# 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, tasks, problems
######################################################################
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
-parser.add_argument(
- "--task",
- type=str,
- default="sandbox",
- help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
-)
+parser.add_argument("--task", type=str, default="world", help="world")
parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
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-4)
-parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
-
-parser.add_argument("--dim_model", type=int, default=512)
-
-parser.add_argument("--dim_keys", type=int, default=64)
-
-parser.add_argument("--dim_hidden", type=int, default=2048)
-
-parser.add_argument("--nb_heads", type=int, default=8)
-
-parser.add_argument("--nb_blocks", type=int, default=12)
-
-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("--sandbox_level", type=int, default=0)
-
-parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
-
-parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
-
-parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
-
-##############################
-# 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("--picocvlr_prune_properties", type=str, default="none")
+########################################
-##############################
-# Maze options
-
-parser.add_argument("--maze_height", type=int, default=23)
-
-parser.add_argument("--maze_width", type=int, default=39)
-
-parser.add_argument("--maze_nb_walls", type=int, default=45)
-
-##############################
-# Snake options
-
-parser.add_argument("--snake_height", type=int, default=6)
-
-parser.add_argument("--snake_width", type=int, default=8)
+parser.add_argument("--model", type=str, default=None)
-parser.add_argument("--snake_nb_colors", type=int, default=5)
+parser.add_argument("--dim_model", type=int, default=None)
-parser.add_argument("--snake_length", type=int, default=200)
+parser.add_argument("--dim_keys", type=int, default=None)
-##############################
-# 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=0.75)
-
-##############################
-# Expr options
+parser.add_argument("--dim_hidden", type=int, default=None)
-parser.add_argument("--expr_nb_variables", type=int, default=5)
+parser.add_argument("--nb_heads", type=int, default=None)
-parser.add_argument("--expr_sequence_length", type=int, default=40)
+parser.add_argument("--nb_blocks", type=int, default=None)
-parser.add_argument("--expr_operand_max", type=int, default=9)
+parser.add_argument("--dropout", type=float, default=0.1)
-parser.add_argument("--expr_result_max", type=int, default=99)
+########################################
-parser.add_argument("--expr_input_file", type=str, default=None)
+parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-##############################
-# World options
+parser.add_argument("--nb_gpts", type=int, default=5)
-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+parser.add_argument("--check", action="store_true", default=False)
######################################################################
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_{args.task}"
######################################################################
-default_args = {
- "sandbox": {
- "nb_epochs": 50,
- "batch_size": 25,
- "nb_train_samples": 100000,
- "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,
+default_task_args = {
+ "world": {
+ "model": "37M",
+ "batch_size": 100,
"nb_train_samples": 250000,
"nb_test_samples": 10000,
},
- "maze": {
- "nb_epochs": 25,
- "batch_size": 5,
- "nb_train_samples": 250000,
- "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)
+
+######################################################################
+
+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.check:
+ args.nb_train_samples = 500
+ 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.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,
+ )
+ 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
-picoclvr_pruner_train = (
- picoclvr_pruner_horizontal_green
- if args.picocvlr_prune_properties in {"train+eval"}
- else None
-)
+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,
+ result_dir=args.result_dir,
+ logger=log_string,
+ device=device,
+ )
+ args.max_percents_of_test_in_train = -1
-picoclvr_pruner_eval = (
- (lambda p: not picoclvr_pruner_horizontal_green(p))
- if args.picocvlr_prune_properties in {"train+eval", "eval"}
