# Written by Francois Fleuret <francois@fleuret.org>
-# 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
import torch, torchvision
from torch import nn
from torch.nn import functional as F
import ffutils
-import mygpt, tasks
+import mygpt
+import sky, reasoning, quizz_machine
+
+# world quizzes vs. culture quizzes
+
+######################################################################
+
+nb_new_c_quizzes_for_train = 1000
+nb_new_c_quizzes_for_test = 100
######################################################################
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
-parser.add_argument(
- "--task",
- type=str,
- default="sandbox",
- help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
-)
-
-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("--max_percents_of_test_in_train", type=int, default=1)
+
+########################################
+
+parser.add_argument("--nb_epochs", type=int, default=10000)
parser.add_argument("--batch_size", type=int, default=None)
+parser.add_argument("--physical_batch_size", type=int, default=None)
+
parser.add_argument("--nb_train_samples", type=int, default=None)
parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--optim", type=str, default="adam")
+parser.add_argument("--learning_rate", type=float, default=1e-3)
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+########################################
-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
+parser.add_argument("--model", type=str, default=None)
-parser.add_argument("--dim_model", type=int, default=512)
+parser.add_argument("--dim_model", type=int, default=None)
-parser.add_argument("--dim_keys", type=int, default=64)
+parser.add_argument("--dim_keys", type=int, default=None)
-parser.add_argument("--dim_hidden", type=int, default=2048)
+parser.add_argument("--dim_hidden", type=int, default=None)
-parser.add_argument("--nb_heads", type=int, default=8)
+parser.add_argument("--nb_heads", type=int, default=None)
-parser.add_argument("--nb_blocks", type=int, default=12)
+parser.add_argument("--nb_blocks", type=int, default=None)
parser.add_argument("--dropout", type=float, default=0.1)
-parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-
-parser.add_argument("--no_checkpoint", action="store_true", default=False)
-
-parser.add_argument("--overwrite_results", action="store_true", default=False)
-
-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
-
-##############################
-# picoclvr options
+########################################
-parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
+parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-parser.add_argument("--picoclvr_height", type=int, default=12)
+parser.add_argument("--problem", type=str, default="sky")
-parser.add_argument("--picoclvr_width", type=int, default=16)
+parser.add_argument("--nb_gpts", type=int, default=5)
-parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
+parser.add_argument("--min_to_validate", type=int, default=None)
-##############################
-# Maze options
+parser.add_argument("--max_to_validate", type=int, default=None)
-parser.add_argument("--maze_height", type=int, default=23)
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
-parser.add_argument("--maze_width", type=int, default=39)
+parser.add_argument("--generation_temperature", type=float, default=2.0)
-parser.add_argument("--maze_nb_walls", type=int, default=45)
-
-##############################
-# Snake options
+parser.add_argument("--deterministic_validation", action="store_true", default=False)
-parser.add_argument("--snake_height", type=int, default=6)
+parser.add_argument("--bidirectional_validation", action="store_true", default=False)
-parser.add_argument("--snake_width", type=int, default=8)
+parser.add_argument("--dirty_debug", action="store_true", default=False)
-parser.add_argument("--snake_nb_colors", type=int, default=5)
-
-parser.add_argument("--snake_length", type=int, default=200)
-
-##############################
-# Stack options
+######################################################################
-parser.add_argument("--stack_nb_steps", type=int, default=100)
+parser.add_argument("--sky_height", type=int, default=6)
-parser.add_argument("--stack_nb_stacks", type=int, default=3)
+parser.add_argument("--sky_width", type=int, default=8)
-parser.add_argument("--stack_nb_digits", type=int, default=3)
