Update.

[culture.git] / main.py

diff --git a/main.py b/main.py
index 73e7ca2..4cf4d59 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -341,7 +341,7 @@ def one_epoch(model, quiz_machine, local_device=None):
 
     train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
 
-    log_string(f"train_perplexity {n_epoch} {train_perplexity}")
+    log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}")
 
     run_tests(model, quiz_machine, deterministic_synthesis=False)
 
@@ -354,9 +354,6 @@ def one_epoch(model, quiz_machine, local_device=None):
 def standard_validity(logproba):
     l = logproba.sort(dim=-1).values
     return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99))
-    # warnings.warn("TEST!!!", RuntimeWarning)
-    # print(l.exp())
-    # return (l[:, 0] < math.log(0.99))
 
 
 def valid_c_quizzes(recorded, criteria):
@@ -452,13 +449,9 @@ for k in range(args.nb_gpts):
     model.id = k
     model.TRAINING_LOCK = threading.Lock()
 
-    model.train_w_quizzes = quiz_machine.generate_token_sequences(
-        args.nb_train_samples
-    ).to(device)
+    model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples)
     quiz_machine.reverse_random_half_in_place(model.train_w_quizzes)
-    model.test_w_quizzes = quiz_machine.generate_token_sequences(
-        args.nb_test_samples
-    ).to(device)
+    model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples)
     quiz_machine.reverse_random_half_in_place(model.test_w_quizzes)
 
     models.append(model)
@@ -532,6 +525,11 @@ if args.dirty_debug:
     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(0.5)
+
+
 ######################################################################
 
 for n_epoch in range(args.nb_epochs):