# Written by Francois Fleuret <francois@fleuret.org>
-import math, sys, argparse, time, tqdm, itertools, os
+import math, sys, argparse, time, tqdm, os, datetime, warnings
import torch, torchvision
from torch import nn
from torch.nn import functional as F
-import mygpt, tensorstack
+import ffutils
-######################################################################
+import mygpt
+import sky, grids, quiz_machine
+
+import threading
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+import torch.multiprocessing as mp
######################################################################
parser = argparse.ArgumentParser(
- description="An implementation of GPT with cache to solve a toy geometric reasoning task."
+ formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("--log_filename", type=str, default="train.log")
-parser.add_argument("--result_dir", type=str, default="results_default")
+parser.add_argument("--result_dir", type=str, default=None)
parser.add_argument("--seed", type=int, default=0)
-parser.add_argument("--nb_epochs", type=int, default=25)
+parser.add_argument("--resume", action="store_true", default=False)
-parser.add_argument("--batch_size", type=int, default=100)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
-parser.add_argument("--data_size", type=int, default=-1)
+########################################
-parser.add_argument("--optim", type=str, default="adam")
+parser.add_argument("--nb_epochs", type=int, default=10000)
-parser.add_argument("--learning_rate", type=float, default=1e-3)
+parser.add_argument("--batch_size", type=int, default=None)
-parser.add_argument(
- "--learning_rate_schedule", type=str, default="10: 2e-4,20: 4e-5,30: 8e-6"
-)
+parser.add_argument("--physical_batch_size", type=int, default=None)
+
+parser.add_argument("--nb_train_samples", type=int, default=None)
-parser.add_argument("--dim_model", type=int, default=512)
+parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--dim_keys", type=int, default=64)
+parser.add_argument("--learning_rate", type=float, default=5e-4)
-parser.add_argument("--dim_hidden", type=int, default=2048)
+########################################
-parser.add_argument("--nb_heads", type=int, default=8)
+parser.add_argument("--model", type=str, default=None)
-parser.add_argument("--nb_blocks", type=int, default=12)
+parser.add_argument("--dim_model", type=int, default=None)
+
+parser.add_argument("--dim_keys", type=int, default=None)
+
+parser.add_argument("--dim_hidden", type=int, default=None)
+
+parser.add_argument("--nb_heads", type=int, default=None)
+
+parser.add_argument("--nb_blocks", type=int, default=None)
parser.add_argument("--dropout", type=float, default=0.1)
-parser.add_argument("--nb_oneshot_blocks", type=int, default=-1)
+########################################
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-parser.add_argument("--no_checkpoint", action="store_true", default=False)
+parser.add_argument("--problem", type=str, default="grids")
+
+parser.add_argument("--nb_threads", type=int, default=1)
+
+parser.add_argument("--gpus", type=str, default="all")
+
+parser.add_argument("--nb_gpts", type=int, default=5)
+
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+
+parser.add_argument("--proba_understands", type=float, default=0.99)
+
+parser.add_argument("--proba_not_understands", type=float, default=0.5)
+
+parser.add_argument("--generation_temperature", type=float, default=2.0)
+
+parser.add_argument("--dirty_debug", action="store_true", default=False)
+
+######################################################################
+
+grids_tasks = ", ".join(
+ [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)
-parser.add_argument("--overwrite_results", action="store_true", default=False)
+parser.add_argument(
+ "--grids_tasks",
+ type=str,
+ default=None,
+ help="A comma-separated subset of: " + grids_tasks + ", or None for all.",
+)
-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
+######################################################################
-##############################
-# picoclvr options
+parser.add_argument("--sky_height", type=int, default=6)
-parser.add_argument("--nb_colors", type=int, default=5)
+parser.add_argument("--sky_width", type=int, default=8)
-parser.add_argument("--height", type=int, default=12)
+parser.add_argument("--sky_nb_birds", type=int, default=3)
-parser.add_argument("--width", type=int, default=16)
+parser.add_argument("--sky_nb_iterations", type=int, default=2)