+parser.add_argument("--sky_nb_birds", type=int, default=3)
-parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
+parser.add_argument("--sky_nb_iterations", type=int, default=2)
-##############################
-# Expr options
+parser.add_argument("--sky_speed", type=int, default=3)
-parser.add_argument("--expr_nb_variables", type=int, default=5)
+######################################################################
-parser.add_argument("--expr_sequence_length", type=int, default=40)
+args = parser.parse_args()
-parser.add_argument("--expr_operand_max", type=int, default=9)
+if args.min_to_validate is None:
+ args.min_to_validate = args.nb_gpts - 1
-parser.add_argument("--expr_result_max", type=int, default=99)
+if args.max_to_validate is None:
+ args.max_to_validate = args.nb_gpts - 1
-parser.add_argument("--expr_input_file", type=str, default=None)
+if args.result_dir is None:
+ args.result_dir = f"results_culture"
-##############################
-# World options
+######################################################################
-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+if args.dirty_debug:
+ args.accuracy_to_make_c_quizzes = 0.0
+ nb_new_c_quizzes_for_train = 100
+ nb_new_c_quizzes_for_test = 10
######################################################################
-args = parser.parse_args()
-
-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
+default_args = {
+ "model": "37M",
+ "batch_size": 100,
+ "nb_train_samples": 100000,
+ "nb_test_samples": 10000,
+}
-if args.result_dir is None:
- args.result_dir = f"results_{args.task}"
+for k, v in default_args.items():
+ if getattr(args, k) is None:
+ setattr(args, k, v)
######################################################################
-default_args = {
- "sandbox": {
- "nb_epochs": 10,
- "batch_size": 25,
- "nb_train_samples": 25000,
- "nb_test_samples": 10000,
- },
- "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": 5,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
+default_model_args = {
+ "17K": {
+ "dim_model": 32,
+ "dim_keys": 32,
+ "dim_hidden": 32,
+ "nb_heads": 2,
+ "nb_blocks": 2,
},
- "snake": {
- "nb_epochs": 5,
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
+ "4M": {
+ "dim_model": 256,
+ "dim_keys": 32,
+ "dim_hidden": 1024,
+ "nb_heads": 4,
+ "nb_blocks": 6,
},
- "stack": {
- "nb_epochs": 5,
- "batch_size": 25,
- "nb_train_samples": 100000,
- "nb_test_samples": 1000,
+ "37M": {
+ "dim_model": 512,
+ "dim_keys": 64,
+ "dim_hidden": 2048,
+ "nb_heads": 8,
+ "nb_blocks": 12,
},
- "expr": {
- "nb_epochs": 40,
- "batch_size": 25,
- "nb_train_samples": 1000000,
- "nb_test_samples": 10000,
+ "122M": {
+ "dim_model": 768,
+ "dim_keys": 64,
+ "dim_hidden": 2048,
+ "nb_heads": 8,
+ "nb_blocks": 24,
},
- "world": {
- "nb_epochs": 10,
- "batch_size": 25,
- "nb_train_samples": 25000,
- "nb_test_samples": 1000,
+ "352M": {
+ "dim_model": 1024,
+ "dim_keys": 64,
+ "dim_hidden": 2048,
+ "nb_heads": 8,
+ "nb_blocks": 48,
},
}
-if args.task in default_args:
- for k, v in default_args[args.task].items():
+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}")
######################################################################
try:
os.mkdir(args.result_dir)
except FileExistsError:
- if not args.overwrite_results:
- print(f"result directory {args.result_dir} already exists")
- exit(1)
+ print(f"result directory {args.result_dir} already exists")
+ exit(1)
log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
sys.stdout.flush()
+log_string(f"argv {' '.join(sys.argv)}")
+
for n in vars(args):
log_string(f"args.{n} {getattr(args, n)}")
######################################################################
+if args.dirty_debug:
+ args.nb_train_samples = 2500
+ args.nb_test_samples = 100
-def picoclvr_pruner_horizontal_green(p):
- return not ("green" in p and ("left" in p or "right" in p))
+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,
+ )
+ back_accuracy = False
+elif args.problem == "reasoning":
+ problem = reasoning.Reasoning(device=device)
+ back_accuracy = True
+else:
+ raise ValueError
+
+quizz_machine = quizz_machine.QuizzMachine(
+ 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=device,
+)
+######################################################################
-picoclvr_pruner_train = (
- picoclvr_pruner_horizontal_green
- if args.picocvlr_prune_properties in {"train+eval"}
- else None
