parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--learning_rate", type=float, default=1e-3)
########################################
default_args = {
"model": "37M",
"batch_size": 100,
- "nb_train_samples": 250000,
+ "nb_train_samples": 100000,
"nb_test_samples": 10000,
}
assert args.nb_test_samples % args.batch_size == 0
quizz_machine = quizz_machine.QuizzMachine(
- sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+ problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
batch_size=args.physical_batch_size,
f"keep c_quizzes {to_keep.size(0)}/{new_c_quizzes.size(0)} ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%) total {sum([ x.size(0) for x in kept])}/{nb_to_generate}"
)
- new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+ new_c_quizzes = torch.cat(kept, dim=0)
+ new_c_quizzes = new_c_quizzes[
+ torch.randperm(new_c_quizzes.size(0))[: nb_for_train + nb_for_test]
+ ]
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)
)
# We keep the first average logits as a reference
- if min_ave_seq_logproba is None:
- min_ave_seq_logproba = ave_seq_logproba
- else:
- log_string(
- f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
- )
+ # if min_ave_seq_logproba is None:
+ # min_ave_seq_logproba = ave_seq_logproba
+ # else:
+ # log_string(
+ # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
+ # )
# We update everyone
for model in models: