Update.
[culture.git] / main.py
diff --git a/main.py b/main.py
index ebecad8..524715a 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -12,7 +12,8 @@ from torch import nn
 from torch.nn import functional as F
 
 import ffutils
 from torch.nn import functional as F
 
 import ffutils
-import mygpt, tasks
+import mygpt
+import sky, quizz_machine
 
 # world quizzes vs. culture quizzes
 
 
 # world quizzes vs. culture quizzes
 
@@ -209,7 +210,8 @@ else:
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
 assert args.nb_train_samples % args.batch_size == 0
 assert args.nb_test_samples % args.batch_size == 0
 
-task = tasks.World(
+quizz_machine = quizz_machine.QuizzMachine(
+    sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     batch_size=args.physical_batch_size,
     nb_train_samples=args.nb_train_samples,
     nb_test_samples=args.nb_test_samples,
     batch_size=args.physical_batch_size,
@@ -222,7 +224,7 @@ task = tasks.World(
 
 log_string(f"device {device}")
 
 
 log_string(f"device {device}")
 
-vocabulary_size = task.vocabulary_size()
+vocabulary_size = quizz_machine.vocabulary_size()
 
 log_string(f"vocabulary_size {vocabulary_size}")
 
 
 log_string(f"vocabulary_size {vocabulary_size}")
 
@@ -231,8 +233,10 @@ log_string(f"vocabulary_size {vocabulary_size}")
 # Compute the entropy of the training tokens
 
 token_count = 0
 # Compute the entropy of the training tokens
 
 token_count = 0
-for input in task.batches(split="train", desc="train-entropy"):
-    token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
+for input in quizz_machine.batches(split="train", desc="train-entropy"):
+    token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum(
+        (0, 1)
+    )
 token_probas = token_count / token_count.sum()
 entropy = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(entropy)
 token_probas = token_count / token_count.sum()
 entropy = -torch.xlogy(token_probas, token_probas).sum()
 train_set_perplexity = math.exp(entropy)
@@ -254,11 +258,11 @@ if args.max_percents_of_test_in_train >= 0:
 
     nb_test, nb_in_train = 0, 0
     for test_subset in subsets_as_tuples(
 
     nb_test, nb_in_train = 0, 0
     for test_subset in subsets_as_tuples(
-        task.batches(split="test", desc="test-check"), 25000
+        quizz_machine.batches(split="test", desc="test-check"), 25000
     ):
         in_train = set()
         for train_subset in subsets_as_tuples(
     ):
         in_train = set()
         for train_subset in subsets_as_tuples(
-            task.batches(split="train", desc="train-check"), 25000
+            quizz_machine.batches(split="train", desc="train-check"), 25000
         ):
             in_train.update(test_subset.intersection(train_subset))
         nb_in_train += len(in_train)
         ):
             in_train.update(test_subset.intersection(train_subset))
         nb_in_train += len(in_train)
@@ -275,14 +279,14 @@ if args.max_percents_of_test_in_train >= 0:
 ##############################
 
 
 ##############################
 
 
-def one_epoch(model, task):
+def one_epoch(model, quizz_machine):
     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 
     model.train()
 
     nb_train_samples, acc_train_loss = 0, 0.0
 
     optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
 
     model.train()
 
     nb_train_samples, acc_train_loss = 0, 0.0
 
-    for input in task.batches(split="train"):
+    for input in quizz_machine.batches(split="train"):
         input = input.to(device)
 
         if nb_train_samples % args.batch_size == 0:
         input = input.to(device)
 
         if nb_train_samples % args.batch_size == 0:
@@ -307,14 +311,14 @@ def one_epoch(model, task):
 ######################################################################
 
 
 ######################################################################
 
 
-def run_tests(model, task, deterministic_synthesis):
+def run_tests(model, quizz_machine, deterministic_synthesis):
     with torch.autograd.no_grad():
         model.eval()
 
         nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 
     with torch.autograd.no_grad():
         model.eval()
 
         nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 
-        for input in task.batches(split="test"):
+        for input in quizz_machine.batches(split="test"):
             input = input.to(device)
 
             bs = model(mygpt.BracketedSequence(input))
             input = input.to(device)
 
             bs = model(mygpt.BracketedSequence(input))
@@ -326,7 +330,7 @@ def run_tests(model, task, deterministic_synthesis):
 
             nb_test_samples += input.size(0)
 
 
             nb_test_samples += input.size(0)
 