-parser.add_argument("--prune_properties", type=str, default="none")
+parser.add_argument("--sky_speed", type=int, default=3)
######################################################################
args = parser.parse_args()
-assert args.prune_properties in {"none", "train+eval", "eval"}
+if args.result_dir is None:
+ args.result_dir = f"results_culture"
-try:
- os.mkdir(args.result_dir)
-except FileExistsError:
- if not args.overwrite_results:
+######################################################################
+
+default_args = {
+ "model": "37M",
+ "batch_size": 25,
+ "nb_train_samples": 100000,
+ "nb_test_samples": 10000,
+}
+
+for k, v in default_args.items():
+ if getattr(args, k) is None:
+ setattr(args, k, v)
+
+######################################################################
+
+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,
+ },
+}
+
+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:
+ assert os.path.isdir(args.result_dir)
+
+else:
+ try:
+ os.mkdir(args.result_dir)
+ except FileExistsError:
print(f"result directory {args.result_dir} already exists")
exit(1)
-log_file = open(os.path.join(args.result_dir, args.log_filename), "w")
+log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
if args.seed >= 0:
# torch.backends.cudnn.deterministic = True
sys.stdout.flush()
+now = time.strftime("%Y%m%d-%H%M%S", time.localtime())
+
+os.system(f"tar zcvf {args.result_dir}/src-{now}.tgz *.py")
+
+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(",")]
-def masked_inplace_autoregression(
- model, batch_size, input, ar_mask, forbidden_tokens=None, device=torch.device("cpu")
-):
-
- for input, ar_mask in zip(input.split(batch_size), ar_mask.split(batch_size)):
- 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]
+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")
-######################################################################
+if args.dirty_debug:
+ args.nb_train_samples = 2500
+ args.nb_test_samples = 100
+if args.physical_batch_size is None:
+ args.physical_batch_size = args.batch_size
+else:
+ assert args.batch_size % args.physical_batch_size == 0
+
+assert args.nb_train_samples % args.batch_size == 0
+assert args.nb_test_samples % args.batch_size == 0
+
+if args.problem == "sky":
+ problem = sky.Sky(
+ height=args.sky_height,
+ width=args.sky_width,
+ nb_birds=args.sky_nb_birds,
+ nb_iterations=args.sky_nb_iterations,
+ speed=args.sky_speed,
+ max_nb_cached_chunks=len(gpus) * args.nb_train_samples // 100,
+ chunk_size=100,
+ nb_threads=args.nb_threads,
+ )
+ back_accuracy = False
+elif args.problem == "grids":
+ 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_tasks,
+ )
+ back_accuracy = True
+else:
+ raise ValueError
+
+problem.save_some_examples(args.result_dir)
+
+quiz_machine = quiz_machine.QuizMachine(
+ problem=problem,
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ back_accuracy=back_accuracy,
+ batch_size=args.physical_batch_size,
+ result_dir=args.result_dir,
+ logger=log_string,
+ device=main_device,
+)
-class Task:
- def batches(self, split="train"):
- pass
+######################################################################
- def vocabulary_size(self):
- pass
+log_string(f"main_device {main_device} gpus {[ str(g) for g in gpus]}")
- def produce_results(self, n_epoch, model):
- pass
+vocabulary_size = quiz_machine.vocabulary_size()
+log_string(f"vocabulary_size {vocabulary_size}")
######################################################################
-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 + ["<nul>"] * (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="<nul>"):
- 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 <img>
- # included, and set to <nul> all the tokens from the beginning of
- # that image to the end
- def excise_last_image(self, input):
- t_img, t_nul = self.token2id["<img>"], self.token2id["<nul>"]
- 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["<nul>"]
- 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["<img>"], self.token2id["<nul>"]
- 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,
- device=self.device,
- )
- model.train(t)
- input, loss_masks = self.trim((input, loss_masks))
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
+ with torch.autograd.no_grad():
+ model.eval().to(local_device)
- return input, loss_masks
+ nb_test_samples, acc_test_loss = 0, 0.0
+ nb_samples_accumulated = 0
- ######################
+ for input in quiz_machine.batches(model, split="test"):
+ input = input.to(local_device)
- def __init__(
- self,
