from torch.nn import functional as F
import ffutils
-import mygpt, tasks, problems
+
+import mygpt
+import sky, grids, quiz_machine
+
+import threading
+
+import torch.multiprocessing as mp
+
+# world quizzes vs. culture quizzes
######################################################################
######################################################################
parser = argparse.ArgumentParser(
- description="An implementation of GPT with cache.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
-parser.add_argument(
- "--task",
- type=str,
- default="twotargets",
- help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
-)
-
-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("--max_percents_of_test_in_train", type=int, default=1)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
########################################
-parser.add_argument("--nb_epochs", type=int, default=50)
+parser.add_argument("--nb_epochs", type=int, default=10000)
parser.add_argument("--batch_size", 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-4)
-
-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
+parser.add_argument("--learning_rate", type=float, default=5e-4)
########################################
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-parser.add_argument("--no_checkpoint", action="store_true", default=False)
-
-parser.add_argument("--resume", action="store_true", default=False)
-
-parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
-
-##############################
-# filetask
-
-parser.add_argument("--filetask_train_file", type=str, default=None)
-
-parser.add_argument("--filetask_test_file", type=str, default=None)
-
-##############################
-# rpl options
-
-parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
-
-parser.add_argument("--rpl_max_input", type=int, default=9)
-
-parser.add_argument("--rpl_prog_len", type=int, default=8)
-
-parser.add_argument("--rpl_nb_runs", type=int, default=5)
-
-parser.add_argument("--rpl_no_prog", action="store_true", default=False)
-
-##############################
-# grid options
-
-parser.add_argument("--grid_size", type=int, default=6)
-
-parser.add_argument("--grid_fraction_play", type=float, default=0)
-
-##############################
-# picoclvr options
-
-parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
-
-parser.add_argument("--picoclvr_height", type=int, default=12)
-
-parser.add_argument("--picoclvr_width", type=int, default=16)
+parser.add_argument("--problem", type=str, default="grids")
-parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
+parser.add_argument("--nb_threads", type=int, default=1)
-##############################
-# Maze options
+parser.add_argument("--nb_gpus", type=int, default=1)
-parser.add_argument("--maze_height", type=int, default=13)
+parser.add_argument("--nb_gpts", type=int, default=5)
-parser.add_argument("--maze_width", type=int, default=21)
+parser.add_argument("--min_to_validate", type=int, default=None)
-parser.add_argument("--maze_nb_walls", type=int, default=15)
+parser.add_argument("--max_to_validate", type=int, default=None)
-##############################
-# Snake options
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
-parser.add_argument("--snake_height", type=int, default=9)
+parser.add_argument("--generation_temperature", type=float, default=2.0)
-parser.add_argument("--snake_width", type=int, default=12)
+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)
-
-##############################
-# ByHeart options
-
-parser.add_argument("--byheart_separation", type=int, default=1)
-
-##############################
-# Stack options
-
-parser.add_argument("--stack_nb_steps", type=int, default=100)
-
-parser.add_argument("--stack_nb_stacks", type=int, default=3)
-
-parser.add_argument("--stack_nb_digits", type=int, default=3)
-
-parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
-
-##############################
-# Expr options
-
-parser.add_argument("--expr_nb_variables", type=int, default=5)
-
-parser.add_argument("--expr_sequence_length", type=int, default=40)
-
-parser.add_argument("--expr_operand_max", type=int, default=9)
-
-parser.add_argument("--expr_result_max", type=int, default=99)
-
-parser.add_argument("--expr_input_file", type=str, default=None)
-
-##############################
-# Mixing
+######################################################################
-parser.add_argument("--mixing_hard", action="store_true", default=False)
+grids_tasks = ", ".join(
+ [x.__name__.removeprefix("task_") for x in grids.Grids().all_tasks]
+)
-parser.add_argument("--mixing_deterministic_start", 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.",
+)
-##############################
-# greed options
+######################################################################
-parser.add_argument("--greed_height", type=int, default=5)
+parser.add_argument("--sky_height", type=int, default=6)
-parser.add_argument("--greed_width", type=int, default=7)
+parser.add_argument("--sky_width", type=int, default=8)
-parser.add_argument("--greed_T", type=int, default=25)
+parser.add_argument("--sky_nb_birds", type=int, default=3)
-parser.add_argument("--greed_nb_walls", type=int, default=5)
+parser.add_argument("--sky_nb_iterations", type=int, default=2)
-parser.add_argument("--greed_nb_coins", type=int, default=2)
+parser.add_argument("--sky_speed", type=int, default=3)
######################################################################
args = parser.parse_args()
-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
+if args.min_to_validate is None:
+ args.min_to_validate = args.nb_gpts - 1
+
