projects
/
culture.git
/ commitdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
| commitdiff |
tree
raw
|
patch
|
inline
| side by side (parent:
f2ab5fd
)
Update.
author
François Fleuret
<francois@fleuret.org>
Fri, 12 Jul 2024 10:40:27 +0000
(12:40 +0200)
committer
François Fleuret
<francois@fleuret.org>
Fri, 12 Jul 2024 10:40:27 +0000
(12:40 +0200)
main.py
patch
|
blob
|
history
diff --git
a/main.py
b/main.py
index
fc55b9c
..
63819f2
100755
(executable)
--- a/
main.py
+++ b/
main.py
@@
-24,14
+24,6
@@
import torch.multiprocessing as mp
######################################################################
######################################################################
-if torch.cuda.is_available():
- device = torch.device("cuda")
- torch.backends.cuda.matmul.allow_tf32 = True
-else:
- device = torch.device("cpu")
-
-######################################################################
-
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
@@
-82,7
+74,7
@@
parser.add_argument("--problem", type=str, default="grids")
parser.add_argument("--nb_threads", type=int, default=1)
parser.add_argument("--nb_threads", type=int, default=1)
-parser.add_argument("--
nb_gpus", type=int, default=1
)
+parser.add_argument("--
gpus", type=str, default="all"
)
parser.add_argument("--nb_gpts", type=int, default=5)
parser.add_argument("--nb_gpts", type=int, default=5)
@@
-92,6
+84,10
@@
parser.add_argument("--max_to_validate", type=int, default=None)
parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+parser.add_argument("--proba_understands", type=float, default=0.99)
+
+parser.add_argument("--proba_not_understands", type=float, default=0.5)
+
parser.add_argument("--generation_temperature", type=float, default=2.0)
parser.add_argument("--dirty_debug", action="store_true", default=False)
parser.add_argument("--generation_temperature", type=float, default=2.0)
parser.add_argument("--dirty_debug", action="store_true", default=False)
@@
-234,6
+230,19
@@
for n in vars(args):
######################################################################
######################################################################
+if args.gpus == "all":
+ gpus_idx = range(torch.cuda.device_count())
+else:
+ gpus_idx = [int(k) for k in args.gpus.split(",")]
+
+gpus = [torch.device(f"cuda:{n}") for n in gpus_idx]
+
+if torch.cuda.is_available():
+ main_device = gpus[0]
+else:
+ assert len(gpus) == 0
+ main_device = torch.device("cpu")
+
if args.dirty_debug:
args.nb_train_samples = 2500
args.nb_test_samples = 100
if args.dirty_debug:
args.nb_train_samples = 2500
args.nb_test_samples = 100
@@
-253,14
+262,14
@@
if args.problem == "sky":
nb_birds=args.sky_nb_birds,
nb_iterations=args.sky_nb_iterations,
speed=args.sky_speed,
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,
+ max_nb_cached_chunks=
len(gpus)
* args.nb_train_samples // 100,
chunk_size=100,
nb_threads=args.nb_threads,
)
back_accuracy = False
elif args.problem == "grids":
problem = grids.Grids(
chunk_size=100,
nb_threads=args.nb_threads,
)
back_accuracy = False
elif args.problem == "grids":
problem = grids.Grids(
- max_nb_cached_chunks=
args.nb_gpus
* args.nb_train_samples // 100,
+ max_nb_cached_chunks=
len(gpus)
* args.nb_train_samples // 100,
chunk_size=100,
nb_threads=args.nb_threads,
tasks=args.grids_tasks,
chunk_size=100,
nb_threads=args.nb_threads,
tasks=args.grids_tasks,
@@
-279,12
+288,12
@@
quiz_machine = quiz_machine.QuizMachine(
batch_size=args.physical_batch_size,
result_dir=args.result_dir,
logger=log_string,
batch_size=args.physical_batch_size,
result_dir=args.result_dir,
logger=log_string,
- device=device,
+ device=
main_
device,
)
######################################################################
)
######################################################################
-log_string(f"
device {device
}")
+log_string(f"
main_device {main_device} gpus {[ str(g) for g in gpus]
}")
vocabulary_size = quiz_machine.vocabulary_size()
vocabulary_size = quiz_machine.vocabulary_size()
@@
-293,13
+302,7
@@
log_string(f"vocabulary_size {vocabulary_size}")
######################################################################
######################################################################
-######################################################################
-
-
-def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
- if local_device is None:
- local_device = device
-
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=main_device):
with torch.autograd.no_grad():
model.eval().to(local_device)
with torch.autograd.no_grad():
model.eval().to(local_device)
@@
-330,10
+333,7
@@
def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
)
)
-def one_epoch(model, quiz_machine, local_device=None):
- if local_device is None:
- local_device = device
-
+def one_epoch(model, quiz_machine, local_device=main_device):
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
model.to(local_device).train()
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
model.to(local_device).train()
@@
-369,7
+369,9
@@
def one_epoch(model, quiz_machine, local_device=None):
def standard_validity(logproba):
l = logproba.sort(dim=-1).values
def standard_validity(logproba):
l = logproba.sort(dim=-1).values
- return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
+ return (l[:, 0] < math.log(args.proba_not_understands)) & (
+ l[:, 1] > math.log(args.proba_understands)
+ )
def valid_c_quizzes(recorded, criteria):
def valid_c_quizzes(recorded, criteria):
@@
-459,7
+461,7
@@
for k in range(args.nb_gpts):
nb_blocks=args.nb_blocks,
causal=True,
dropout=args.dropout,
nb_blocks=args.nb_blocks,
causal=True,
dropout=args.dropout,
- ).to(device)
+ ).to(
main_
device)
model.main_test_accuracy = 0.0
model.id = k
model.main_test_accuracy = 0.0
model.id = k
@@
-558,15
+560,15
@@
for n_epoch in range(args.nb_epochs):
ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
- weakest_models = ranked_models[:
args.nb_gpus
]
+ weakest_models = ranked_models[:
len(gpus)
]
threads = []
threads = []
- for gpu
_id, model in enumerate(
weakest_models):
+ for gpu
, model in zip(gpus,
weakest_models):
log_string(f"training model {model.id}")
t = threading.Thread(
log_string(f"training model {model.id}")
t = threading.Thread(
- target=one_epoch, daemon=True, args=(model, quiz_machine,
f"cuda:{gpu_id}"
)
+ target=one_epoch, daemon=True, args=(model, quiz_machine,
gpu
)
)
threads.append(t)
)
threads.append(t)