- 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,
- )
+ bs = model(mygpt.BracketedSequence(input))
+ output = bs.x
+
+ loss = F.cross_entropy(output.transpose(1, 2), input)
+
+ acc_test_loss += loss.item() * input.size(0)
+
+ nb_test_samples += input.size(0)
+
+ test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
- self.height = height
- self.width = width
- self.batch_size = batch_size
- self.device = device
- nb = args.data_size if args.data_size > 0 else 250000
- self.pruner_train = pruner_train
- self.pruner_eval = pruner_eval
-
- param = {
- "nb": nb,
- "height": height,
- "width": width,
- "nb_colors": nb_colors,
- "batch_size": batch_size,
- "rng_state": list(torch.get_rng_state()),
- }
-
- log_string(f"generating {nb} samples (can take some time)")
- self.train_descr = generate_descr(
- (nb * 4) // 5, "train", pruner=self.pruner_train
+ log_string(f"test_perplexity {n_epoch} model {model.id} {test_perplexity}")
+
+ model.main_test_accuracy = quiz_machine.produce_results(
+ n_epoch=n_epoch,
+ model=model,
+ result_dir=args.result_dir,
+ deterministic_synthesis=deterministic_synthesis,
)
- self.test_descr = generate_descr((nb * 1) // 5, "test", pruner=None)
-
- # Build the tokenizer
- tokens = {"<nul>", "<img>"}
- 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):
+def one_epoch(model, quiz_machine, local_device=main_device):
+ model.to(local_device).train()
- acc_nb_requested_properties = []
- acc_nb_missing_properties = []
- acc_nb_results = 0
+ optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
- 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_train_samples, acc_train_loss = 0, 0.0
- nb_requested_properties = sum(acc_nb_requested_properties)
- nb_missing_properties = sum(acc_nb_missing_properties)
+ for input in quiz_machine.batches(model, split="train"):
+ input = input.to(local_device)
- 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}%"
- )
+ if nb_train_samples % args.batch_size == 0:
+ optimizer.zero_grad()
- ######################################################################
+ output = model(mygpt.BracketedSequence(input)).x
+ loss = F.cross_entropy(output.transpose(1, 2), input)
+ acc_train_loss += loss.item() * input.size(0)
- def produce_results(self, n_epoch, model):
+ nb_train_samples += input.size(0)
- self.compute_missing_properties(n_epoch, model)
+ loss.backward()
- if self.pruner_eval is not None:
- self.compute_missing_properties(n_epoch, model, self.pruner_eval)
+ if nb_train_samples % args.batch_size == 0:
+ optimizer.step()
- nb_tokens_to_generate = self.height * self.width + 3
- result_descr = []
- nb_per_primer = 8
- primer = []
+ train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- for primer_descr in [
- "red above green <sep> green top <sep> blue right of red",
- "there is red <sep> there is yellow <sep> there is blue",
- "red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left",
- "green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top",
- ]:
- primer += [primer_descr] * nb_per_primer
+ log_string(f"train_perplexity {n_epoch} model {model.id} {train_perplexity}")
- tape = self.tensorize(primer)
- loss_masks = 1 - (tape == self.token2id["<nul>"]).long()
- tape, loss_masks = self.add_generated_image(tape, loss_masks, model)
- result_descr = self.detensorize(tape)
+ run_tests(model, quiz_machine, deterministic_synthesis=False)
- np = picoclvr.nb_properties(result_descr, height=self.height, width=self.width)
+ model.to(main_device)
- 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)
+######################################################################
+
+
+def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return (l[:, 0] < math.log(args.proba_not_understands)) & (
+ l[:, 1] > math.log(args.proba_understands)
+ )
- 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}%"
- )
- img = picoclvr.descr2img(result_descr, height=self.height, width=self.width)
-
- 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"result_{n_epoch:04d}.png")
- torchvision.utils.save_image(
- img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
+def valid_quizzes_and_logprobas(recorded, criteria):
+ validated_quizzes, validated_logprobas = [], []
+ for q, lp in recorded:
+ validated_indices = criteria(lp)
+ validated_quizzes.append(q[validated_indices])
+ validated_logprobas.append(lp[validated_indices])
+
+ if len(validated_quizzes) > 0:
+ return torch.cat(validated_quizzes, dim=0), torch.cat(
+ validated_logprobas, dim=0
)
- log_string(f"wrote {image_name}")
+ else:
+ return None, None