+if args.max_to_validate is None:
+ args.max_to_validate = args.nb_gpts - 1
if args.result_dir is None:
- args.result_dir = f"results_{args.task}"
+ args.result_dir = f"results_culture"
######################################################################
-default_task_args = {
- "file": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
- },
- "addition": {
- "model": "352M",
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
- },
- "world": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 50000,
- "nb_test_samples": 10000,
- },
- "byheart": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 50000,
- "nb_test_samples": 10000,
- },
- "expr": {
- "model": "352M",
- "batch_size": 25,
- "nb_train_samples": 2500000,
- "nb_test_samples": 10000,
- },
- "grid": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
- },
- "qmlp": {
- "model": "37M",
- "batch_size": 10,
- "nb_train_samples": 100000,
- "nb_test_samples": 1000,
- },
- "guessop": {
- "model": "352M",
- "batch_size": 25,
- "nb_train_samples": 1000000,
- "nb_test_samples": 10000,
- },
- "learnop": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 50000,
- "nb_test_samples": 10000,
- },
- "maze": {
- "model": "37M",
- "batch_size": 5,
- "nb_train_samples": 100000,
- "nb_test_samples": 10000,
- },
- "picoclvr": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
- },
- "rpl": {
- "model": "352M",
- "batch_size": 5,
- "nb_train_samples": 2500000,
- "nb_test_samples": 10000,
- },
- "snake": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
- },
- "stack": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 100000,
- "nb_test_samples": 1000,
- },
- "twotargets": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 50000,
- "nb_test_samples": 10000,
- },
- "memory": {
- "model": "37M",
- "batch_size": 100,
- "nb_train_samples": 25000,
- "nb_test_samples": 1000,
- },
- "mixing": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 250000,
- "nb_test_samples": 10000,
- },
- "mnist": {
- "model": "37M",
- "batch_size": 10,
- "nb_train_samples": 60000,
- "nb_test_samples": 10000,
- },
- "greed": {
- "model": "37M",
- "batch_size": 25,
- "nb_train_samples": 25000,
- "nb_test_samples": 10000,
- },
+default_args = {
+ "model": "37M",
+ "batch_size": 25,
+ "nb_train_samples": 100000,
+ "nb_test_samples": 10000,
}
-if args.task in default_task_args:
- for k, v in default_task_args[args.task].items():
- if getattr(args, k) is None:
- setattr(args, k, v)
+for k, v in default_args.items():
+ if getattr(args, k) is None:
+ setattr(args, k, v)
######################################################################
try:
os.mkdir(args.result_dir)
except FileExistsError:
- if not args.resume:
- 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")
######################################################################
-
-def picoclvr_pruner_horizontal_green(p):
- return not ("green" in p and ("left" in p or "right" in p))
-
-
-picoclvr_pruner_train = (
- picoclvr_pruner_horizontal_green
- if args.picocvlr_prune_properties in {"train+eval"}
- else None
-)
-
-picoclvr_pruner_eval = (
- (lambda p: not picoclvr_pruner_horizontal_green(p))
- if args.picocvlr_prune_properties in {"train+eval", "eval"}
- else None
-)
-
-######################################################################
+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
assert args.nb_train_samples % args.batch_size == 0
assert args.nb_test_samples % args.batch_size == 0
-if args.task == "file":
- assert (
- args.filetask_train_file is not None and args.filetask_test_file is not None
- ), "You have to specify the task train and test files"
- task = tasks.TaskFromFile(
- args.filetask_train_file,
- args.filetask_test_file,
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- shuffle=True,
- device=device,
+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=args.nb_gpus * args.nb_train_samples // 100,
+ chunk_size=100,
+ nb_threads=args.nb_threads,
)
- args.max_percents_of_test_in_train = 0
-
-elif args.task == "byheart":
- task = tasks.SandBox(
- problem=problems.ProblemByHeart(separation=args.byheart_separation),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
- args.max_percents_of_test_in_train = -1
-
-elif args.task == "world":
- task = tasks.World(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
- args.max_percents_of_test_in_train = -1
-
-elif args.task == "learnop":
- task = tasks.SandBox(
- problem=problems.ProblemLearnOperator(),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
+ back_accuracy = False
+elif args.problem == "grids":
+ problem = grids.Grids(
+ max_nb_cached_chunks=args.nb_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=device,
+)
+######################################################################
-elif args.task == "guessop":
- task = tasks.SandBox(
- problem=problems.ProblemGuessOperator(),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
+log_string(f"device {device}")
+vocabulary_size = quiz_machine.vocabulary_size()
-elif args.task == "twotargets":
- task = tasks.SandBox(
- problem=problems.ProblemTwoTargets(),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
+log_string(f"vocabulary_size {vocabulary_size}")
-elif args.task == "memory":
- task = tasks.SandBox(
- problem=problems.ProblemMemory(),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
+######################################################################
-elif args.task == "mixing":
- task = tasks.SandBox(
- problem=problems.ProblemMixing(
- hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
- ),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
-elif args.task == "addition":
- task = tasks.SandBox(