-        main_test_accuracy = task.produce_results(
+        main_test_accuracy = quizz_machine.produce_results(
             n_epoch=n_epoch,
             model=model,
             result_dir=args.result_dir,
             n_epoch=n_epoch,
             model=model,
             result_dir=args.result_dir,
@@ -347,10 +351,10 @@ def run_tests(model, task, deterministic_synthesis):
 def create_c_quizzes(
     model,
     other_models,
 def create_c_quizzes(
     model,
     other_models,
-    task,
+    quizz_machine,
     nb_for_train=1000,
     nb_for_test=100,
     nb_for_train=1000,
     nb_for_test=100,
-    desired_average_logits=None,
+    min_ave_seq_logproba=None,
 ):
     kept = []
 
 ):
     kept = []
 
@@ -359,17 +363,17 @@ def create_c_quizzes(
     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
         nb_to_generate = 4 * (nb_for_train + nb_for_test)
 
     while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
         nb_to_generate = 4 * (nb_for_train + nb_for_test)
 
-        new_c_quizzes, nb_correct, average_logits = task.create_c_quizzes(
+        new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,
             nb=nb_to_generate,
             model=model,
             other_models=other_models,
             n_epoch=n_epoch,
             result_dir=args.result_dir,
             logger=log_string,
             nb=nb_to_generate,
             model=model,
             other_models=other_models,
-            desired_average_logits=desired_average_logits,
+            min_ave_seq_logproba=min_ave_seq_logproba,
         )
 
         )
 
-        sum_logits += new_c_quizzes.size(0) * average_logits
+        sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
         sum_nb_c_quizzes += new_c_quizzes.size(0)
 
         to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
         sum_nb_c_quizzes += new_c_quizzes.size(0)
 
         to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
@@ -385,10 +389,10 @@ def create_c_quizzes(
 
     new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 
 
     new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
 
-    task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
-    task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
+    quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
+    quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
 
 
-    task.save_quizzes(
+    quizz_machine.problem.save_quizzes(
         new_c_quizzes[:72],
         args.result_dir,
         f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
         new_c_quizzes[:72],
         args.result_dir,
         f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
@@ -425,7 +429,7 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 
 ######################################################################
 
 
 ######################################################################
 
-desired_average_logits = None
+min_ave_seq_logproba = None
 
 for n_epoch in range(args.nb_epochs):
     log_string(f"--- epoch {n_epoch} ----------------------------------------")
 
 for n_epoch in range(args.nb_epochs):
     log_string(f"--- epoch {n_epoch} ----------------------------------------")
@@ -443,45 +447,45 @@ for n_epoch in range(args.nb_epochs):
     )
 
     # improve it
     )
 
     # improve it
-    one_epoch(model, task)
+    one_epoch(model, quizz_machine)
 
 
-    task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+    quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 
     log_string(
 
     log_string(
-        f"train_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
+        f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 
     # test it
     )
 
     # test it
-    run_tests(model, task, deterministic_synthesis=False)
+    run_tests(model, quizz_machine, deterministic_synthesis=False)
 
     log_string(
 
     log_string(
-        f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
+        f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 
     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
         other_models = models.copy()
         other_models.remove(model)
 
     )
 
     if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
         other_models = models.copy()
         other_models.remove(model)
 
-        average_logits = create_c_quizzes(
+        ave_seq_logproba = create_c_quizzes(
             model,
             other_models,
             model,
             other_models,
-            task,
+            quizz_machine,
             nb_for_train=nb_new_c_quizzes_for_train,
             nb_for_test=nb_new_c_quizzes_for_test,
             nb_for_train=nb_new_c_quizzes_for_train,
             nb_for_test=nb_new_c_quizzes_for_test,
-            desired_average_logits=desired_average_logits,
+            min_ave_seq_logproba=min_ave_seq_logproba,
         )
 
         # We keep the first average logits as a reference
         )
 
         # We keep the first average logits as a reference
-        if desired_average_logits is None:
-            desired_average_logits = average_logits
+        if min_ave_seq_logproba is None:
+            min_ave_seq_logproba = ave_seq_logproba
         else:
             log_string(
         else:
             log_string(
-                f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
+                f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
             )
 
         # We update everyone
         for model in models:
             )
 
         # We update everyone
         for model in models:
-            run_tests(model, task, deterministic_synthesis=False)
+            run_tests(model, quizz_machine, deterministic_synthesis=False)
 
 
 ######################################################################
 
 
 ######################################################################