-)
+log_string(f"device {device}")
-picoclvr_pruner_eval = (
- (lambda p: not picoclvr_pruner_horizontal_green(p))
- if args.picocvlr_prune_properties in {"train+eval", "eval"}
- else None
-)
+vocabulary_size = quizz_machine.vocabulary_size()
+
+log_string(f"vocabulary_size {vocabulary_size}")
######################################################################
-if args.task == "sandbox":
- task = tasks.SandBox(
- tasks.ProblemLevel1(),
- # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
- logger=log_string,
- device=device,
- )
+# Compute the entropy of the training tokens
-elif args.task == "picoclvr":
- task = tasks.PicoCLVR(
- 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,
- logger=log_string,
- device=device,
- pruner_train=picoclvr_pruner_train,
- pruner_eval=picoclvr_pruner_eval,
+token_count = 0
+for input in quizz_machine.batches(split="train", desc="train-entropy"):
+ token_count += F.one_hot(input, num_classes=quizz_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)
-elif args.task == "mnist":
- task = tasks.MNIST(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
- device=device,
- )
+######################################################################
+# A bit of paranoia never hurts
-elif args.task == "maze":
- task = tasks.Maze(
- 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,
+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(
+ quizz_machine.batches(split="test", desc="test-check"), 25000
+ ):
+ in_train = set()
+ for train_subset in subsets_as_tuples(
+ quizz_machine.batches(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)
+
+ 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"
)
-elif args.task == "snake":
- task = tasks.Snake(
- 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,
- )
+ 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"
-elif args.task == "stack":
- task = tasks.Stack(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
- logger=log_string,
- 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,
- )
+##############################
-elif args.task == "expr":
- task = tasks.Expr(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- nb_variables=args.expr_nb_variables,
- sequence_length=args.expr_sequence_length,
- operand_max=args.expr_operand_max,
- result_max=args.expr_result_max,
- batch_size=args.batch_size,
- device=device,
- )
-elif args.task == "world":
- task = tasks.World(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
- vqae_nb_epochs=args.world_vqae_nb_epochs,
- logger=log_string,
- device=device,
- )
+def one_epoch(model, quizz_machine):
+ optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
-else:
- raise ValueError(f"Unknown task {args.task}")
+ model.train()
-######################################################################
+ nb_train_samples, acc_train_loss = 0, 0.0
-log_string(f"device {device}")
+ for input in quizz_machine.batches(split="train"):
+ input = input.to(device)
-vocabulary_size = task.vocabulary_size()
+ if nb_train_samples % args.batch_size == 0:
+ optimizer.zero_grad()
-log_string(f"vocabulary_size {vocabulary_size}")
+ 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)
-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,
-)
+ loss.backward()
-model.to(device)
+ if nb_train_samples % args.batch_size == 0:
+ optimizer.step()
+
+ train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+ log_string(f"train_perplexity {n_epoch} {train_perplexity}")
-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
-if args.no_checkpoint:
- log_string(f"not trying to load checkpoint.")
+def run_tests(model, quizz_machine, deterministic_synthesis):
+ with torch.autograd.no_grad():
+ model.eval()
-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"])
+ nb_test_samples, acc_test_loss = 0, 0.0
+ nb_samples_accumulated = 0
+
+ for input in quizz_machine.batches(split="test"):
+ input = input.to(device)
+
+ 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))
- log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
+ log_string(f"test_perplexity {n_epoch} {test_perplexity}")
- except FileNotFoundError:
- log_string("starting from scratch.")
+ model.main_test_accuracy = quizz_machine.produce_results(
+ n_epoch=n_epoch,
+ model=model,
+ result_dir=args.result_dir,
+ deterministic_synthesis=deterministic_synthesis,
+ )
- except:
- log_string("error when loading the checkpoint.")
