from torch import nn
from torch.nn import functional as F
+import mygpt
from mygpt import BracketedSequence
######################################################################
+# 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.
+
+
+def one_batch_masked_inplace_autoregression(
+ model,
+ input,
+ ar_mask,
+ seq_logproba,
+ temperature=1.0,
+ deterministic_synthesis=False,
+ forbidden_tokens=None,
+ forced_biases=None,
+):
+ to_generate = (ar_mask.sum(0) > 0).nonzero()
+
+ if to_generate.min() > 0:
+ model(
+ BracketedSequence(input, 0, to_generate.min())
+ ) # Needed to initialize the model's cache
+ for s in range(to_generate.min(), to_generate.max() + 1):
+ output = model(BracketedSequence(input, s, 1)).x
+
+ logits = output[:, s]
+
+ 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:
+ dist = torch.distributions.categorical.Categorical(logits=logits)
+ t_next = dist.sample()
+
+ all_n = torch.arange(t_next.size(0))
+ seq_logproba += logits[all_n, t_next].sum(dim=-1)
+
+ input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
+
def masked_inplace_autoregression(
model,
model.eval()
for input, ar_mask, seq_logproba in batches:
- model.masked_inplace_autoregression(
+ one_batch_masked_inplace_autoregression(
+ model=model,
input=input,
ar_mask=ar_mask,
seq_logproba=seq_logproba,
nb_train_samples,
nb_test_samples,
batch_size,
- result_dir=None,
- logger=None,
+ result_dir,
+ logger,
device=torch.device("cpu"),
):
super().__init__()
self.problem = problem
self.batch_size = batch_size
self.device = device
+ self.logger = logger
- self.train_w_quizzes = self.problem.generate_seq(nb_train_samples).to(device)
- self.test_w_quizzes = self.problem.generate_seq(nb_test_samples).to(device)
+ 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(
+ device
+ )
self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
return self.nb_codes
def produce_results(
- self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
+ self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
):
- def compute_accuracy(input, logger=None):
+ def compute_accuracy(input):
input = input[:nmax]
ar_mask = self.make_ar_mask(input)
result = input.clone() * (1 - ar_mask)
train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
- logger(
+ self.logger(
f"accuracy_train {n_epoch} nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
)
- test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
+ test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
- logger(
+ self.logger(
f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
)
main_test_accuracy = test_nb_correct / test_nb_total
- logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
+ self.logger(f"main_test_accuracy {n_epoch} {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_seq(nb).to(self.device)
+ input[-nb:] = self.problem.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 create_c_quizzes(
- self,
- nb,
- model_for_generation,
- models_for_validation,
- min_ave_seq_logproba,
- n_epoch,
- result_dir,
- logger,
- ):
- ###############################################################
- # Generate quizzes with model
-
- c_quizzes = torch.empty(
- nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
- )
-
- ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
- seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
-
- temperature = 1
- d_temperature = 1 / 3
-
- while True:
- seq_logproba[...] = 0
-
- masked_inplace_autoregression(
- model=model_for_generation,
- batch_size=self.batch_size,
- input=c_quizzes,
- ar_mask=ar_mask,
- seq_logproba=seq_logproba,
- temperature=temperature,
- deterministic_synthesis=False,
- # progress_bar_desc="sampling c_quizzes",
- device=self.device,
- )
-
- ave_seq_logproba = seq_logproba.mean()
-
- 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
-
- logger(f"changing temperature to {temperature}")
-
- ###############################################################
- # 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)
- ###############################################################
- # Check how many of the other models can solve them in both
- # directions
+ # Check how many of models can solve the quizzes in both directions
- nb_correct = []
+ nb_correct = 0
for model in models_for_validation:
result = c_quizzes.clone()
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.append((correct * reverse_correct)[None, :])
+ nb_correct += correct * reversed_correct
+
+ return nb_correct
+
+ ###############################################################
- nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
+ def generate_quizzes(self, nb, model_for_generation, 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_solve = 1 - ar_mask_prompt
+ seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
+
+ if reverse_cleanup:
+ warnings.warn("very high temperature with reversed cleanup", RuntimeWarning)
+ temperature = 10.0
+ else:
+ temperature = 1.0
+
+ # warnings.warn("noise injection", RuntimeWarning)
+ # noise_std = torch.rand(1).item()
+ # self.logger(f"{noise_std=}")
+
+ # 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,
+ device=self.device,
+ )
+
+ # 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,
+ 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,
+ input=c_quizzes,
+ ar_mask=ar_mask_solve,
+ seq_logproba=seq_logproba,
+ temperature=temperature,
+ deterministic_synthesis=True,
+ device=self.device,
+ )
- return c_quizzes, nb_correct, seq_logproba.mean()
+ return c_quizzes, seq_logproba.mean()