- problem=problems.ProblemAddition(),
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- logger=log_string,
- device=device,
- )
+######################################################################
-elif args.task == "picoclvr":
- task = tasks.PicoCLVR(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_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,
- )
-elif args.task == "mnist":
- task = tasks.MNIST(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- device=device,
- )
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
+ if local_device is None:
+ local_device = device
-elif args.task == "maze":
- task = tasks.Maze(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- height=args.maze_height,
- width=args.maze_width,
- nb_walls=args.maze_nb_walls,
- device="cpu",
- )
+ with torch.autograd.no_grad():
+ model.eval().to(local_device)
-elif args.task == "snake":
- task = tasks.Snake(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_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,
- )
+ nb_test_samples, acc_test_loss = 0, 0.0
+ nb_samples_accumulated = 0
-elif args.task == "stack":
- task = tasks.Stack(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_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,
- )
+ for input in quiz_machine.batches(model, split="test"):
+ input = input.to(local_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.physical_batch_size,
- device=device,
- )
+ bs = model(mygpt.BracketedSequence(input))
+ output = bs.x
-elif args.task == "rpl":
- task = tasks.RPL(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- nb_starting_values=args.rpl_nb_starting_values,
- max_input=args.rpl_max_input,
- prog_len=args.rpl_prog_len,
- nb_runs=args.rpl_nb_runs,
- no_prog=args.rpl_no_prog,
- logger=log_string,
- device=device,
- )
+ loss = F.cross_entropy(output.transpose(1, 2), input)
-elif args.task == "grid":
- task = tasks.Grid(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- size=args.grid_size,
- fraction_play=args.grid_fraction_play,
- logger=log_string,
- device=device,
- )
+ acc_test_loss += loss.item() * input.size(0)
-elif args.task == "qmlp":
- task = tasks.QMLP(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- result_dir=args.result_dir,
- logger=log_string,
- device=device,
- )
+ nb_test_samples += input.size(0)
-elif args.task == "greed":
- task = tasks.Greed(
- nb_train_samples=args.nb_train_samples,
- nb_test_samples=args.nb_test_samples,
- batch_size=args.physical_batch_size,
- height=args.greed_height,
- width=args.greed_width,
- T=args.greed_T,
- nb_walls=args.greed_nb_walls,
- nb_coins=args.greed_nb_coins,
- logger=log_string,
- device=device,
- )
+ test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
-else:
- raise ValueError(f"Unknown task {args.task}")
+ log_string(f"test_perplexity {n_epoch} {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,
+ )
-log_string(f"device {device}")
-vocabulary_size = task.vocabulary_size()
+def one_epoch(model, quiz_machine, local_device=None):
+ if local_device is None:
+ local_device = device
-log_string(f"vocabulary_size {vocabulary_size}")
+ optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
-##############################
-
-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,
-)
+ model.to(local_device).train()
-model.to(device)
+ nb_train_samples, acc_train_loss = 0, 0.0
+
+ for input in quiz_machine.batches(model, split="train"):
+ input = input.to(local_device)
+
+ 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)
+
+ nb_train_samples += input.size(0)
+
+ loss.backward()
+
+ 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} model.id {model.id} {train_perplexity}")
+
+ run_tests(model, quiz_machine, deterministic_synthesis=False)
-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 standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
-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"])
- log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
+def valid_c_quizzes(recorded, criteria):
+ result = [q[criteria(lp)] for q, lp in recorded]
+ return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([])
+
+
+######################################################################
+
+
+def create_c_quizzes(
+ models,
+ quiz_machine,
+ nb_for_train=1000,
+ nb_for_test=100,
+):
+ quizzes_and_logproba_records = []
+
+ nb_to_create = nb_for_train + nb_for_test
+
+ # ------------------------------------------------------------
+
+ 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(quizzes_and_logproba_records, 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,))]
+
+ c_quizzes = quiz_machine.generate_quizzes(
+ nb_to_create,
+ model_for_generation=model_for_generation,
+ temperature=args.generation_temperature,
+ )
+
+ c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
+
+ if c_quizzes.size(0) > 0:
+ logproba = quiz_machine.logproba_of_solutions(models, c_quizzes)
+ for l in logproba:
+ s = " ".join([str(x.item()) for x in l])
+ logp_file.write(s + "\n")
+ quizzes_and_logproba_records.append((c_quizzes, logproba))
+
+ nb_validated = valid_c_quizzes(
+ quizzes_and_logproba_records, standard_validity
+ ).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
+
+ new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_records, standard_validity)
+
+ quiz_machine.reverse_random_half_in_place(new_c_quizzes)
+
+ quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
+ quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+
+ # save a bunch of images to investigate what quizzes with a
+ # certain nb of correct predictions look like
- except FileNotFoundError:
- log_string("starting from scratch.")