- else None
-)
+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,
+ )
-######################################################################
-if args.task == "sandbox":
- if args.sandbox_level == 0:
- problem = tasks.ProblemLevel0(
- nb_sentences=args.sandbox_levels_nb_items,
- len_prompt=args.sandbox_levels_len_source,
- len_result=args.sandbox_levels_len_result,
- )
- elif args.sandbox_level == 1:
- problem = tasks.ProblemLevel1(
- nb_operators=args.sandbox_levels_nb_items,
- len_source=args.sandbox_levels_len_source,
- len_result=args.sandbox_levels_len_result,
- )
- elif args.sandbox_level == 2:
- problem = tasks.ProblemLevel2(
- len_source=args.sandbox_levels_len_source,
- len_result=args.sandbox_levels_len_result,
- )
- else:
- raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
+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,
+ )
+
+elif args.task == "twotargets":
task = tasks.SandBox(
- problem,
- # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
+ problem=problems.ProblemTwoTargets(),
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
+ batch_size=args.physical_batch_size,
+ logger=log_string,
+ device=device,
+ )
+
+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,
)
task = tasks.PicoCLVR(
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
+ batch_size=args.physical_batch_size,
height=args.picoclvr_height,
width=args.picoclvr_width,
nb_colors=args.picoclvr_nb_colors,
task = tasks.MNIST(
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
+ batch_size=args.physical_batch_size,
device=device,
)
task = tasks.Maze(
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
+ batch_size=args.physical_batch_size,
height=args.maze_height,
width=args.maze_width,
nb_walls=args.maze_nb_walls,
- device=device,
+ device="cpu",
)
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,
+ batch_size=args.physical_batch_size,
height=args.snake_height,
width=args.snake_width,
nb_colors=args.snake_nb_colors,
task = tasks.Stack(
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
- batch_size=args.batch_size,
+ batch_size=args.physical_batch_size,
logger=log_string,
nb_steps=args.stack_nb_steps,
nb_stacks=args.stack_nb_stacks,
sequence_length=args.expr_sequence_length,
operand_max=args.expr_operand_max,
result_max=args.expr_result_max,
- batch_size=args.batch_size,
+ batch_size=args.physical_batch_size,
device=device,
)
-elif args.task == "world":
- task = tasks.World(
+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,
+ )
+
+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,
+ )
+
+elif args.task == "qmlp":
+ task = tasks.QMLP(
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,
+ batch_size=args.physical_batch_size,
+ result_dir=args.result_dir,
+ logger=log_string,
+ device=device,
+ )
+
+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,
)
log_string(f"vocabulary_size {vocabulary_size}")
+######################################################################
+
+# 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))
+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(
+ task.batches(split="test", desc="test-check"), 25000
+ ):
+ in_train = set()
+ for train_subset in subsets_as_tuples(
+ task.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"
+ )
+
+ 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"
+
##############################
-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(device)
+def one_epoch(model, task):
+ optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
-nb_parameters = sum(p.numel() for p in model.parameters())
-log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
+ model.train()
-######################################################################
+ nb_train_samples, acc_train_loss = 0, 0.0
-nb_epochs_finished = 0
+ for input in task.batches(split="train"):
+ input = input.to(device)
-if args.no_checkpoint:
- log_string(f"not trying to load checkpoint.")
+ if nb_train_samples % args.batch_size == 0:
+ optimizer.zero_grad()
-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"])
+ output = model(mygpt.BracketedSequence(input)).x
+ loss = F.cross_entropy(output.transpose(1, 2), input)
+ acc_train_loss += loss.item() * input.size(0)
- log_string(f"checkpoint loaded with {nb_epochs_finished} epochs finished.")
+ nb_train_samples += input.size(0)
- except FileNotFoundError:
- log_string("starting from scratch.")
+ loss.backward()
- except:
- log_string("error when loading the checkpoint.")