######################################################################
-log_string(f"device {device}")
+def create_c_quizzes(models, quiz_machine, nb_for_train=1000, nb_for_test=100):
+ nb_to_create = nb_for_train + nb_for_test
-def pruner_horizontal_green(p):
- return not ("green" in p and ("left" in p or "right" in p))
+ recorded_quizzes_logprobas = []
+ nb_validated = 0
-task = TaskPicoCLVR(
- batch_size=args.batch_size,
- height=args.height,
- width=args.width,
- nb_colors=args.nb_colors,
- device=device,
- pruner_train=pruner_horizontal_green
- if args.prune_properties in {"train+eval"}
- else None,
- pruner_eval=(lambda p: not pruner_horizontal_green(p))
- if args.prune_properties in {"train+eval", "eval"}
- else None,
-)
+ while nb_validated < nb_to_create:
+ model_for_generation = models[torch.randint(len(models), (1,))]
-vocabulary_size = task.vocabulary_size()
+ c_quizzes = quiz_machine.generate_quizzes(
+ nb_to_create,
+ model_for_generation=model_for_generation,
+ temperature=args.generation_temperature,
+ )
-log_string(f"vocabulary_size {vocabulary_size}")
+ c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
-##############################
-
-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,
-)
+ if c_quizzes.size(0) > 0:
+ logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
+ recorded_quizzes_logprobas.append((c_quizzes, logproba))
-model.to(device)
+ validated_quizzes, validated_logprobas = valid_quizzes_and_logprobas(
+ recorded_quizzes_logprobas, standard_validity
+ )
+
+ if validated_quizzes is not None:
+ nb_validated = validated_quizzes.size(0)
+
+ log_string(
+ f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}"
+ )
+
+ # store the new c_quizzes which have been validated
+
+ quiz_machine.reverse_random_half_in_place(validated_quizzes)
+ quiz_machine.store_c_quizzes(validated_quizzes[:nb_for_train], for_train=True)
+ quiz_machine.store_c_quizzes(
+ validated_quizzes[nb_for_train:nb_to_create], for_train=False
+ )
+
+ ######################################################################
+ # save the log probas
+
+ file_name = os.path.join(
+ args.result_dir, f"culture_c_quiz_all_{n_epoch:04d}_logp.dat"
+ )
+
+ with open(file_name, "w") as logp_file:
+ for _, ll in recorded_quizzes_logprobas:
+ for l in ll:
+ s = " ".join([str(x.item()) for x in l])
+ logp_file.write(s + "\n")
+
+ ######################################################################
+ # save images with their logprobas
+
+ vq = validated_quizzes[:72]
+ vl = validated_logprobas[:72]
+
+ if vq.size(0) > 0:
+ prefix = f"culture_c_quiz_{n_epoch:04d}"
+
+ file_name = os.path.join(args.result_dir, prefix + "_logp.dat")
+ with open(file_name, "w") as logp_file:
+ for l in vl:
+ s = " ".join([str(x.item()) for x in l])
+ logp_file.write(s + "\n")
+
+ quiz_machine.save_quiz_illustrations(args.result_dir, prefix, vq)
-nb_parameters = sum(p.numel() for p in model.parameters())
-log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
######################################################################
-nb_epochs_finished = 0
+models = []
-if args.no_checkpoint:
- log_string(f"not trying to load checkpoint.")
+for k in range(args.nb_gpts):
+ log_string(f"creating model {k} and its w_quizzes")
+ 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,
+ ).to(main_device)
-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"])
+ model.main_test_accuracy = 0.0
+ model.id = k
+
+ model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
+ quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
+ model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
+ quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
- log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
+ models.append(model)
+
+######################################################################
- except FileNotFoundError:
- log_string("starting from scratch.")
+if args.resume:
+ try:
+ for model in models:
+ filename = f"gpt_{model.id:03d}.pth"
+
+ try:
+ d = torch.load(os.path.join(args.result_dir, filename))
+ model.load_state_dict(d[0])
+ model.main_test_accuracy = d[1]
+ log_string(f"successfully loaded {filename}")
+ except FileNotFoundError:
+ log_string(f"cannot find {filename}")
+ pass
+
+ try:
+ filename = "c_quizzes.pth"
+ quiz_machine.load_c_quizzes(os.path.join(args.result_dir, filename))
+ log_string(f"successfully loaded {filename}")
+ except FileNotFoundError:
+ log_string(f"cannot find {filename}")
+ pass
except:
- log_string("error when loading the checkpoint.")
+ log_string(f"error when loading {filename}.")
exit(1)
######################################################################
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
+
+######################################################################
+
+# Compute the entropy of the training tokens
token_count = 0
-for input in task.batches(split="train"):
- token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
+for input in quiz_machine.batches(models[0], split="train", desc="train-entropy"):
+ token_count += F.one_hot(input, num_classes=quiz_machine.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)
-##############################
+######################################################################
+# A bit of paranoia never hurts
+
+if args.max_percents_of_test_in_train >= 0:
+
+ def subsets_as_tuples(batches, cs):
+ s = set()
+ for batch in batches:
+ for x in batch:
+ s.add(tuple([v.item() for v in x]))
+ if len(s) == cs:
+ yield s
+ s = set()
+ yield s
+
+ nb_test, nb_in_train = 0, 0
+ for test_subset in subsets_as_tuples(
+ quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
+ ):
+ in_train = set()
+ for train_subset in subsets_as_tuples(
+ quiz_machine.batches(models[0], split="train", desc="train-check"), 25000
+ ):
+ in_train.update(test_subset.intersection(train_subset))
+ nb_in_train += len(in_train)
+ nb_test += len(test_subset)
-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"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
+ )
- 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
+ assert (
+ nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
+ ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+######################################################################
-##############################
+nb_new_c_quizzes_for_train = args.nb_train_samples // 50
+nb_new_c_quizzes_for_test = args.nb_test_samples // 50
-nb_samples_seen = 0
+log_string(
+ f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
+)
-if nb_epochs_finished >= nb_epochs:
- task.produce_results(nb_epochs_finished, model)
+######################################################################
-for n_epoch in range(nb_epochs_finished, nb_epochs):
+if args.dirty_debug:
+ args.accuracy_to_make_c_quizzes = 0.0
+ args.nb_gpts = 2
+ nb_new_c_quizzes_for_train = 100
+ nb_new_c_quizzes_for_test = 10
- learning_rate = learning_rate_schedule[n_epoch]
+ def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return l[:, 0] < math.log(0.5)
- 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}.")
+######################################################################
- model.train()
+for n_epoch in range(args.nb_epochs):
+ log_string(f"--- epoch {n_epoch} ----------------------------------------")
- nb_train_samples, acc_train_loss = 0, 0.0
+ cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
+ log_string(f"current_test_accuracies {cta}")
- 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)
+ ##################################################
+ # Select, improve, and eval the worst model
- optimizer.zero_grad()
- loss.backward()
- optimizer.step()
+ ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
- with torch.autograd.no_grad():
+ weakest_models = ranked_models[: len(gpus)]
- model.eval()
+ threads = []
- nb_test_samples, acc_test_loss = 0, 0.0
+ for gpu, model in zip(gpus, weakest_models):
+ log_string(f"training model {model.id}")
+
+ t = threading.Thread(
+ target=one_epoch, daemon=True, args=(model, quiz_machine, gpu)
+ )
- for input in task.batches(split="test"):
- input = input.to(device)
+ threads.append(t)
- # input, loss_masks, true_images = task.excise_last_image(input)
- # input, loss_masks = task.add_true_image(input, true_images, loss_masks)
+ t.start()
- 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)
+ for t in threads:
+ t.join()
- train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+ # Save the models to disk
- log_string(
- f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
+ for model in weakest_models:
+ filename = f"gpt_{model.id:03d}.pth"
+ torch.save(
+ (model.state_dict(), model.main_test_accuracy),
+ os.path.join(args.result_dir, filename),
)
+ log_string(f"wrote {filename}")
- task.produce_results(n_epoch, model)
+ # Renew the training samples
- checkpoint = {
- "nb_epochs_finished": n_epoch + 1,
- "model_state": model.state_dict(),
- "rng_state": torch.get_rng_state(),
- }
+ for model in weakest_models:
+ quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
- if torch.cuda.is_available():
- checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+ ##################################################
+ # If all the models are good enough, generate new quizzes and
+ # re-compute the test errors
+
+ if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
+ create_c_quizzes(
+ models,
+ quiz_machine,
+ nb_for_train=nb_new_c_quizzes_for_train,
+ nb_for_test=nb_new_c_quizzes_for_test,
+ )
- checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
- torch.save(checkpoint, checkpoint_name)
- log_string(f"saved checkpoint {checkpoint_name}")
+ filename = "c_quizzes.pth"
+ quiz_machine.save_c_quizzes(os.path.join(args.result_dir, filename))
+ log_string(f"wrote {filename}")
######################################################################