+ q = new_c_quizzes[:72]
+
+ if q.size(0) > 0:
+ quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
- 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(
- n_epoch=nb_epochs_finished,
- model=model,
- result_dir=args.result_dir,
- logger=log_string,
- deterministic_synthesis=args.deterministic_synthesis,
- input_file=args.expr_input_file,
- )
+models = []
+
+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(device)
+
+ 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)
+
+ models.append(model)
+
- exit(0)
+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", desc="train-entropy"):
- 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)
nb_test, nb_in_train = 0, 0
for test_subset in subsets_as_tuples(
- task.batches(split="test", desc="test-check"), 25000
+ quiz_machine.batches(models[0], split="test", desc="test-check"), 25000
):
in_train = set()
for train_subset in subsets_as_tuples(
- task.batches(split="train", desc="train-check"), 25000
+ 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_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"
-##############################
-
-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}")
-
-##############################
-
-if nb_epochs_finished >= args.nb_epochs:
- task.produce_results(
- n_epoch=nb_epochs_finished,
- model=model,
- result_dir=args.result_dir,
- logger=log_string,
- deterministic_synthesis=args.deterministic_synthesis,
- )
-
-time_pred_result = None
-
######################################################################
+nb_new_c_quizzes_for_train = args.nb_train_samples // 50
+nb_new_c_quizzes_for_test = args.nb_test_samples // 50
-def one_epoch(model, task, learning_rate):
- 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()
-
- nb_train_samples, acc_train_loss = 0, 0.0
-
- for input in task.batches(split="train"):
- input = input.to(device)
-
- 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)
-
- nb_train_samples += input.size(0)
-
- loss.backward()
+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_train_samples % args.batch_size == 0:
- optimizer.step()
+######################################################################
- train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+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
- log_string(f"train)perplexity {n_epoch} {train_perplexity}")
+ def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return l[:, 0] < math.log(0.5)
######################################################################
+for n_epoch in range(args.nb_epochs):
+ log_string(f"--- epoch {n_epoch} ----------------------------------------")
-def run_tests(model, task):
- with torch.autograd.no_grad():
- model.eval()
+ cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
+ log_string(f"current_test_accuracies {cta}")
- nb_test_samples, acc_test_loss = 0, 0.0
- nb_samples_accumulated = 0
+ ##################################################
+ # Select, improve, and eval the worst model
- for input in task.batches(split="test"):
- input = input.to(device)
+ ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
- bs = model(mygpt.BracketedSequence(input))
- output = bs.x
+ weakest_models = ranked_models[: args.nb_gpus]
- loss = F.cross_entropy(output.transpose(1, 2), input)
+ threads = []
- acc_test_loss += loss.item() * input.size(0)
+ for gpu_id, model in enumerate(weakest_models):
+ log_string(f"training model {model.id}")
- nb_test_samples += input.size(0)
-
- task.produce_results(
- n_epoch=n_epoch,
- model=model,
- result_dir=args.result_dir,
- logger=log_string,
- deterministic_synthesis=args.deterministic_synthesis,
+ t = threading.Thread(
+ target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
)
- test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
- log_string(f"test)perplexity {n_epoch} {test_perplexity}")
+ threads.append(t)
+ t.start()
-######################################################################
+ for t in threads:
+ t.join()
-for n_epoch in range(nb_epochs_finished, args.nb_epochs):
- learning_rate = learning_rate_schedule[n_epoch]
+ ##################################################
+ # Replace a fraction of the w_quizzes with fresh ones
- one_epoch(model, task, learning_rate)
-
- run_tests(model, task)
+ log_string(
+ f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
+ )
- time_current_result = datetime.datetime.now()
- if time_pred_result is not None:
- log_string(
- f"next_result {time_current_result + (time_current_result - time_pred_result)}"
- )
- time_pred_result = time_current_result
+ # Renew entirely the train set
- 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
- checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
- torch.save(checkpoint, checkpoint_name)
- log_string(f"saved checkpoint {checkpoint_name}")
+ 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,
+ )
######################################################################