):
break
+ def compute_distance(self, walls, goal_i, goal_j, start_i, start_j):
+ max_length = walls.numel()
+ dist = torch.full_like(walls, max_length)
+
+ dist[goal_i, goal_j] = 0
+ pred_dist = torch.empty_like(dist)
+
+ while True:
+ pred_dist.copy_(dist)
+ d = (
+ torch.cat(
+ (
+ dist[None, 1:-1, 0:-2],
+ dist[None, 2:, 1:-1],
+ dist[None, 1:-1, 2:],
+ dist[None, 0:-2, 1:-1],
+ ),
+ 0,
+ ).min(dim=0)[0]
+ + 1
+ )
+
+ dist[1:-1, 1:-1].minimum_(d) # = torch.min(dist[1:-1, 1:-1], d)
+ dist = walls * max_length + (1 - walls) * dist
+
+ if dist[start_i, start_j] < max_length or dist.equal(pred_dist):
+ return dist * (1 - walls)
+
# @torch.compile
- def task_islands(self, A, f_A, B, f_B):
- pass
+ def task_path(self, A, f_A, B, f_B):
+ c = torch.randperm(len(self.colors) - 1)[:3] + 1
+ dist = torch.empty(self.height + 2, self.width + 2)
+ for X, f_X in [(A, f_A), (B, f_B)]:
+ nb_rec = torch.randint(3, (1,)) + 1
+ while True:
+ r = self.rec_coo(nb_rec, prevent_overlap=True)
+ X[...] = 0
+ f_X[...] = 0
+ for n in range(nb_rec):
+ i1, j1, i2, j2 = r[n]
+ X[i1:i2, j1:j2] = c[0]
+ f_X[i1:i2, j1:j2] = c[0]
+ while True:
+ i0, j0 = torch.randint(self.height, (1,)), torch.randint(
+ self.width, (1,)
+ )
+ if X[i0, j0] == 0:
+ break
+ while True:
+ i1, j1 = torch.randint(self.height, (1,)), torch.randint(
+ self.width, (1,)
+ )
+ if X[i1, j1] == 0:
+ break
+ dist[...] = 1
+ dist[1:-1, 1:-1] = (X != 0).long()
+ dist[...] = self.compute_distance(dist, i1 + 1, j1 + 1, i0 + 1, j0 + 1)
+ if dist[i0 + 1, j0 + 1] >= 1 and dist[i0 + 1, j0 + 1] < self.height * 4:
+ break
+
+ dist[1:-1, 1:-1] += (X != 0).long() * self.height * self.width
+ dist[0, :] = self.height * self.width
+ dist[-1, :] = self.height * self.width
+ dist[:, 0] = self.height * self.width
+ dist[:, -1] = self.height * self.width
+ # dist += torch.rand(dist.size())
+
+ i, j = i0 + 1, j0 + 1
+ while i != i1 + 1 or j != j1 + 1:
+ f_X[i - 1, j - 1] = c[2]
+ r, s, t, u = (
+ dist[i - 1, j],
+ dist[i, j - 1],
+ dist[i + 1, j],
+ dist[i, j + 1],
+ )
+ m = min(r, s, t, u)
+ if r == m:
+ i = i - 1
+ elif t == m:
+ i = i + 1
+ elif s == m:
+ j = j - 1
+ else:
+ j = j + 1
+
+ X[i0, j0] = c[2]
+ # f_X[i0, j0] = c[1]
+
+ X[i1, j1] = c[1]
+ f_X[i1, j1] = c[1]
# for X, f_X in [(A, f_A), (B, f_B)]:
# n = torch.arange(self.height * self.width).reshape(self.height, self.width)
self.task_scale,
self.task_symbols,
self.task_ortho,
- # self.task_islands,
+ # self.task_path,
]
def trivial_prompts_and_answers(self, prompts, answers):
# exit(0)
# if True:
- # nb,nrow = 72,4
- nb, nrow = 8, 2
+ nb, nrow = 72, 4
+ # nb, nrow = 8, 2
- for t in grids.all_tasks():
- # for t in [grids.task_replace_color]:
+ # for t in grids.all_tasks():
+ for t in [grids.task_path]:
print(t.__name__)
prompts, answers = grids.generate_prompts_and_answers_(nb, tasks=[t])
grids.save_quizzes("/tmp", t.__name__, prompts[:nb], answers[:nb], nrow=nrow)
+ # exit(0)
+
nb = 1000
for t in grids.all_tasks():
parser.add_argument("--max_to_validate", type=int, default=None)
-parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.9)
parser.add_argument("--generation_temperature", type=float, default=2.0)
-parser.add_argument("--deterministic_validation", action="store_true", default=False)
-
-parser.add_argument("--bidirectional_validation", action="store_true", default=False)
-
parser.add_argument("--dirty_debug", action="store_true", default=False)
######################################################################
quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q)
-######################################################################
-
-
-def create_c_quizzes_(
- models,
- quiz_machine,
- nb_for_train=1000,
- nb_for_test=100,
