X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=6b00bbfd991178841468b75006f94121668c2b4f;hb=HEAD;hp=4770a125172cf77a41aada959ef5f18b44517be4;hpb=757876d57637e0da35f3680ec6ac9573b91f902a;p=culture.git diff --git a/main.py b/main.py index 4770a12..5dceefc 100755 --- a/main.py +++ b/main.py @@ -5,130 +5,179 @@ # Written by Francois Fleuret -# torch.backends.cuda.matmul.allow_tf23 -# torch.autocast(torch.bfloat16) - -import math, sys, argparse, time, tqdm, os +import math, sys, argparse, time, tqdm, os, datetime, warnings, copy import torch, torchvision from torch import nn from torch.nn import functional as F -import mygpt, tensorstack +import ffutils, grids, attae -###################################################################### +import threading, subprocess -if torch.cuda.is_available(): - device = torch.device("cuda") - torch.backends.cuda.matmul.allow_tf32 = True -else: - device = torch.device("cpu") +# import torch.multiprocessing as mp + +torch.set_float32_matmul_precision("high") + +# torch.set_default_dtype(torch.bfloat16) ###################################################################### parser = argparse.ArgumentParser( - description="An implementation of GPT with cache.", formatter_class=argparse.ArgumentDefaultsHelpFormatter, ) -parser.add_argument( - "--task", type=str, default="picoclvr", help="picoclvr, mnist, maze, snake, stack" -) - -parser.add_argument("--log_filename", type=str, default="train.log", help=" ") +parser.add_argument("--log_filename", type=str, default="train.log") 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=None) +parser.add_argument("--resume", action="store_true", default=False) -parser.add_argument("--batch_size", type=int, default=None) +# ---------------------------------- -parser.add_argument("--nb_train_samples", type=int, default=None) +parser.add_argument("--nb_epochs", type=int, default=10000) -parser.add_argument("--nb_test_samples", type=int, default=None) +parser.add_argument("--batch_size", type=int, default=25) -parser.add_argument("--optim", type=str, default="adam") +parser.add_argument("--train_batch_size", type=int, default=None) -parser.add_argument("--learning_rate", type=float, default=1e-4) +parser.add_argument("--eval_batch_size", type=int, default=25) -parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6") +parser.add_argument("--nb_train_samples", type=int, default=50000) -parser.add_argument("--dim_model", type=int, default=512) +parser.add_argument("--nb_test_samples", type=int, default=2500) -parser.add_argument("--dim_keys", type=int, default=64) +parser.add_argument("--nb_c_quizzes", type=int, default=5000) -parser.add_argument("--dim_hidden", type=int, default=2048) +parser.add_argument("--c_quiz_multiplier", type=int, default=1) -parser.add_argument("--nb_heads", type=int, default=8) +parser.add_argument("--learning_rate", type=float, default=5e-4) -parser.add_argument("--nb_blocks", type=int, default=12) +parser.add_argument("--nb_have_to_be_correct", type=int, default=3) -parser.add_argument("--dropout", type=float, default=0.1) +parser.add_argument("--nb_have_to_be_wrong", type=int, default=1) -parser.add_argument("--deterministic_synthesis", action="store_true", default=False) +parser.add_argument("--nb_mistakes_to_be_wrong", type=int, default=5) -parser.add_argument("--no_checkpoint", action="store_true", default=False) +# ---------------------------------- -parser.add_argument("--overwrite_results", action="store_true", default=False) +parser.add_argument("--model_type", type=str, default="standard") -parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth") +parser.add_argument("--model", type=str, default="37M") -############################## -# picoclvr options +parser.add_argument("--dim_model", type=int, default=None) -parser.add_argument("--picoclvr_nb_colors", type=int, default=5) +parser.add_argument("--dim_keys", type=int, default=None) -parser.add_argument("--picoclvr_height", type=int, default=12) +parser.add_argument("--dim_hidden", type=int, default=None) -parser.add_argument("--picoclvr_width", type=int, default=16) +parser.add_argument("--nb_heads", type=int, default=None) -parser.add_argument("--picocvlr_prune_properties", type=str, default="none") +parser.add_argument("--nb_blocks", type=int, default=None) -############################## -# Maze options +parser.add_argument("--dropout", type=float, default=0.5) -parser.add_argument("--maze_height", type=int, default=13) +# ---------------------------------- -parser.add_argument("--maze_width", type=int, default=21) +parser.add_argument("--nb_threads", type=int, default=1) -parser.add_argument("--maze_nb_walls", type=int, default=15) +parser.add_argument("--gpus", type=str, default="all") -############################## -# Snake options +# ---------------------------------- -parser.add_argument("--snake_height", type=int, default=6) +parser.add_argument("--nb_models", type=int, default=5) -parser.add_argument("--snake_width", type=int, default=8) +parser.add_argument("--diffusion_nb_iterations", type=int, default=25) -parser.add_argument("--snake_nb_colors", type=int, default=5) +parser.add_argument("--diffusion_proba_corruption", type=float, default=0.05) -parser.add_argument("--snake_length", type=int, default=200) +parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.95) -############################## -# Snake