Update.

diff --git a/main.py b/main.py
index d194a8d..d63398c 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -79,9 +79,7 @@ parser.add_argument("--dropout", type=float, default=0.1)
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 
 
 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
 

-parser.add_argument("--reverse_cleanup", action="store_true", default=True)
-
-parser.add_argument("--validation_forward_only", action="store_true", default=False)
+parser.add_argument("--both_directions", action="store_true", default=False)

 
 parser.add_argument("--problem", type=str, default="sky")
 
 
 parser.add_argument("--problem", type=str, default="sky")
 


@@ -409,11 +407,10 @@ def create_c_quizzes(
         c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes(
             nb_to_create,
             model_for_generation=model_for_generation,
         c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes(
             nb_to_create,
             model_for_generation=model_for_generation,

-            reverse_cleanup=args.reverse_cleanup,

         )
 
         nb_correct = quizz_machine.compute_correctness(
         )
 
         nb_correct = quizz_machine.compute_correctness(

-            c_quizzes, models, both_directions=not args.validation_forward_only
+            c_quizzes, models, both_directions=args.both_directions

         )
 
         if args.dirty_debug:
         )
 
         if args.dirty_debug:


@@ -429,7 +426,7 @@ def create_c_quizzes(
         nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
 
         log_string(
         nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
 
         log_string(

-            f"keep c_quizzes kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
+            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
         )
 
     # store the new c_quizzes which have been validated


@@ -487,20 +484,20 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
 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} ----------------------------------------")
 

+    # Select, improve, and eval the worst model
+

     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 
     log_string(
         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
     )
 
     weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
 
     log_string(
         f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
     )
 

-    # improve it

     one_epoch(weakest_model, quizz_machine)
 
     log_string(
         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 
     one_epoch(weakest_model, quizz_machine)
 
     log_string(
         f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
     )
 

-    # test it

     run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
 
     log_string(
     run_tests(weakest_model, quizz_machine, deterministic_synthesis=False)
 
     log_string(


@@ -510,9 +507,13 @@ for n_epoch in range(args.nb_epochs):
     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
     log_string(f"current_test_accuracies {cta}")
 
     cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
     log_string(f"current_test_accuracies {cta}")
 

-    # replace a fraction of the w_quizzes with fresh ones
+    # Replace a fraction of the w_quizzes with fresh ones
+

     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 
     quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
 

+    # If all the models are good enough, generate new quizzes and
+    # re-compute the test errors
+

     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
         create_c_quizzes(
             models,
     if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
         create_c_quizzes(
             models,


@@ -521,7 +522,6 @@ for n_epoch in range(args.nb_epochs):
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 

-        # We update everyone

         for model in models:
             run_tests(model, quizz_machine, deterministic_synthesis=False)
 
         for model in models:
             run_tests(model, quizz_machine, deterministic_synthesis=False)
 

diff --git a/quizz_machine.py b/quizz_machine.py
index 5807b66..0d6d8f5 100755 (executable)
--- a/quizz_machine.py
+++ b/quizz_machine.py
@@ -333,7 +333,7 @@ class QuizzMachine:
         )
 
     def compute_correctness(
         )
 
     def compute_correctness(

-        self, c_quizzes, models_for_validation, both_directions=True
+        self, c_quizzes, models_for_validation, both_directions=False

     ):
         reversed_c_quizzes = self.reverse_time(c_quizzes)
 
     ):
         reversed_c_quizzes = self.reverse_time(c_quizzes)
 


@@ -390,7 +390,7 @@ class QuizzMachine:
 
     ###############################################################
 
 
     ###############################################################
 

-    def generate_quizzes(self, nb, model_for_generation, reverse_cleanup=False):
+    def generate_quizzes(self, nb, model_for_generation):

         c_quizzes = torch.empty(
             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
         )
         c_quizzes = torch.empty(
             nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
         )


@@ -403,10 +403,7 @@ class QuizzMachine:
 
         seq_logproba = torch.empty(ar_mask_first.size(0), device=self.device)
 
 
         seq_logproba = torch.empty(ar_mask_first.size(0), device=self.device)
 

-        if reverse_cleanup:
-            temperature = 10.0
-        else:
-            temperature = 1.0
+        temperature = 10.0

 
         # First, we generate the answer at high temperature
 
 
         # First, we generate the answer at high temperature
 


@@ -433,7 +430,7 @@ class QuizzMachine:
             input=c_quizzes,
             ar_mask=ar_mask_second,
             seq_logproba=seq_logproba,
             input=c_quizzes,
             ar_mask=ar_mask_second,
             seq_logproba=seq_logproba,

-            temperature=temperature,
+            temperature=1.0,

             deterministic_synthesis=True,
             device=self.device,
         )
             deterministic_synthesis=True,
             device=self.device,
         )

diff --git a/sky.py b/sky.py
index d2a4568..2183cf1 100755 (executable)
--- a/sky.py
+++ b/sky.py
@@ -165,7 +165,7 @@ class Sky(problem.Problem):
     ######################################################################
 
     def frame2img(self, x, scale=15):
     ######################################################################
 
     def frame2img(self, x, scale=15):

-        x = x.reshape(-1, self.height, self.width)
+        x = x.reshape(x.size(0), self.height, -1)

         m = torch.logical_and(
             x >= 0, x < self.first_bird_token + self.nb_bird_tokens
         ).long()
         m = torch.logical_and(
             x >= 0, x < self.first_bird_token + self.nb_bird_tokens
         ).long()


@@ -274,7 +274,7 @@ class Sky(problem.Problem):
 if __name__ == "__main__":
     import time
 
 if __name__ == "__main__":
     import time
 

-    sky = Sky(height=6, width=8, speed=4, nb_iterations=2)
+    sky = Sky(height=6, width=8, speed=1, nb_iterations=4)

 
     prompts, answers = sky.generate_prompts_and_answers(4)
 
 
     prompts, answers = sky.generate_prompts_and_answers(4)