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.nb_models_for_generation = nb_models_for_generation
- self.mode = mode
-
- def forward(self, bs):
- # If first = 0, we are re-starting an auto-regressive process,
- # that's the right moment to randomize who gonna do it
- if bs.first == 0:
- self.models_to_use = [
- self.models[k]
- for k in torch.randperm(len(self.models))[
- : self.nb_models_for_generation
- ]
- ]
-
- all_the_logits = torch.cat(
- [model(bs).x[None] for model in self.models_to_use], dim=0
- )
-
- if self.mode == "groupthink":
- y = all_the_logits.mean(dim=0)
- elif self.mode == "groupwork":
- m = torch.rand(all_the_logits.size(), device=all_the_logits.device)
- m = (m.sort(dim=0).indices == 0).long()
- y = (y * m).sum(dim=0)
- else:
- raise ValueError(f"Invalid mode {self.mode}")
-
- return BracketedSequence(y, bs.first, bs.nb)
-
-
-######################################################################
-
# ar_mask is a tensor with 0s and 1s, of same shape as input, with
# 1s where tokens should be generated. The others are kept
# unchanged.
seq_logproba,
temperature=1.0,
deterministic_synthesis=False,
- forbidden_tokens=None,
- forced_biases=None,
):
to_generate = (ar_mask.sum(0) > 0).nonzero()
logits = (logits / temperature).log_softmax(dim=-1)
- if forbidden_tokens is not None:
- logits = logits.masked_fill(forbidden_tokens, float("-inf"))
-
- if forced_biases is not None:
- logits = logits + forced_biases[None, :]
-
if deterministic_synthesis:
t_next = logits.argmax(-1)
else:
seq_logproba=seq_logproba,
temperature=temperature,
deterministic_synthesis=deterministic_synthesis,
- forbidden_tokens=forbidden_tokens,
- forced_biases=logit_biases,
)
model.train(t)
class QuizzMachine:
- def make_ar_mask(self, input):
- b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
- return b.long()[None, :].expand_as(input)
+ def indices_forward_and_backward(self, quizzes):
+ i_forward = quizzes[:, 0] == self.token_forward
+ j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
+ i_backward = quizzes[:, 0] == self.token_backward
+ j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
+ assert torch.logical_or(
+ torch.logical_and(i_forward, j_forward),
+ torch.logical_and(i_backward, j_backward),
+ ).all()
+ return i_forward, i_backward
+
+ def reverse_time(self, quizzes):
+ i_forward, i_backward = self.indices_forward_and_backward(quizzes)
+
+ forward_to_backward = torch.cat(
+ [
+ quizzes[:, 0:1],
+ quizzes[:, 2 + self.prompt_len :],
+ quizzes[:, 1 + self.prompt_len : 2 + self.prompt_len],
+ quizzes[:, 1 : 1 + self.prompt_len],
+ ],
+ dim=1,
+ )
+ forward_to_backward[:, 0] = self.token_backward
+ forward_to_backward[:, 1 + self.answer_len] = self.token_backward
+
+ backward_to_forward = torch.cat(
+ [
+ quizzes[:, 0:1],
+ quizzes[:, 2 + self.answer_len :],
+ quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
+ quizzes[:, 1 : 1 + self.answer_len],
+ ],
+ dim=1,
+ )
+
+ backward_to_forward[:, 0] = self.token_forward
+ backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
+
+ m = i_forward.long()[:, None]
+
+ return m * forward_to_backward + (1 - m) * backward_to_forward
+
+ def make_ar_mask(self, quizzes, first=False):
+ i_forward, i_backward = self.indices_forward_and_backward(quizzes)
+
+ t = torch.arange(quizzes.size(1), device=quizzes.device)
+
+ if first:
+ m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
+ m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
+ else:
+ m_forward = (t >= 2 + self.prompt_len).long()
+ m_backward = (t >= 2 + self.answer_len).long()
+
+ m = i_forward.long()[:, None]
+
+ return m * m_forward + (1 - m) * m_backward
+
+ def generate_token_sequences(self, nb):
+ prompts, answers = self.problem.generate_prompts_and_answers(nb)
+
+ if self.prompt_len is None:
+ self.prompt_len = prompts.size(1)
+
+ if self.answer_len is None:
+ self.answer_len = answers.size(1)
+
+ assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
+
+ result = []
+
+ for prompt, answer in zip(prompts, answers):
+ if torch.rand(1) < 0.5:
+ a = [
+ torch.tensor([self.token_forward]),
+ prompt,
+ torch.tensor([self.token_forward]),
+ answer,
+ ]
+ else:
+ a = [
+ torch.tensor([self.token_backward]),
+ answer,
+ torch.tensor([self.token_backward]),
+ prompt,
+ ]
+
+ result.append(torch.cat(a, dim=0)[None, :])
+
+ return torch.cat(result, dim=0)
def __init__(
self,
):
super().__init__()
+ v = problem.nb_token_values()
+ self.token_forward = v
+ self.token_backward = v + 1
+ self.nb_token_values = v + 2
+
self.problem = problem
self.batch_size = batch_size
self.device = device
self.logger = logger
+ self.prompt_len = None
+ self.answer_len = None
- self.train_w_quizzes = self.problem.generate_token_sequences(
- nb_train_samples
- ).to(device)
- self.test_w_quizzes = self.problem.generate_token_sequences(nb_test_samples).to(
+ self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
device
)
- self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
+ self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
self.train_c_quizzes = []
self.test_c_quizzes = []
if result_dir is not None:
- self.problem.save_quizzes(
- self.train_w_quizzes[:72], result_dir, "culture_w_quizzes"
+ self.save_quizzes(
+ result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
)
+ # toto = self.reverse_time(self.train_w_quizzes[:72])
+ # self.save_quizzes(result_dir, "toto", toto)
+ # exit(0)
+
+ def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
+ forward = quizzes[quizzes[:, 0] == self.token_forward]
+ ib = quizzes[:, 0] == self.token_backward
+ backward = quizzes[ib]
+ assert forward.size(0) + backward.size(0) == quizzes.size(0)
+ quizzes[ib] = self.reverse_time(quizzes[ib])
+
+ if prediction:
+ predicted_prompts = ib
+ predicted_answers = torch.logical_not(ib)
+ else:
+ predicted_prompts = None
+ predicted_answers = None
+
+ self.problem.save_quizzes(
+ result_dir,
+ filename_prefix,
+ quizzes[:, 1 : 1 + self.prompt_len],
+ quizzes[:, 2 + self.prompt_len :],
+ predicted_prompts,
+ predicted_answers,
+ )
+
def batches(self, split="train", desc=None):
assert split in {"train", "test"}
if split == "train":
yield batch
def vocabulary_size(self):
- return self.nb_codes
+ return self.nb_token_values
def produce_results(
self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
device=self.device,
)
- nb_total, nb_correct = (
- input.size(0),
- (input == result).long().min(dim=1).values.sum(),
- )
+ nb_total = input.size(0)
+ nb_correct = (input == result).long().min(dim=1).values.sum()
return nb_total, nb_correct
device=self.device,
)
- self.problem.save_quizzes(
- result[:72], result_dir, f"culture_prediction_{n_epoch:04d}_{model.id:02d}"
+ self.save_quizzes(
+ result_dir,
+ f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
+ quizzes=result[:72],
+ prediction=True,
)
return main_test_accuracy
input = self.train_w_quizzes if for_train else self.test_w_quizzes
nb = min(nb, input.size(0))
input[:-nb] = input[nb:].clone()
- input[-nb:] = self.problem.generate_token_sequences(nb).to(self.device)
+ input[-nb:] = self.generate_token_sequences(nb).to(self.device)
def store_c_quizzes(self, new_c_quizzes, for_train=True):
if for_train:
else:
self.test_c_quizzes.append(new_c_quizzes)
- def comput_correctness(self, c_quizzes, models_for_validation):
- # Create the reverse quizzes
-
- token_forward, token_backward = self.problem.direction_tokens()
+ def compute_correctness(
+ self,
+ c_quizzes,
+ models_for_validation,
+ bidirectional_validation=False,
+ deterministic_validation=True,
+ ):
+ if bidirectional_validation:
+ backward_c_quizzes = self.forward_to_backward(c_quizzes)
- l = (c_quizzes.size(1) - 1) // 2
- direction = c_quizzes[:, l : l + 1]
- 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
+ seq_logproba = torch.zeros(
+ c_quizzes.size(0),
+ max([m.id for m in models_for_validation]) + 1,
+ device=self.device,
)
- ar_mask = self.make_ar_mask(c_quizzes)
- seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
-
- # Check how many of models can solve the quizzes in both directions
-
nb_correct = 0
for model in models_for_validation:
result = c_quizzes.clone()
