import mygpt
from mygpt import BracketedSequence
+import threading
+
######################################################################
# ar_mask is a tensor with 0s and 1s, of same shape as input, with
self.prompt_len = None
self.answer_len = None
- # self.train_w_quizzes = self.generate_token_sequences(nb_train_samples)
- # self.reverse_random_half_in_place(self.train_w_quizzes)
-
- # self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
- # self.reverse_random_half_in_place(self.test_w_quizzes)
-
+ self.LOCK_C_QUIZZES = threading.Lock()
self.train_c_quizzes = []
self.test_c_quizzes = []
- # if result_dir is not None:
- # self.save_quizzes(
- # result_dir,
- # "culture_w_quizzes",
- # self.train_w_quizzes[:72],
- # )
-
def save_quizzes(
self,
result_dir,
def batches(self, model, split="train", desc=None):
assert split in {"train", "test"}
- if split == "train":
- w_quizzes = model.train_w_quizzes
- c_quizzes = self.train_c_quizzes
- else:
- w_quizzes = model.test_w_quizzes
- c_quizzes = self.test_c_quizzes
- if len(c_quizzes) > 0:
- c_quizzes = torch.cat(c_quizzes, dim=0)
- if c_quizzes.size(0) > w_quizzes.size(0) // 2:
- i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
- c_quizzes = c_quizzes[i]
+ with self.LOCK_C_QUIZZES:
+ if split == "train":
+ w_quizzes = model.train_w_quizzes
+ c_quizzes = self.train_c_quizzes
+ else:
+ w_quizzes = model.test_w_quizzes
+ c_quizzes = self.test_c_quizzes
+
+ if len(c_quizzes) > 0:
+ c_quizzes = torch.cat(c_quizzes, dim=0)
+ if c_quizzes.size(0) > w_quizzes.size(0) // 2:
+ i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
+ c_quizzes = c_quizzes[i]
+
+ i = torch.randperm(w_quizzes.size(0))[
+ : w_quizzes.size(0) - c_quizzes.size(0)
+ ]
+ w_quizzes = w_quizzes[i]
- i = torch.randperm(w_quizzes.size(0))[
- : w_quizzes.size(0) - c_quizzes.size(0)
- ]
- w_quizzes = w_quizzes[i]
+ self.nb_batch_w_quizzes = w_quizzes.size(0)
+ self.nb_batch_c_quizzes = c_quizzes.size(0)
- self.nb_batch_w_quizzes = w_quizzes.size(0)
- self.nb_batch_c_quizzes = c_quizzes.size(0)
-
- input = torch.cat([w_quizzes, c_quizzes], dim=0)
- else:
- input = w_quizzes
- self.nb_batch_w_quizzes = w_quizzes.size(0)
- self.nb_batch_c_quizzes = 0
+ input = torch.cat([w_quizzes, c_quizzes], dim=0)
+ else:
+ input = w_quizzes
+ self.nb_batch_w_quizzes = w_quizzes.size(0)
+ self.nb_batch_c_quizzes = 0
# Shuffle
input = input[torch.randperm(input.size(0))]
######################################################################
def store_c_quizzes(self, new_c_quizzes, for_train=True):
- if for_train:
- self.train_c_quizzes.append(new_c_quizzes)
- else:
- self.test_c_quizzes.append(new_c_quizzes)
+ with self.LOCK_C_QUIZZES:
+ if for_train:
+ self.train_c_quizzes.append(new_c_quizzes)
+ else:
+ self.test_c_quizzes.append(new_c_quizzes)
######################################################################