- exit(1)
######################################################################
-if args.task == "expr" and args.expr_input_file is not None:
- task.produce_results(
- nb_epochs_finished,
- model,
- args.result_dir,
- log_string,
- args.deterministic_synthesis,
- args.expr_input_file,
- )
- exit(0)
+def valid_c_quizzes(recorded, criteria):
+ result = [q[criteria(c)] for q, c in recorded]
+ return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
+
######################################################################
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
-# Compute the entropy of the training tokens
+def create_c_quizzes(
+ models,
+ quizz_machine,
+ nb_for_train=1000,
+ nb_for_test=100,
+):
+ recorded = []
-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)
+ nb_to_create = nb_for_train + nb_for_test
-##############################
+ # ------------------------------------------------------------
-# A bit of paranoia never hurts
+ standard_validity = lambda nb_correct: torch.logical_and(
+ nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate
+ )
-train_examples = {}
+ file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
+ with open(file_name, "w") as logp_file:
+ while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create:
+ # Select a model at random to generate the new quizzes
+ model_for_generation = models[torch.randint(len(models), (1,))]
-for input in task.batches(split="train"):
- assert input.dim() == 2 and input.dtype == torch.int64
- for x in input:
- train_examples[x.sum().item()] = x
+ c_quizzes = quizz_machine.generate_quizzes(
+ nb_to_create,
+ model_for_generation=model_for_generation,
+ temperature=args.generation_temperature,
+ )
-nb_total, nb_collisions = 0, 0
-for input in task.batches(split="test"):
- assert input.dim() == 2 and input.dtype == torch.int64
- for x in input:
- nb_total += 1
- y = train_examples.get(x.sum().item())
- if y is not None:
- if x.size() == y.size() and (x - y).abs().sum() == 0:
- nb_collisions += 1
+ nb_correct, seq_logproba = quizz_machine.compute_correctness(
+ c_quizzes,
+ models,
+ bidirectional_validation=args.bidirectional_validation,
+ deterministic_validation=args.deterministic_validation,
+ )
-del train_examples
+ for n, l in zip(nb_correct, seq_logproba):
+ s = " ".join([str(x.item()) for x in l])
+ logp_file.write(f"{n} {s}\n")
-log_string(
- f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
-)
+ if args.dirty_debug:
+ nb_correct = torch.randint(
+ len(models) + 1, nb_correct.size(), device=c_quizzes.device
+ )
-##############################
+ recorded.append((c_quizzes, nb_correct))
-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(",")
- ]
- }
-
- 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
-
-log_string(f"learning_rate_schedule {learning_rate_schedule}")
+ nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
+ nv = " ".join([str(x.item()) for x in nv])
-##############################
+ nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
-nb_samples_seen = 0
+ log_string(
+ f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
+ )
-if nb_epochs_finished >= nb_epochs:
- task.produce_results(
- nb_epochs_finished,
- model,
- args.result_dir,
- log_string,
- args.deterministic_synthesis,
- )
+ # store the new c_quizzes which have been validated
-for n_epoch in range(nb_epochs_finished, nb_epochs):
- learning_rate = learning_rate_schedule[n_epoch]
+ new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
- log_string(f"learning_rate {learning_rate}")
+ quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
+ quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
- 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}.")
+ # save a bunch of images to investigate what quizzes with a
+ # certain nb of correct predictions look like
- model.train()
+ for n in range(len(models) + 1):
+ s = (
+ "_validated"
+ if n >= args.min_to_validate and n <= args.max_to_validate
+ else ""
+ )
- nb_train_samples, acc_train_loss = 0, 0.0
+ q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
- 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)
+ if q.size(0) > 0:
+ quizz_machine.save_quizzes(
+ args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
+ )
- optimizer.zero_grad()
- loss.backward()
- optimizer.step()
- with torch.autograd.no_grad():
- model.eval()
+######################################################################
- nb_test_samples, acc_test_loss = 0, 0.0
+models = []
- for input in task.batches(split="test"):
- input = input.to(device)
+for k in range(args.nb_gpts):
+ 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(device)
- 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)
+ model.main_test_accuracy = 0.0
+ model.id = k
- train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+ models.append(model)
- log_string(
- f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
- )
- task.produce_results(
- n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
- )
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
- checkpoint = {
- "nb_epochs_finished": n_epoch + 1,
- "model_state": model.state_dict(),
- "rng_state": torch.get_rng_state(),
- }
+######################################################################
- if torch.cuda.is_available():
- checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+for n_epoch in range(args.nb_epochs):
+ log_string(f"--- epoch {n_epoch} ----------------------------------------")
+
+ cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
+ log_string(f"current_test_accuracies {cta}")
+
+ # Select, improve, and eval the worst model
+
+ weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
+
+ log_string(
+ f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
+ )
+
+ one_epoch(weakest_model, quizz_machine)
+
+ log_string(
+ f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+ )
+
+ run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
+
+ log_string(
+ f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
+ )
+
+ # Replace a fraction of the w_quizzes with fresh ones
+
+ quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+
+ # 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,
+ quizz_machine,
+ nb_for_train=nb_new_c_quizzes_for_train,
+ nb_for_test=nb_new_c_quizzes_for_test,
+ )
+
+ for model in models:
+ run_tests(model, quizz_machine, deterministic_synthesis=False)
- checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
- torch.save(checkpoint, checkpoint_name)
- log_string(f"saved checkpoint {checkpoint_name}")
######################################################################