+ seq_logproba[...] = 0.0
+
+ ar_mask = self.make_ar_mask(result)
+
masked_inplace_autoregression(
model=model,
batch_size=self.batch_size,
input=result,
ar_mask=ar_mask,
- seq_logproba=seq_logproba,
+ seq_logproba=seq_logproba[:, model.id],
temperature=1.0,
- deterministic_synthesis=True,
+ deterministic_synthesis=deterministic_validation,
# progress_bar_desc="solving c_quizzes",
device=self.device,
)
correct = (c_quizzes == result).long().min(dim=-1).values
- reverse_result = reverse_c_quizzes.clone()
+ if bidirectional_validation:
+ backward_result = backward_c_quizzes.clone()
- masked_inplace_autoregression(
- model=model,
- batch_size=self.batch_size,
- input=reverse_result,
- ar_mask=ar_mask,
- seq_logproba=seq_logproba,
- temperature=1.0,
- deterministic_synthesis=True,
- # progress_bar_desc="solving reversed c_quizzes",
- device=self.device,
- )
+ ar_mask = self.make_ar_mask(backward_result)
- reverse_correct = (
- (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
- )
+ masked_inplace_autoregression(
+ model=model,
+ batch_size=self.batch_size,
+ input=backward_result,
+ ar_mask=ar_mask,
+ seq_logproba=seq_logproba[:, model.id],
+ temperature=1.0,
+ deterministic_synthesis=deterministic_validation,
+ # progress_bar_desc="solving backward c_quizzes",
+ device=self.device,
+ )
+
+ backward_correct = (
+ (backward_c_quizzes == backward_result).long().min(dim=-1).values
+ )
+
+ correct *= backward_correct
+
+ # endif
- nb_correct += correct * reverse_correct
+ nb_correct += correct
- return nb_correct
+ return nb_correct, seq_logproba
###############################################################
- def generate_quizzes(self, nb, model_for_generation, min_ave_seq_logproba):
+ def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
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_solve = 1 - ar_mask_prompt
- seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
+ seq_logproba = torch.zeros(nb, device=self.device)
- # bracketing of the temperature to get the target logproba
+ # First, we generate the answer at high temperature
- warnings.warn("high temperature!", RuntimeWarning)
- temperature = 2
- d_temperature = 1 / 3
+ c_quizzes[:, 0] = self.token_backward
+ c_quizzes[:, 1 + self.answer_len] = self.token_backward
- while True:
- seq_logproba[...] = 0
-
- 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()
-
- 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 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()
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=self.make_ar_mask(c_quizzes, first=True),
+ seq_logproba=seq_logproba,
+ temperature=temperature,
+ deterministic_synthesis=False,
+ device=self.device,
+ )
- ######################################################################
+ # Then, we generate the prompt at low temperature
- def create_c_quizzes(
- self,
- nb,
- model_for_generation,
- models_for_validation,
- min_ave_seq_logproba,
- n_epoch,
- result_dir,
- ):
- c_quizzes, ave_seq_logproba = self.generate_quizzes(
- nb, model_for_generation, min_ave_seq_logproba
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=self.make_ar_mask(c_quizzes),
+ seq_logproba=seq_logproba,
+ temperature=1 / temperature,
+ deterministic_synthesis=False,
+ device=self.device,
)
- nb_correct = self.comput_correctness(c_quizzes, models_for_validation)
+ # Then we return the quizz, and re-generate the response, now
+ # at low temperature
- return c_quizzes, nb_correct, ave_seq_logproba
+ c_quizzes = self.reverse_time(c_quizzes)
- ######################################################################
-
- def gang_create_c_quizzes(
- self,
- nb,
- nb_models_for_generation,
- models,
- mode,
- min_ave_seq_logproba,
- 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,
+ masked_inplace_autoregression(
+ model=model_for_generation,
+ batch_size=self.batch_size,
+ input=c_quizzes,
+ ar_mask=self.make_ar_mask(c_quizzes),
+ seq_logproba=seq_logproba,
+ temperature=1 / temperature,
+ deterministic_synthesis=False,
+ device=self.device,
)
+
+ return c_quizzes