Update.
authorFrançois Fleuret <francois@fleuret.org>
Wed, 10 Jul 2024 18:09:50 +0000 (20:09 +0200)
committerFrançois Fleuret <francois@fleuret.org>
Wed, 10 Jul 2024 18:09:50 +0000 (20:09 +0200)
main.py
quiz_machine.py

diff --git a/main.py b/main.py
index 0c193f7..1ef01e9 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -16,6 +16,8 @@ import ffutils
 import mygpt
 import sky, grids, quiz_machine
 
+import threading
+
 # world quizzes vs. culture quizzes
 
 ######################################################################
@@ -38,7 +40,7 @@ parser.add_argument("--result_dir", type=str, default=None)
 
 parser.add_argument("--seed", type=int, default=0)
 
-parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
+parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1)
 
 ########################################
 
@@ -78,6 +80,8 @@ parser.add_argument("--problem", type=str, default="grids")
 
 parser.add_argument("--nb_threads", type=int, default=1)
 
+parser.add_argument("--nb_gpus", type=int, default=1)
+
 parser.add_argument("--nb_gpts", type=int, default=5)
 
 parser.add_argument("--min_to_validate", type=int, default=None)
@@ -238,14 +242,14 @@ if args.problem == "sky":
         nb_birds=args.sky_nb_birds,
         nb_iterations=args.sky_nb_iterations,
         speed=args.sky_speed,
-        max_nb_cached_chunks=args.nb_train_samples // 100,
+        max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
         chunk_size=100,
         nb_threads=args.nb_threads,
     )
     back_accuracy = False
 elif args.problem == "grids":
     problem = grids.Grids(
-        max_nb_cached_chunks=args.nb_train_samples // 100,
+        max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100,
         chunk_size=100,
         nb_threads=args.nb_threads,
     )
@@ -273,50 +277,23 @@ vocabulary_size = quiz_machine.vocabulary_size()
 log_string(f"vocabulary_size {vocabulary_size}")
 
 ######################################################################
-##############################
-
-
-def one_epoch(model, quiz_machine):
-    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
-
-    model.train()
-
-    nb_train_samples, acc_train_loss = 0, 0.0
-
-    for input in quiz_machine.batches(model, split="train"):
-        input = input.to(device)
-
-        if nb_train_samples % args.batch_size == 0:
-            optimizer.zero_grad()
-
-        output = model(mygpt.BracketedSequence(input)).x
-        loss = F.cross_entropy(output.transpose(1, 2), input)
-        acc_train_loss += loss.item() * input.size(0)
-
-        nb_train_samples += input.size(0)
-
-        loss.backward()
-
-        if nb_train_samples % args.batch_size == 0:
-            optimizer.step()
-
-    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
-
-    log_string(f"train_perplexity {n_epoch} {train_perplexity}")
 
 
 ######################################################################
 
 
-def run_tests(model, quiz_machine, deterministic_synthesis):
+def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None):
+    if local_device is None:
+        local_device = device
+
     with torch.autograd.no_grad():
-        model.eval()
+        model.eval().to(local_device)
 
         nb_test_samples, acc_test_loss = 0, 0.0
         nb_samples_accumulated = 0
 
         for input in quiz_machine.batches(model, split="test"):
-            input = input.to(device)
+            input = input.to(local_device)
 
             bs = model(mygpt.BracketedSequence(input))
             output = bs.x
@@ -339,6 +316,42 @@ def run_tests(model, quiz_machine, deterministic_synthesis):
         )
 