- exit(1)
+ if nb_train_samples % args.batch_size == 0:
+ optimizer.step()
-######################################################################
+ train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-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,
- )
+ log_string(f"train_perplexity {n_epoch} {train_perplexity}")
- exit(0)
######################################################################
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
-# Compute the entropy of the training tokens
+def run_tests(model, task, deterministic_synthesis):
+ with torch.autograd.no_grad():
+ model.eval()
-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_test_samples, acc_test_loss = 0, 0.0
+ nb_samples_accumulated = 0
-##############################
+ for input in task.batches(split="test"):
+ input = input.to(device)
-# A bit of paranoia never hurts
+ bs = model(mygpt.BracketedSequence(input))
+ output = bs.x
-train_examples = {}
+ loss = F.cross_entropy(output.transpose(1, 2), input)
+ acc_test_loss += loss.item() * input.size(0)
-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
+ nb_test_samples += input.size(0)
-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
+ main_test_accuracy = task.produce_results(
+ n_epoch=n_epoch,
+ model=model,
+ result_dir=args.result_dir,
+ logger=log_string,
+ deterministic_synthesis=deterministic_synthesis,
+ )
-del train_examples
+ test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
-log_string(
- f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
-)
+ log_string(f"test_perplexity {n_epoch} {test_perplexity}")
-##############################
+ model.main_test_accuracy = main_test_accuracy
-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}")
-##############################
+######################################################################
+
+
+def create_quizzes(
+ model,
+ other_models,
+ task,
+ nb_for_train=1000,
+ nb_for_test=100,
+):
+ kept = []
+
+ while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
+ new_quizzes, nb_correct = task.create_new_quizzes(
+ n_epoch=n_epoch,
+ result_dir=args.result_dir,
+ logger=log_string,
+ nb=4 * (nb_for_train + nb_for_test),
+ model=model,
+ other_models=other_models,
+ )
+
+ to_keep = new_quizzes[nb_correct == len(other_models) - 1]
+ log_string(f"keep {to_keep.size(0)} quizzes")
+ kept.append(to_keep)
+
+ new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
-nb_samples_seen = 0
+ task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
+ task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
-if nb_epochs_finished >= nb_epochs:
- task.produce_results(
- nb_epochs_finished,
- model,
+ task.save_image(
+ new_quizzes[:96],
args.result_dir,
+ f"world_quiz_{n_epoch:04d}_{model.id:02d}.png",
log_string,
- args.deterministic_synthesis,
)
-for n_epoch in range(nb_epochs_finished, nb_epochs):
- learning_rate = learning_rate_schedule[n_epoch]
- 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}.")
+models = []
- model.train()
+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)
- nb_train_samples, acc_train_loss = 0, 0.0
+ model.main_test_accuracy = 0.0
+ model.id = k
- 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)
+ models.append(model)
- optimizer.zero_grad()
- loss.backward()
- optimizer.step()
- with torch.autograd.no_grad():
- model.eval()
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
- nb_test_samples, acc_test_loss = 0, 0.0
+######################################################################
- for input in task.batches(split="test"):
- input = input.to(device)
+accuracy_to_make_quizzes = 0.975
+nb_new_quizzes_for_train = 1000
+nb_new_quizzes_for_test = 100
- 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)
+if args.check:
+ accuracy_to_make_quizzes = 0.0
+ nb_new_quizzes_for_train = 10
+ nb_new_quizzes_for_test = 10
- train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
- test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
+for n_epoch in range(args.nb_epochs):
+ # select the model with lowest accuracy
+ models.sort(key=lambda model: model.main_test_accuracy)
+ model = models[0]
- log_string(
- f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
- )
+ log_string(
+ f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+ )
- task.produce_results(
- n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
- )
+ # improve it
+ one_epoch(model, task)
- checkpoint = {
- "nb_epochs_finished": n_epoch + 1,
- "model_state": model.state_dict(),
- "rng_state": torch.get_rng_state(),
- }
+ log_string(
+ f"train_set_composition world {task.nb_batch_samples_world} quizzes {task.nb_batch_samples_quizzes}"
+ )
- if torch.cuda.is_available():
- checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+ # test it
+ run_tests(model, task, deterministic_synthesis=False)
+
+ if model.main_test_accuracy >= accuracy_to_make_quizzes:
+ other_models = models.copy()
+ other_models.remove(model)
+
+ create_quizzes(
+ model,
+ other_models,
+ task,
+ nb_for_train=nb_new_quizzes_for_train,
+ nb_for_test=nb_new_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}")
######################################################################