from torch.nn import functional as F
import ffutils
-import mygpt, tasks
+import mygpt, tasks, problems
######################################################################
"--task",
type=str,
default="sandbox",
- help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
+ help="byheart, learnop, guessop, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl",
)
parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
parser.add_argument("--seed", type=int, default=0)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+
+########################################
+
parser.add_argument("--nb_epochs", type=int, default=None)
parser.add_argument("--batch_size", type=int, default=None)
parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
+########################################
+
parser.add_argument("--model", type=str, default="37M")
parser.add_argument("--dim_model", type=int, default=None)
parser.add_argument("--dropout", type=float, default=0.1)
+########################################
+
parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
parser.add_argument("--no_checkpoint", action="store_true", default=False)
parser.add_argument("--rpl_no_prog", action="store_true", default=False)
-##############################
-# sandbox options
-
-parser.add_argument("--sandbox_level", type=int, default=0)
-
-parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
-
-parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
-
-parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
-
##############################
# picoclvr options
######################################################################
-if args.task == "sandbox":
- if args.sandbox_level == 0:
- problem = tasks.ProblemLevel0(
- nb_sentences=args.sandbox_levels_nb_items,
- len_prompt=args.sandbox_levels_len_source,
- len_result=args.sandbox_levels_len_result,
- )
- elif args.sandbox_level == 1:
- problem = tasks.ProblemLevel1(
- nb_operators=args.sandbox_levels_nb_items,
- len_source=args.sandbox_levels_len_source,
- len_result=args.sandbox_levels_len_result,
- )
- elif args.sandbox_level == 2:
- problem = tasks.ProblemLevel2(
- len_source=args.sandbox_levels_len_source,
- len_result=args.sandbox_levels_len_result,
- )
- else:
- raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
+if args.task == "byheart":
+ task = tasks.SandBox(
+ problem=problems.ProblemByHeart(),
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ logger=log_string,
+ device=device,
+ )
+
+
+elif args.task == "learnop":
+ task = tasks.SandBox(
+ problem=problems.ProblemLearnOperator(),
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ logger=log_string,
+ device=device,
+ )
+
+
+elif args.task == "guessop":
+ task = tasks.SandBox(
+ problem=problems.ProblemGuessOperator(),
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ logger=log_string,
+ device=device,
+ )
+
+elif args.task == "twotargets":
task = tasks.SandBox(
- problem,
- # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
+ problem=problems.ProblemTwoTargets(),
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ logger=log_string,
+ device=device,
+ )
+
+elif args.task == "addition":
+ task = tasks.SandBox(
+ problem=problems.ProblemAddition(),
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
batch_size=args.batch_size,
entropy = -torch.xlogy(token_probas, token_probas).sum()
train_set_perplexity = math.exp(entropy)
-##############################
-
+######################################################################
# A bit of paranoia never hurts
-train_examples = {}
+def subsets_as_tuples(batches, cs):
+ s = set()
+ for batch in batches:
+ for x in batch:
+ s.add(tuple([v.item() for v in x]))
+ if len(s) == cs:
+ yield s
+ s = set()
+ yield s
-for input in task.batches(split="train"):
- assert input.dim() == 2 and input.dtype == torch.int64
- for x in input:
- train_examples[x.sum().item()] = x
-
-nb_total, nb_collisions = 0, 0
-for input in task.batches(split="test"):
- assert input.dim() == 2 and input.dtype == torch.int64
- for x in input:
- nb_total += 1
- y = train_examples.get(x.sum().item())
- if y is not None:
- if x.size() == y.size() and (x - y).abs().sum() == 0:
- nb_collisions += 1
-
-del train_examples
+
+nb_test, nb_in_train = 0, 0
+for test_subset in subsets_as_tuples(task.batches(split="test"), 25000):
+ in_train = set()
+ for train_subset in subsets_as_tuples(task.batches(split="train"), 25000):
+ in_train.update(test_subset.intersection(train_subset))
+ nb_in_train += len(in_train)
+ nb_test += len(test_subset)
log_string(
- f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
+ f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
)
+assert (
+ nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
+), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
+
##############################
if args.learning_rate_schedule == "cos":
projects / picoclvr.git / blobdiff