes.append(new_quizzes) -class Stack(Task): - def __init__( + def create_new_quizzes( self, - nb_train_samples, - nb_test_samples, - batch_size, + n_epoch, + result_dir, logger, - nb_steps, - nb_stacks, - nb_digits, - fraction_values_for_train=None, - device=torch.device("cpu"), + nb, + model, + other_models, + desired_average_logits=None, ): - 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, - ) + ############################################################### + # Generate quizzes with model - self.test_input, self.test_stack_counts = stack.generate_sequences( - nb_test_samples, - nb_steps, - nb_stacks, - nb_digits, - values_for_test, - self.device, + quizzes = torch.empty( + nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64 ) - 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) - logger(f"test_pop_stack_counts {counts}") + ar_mask = torch.full(quizzes.size(), 1, device=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 + temperature = 1 + d_temperature = 1 - def vocabulary_size(self): - return self.nb_codes - - def produce_results( - self, n_epoch, model, result_dir, logger, deterministic_synthesis - ): - 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, - deterministic_synthesis, - 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]) - - logger( - 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}%" + while True: + sum_logits = masked_inplace_autoregression( + model=model, + batch_size=self.batch_size, + input=quizzes, + ar_mask=ar_mask, + temperature=temperature, + deterministic_synthesis=False, + progress_bar_desc="creating quizzes", + device=self.device, ) - ############################################################## - # 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)): - logger( - 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, - deterministic_synthesis, - device=self.device, - ) - for n in range(result.size(0)): - logger( - f"test_after {stack.seq_to_str(result[n],nb_stacks=self.nb_stacks,nb_digits=self.nb_digits)}" - ) - ############################################################## - - model.train(t) + average_logits = sum_logits / quizzes.size(0) + logger(f"{average_logits=} {desired_average_logits=}") -###################################################################### + if desired_average_logits is None: + break + # Oh man that's ugly + if average_logits > desired_average_logits: + if d_temperature < 0: + d_temperature *= -0.5 + temperature += d_temperature + else: + if d_temperature > 0: + d_temperature *= -0.5 + temperature += d_temperature + + logger(f"chaging temperature to {temperature}") + + ############################################################### + # Create the reverse quizzes + + l = self.height * self.width + direction = quizzes[:, l : l + 1] + direction = world.token_forward * ( + direction == world.token_backward + ) + world.token_backward * (direction == world.token_forward) + reverse_quizzes = torch.cat( + [quizzes[:, l + 1 :], direction, quizzes[:, :l]], dim=1 + ) -import expr + ar_mask = self.make_ar_mask(quizzes) + ############################################################### + # Check how many of the other models can solve them in both + # directions -class Expr(Task): - def __init__( - self, - nb_train_samples, - nb_test_samples, - nb_variables, - sequence_length, - batch_size, - device=torch.device("cpu"), - ): - self.batch_size = batch_size - self.device = device - - train_sequences = expr.generate_sequences( - nb_train_samples, - nb_variables=nb_variables, - length=sequence_length, - # length=2 * sequence_length, - # randomize_length=True, - ) - test_sequences = expr.generate_sequences( - nb_test_samples, - nb_variables=nb_variables, - length=sequence_length, - ) - self.char2id = dict( - [ - (c, n) - for n, c in enumerate( - set("#" + "".join(train_sequences + test_sequences)) - ) - ] - ) - self.id2char = dict([(n, c) for c, n in self.char2id.items()]) - - self.filler, self.space = self.char2id["#"], self.char2id[" "] - - len_max = max([len(x) for x in train_sequences]) - self.train_input = torch.cat( - [ - torch.tensor( - [ - [self.char2id[c] for c in s + "#" * (len_max - len(s))] - for s in train_sequences - ] - ) - ], - 0, - ).to(device) + nb_correct = [] - len_max = max([len(x) for x in test_sequences]) - self.test_input = torch.cat( - [ - torch.tensor( - [ - [self.char2id[c] for c in s + "#" * (len_max - len(s))] - for s in test_sequences - ] - ) - ], - 0, - ).to(device) + for m in other_models: + result = quizzes.clone() - 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 - ): - if split == "train": - last = (batch != self.filler).max(0).values.nonzero().max() + 3 - batch = batch[:, :last] - yield batch + masked_inplace_autoregression( + model=m, + batch_size=self.batch_size, + input=result, + ar_mask=ar_mask, + temperature=1.0, + deterministic_synthesis=True, + progress_bar_desc="solving quizzes", + device=self.device, + ) - def vocabulary_size(self): - return self.nb_codes + correct = (quizzes == result).long().min(dim=-1).values - def seq2str(self, s): - return "".join([self.id2char[k.item()] for k in s]) + reverse_result = reverse_quizzes.clone() - def produce_results( - self, n_epoch, model, result_dir, logger, deterministic_synthesis - ): - with torch.autograd.no_grad(): - t = model.training - model.eval() - - def compute_nb_correct(input): - result = input.clone() - ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1) - result = (1 - ar_mask) * result + ar_mask * self.filler - masked_inplace_autoregression( - model, - self.batch_size, - result, - ar_mask, - deterministic_synthesis, - device=self.device, - ) - - nb_total = input.size(0) - nb_correct = (input == result).long().min(1).values.sum() - - ####################################################################### - # Comput predicted vs. true variable values - - nb_delta = torch.zeros(5, dtype=torch.int64) - nb_missed = 0 - - values_input = expr.extract_results([self.seq2str(s) for s in input]) - values_result = expr.extract_results([self.seq2str(s) for s in result]) - - for i, r in zip(values_input, values_result): - for n, vi in i.items(): - vr = r.get(n) - if vr is None or vr < 0: - nb_missed += 1 - else: - d = abs(vr - vi) - if d >= nb_delta.size(0): - nb_missed += 1 - else: - nb_delta[d] += 1 - - ###################################################################### - - return nb_total, nb_correct, nb_delta, nb_missed - - ( - test_nb_total, - test_nb_correct, - test_nb_delta, - test_nb_missed, - ) = compute_nb_correct(self.test_input[:1000]) - - logger( - 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}%" + masked_inplace_autoregression( + model=m, + batch_size=self.batch_size, + input=reverse_result, + ar_mask=ar_mask, + temperature=1.0, + deterministic_synthesis=True, + progress_bar_desc="solving reversed quizzes", + device=self.device, ) - nb_total = test_nb_delta.sum() + test_nb_missed - for d in range(test_nb_delta.size(0)): - logger( - f"error_value {n_epoch} delta {d} {test_nb_delta[d]} {test_nb_delta[d]*100/nb_total:.02f}%" - ) - logger( - f"error_value {n_epoch} missed {test_nb_missed} {test_nb_missed*100/nb_total:.02f}%" + reverse_correct = ( + (reverse_quizzes == reverse_result).long().min(dim=-1).values ) - ############################################################## - # Log a few generated sequences - input = self.test_input[:10] - result = input.clone() - ar_mask = (result == self.space).long().cumsum(dim=1).clamp(max=1) - result = (1 - ar_mask) * result + ar_mask * self.filler - for n in range(result.size(0)): - logger(f"test_before {self.seq2str(result[n])}") - masked_inplace_autoregression( - model, - self.batch_size, - result, - ar_mask, - deterministic_synthesis, - device=self.device, - ) - correct = (1 - ar_mask) * self.space + ar_mask * input - for n in range(result.size(0)): - comment = "GOOD" if (result[n] - input[n]).abs().max() == 0 else "" - logger(f"test_after {self.seq2str(result[n])} {comment}") - logger(f"correct {self.seq2str(correct[n])}") - ############################################################## - - model.train(t) + nb_correct.append((correct * reverse_correct)[None, :]) + nb_correct = torch.cat(nb_correct, dim=0) -###################################################################### + # filename = os.path.join(result_dir, "correct_{n_epoch:04d}.dat") + # with open(filename, "w") as f: + # for k in nb_correct: + # f.write(f"{k}\n") + + return quizzes, nb_correct.sum(dim=0), sum_logits