from torch import nn
from torch.nn import functional as F
+import mygpt
from mygpt import BracketedSequence
######################################################################
class Gang(nn.Module):
def __init__(self, models, nb_models_for_generation, mode="groupthink"):
super().__init__()
- self.models = models
+ self.models = nn.ModuleList(models)
self.nb_models_for_generation = nb_models_for_generation
self.mode = mode
else:
self.test_c_quizzes.append(new_c_quizzes)
- def comput_correctness(self, c_quizzes, models_for_validation):
- # Create the reverse quizzes
-
+ def reverse_time(self, c_quizzes):
token_forward, token_backward = self.problem.direction_tokens()
l = (c_quizzes.size(1) - 1) // 2
direction = self.problem.token_forward * (
direction == self.problem.token_backward
) + self.problem.token_backward * (direction == self.problem.token_forward)
- reverse_c_quizzes = torch.cat(
- [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
- )
+
+ return torch.cat([c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1)
+
+ def comput_correctness(self, c_quizzes, models_for_validation):
+ reversed_c_quizzes = self.reverse_time(c_quizzes)
ar_mask = self.make_ar_mask(c_quizzes)
seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
correct = (c_quizzes == result).long().min(dim=-1).values
- reverse_result = reverse_c_quizzes.clone()
+ reversed_result = reversed_c_quizzes.clone()
masked_inplace_autoregression(
model=model,
batch_size=self.batch_size,
- input=reverse_result,
+ input=reversed_result,
ar_mask=ar_mask,
seq_logproba=seq_logproba,
temperature=1.0,
device=self.device,
)
- reverse_correct = (
- (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
+ reversed_correct = (
+ (reversed_c_quizzes == reversed_result).long().min(dim=-1).values
)
- nb_correct += correct * reverse_correct
+ nb_correct += correct * reversed_correct
return nb_correct
###############################################################
- def generate_quizzes(self, nb, model_for_generation, min_ave_seq_logproba):
+ def generate_quizzes(
+ self, nb, model_for_generation, min_ave_seq_logproba, reverse_cleanup=False
+ ):
c_quizzes = torch.empty(
nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
)
ar_mask_prompt = torch.zeros(c_quizzes.size(), device=self.device)
- ar_mask_prompt[:, ar_mask_prompt.size(1) // 2 + 1] = 1
+ ar_mask_prompt[:, : ar_mask_prompt.size(1) // 2 + 1] = 1
ar_mask_solve = 1 - ar_mask_prompt
- seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
+ seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
- # bracketing of the temperature to get the target logproba
+ if reverse_cleanup:
+ warnings.warn("very high temperature with reversed cleanup", RuntimeWarning)
+ temperature = 10.0
+ else:
+ temperature = 1.0
- temperature = 1
- d_temperature = 1 / 3
+ # warnings.warn("noise injection", RuntimeWarning)
+ # noise_std = torch.rand(1).item()
+ # self.logger(f"{noise_std=}")
- while True:
- seq_logproba[...] = 0
+ # mygpt.set_noise_injection(model_for_generation, noise_std)
- masked_inplace_autoregression(
- model=model_for_generation,
- batch_size=self.batch_size,
- input=c_quizzes,
- ar_mask=ar_mask_prompt,
- seq_logproba=seq_logproba,
- temperature=temperature,
- deterministic_synthesis=False,
- # progress_bar_desc="sampling c_quizzes",
- device=self.device,
- )
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=ar_mask_prompt,
+ seq_logproba=seq_logproba,
+ temperature=temperature,
+ deterministic_synthesis=False,
+ # progress_bar_desc="sampling c_quizzes",
+ device=self.device,
+ )
- ave_seq_logproba = seq_logproba.mean()
+ # mygpt.set_noise_injection(model_for_generation, 0.0)
+ ave_seq_logproba = seq_logproba.mean()
+
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=ar_mask_solve,
+ seq_logproba=seq_logproba,
+ temperature=temperature,
+ deterministic_synthesis=True,
+ # progress_bar_desc="sampling c_quizzes",
+ device=self.device,
+ )
+
+ if reverse_cleanup:
+ c_quizzes = self.reverse_time(c_quizzes)
masked_inplace_autoregression(
model=model_for_generation,
batch_size=self.batch_size,
device=self.device,
)
- # If we do not have target logprobs, get out now
- if min_ave_seq_logproba is None:
- break
-
- # Oh man that's ugly
- if ave_seq_logproba < min_ave_seq_logproba:
- if d_temperature > 0:
- d_temperature *= -1 / 3
- temperature += d_temperature
- elif ave_seq_logproba > min_ave_seq_logproba * 0.99:
- if d_temperature < 0:
- d_temperature *= -1 / 3
- temperature += d_temperature
- else:
- break
-
- self.logger(f"changing temperature to {temperature}")
-
return c_quizzes, seq_logproba.mean()
######################################################################
model_for_generation,
models_for_validation,
min_ave_seq_logproba,
+ reverse_cleanup,
n_epoch,
result_dir,
):
c_quizzes, ave_seq_logproba = self.generate_quizzes(
- nb, model_for_generation, min_ave_seq_logproba
+ nb,
+ model_for_generation=model_for_generation,
+ min_ave_seq_logproba=min_ave_seq_logproba,
+ reverse_cleanup=reverse_cleanup,
)
nb_correct = self.comput_correctness(c_quizzes, models_for_validation)
models,
mode,
min_ave_seq_logproba,
+ reverse_cleanup,
n_epoch,
result_dir,
):
model_for_generation = Gang(models, nb_models_for_generation, mode)
models_for_validation = models
return self.create_c_quizzes(
- nb,
- model_for_generation,
- models_for_validation,
- min_ave_seq_logproba,
- n_epoch,
- result_dir,
+ nb=nb,
+ model_for_generation=model_for_generation,
+ models_for_validation=models_for_validation,
+ min_ave_seq_logproba=min_ave_seq_logproba,
+ reverse_cleanup=reverse_cleanup,
+ n_epoch=n_epoch,
+ result_dir=result_dir,
)