options +parser.add_argument("--proba_prompt_noise", type=float, default=0.05) -parser.add_argument("--stack_nb_steps", type=int, default=100) +parser.add_argument("--proba_hint", type=float, default=0.25) -parser.add_argument("--stack_nb_stacks", type=int, default=1) +parser.add_argument("--quizzes", type=str, default=None) -parser.add_argument("--stack_nb_digits", type=int, default=3) +###################################################################### -parser.add_argument("--stack_fraction_values_for_train", type=float, default=None) +grids_tasks = ", ".join( + [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks] +) + +parser.add_argument( + "--grids_world_tasks", + type=str, + default="replace_color,translate,grow,frame", + help="A comma-separated subset of: " + grids_tasks + ".", +) ###################################################################### 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_culture" + +###################################################################### -if args.result_dir is None: args.result_dir=f"results_{args.task}" +default_model_args = { + "17K": { + "dim_model": 32, + "dim_keys": 32, + "dim_hidden": 32, + "nb_heads": 2, + "nb_blocks": 2, + }, + "4M": { + "dim_model": 256, + "dim_keys": 32, + "dim_hidden": 1024, + "nb_heads": 4, + "nb_blocks": 6, + }, + "37M": { + "dim_model": 512, + "dim_keys": 64, + "dim_hidden": 2048, + "nb_heads": 8, + "nb_blocks": 12, + }, + "122M": { + "dim_model": 768, + "dim_keys": 64, + "dim_hidden": 2048, + "nb_heads": 8, + "nb_blocks": 24, + }, + "352M": { + "dim_model": 1024, + "dim_keys": 64, + "dim_hidden": 2048, + "nb_heads": 8, + "nb_blocks": 48, + }, +} -try: - os.mkdir(args.result_dir) -except FileExistsError: - if not args.overwrite_results: +if args.model in default_model_args: + for k, v in default_model_args[args.model].items(): + if getattr(args, k) is None: + setattr(args, k, v) +else: + raise ValueError(f"Unknown model {args.model}") + +###################################################################### + +if args.resume: + if not os.path.isdir(args.result_dir): + print(f"Trying to resume from a non-existing result dir {args.result_dir}.") + exit(1) +else: + try: + os.mkdir(args.result_dir) + except FileExistsError: print(f"result directory {args.result_dir} already exists") exit(1) @@ -144,48 +193,11 @@ if args.seed >= 0: ###################################################################### -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) - -###################################################################### - def log_string(s): + """print the given string prefixed with a time stamps, and log it + into log_file is not None""" + t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime()) if log_file is not None: @@ -196,1038 +208,793 @@ def log_string(s): sys.stdout.flush() +###################################################################### +# Create a time-stamped archive of the source code + +with open("this_run.sh", "w") as f: + f.write(f"{' '.join(sys.argv)}\n") + +now = time.strftime("%Y%m%d-%H%M%S", time.localtime()) + +os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py *.sh") + +###################################################################### + +log_string(f"argv {' '.join(sys.argv)}") + for n in vars(args): log_string(f"args.{n} {getattr(args, n)}") + ###################################################################### +if args.gpus == "all": + gpus_idx = range(torch.cuda.device_count()) +else: + gpus_idx = [int(k) for k in args.gpus.split(",")] -# ra_mask is boolean, with 1s on the values to generate +gpus = [torch.device(f"cuda:{n}") for n in gpus_idx] +if torch.cuda.is_available(): + main_device = gpus[0] +else: + assert len(gpus) == 0 + main_device = torch.device("cpu") -def masked_inplace_autoregression( - model, - batch_size, - input, - ar_mask, - forbidden_tokens=None, - progress_bar_desc="autoregression", - device=torch.device("cpu"), -): - # 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( - mygpt.BracketedSequence(input, 0, i.min()) - ) # Needed to initialize the model's cache - for s in range(i.min(), i.max() + 1): - output = model(mygpt.BracketedSequence(input, s, 1)).x - logits = output[:, s] - if forbidden_tokens is not None: - logits = logits.masked_fill(forbidden_tokens, float("-inf")) - if args.deterministic_synthesis: - t_next = logits.argmax(1) - else: - dist = torch.distributions.categorical.Categorical(logits=logits) - t_next = dist.sample() - input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s] +if args.train_batch_size is None: + args.train_batch_size = args.batch_size +else: + assert args.batch_size % args.train_batch_size == 0 + +assert args.nb_train_samples % args.batch_size == 0 +assert args.nb_test_samples % args.batch_size == 0 + +###################################################################### + + +def optimizer_to(optim, device): + """Move the optimizer optim to the device""" + for param in optim.state.values(): + # Not sure there are any global tensors in the state dict + if