-):
- quizzes_and_nb_correct_records = []
-
- nb_to_create = nb_for_train + nb_for_test
-
- # ------------------------------------------------------------
-
- standard_validity = lambda nb_correct: (nb_correct >= args.min_to_validate) & (
- nb_correct <= args.max_to_validate
- )
-
- file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat")
-
- with open(file_name, "w") as logp_file:
- while (
- valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0)
- < nb_to_create
- ):
- # Select a model at random to generate the new quizzes
-
- model_for_generation = models[torch.randint(len(models), (1,))]
-
- c_quizzes = quiz_machine.generate_quizzes(
- nb_to_create,
- model_for_generation=model_for_generation,
- temperature=args.generation_temperature,
- )
-
- # if args.prediction_correctness:
-
- # else:
- # logproba = quiz_machine.new(quiz_machine.size(0), len(models))
- # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)):
- # for model in models:
- # l[...] = F.cross_entropy(model(q))
-
- c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)]
-
- if c_quizzes.size(0) > 0:
- nb_correct, seq_logproba = quiz_machine.compute_correctness(
- c_quizzes,
- models,
- bidirectional_validation=args.bidirectional_validation,
- deterministic_validation=args.deterministic_validation,
- )
-
- for n, l in zip(nb_correct, seq_logproba):
- s = " ".join([str(x.item()) for x in l])
- logp_file.write(f"{n} {s}\n")
-
- if args.dirty_debug:
- nb_correct = torch.randint(
- len(models) + 1, nb_correct.size(), device=c_quizzes.device
- )
-
- quizzes_and_nb_correct_records.append((c_quizzes, nb_correct))
-
- nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0)
- nv = " ".join([str(x.item()) for x in nv])
-
- nb_validated = valid_c_quizzes(
- quizzes_and_nb_correct_records, standard_validity
- ).size(0)
-
- log_string(
- f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
- )
-
- # store the new c_quizzes which have been validated
-
- new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity)
-
- quiz_machine.reverse_random_half_in_place(new_c_quizzes)
-
- quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
- quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
-
- # save a bunch of images to investigate what quizzes with a
- # certain nb of correct predictions look like
-
- for n in range(len(models) + 1):
- s = (
- "_validated"
- if n >= args.min_to_validate and n <= args.max_to_validate
- else ""
- )
-
- q = valid_c_quizzes(
- quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n
- )[:72]
-
- quiz_machine.reverse_random_half_in_place(q)
-
- if q.size(0) > 0:
- quiz_machine.save_quizzes(
- args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q
- )
-
-
######################################################################
models = []
nb_new_c_quizzes_for_train = 100
nb_new_c_quizzes_for_test = 10
+ def standard_validity(logproba):
+ l = logproba.sort(dim=-1).values
+ return l[:, 0] < math.log(0.99)
+
+
######################################################################
for n_epoch in range(args.nb_epochs):
log_string(f"current_test_accuracies {cta}")
##################################################
- # Select, improve, and eval the worst model
+ # Select, improve, and eval the worst models
ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
model.TRAINING_LOCK.release()
##################################################
- # Replace a fraction of the w_quizzes with fresh ones
+ # Renew the train sets
log_string(
f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
)
- # Renew entirely the train set
-
for model in weakest_models:
quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
projects / culture.git / commitdiff