# Written by Francois Fleuret <francois@fleuret.org>
-import math, sys, argparse, time, tqdm, os, datetime
+import math, sys, argparse, time, tqdm, os, datetime, warnings
import torch, torchvision
from torch import nn
########################################
-parser.add_argument("--nb_epochs", type=int, default=25)
+parser.add_argument("--nb_epochs", type=int, default=50)
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("--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")
-
##############################
# filetask
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_digits", type=int, default=3)
-parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.75)
+parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
##############################
# Expr options
######################################################################
default_task_args = {
+ "world": {
+ "model": "37M",
+ "batch_size": 100,
+ "nb_train_samples": 250000,
+ "nb_test_samples": 10000,
+ },
"file": {
"model": "37M",
"batch_size": 25,
try:
os.mkdir(args.result_dir)
except FileExistsError:
- if not args.overwrite_results:
+ if not args.resume:
print(f"result directory {args.result_dir} already exists")
exit(1)
######################################################################
+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
args.filetask_test_file,
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,
shuffle=True,
device=device,
)
elif args.task == "byheart":
task = tasks.SandBox(
- problem=problems.ProblemByHeart(),
+ 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.batch_size,
+ batch_size=args.physical_batch_size,
+ result_dir=args.result_dir,
logger=log_string,
device=device,
)
problem=problems.ProblemLearnOperator(),
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,
)
problem=problems.ProblemGuessOperator(),
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,
)
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,
)
problem=problems.ProblemMemory(),
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,
)
),
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,
)
problem=problems.ProblemAddition(),
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,
)
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,
)
task = tasks.RPL(
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,
nb_starting_values=args.rpl_nb_starting_values,
max_input=args.rpl_max_input,
prog_len=args.rpl_prog_len,
task = tasks.Grid(
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,
size=args.grid_size,
fraction_play=args.grid_fraction_play,
logger=log_string,
task = tasks.QMLP(
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,
result_dir=args.result_dir,
logger=log_string,
device=device,
task = tasks.Greed(
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.greed_height,
width=args.greed_width,
T=args.greed_T,
##############################
-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,
-)
+models = []
+
+for k in range(2):
+ models.append(
+ 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.to(device)
-nb_parameters = sum(p.numel() for p in model.parameters())
+nb_parameters = sum(p.numel() for p in models[0].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.")
-
-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.")
-
- except FileNotFoundError:
- log_string("starting from scratch.")
-
- 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,
- )
-
- exit(0)
-
-######################################################################
-
# Compute the entropy of the training tokens
token_count = 0
log_string(f"learning_rate_schedule {learning_rate_schedule}")
-##############################
-
-nb_samples_seen = 0
-
-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
-for n_epoch in range(nb_epochs_finished, args.nb_epochs):
- learning_rate = learning_rate_schedule[n_epoch]
+######################################################################
+
+def one_epoch(model, task, learning_rate):
log_string(f"learning_rate {learning_rate}")
if args.optim == "sgd":
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)
- nb_samples_seen += input.size(0)
- optimizer.zero_grad()
loss.backward()
- optimizer.step()
+ 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}")
+
+
+######################################################################
+
+
+def run_tests(model, task, deterministic_synthesis):
with torch.autograd.no_grad():
model.eval()
nb_test_samples, acc_test_loss = 0, 0.0
+ nb_samples_accumulated = 0
for input in task.batches(split="test"):
input = input.to(device)
- output = model(mygpt.BracketedSequence(input)).x
+ 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)
- train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+ main_test_accuracy = task.produce_results(
+ n_epoch=n_epoch,
+ model=model,
+ result_dir=args.result_dir,
+ logger=log_string,
+ deterministic_synthesis=deterministic_synthesis,
+ )
+
test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
- log_string(
- f"perplexity {n_epoch} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}"
- )
+ log_string(f"test_perplexity {n_epoch} {test_perplexity}")
- task.produce_results(
+ return main_test_accuracy
+
+
+######################################################################
+
+
+def create_quizzes(
+ model,
+ other_models,
+ task,
+ nb_for_train=1000,
+ nb_for_test=100,
+ nb_runs=10,
+ nb_min_correct=9,
+ nb_max_correct=9,
+):
+ 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,
- model=model,
result_dir=args.result_dir,
logger=log_string,
- deterministic_synthesis=args.deterministic_synthesis,
+ nb=4 * (nb_for_train + nb_for_test),
+ model=model,
+ other_models=other_models,
+ nb_runs=nb_runs,
)
- 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)}"
+ to_keep = new_quizzes[
+ torch.logical_and(
+ nb_correct >= nb_min_correct, nb_correct <= nb_max_correct
)
- time_pred_result = time_current_result
+ ]
+ log_string(f"keep {to_keep.size(0)} quizzes")
+ kept.append(to_keep)
- checkpoint = {
- "nb_epochs_finished": n_epoch + 1,
- "model_state": model.state_dict(),
- "rng_state": torch.get_rng_state(),
- }
+ new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
- if torch.cuda.is_available():
- checkpoint["cuda_rng_state"] = torch.cuda.get_rng_state()
+ task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
+ task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
+
+ task.save_image(
+ new_quizzes[:96],
+ args.result_dir,
+ f"world_new_{n_epoch:04d}.png",
+ log_string,
+ )
- checkpoint_name = os.path.join(args.result_dir, args.checkpoint_name)
- torch.save(checkpoint, checkpoint_name)
- log_string(f"saved checkpoint {checkpoint_name}")
+
+######################################################################
+
+accuracy_to_make_quizzes = 0.95
+
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
+ learning_rate = learning_rate_schedule[n_epoch]
+
+ for m in models:
+ one_epoch(m, task, learning_rate)
+ test_accuracy = run_tests(m, task, deterministic_synthesis=False)
+
+ if test_accuracy >= accuracy_to_make_quizzes:
+ other_models = models.copy()
+ other_models.remove(model)
+ create_quizzes(other_models, task)
+
+ # --------------------------------------------
+
+ 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
######################################################################
projects / culture.git / blobdiff