isinstance(param, torch.Tensor): + param.data = param.data.to(device) + if param._grad is not None: + param._grad.data = param._grad.data.to(device) + elif isinstance(param, dict): + for subparam in param.values(): + if isinstance(subparam, torch.Tensor): + subparam.data = subparam.data.to(device) + if subparam._grad is not None: + subparam._grad.data = subparam._grad.data.to(device) ###################################################################### -class Task: - def batches(self, split="train"): - pass +def generate_quiz_set(nb_samples, c_quizzes, c_quiz_multiplier=1): + if c_quizzes is None: + quizzes = problem.generate_w_quizzes(nb_samples) + nb_w_quizzes = quizzes.size(0) + nb_c_quizzes = 0 + else: + if c_quiz_multiplier > 1: + n = min(c_quiz_multiplier, (nb_samples // 2) // c_quizzes.size(0)) + body = c_quizzes.repeat(n, 1) + if n < c_quiz_multiplier: + tail = c_quizzes[ + torch.randperm(c_quizzes.size(0))[: nb_samples // 2 - body.size(0)] + ] + c_quizzes = torch.cat([body, tail], dim=0) + else: + c_quizzes = body - def vocabulary_size(self): - pass + if c_quizzes.size(0) > nb_samples // 2: + i = torch.randperm(c_quizzes.size(0))[: nb_samples // 2] + c_quizzes = c_quizzes[i] - def produce_results(self, n_epoch, model): - pass + w_quizzes = problem.generate_w_quizzes(nb_samples - c_quizzes.size(0)) + + quizzes = torch.cat([w_quizzes, c_quizzes], dim=0) + nb_w_quizzes = w_quizzes.size(0) + nb_c_quizzes = c_quizzes.size(0) + + i = torch.randperm(quizzes.size(0), device=quizzes.device) + quizzes = quizzes[i].contiguous() + + log_string(f"quiz_set nb_w_quizzes {nb_w_quizzes} nb_c_quizzes {nb_c_quizzes}") + + return quizzes ###################################################################### -import picoclvr - - -class TaskPicoCLVR(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): - 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, - forbidden_tokens, - progress_bar_desc=None, - device=self.device, - ) - model.train(t) - input, loss_masks = self.trim((input, loss_masks)) +def add_hints_imt(imt_set): + """Set every component of the mask to zero with probability + args.proba_hint, and for each component set to zero, copy the + corresponding value from the target into the input + + """ + input, masks, targets = imt_set.unbind(dim=1) + # h = torch.rand(masks.size(), device=masks.device) - masks + # t = h.sort(dim=1).values[:, args.nb_hints, None] + # mask_hints = (h < t).long() + mask_hints = ( + torch.rand(input.size(), device=input.device) < args.proba_hint + ).long() * masks + masks = (1 - mask_hints) * masks + input = (1 - mask_hints) * input + mask_hints * targets + return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1) + + +def add_noise_imt(imt_set): + """Replace every component of the input by a random value with + probability args.proba_prompt_noise.""" + input, masks, targets = imt_set.unbind(dim=1) + noise = problem.pure_noise(input.size(0), input.device) + change = (1 - masks) * ( + torch.rand(input.size(), device=input.device) < args.proba_prompt_noise + ).long() + input = (1 - change) * input + change * noise + return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1) - return input, loss_masks - ###################### +###################################################################### +# Prediction - def __init__( - self, - nb_train_samples, - nb_test_samples, - batch_size, - height, - width, - nb_colors=5, - 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, - ) - 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()), - } - - log_string( - 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, 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) - 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_" - log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}") - log_string( - f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}" - ) - log_string( - f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%" - ) +def samples_for_prediction_imt(input): + nb = input.size(0) + masks = input.new_zeros(input.size()) + u = F.one_hot(torch.randint(4, (nb,), device=masks.device), num_classes=4) + masks.view(nb, 4, -1)[...] = u[:, :, None] + targets = input + input = (1 - masks) * targets - ###################################################################### + return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1) - def produce_results(self, n_epoch, model): - self.compute_missing_properties(n_epoch, model) - if self.pruner_eval is not None: - self.compute_missing_properties(n_epoch, model, self.pruner_eval) +def ae_predict(model, imt_set, local_device=main_device): + model.eval().to(local_device) - nb_tokens_to_generate = self.height * self.width + 3 - result_descr = [] - nb_per_primer = 8 - primer = [] + record = [] - 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 + src = tqdm.tqdm( + imt_set.split(args.eval_batch_size), + dynamic_ncols=True, + desc="predict", + total=imt_set.size(0) // args.eval_batch_size, + delay=10, + ) - tape = self.tensorize(primer) - loss_masks = 1 - (tape == self.token2id[""]).long() - tape, loss_masks = self.add_generated_image(tape, loss_masks, model) - result_descr = self.detensorize(tape) + for imt in src: + # some paranoia + imt = imt.clone() + imt[:, 0] = imt[:, 0] * (1 - imt[:, 1]) - np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width) + with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = model(imt[:, 0] * 2 + imt[:, 1]) + dist = torch.distributions.categorical.Categorical(logits=logits) + result = (1 - imt[:, 1]) * imt[:, 0] + imt[:, 1] * dist.sample() + record.append(result) - acc_nb_requested_properties, _, acc_nb_missing_properties = zip(*np) - acc_nb_results = len(result_descr) + return torch.cat(record) - nb_requested_properties = sum(acc_nb_requested_properties) - nb_missing_properties = sum(acc_nb_missing_properties) - prefix = "demo_" - log_string(f"nb_{prefix}samples {n_epoch} {acc_nb_results}") - log_string( - f"property_{prefix}nb {n_epoch} requested {sum(acc_nb_requested_properties)} missing {sum(acc_nb_missing_properties)}" - ) - log_string( - f"property_{prefix}miss {n_epoch} {100*nb_missing_properties/nb_requested_properties:.02f}%" - ) +def predict_the_four_grids( + model, input, with_noise=False, with_hints=False, local_device=main_device +): + input = input[:, None, :].expand(-1, 4, -1).reshape(-1, input.size(1)) + nb = input.size(0) + masks = input.new_zeros(input.size()) + u = F.one_hot(torch.arange(nb, device=masks.device) % 4, num_classes=4) + masks.view(nb, 4, -1)[...] = u[:, :, None] + targets = input + input = (1 - masks) * targets + imt_set = torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1) - img = picoclvr.descr2img(result_descr, height=self.height, width=self.width) + if with_hints: + imt_set = add_hints_imt(imt_set) - 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(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=0.0 - ) - log_string(f"wrote {image_name}") + if with_noise: + imt_set = add_noise_imt(imt_set) + + result = ae_predict(model, imt_set, local_device=local_device) + result = (result * masks).reshape(-1, 4, result.size(1)).sum(dim=1) + + return result ###################################################################### -class TaskMNIST(Task): - def __init__(self, batch_size, device=torch.device("cpu")): - self.device = device - self.batch_size = batch_size +def samples_for_generation_imt(input): + nb = input.size(0) + probs_iterations = 0.1 ** torch.linspace( + 0, 1, args.diffusion_nb_iterations, device=input.device + ) + probs_iterations = probs_iterations[None, :] / probs_iterations.sum() + probs_iterations = probs_iterations.expand(nb, -1) + dist = torch.distributions.categorical.Categorical(probs=probs_iterations) + t = dist.sample() + 1 + r = torch.rand(input.size(), device=input.device) + proba_erased = 1 - (1 - args.diffusion_proba_corruption) ** t + mask_erased = (r <= proba_erased[:, None]).long() - 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, + noise = problem.pure_noise(nb, input.device) + targets = input + input = (1 - mask_erased) * input + mask_erased * noise + masks = input.new_full(input.size(), 1) + + return torch.cat([input[:, None], masks[:, None], targets[:, None]], dim=1) + + +def prioritized_rand(low): + x = torch.rand(low.size(), device=low.device).sort(dim=1, descending=True).values + k = torch.rand(low.size(), device=low.device) + low.long() + k = k.sort(dim=1).indices + y = x.new(x.size()) + y.scatter_(dim=1, index=k, src=x) + return y + + +def ae_generate(model, nb, local_device=main_device): + model.eval().to(local_device) + + # We loop through the iterations first and through the + # mini-batches second so that we keep only the samples that have + # not stabilized + + all_input = problem.pure_noise(nb, local_device) + all_masks = all_input.new_full(all_input.size(), 1) + all_changed = torch.full((all_input.size(0),), True, device=all_input.device) + + for it in range(args.diffusion_nb_iterations): + # log_string(f"nb_changed {all_changed.long().sum().item()}") + + if not all_changed.any(): + break + + sub_input = all_input[all_changed].clone() + sub_masks = all_masks[all_changed].clone() + sub_changed = all_changed[all_changed].clone() + + src = zip( + sub_input.split(args.eval_batch_size), + sub_masks.split(args.eval_batch_size), + sub_changed.split(args.eval_batch_size), ) - log_string(f"wrote {image_name}") + + for input, masks, changed in src: + with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = model(input * 2 + masks) + dist = torch.distributions.categorical.Categorical(logits=logits) + output = dist.sample() + r = prioritized_rand(input != output) + mask_changes = (r <= args.diffusion_proba_corruption).long() * masks + update = (1 - mask_changes) * input + mask_changes * output + changed[...] = changed & (update != input).max(dim=1).values + input[...] = update + + a = all_changed.clone() + all_input[a] = sub_input + all_masks[a] = sub_masks + all_changed[a] = sub_changed + + return all_input ###################################################################### -import maze +def one_epoch(model, n_epoch, c_quizzes, train=True, local_device=main_device): + quizzes = generate_quiz_set( + args.nb_train_samples if train else args.nb_test_samples, + c_quizzes, + args.c_quiz_multiplier, + ) + + q_p, q_g = quizzes.to(local_device).chunk(2) + + # Half of the samples train the prediction, and we inject noise in + # all, and hints in half + b_p = samples_for_prediction_imt(q_p) + b_p = add_noise_imt(b_p) + half = torch.rand(b_p.size(0)) < 0.5 + b_p[half] = add_hints_imt(b_p[half]) -class TaskMaze(Task): - def map2seq(self, *m): - return torch.cat([x.flatten(1) for x in m], 1) + # The other half are denoising examples for the generation + b_g = samples_for_generation_imt(q_g) - 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))) + imt_set = torch.cat([b_p, b_g]) + imt_set = imt_set[torch.randperm(imt_set.size(0), device=imt_set.device)] - def __init__( - self, - nb_train_samples, - nb_test_samples, - batch_size, - height, - width, - nb_walls, - device=torch.device("cpu"), + if train: + label = "train" + model.train().to(local_device) + optimizer_to(model.optimizer, local_device) + batch_size = args.train_batch_size + else: + label = "test" + model.eval().to(local_device) + batch_size = args.eval_batch_size + + nb_samples, acc_loss = 0, 0.0 + + for imt in tqdm.tqdm( + imt_set.split(batch_size), + dynamic_ncols=True, + desc=label, + total=quizzes.size(0) // batch_size, + delay=10, ): - self.batch_size = batch_size - self.height = height - self.width = width - self.device = device - - 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)) - - 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 compute_error(self, model, split="train", nb_to_use=-1): - nb_total, nb_correct = 0, 0 - count = torch.zeros( - self.width * self.height, - self.width * self.height, - device=self.device, - dtype=torch.int64, + input, masks, targets = imt.unbind(dim=1) + if train and nb_samples % args.batch_size == 0: + model.optimizer.zero_grad() + + with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = model(input * 2 + masks) + + loss_per_token = F.cross_entropy( + logits.transpose(1, 2), targets, reduction="none" ) - 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, - 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) + loss = (loss_per_token * masks).mean() + acc_loss += loss.item() * imt.size(0) + nb_samples += imt.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): - 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 - ) - 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}%" - ) + if train: + loss.backward() - 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}%" - ) + if nb_samples % args.batch_size == 0: + model.optimizer.step() - 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()) - ar_mask[:, self.height * self.width :] = 1 - result *= 1 - ar_mask - masked_inplace_autoregression( - model, self.batch_size, result, ar_mask, device=self.device - ) + log_string(f"{label}_loss {n_epoch} model {model.id} {acc_loss/nb_samples}") - mazes, paths = self.seq2map(input) - _, predicted_paths = self.seq2map(result) - - filename = os.path.join(args.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), - ) - log_string(f"wrote {filename}") - model.train(t) +###################################################################### + + +def save_inference_images(model, n_epoch, c_quizzes, c_quiz_multiplier, local_device): + # Save some images of the prediction results + + quizzes = generate_quiz_set(150, c_quizzes, args.c_quiz_multiplier) + imt_set = samples_for_prediction_imt(quizzes.to(local_device)) + result = ae_predict(model, imt_set, local_device=local_device).to("cpu") + masks = imt_set[:, 1].to("cpu") + + correct = (quizzes == result).min(dim=1).values.long() + correct_parts = (2 * correct - 1)[:, None] * masks.reshape(masks.size(0), 4, -1)[ + :, :, 1 + ] + predicted_parts = correct_parts.abs() + + problem.save_quizzes_as_image( + args.result_dir, + f"culture_prediction_{n_epoch}_{model.id}.png", + quizzes=result[:128], + predicted_parts=predicted_parts[:128], + correct_parts=correct_parts[:128], + ) + + # Save some images of the ex nihilo generation of the four grids + + result = ae_generate(model, 150, local_device=local_device).to("cpu") + problem.save_quizzes_as_image( + args.result_dir, + f"culture_generation_{n_epoch}_{model.id}.png", + quizzes=result[:128], + ) ###################################################################### -import snake +def one_complete_epoch( + model, n_epoch, train_c_quizzes, test_c_quizzes, local_device=main_device +): + one_epoch(model, n_epoch, train_c_quizzes, train=True, local_device=local_device) + + one_epoch(model, n_epoch, test_c_quizzes, train=False, local_device=local_device) + # Compute the test accuracy -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, - ) + quizzes = generate_quiz_set(args.nb_test_samples, c_quizzes, args.c_quiz_multiplier) + imt_set = samples_for_prediction_imt(quizzes.to(local_device)) + result = ae_predict(model, imt_set, local_device=local_device).to("cpu") + correct = (quizzes == result).min(dim=1).values.long() - 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] - ) + nb_correct, nb_total = correct.sum().item(), quizzes.size(0) + model.test_accuracy = nb_correct / nb_total - 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}%" - ) + log_string( + f"test_accuracy {n_epoch} model {model.id} nb_correct {nb_correct} / {nb_total} ({model.test_accuracy*100:.02f}%)" + ) - model.train(t) + save_inference_images( + model, n_epoch, c_quizzes, args.c_quiz_multiplier, local_device=local_device + ) ###################################################################### -import stack +def max_nb_mistakes_on_one_grid(quizzes, prediction): + return ( + (prediction != quizzes) + .long() + .reshape(quizzes.size(0), 4, -1) + .sum(dim=2) + .max(dim=1) + .values + ) -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, +def evaluate_quizzes(quizzes, models, with_hints, local_device): + nb_correct, nb_wrong = 0, 0 + + for model in models: + model = copy.deepcopy(model).to(local_device).eval() + predicted = predict_the_four_grids( + model=model, + input=quizzes, + with_noise=False, + with_hints=with_hints, + local_device=local_device, ) + nb_mistakes = max_nb_mistakes_on_one_grid(quizzes, predicted) + nb_correct += (nb_mistakes == 0).long() + nb_wrong += (nb_mistakes >= args.nb_mistakes_to_be_wrong).long() + + # print("\n\n", nb_correct, nb_wrong) + + return nb_correct, nb_wrong + + +###################################################################### + + +def identity_quizzes(quizzes): + quizzes = quizzes.reshape(quizzes.size(0), 4, -1) + return (quizzes[:, 0] == quizzes[:, 1]).min(dim=1).values | ( + quizzes[:, 2] == quizzes[:, 3] + ).min(dim=1).values + + +def generate_c_quizzes(models, nb_to_generate, local_device=main_device): + record = [] + nb_validated = 0 + + start_time = time.perf_counter() + last_log = -1 - self.test_input, self.test_stack_counts = stack.generate_sequences( - nb_test_samples, - nb_steps, - nb_stacks, - nb_digits, - values_for_test, - self.device, + while nb_validated < nb_to_generate: + # Generate new quizzes + + model = models[torch.randint(len(models), (1,)).item()] + model = copy.deepcopy(model).to(local_device).eval() + generator_id = model.id + + c_quizzes = ae_generate( + model=model, nb=args.eval_batch_size * 10, local_device=local_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]) + c_quizzes = c_quizzes[identity_quizzes(c_quizzes) == False] - 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}%" + if c_quizzes.size(0) > 0: + # Select the ones that are solved properly by some models and + # not understood by others + + nb_correct, nb_wrong = evaluate_quizzes( + quizzes=c_quizzes, + models=models, + with_hints=True, + local_device=local_device, ) - #!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! - 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 + to_keep = (nb_correct >= args.nb_have_to_be_correct) & ( + nb_wrong >= args.nb_have_to_be_wrong + ) + + nb_validated += to_keep.long().sum().item() + record.append(c_quizzes[to_keep]) + + ##################### + + duration = time.perf_counter() - start_time + + if last_log < 0 or duration > last_log + 10: + last_log = duration + if nb_validated > 0: + if nb_validated < nb_to_generate: + d = (nb_to_generate - nb_validated) * duration / nb_validated + e = ( + datetime.datetime.now() + datetime.timedelta(seconds=d) + ).strftime("%a %H:%M") + else: + e = "now!" + else: + e = "???" + + log_string( + f"nb_validated {nb_validated} model {generator_id} (finishes {e} -- {int((nb_validated * 3600)/duration)}/h)" ) - 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) + ##################### + + duration = time.perf_counter() - start_time + + log_string(f"generate_c_quizz_speed {int(3600 * nb_validated / duration)}/h") + + return torch.cat(record).to("cpu") ###################################################################### -def picoclvr_pruner_horizontal_green(p): - return not ("green" in p and ("left" in p or "right" in p)) +def multithread_execution(fun, arguments): + # Single instance, no thread + if len(arguments) == 1: + return fun(*(arguments[0])) + records, threads = [], [] -picoclvr_pruner_train = ( - picoclvr_pruner_horizontal_green - if args.picocvlr_prune_properties in {"train+eval"} - else None -) + def threadable_fun(*args): + r = fun(*args) + if type(r) is not tuple: + r = (r,) + records.append(r) + + for args in arguments: + # To get a different sequence between threads + log_string(f"dummy_rand {torch.rand(1)}") + # torch.rand(1) + t = threading.Thread(target=threadable_fun, daemon=True, args=args) + threads.append(t) + t.start() + + for t in threads: + t.join() + + if records[0] == (None,): + return + else: + return [ + torch.cat([x[k] for x in records], dim=0) for k in range(len(records[0])) + ] -picoclvr_pruner_eval = ( - (lambda p: not picoclvr_pruner_horizontal_green(p)) - if args.picocvlr_prune_properties in {"train+eval", "eval"} - else None -) ###################################################################### -if args.task == "picoclvr": - task = TaskPicoCLVR( - nb_train_samples=args.nb_train_samples, - nb_test_samples=args.nb_test_samples, - batch_size=args.batch_size, - height=args.picoclvr_height, - width=args.picoclvr_width, - nb_colors=args.picoclvr_nb_colors, - device=device, - pruner_train=picoclvr_pruner_train, - pruner_eval=picoclvr_pruner_eval, - ) -elif args.task == "mnist": - task = TaskMNIST( - batch_size=args.batch_size, - device=device, - ) +def save_models(models, suffix=""): + if suffix != "": + suffix = "_" + suffix -elif args.task == "maze": - task = TaskMaze( - nb_train_samples=args.nb_train_samples, - nb_test_samples=args.nb_test_samples, - batch_size=args.batch_size, - height=args.maze_height, - width=args.maze_width, - nb_walls=args.maze_nb_walls, - device=device, - ) + for model in models: + filename = f"ae_{model.id:03d}{suffix}.pth" + torch.save( + { + "state_dict": model.state_dict(), + "optimizer_state_dict": model.optimizer.state_dict(), + "test_accuracy": model.test_accuracy, + }, + os.path.join(args.result_dir, filename), + ) -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, + log_string(f"wrote ae_*{suffix}.pth") + + +###################################################################### + + +def save_quiz_image(models, c_quizzes, filename, local_device=main_device): + c_quizzes = c_quizzes.to(local_device) + + nb_correct, nb_wrong = evaluate_quizzes( + quizzes=c_quizzes, + models=models, + with_hints=False, + local_device=local_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, + comments = [f"nb_correct {c} nb_wrong {w}" for c, w in zip(nb_correct, nb_wrong)] + + problem.save_quizzes_as_image( + args.result_dir, + filename, + quizzes=c_quizzes, + comments=comments, + delta=True, + nrow=8, ) -else: - raise ValueError(f"Unknown task {args.task}") + log_string(f"wrote {filename}") + ###################################################################### -log_string(f"device {device}") +problem = grids.Grids( + max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100, + chunk_size=100, + nb_threads=args.nb_threads, + tasks=args.grids_world_tasks, +) -vocabulary_size = task.vocabulary_size() +if not args.resume: + problem.save_some_examples(args.result_dir) -log_string(f"vocabulary_size {vocabulary_size}") -############################## - -model = mygpt.MyGPT( - vocabulary_size=vocabulary_size, - dim_model=args.dim_model, - dim_keys=args.dim_keys, - dim_hidden=args.dim_hidden, - nb_heads=args.nb_heads, - nb_blocks=args.nb_blocks, - causal=True, - dropout=args.dropout, -) +log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}") -model.to(device) +vocabulary_size = problem.vocabulary_size() -nb_parameters = sum(p.numel() for p in model.parameters()) -log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") +log_string(f"vocabulary_size {vocabulary_size}") ###################################################################### -nb_epochs_finished = 0 - -if args.no_checkpoint: - log_string(f"not trying to load checkpoint.") +models = [] +if args.model_type == "standard": + model_constructor = attae.AttentionAE +elif args.model_type == "functional": + model_constructor = attae.FunctionalAttentionAE else: - try: - checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name) - checkpoint = torch.load(checkpoint_name) - nb_epochs_finished = checkpoint["nb_epochs_finished"] - model.load_state_dict(checkpoint["model_state"]) - torch.set_rng_state(checkpoint["rng_state"]) - if torch.cuda.is_available(): - torch.cuda.set_rng_state(checkpoint["cuda_rng_state"]) + raise ValueError(f"Unknown model type {args.model_type}") + + +for i in range(args.nb_models): + model = model_constructor( + vocabulary_size=vocabulary_size * 2, + dim_model=args.dim_model, + dim_keys=args.dim_keys, + dim_hidden=args.dim_hidden, + nb_heads=args.nb_heads, + nb_blocks=args.nb_blocks, + dropout=args.dropout, + ) - log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.") + # model = torch.compile(model) - except FileNotFoundError: - log_string("starting from scratch.") + model.id = i + model.test_accuracy = 0.0 + model.optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate) - except: - log_string("error when loading the checkpoint.") - exit(1) + models.append(model) ###################################################################### -nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default +current_epoch = 0 -token_count = 0 -for input in task.batches(split="train"): - token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1)) -token_probas = token_count / token_count.sum() -entropy = -torch.xlogy(token_probas, token_probas).sum() -train_set_perplexity = math.exp(entropy) +if args.resume: + for model in models: + filename = f"ae_{model.id:03d}.pth" -############################## + d = torch.load( + os.path.join(args.result_dir, filename), + map_location="cpu", + weights_only=False, + ) + model.load_state_dict(d["state_dict"]) + model.optimizer.load_state_dict(d["optimizer_state_dict"]) + model.test_accuracy = d["test_accuracy"] + log_string(f"successfully loaded {filename}") + + filename = "state.pth" + state = torch.load( + os.path.join(args.result_dir, filename), + map_location="cpu", + weights_only=False, + ) -if args.learning_rate_schedule == "cos": - learning_rate_schedule = {} - for n_epoch in range(args.nb_epochs): - u = n_epoch / args.nb_epochs * math.pi - learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u)) -else: - u = { - int(k): float(v) - for k, v in [ - tuple(x.split(":")) for x in args.learning_rate_schedule.split(",") - ] - } + log_string(f"successfully loaded {filename}") - learning_rate_schedule = {} - learning_rate = args.learning_rate - for n_epoch in range(args.nb_epochs): - if n_epoch in u: - learning_rate = u[n_epoch] - learning_rate_schedule[n_epoch] = learning_rate + current_epoch = state["current_epoch"] + train_c_quizzes = state["train_c_quizzes"] + test_c_quizzes = state["test_c_quizzes"] -log_string(f"learning_rate_schedule {learning_rate_schedule}") +###################################################################### -############################## +nb_parameters = sum(p.numel() for p in models[0].parameters()) +log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") -nb_samples_seen = 0 -if nb_epochs_finished >= nb_epochs: - task.produce_results(nb_epochs_finished, model) +###################################################################### -for n_epoch in range(nb_epochs_finished, nb_epochs): - learning_rate = learning_rate_schedule[n_epoch] +train_c_quizzes, test_c_quizzes = None, None - log_string(f"learning_rate {learning_rate}") +###################################################################### - if args.optim == "sgd": - optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate) - elif args.optim == "adam": - optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate) - elif args.optim == "adamw": - optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate) - else: - raise ValueError(f"Unknown optimizer {args.optim}.") +for n_epoch in range(current_epoch, args.nb_epochs): + start_time = time.perf_counter() - model.train() + state = { + "current_epoch": n_epoch, + "train_c_quizzes": train_c_quizzes, + "test_c_quizzes": test_c_quizzes, + } - nb_train_samples, acc_train_loss = 0, 0.0 + filename = "state.pth" + torch.save(state, os.path.join(args.result_dir, filename)) + log_string(f"wrote {filename}") - for input in task.batches(split="train"): - input = input.to(device) - output = model(mygpt.BracketedSequence(input)).x - loss = F.cross_entropy(output.transpose(1, 2), input) - acc_train_loss += loss.item() * input.size(0) - nb_train_samples += input.size(0) - nb_samples_seen += input.size(0) + log_string(f"--- epoch {n_epoch} ----------------------------------------") - optimizer.zero_grad() - loss.backward() - optimizer.step() + cta = " ".join([f"{float(m.test_accuracy):.04f}" for m in models]) + log_string(f"current_test_accuracies {cta}") - with torch.autograd.no_grad(): - model.eval() + # -------------------------------------------------------------------- - nb_test_samples, acc_test_loss = 0, 0.0 + lowest_test_accuracy = min([float(m.test_accuracy) for m in models]) - for input in task.batches(split="test"): - input = input.to(device) + if lowest_test_accuracy >= args.accuracy_to_make_c_quizzes: + if train_c_quizzes is None: + save_models(models, "naive") - output = model(mygpt.BracketedSequence(input)).x - loss = F.cross_entropy(output.transpose(1, 2), input) - acc_test_loss += loss.item() * input.size(0) - nb_test_samples += input.size(0) + nb_gpus = len(gpus) + nb_c_quizzes_to_generate = (args.nb_c_quizzes + nb_gpus - 1) // nb_gpus - train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples)) - test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples)) + (new_c_quizzes,) = multithread_execution( + generate_c_quizzes, + [(models, nb_c_quizzes_to_generate, gpu) for gpu in gpus], + ) - log_string( - f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}" + save_quiz_image( + models, new_c_quizzes[:256], f"culture_c_quiz_{n_epoch:04d}.png" ) - task.produce_results(n_epoch, model) + log_string(f"generated_c_quizzes {new_c_quizzes.size()}") - checkpoint = { - "nb_epochs_finished": n_epoch + 1, - "model_state": model.state_dict(), - "rng_state": torch.get_rng_state(), - } + train_c_quizzes = ( + new_c_quizzes + if train_c_quizzes is None + else torch.cat([train_c_quizzes, new_c_quizzes]) + ) + train_c_quizzes = train_c_quizzes[-args.nb_train_samples :] - if torch.cuda.is_available(): - checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state() + nb_correct, _ = evaluate_quizzes( + quizzes=train_c_quizzes, + models=models, + with_hints=False, + local_device=local_device, + ) - checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name) - torch.save(checkpoint, checkpoint_name) - log_string(f"saved checkpoint {checkpoint_name}") + test_c_quizzes = train_c_quizzes[nb_correct >= args.nb_have_to_be_correct] -###################################################################### + for model in models: + model.test_accuracy = 0 + + if train_c_quizzes is None: + log_string("no_c_quiz") + else: + log_string(f"nb_c_quizzes {train_c_quizzes.size(0)}") + + # -------------------------------------------------------------------- + + ranked_models = sorted(models, key=lambda m: float(m.test_accuracy)) + weakest_models = ranked_models[: len(gpus)] + + log_string( + f"weakest_accuracies {[model.test_accuracy for model in weakest_models]}" + ) + + multithread_execution( + one_complete_epoch, + [ + (model, n_epoch, train_c_quizzes, test_c_quizzes, gpu) + for model, gpu in zip(weakest_models, gpus) + ], + ) + + save_models(models) + + # -------------------------------------------------------------------- + + duration = time.perf_counter() - start_time + str_duration = "" + if duration >= 60: + str_duration += f"{int(duration)//60}min" + str_duration += f"{int(duration)%60}s" + str_next = ( + datetime.datetime.now() + datetime.timedelta(seconds=duration) + ).strftime("%H:%M:%S") + log_string(f"epoch_duration {str_duration} next_finish {str_next}")