 
+def one_epoch(model, quiz_machine, local_device=None):
+    if local_device is None:
+        local_device = device
+
+    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
+
+    model.to(local_device).train()
+
+    nb_train_samples, acc_train_loss = 0, 0.0
+
+    for input in quiz_machine.batches(model, split="train"):
+        input = input.to(local_device)
+
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.zero_grad()
+
+        output = model(mygpt.BracketedSequence(input)).x
+        loss = F.cross_entropy(output.transpose(1, 2), input)
+        acc_train_loss += loss.item() * input.size(0)
+
+        nb_train_samples += input.size(0)
+
+        loss.backward()
+
+        if nb_train_samples % args.batch_size == 0:
+            optimizer.step()
+
+    train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
+
+    log_string(f"train_perplexity {n_epoch} {train_perplexity}")
+
+    run_tests(model, quiz_machine, deterministic_synthesis=False)
+
+    model.TRAINING_LOCK.release()
+
+
 ######################################################################
 
 
@@ -548,6 +561,7 @@ for k in range(args.nb_gpts):
 
     model.main_test_accuracy = 0.0
     model.id = k
+    model.TRAINING_LOCK = threading.Lock()
 
     model.train_w_quizzes = quiz_machine.generate_token_sequences(
         args.nb_train_samples
@@ -640,23 +654,24 @@ for n_epoch in range(args.nb_epochs):
     ##################################################
     # Select, improve, and eval the worst model
 
-    weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
+    ranked_models = sorted(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_models = ranked_models[: args.nb_gpus]
 
-    one_epoch(weakest_model, quiz_machine)
+    for gpu_id, model in enumerate(weakest_models):
+        model.TRAINING_LOCK.acquire()
 
-    log_string(
-        f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
-    )
+        log_string(
+            f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
+        )
 
-    run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
+        threading.Thread(
+            target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}")
+        ).start()
 
-    log_string(
-        f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
-    )
+    for model in weakest_models:
+        model.TRAINING_LOCK.acquire()
+        model.TRAINING_LOCK.release()
 
     ##################################################
     # Replace a fraction of the w_quizzes with fresh ones
@@ -667,7 +682,8 @@ for n_epoch in range(args.nb_epochs):
 
     # Renew entirely the train set
 
-    quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
+    for model in weakest_models:
+        quiz_machine.renew_w_quizzes(model, args.nb_train_samples)
 
     ##################################################
     # If all the models are good enough, generate new quizzes and
@@ -681,8 +697,4 @@ for n_epoch in range(args.nb_epochs):
             nb_for_test=nb_new_c_quizzes_for_test,
         )
 
-        for model in models:
-            run_tests(model, quiz_machine, deterministic_synthesis=False)
-
-
 ######################################################################
index 34c09a7..1f1046d 100755 (executable)
@@ -15,6 +15,8 @@ from torch.nn import functional as F
 import mygpt
 from mygpt import BracketedSequence
 
+import threading
+
 ######################################################################
 
 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
@@ -235,22 +237,10 @@ class QuizMachine:
         self.prompt_len = None
         self.answer_len = None
 
-        # self.train_w_quizzes = self.generate_token_sequences(nb_train_samples)
-        # self.reverse_random_half_in_place(self.train_w_quizzes)
-
-        # self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
-        # self.reverse_random_half_in_place(self.test_w_quizzes)
-
+        self.LOCK_C_QUIZZES = threading.Lock()
         self.train_c_quizzes = []
         self.test_c_quizzes = []
 
-        # if result_dir is not None:
-        # self.save_quizzes(
-        # result_dir,
-        # "culture_w_quizzes",
-        # self.train_w_quizzes[:72],
-        # )
-
     def save_quizzes(
         self,
         result_dir,
@@ -292,32 +282,34 @@ class QuizMachine:
 
     def batches(self, model, split="train", desc=None):
         assert split in {"train", "test"}
-        if split == "train":
-            w_quizzes = model.train_w_quizzes
-            c_quizzes = self.train_c_quizzes
-        else:
-            w_quizzes = model.test_w_quizzes
-            c_quizzes = self.test_c_quizzes
 
-        if len(c_quizzes) > 0:
-            c_quizzes = torch.cat(c_quizzes, dim=0)
-            if c_quizzes.size(0) > w_quizzes.size(0) // 2:
-                i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
-                c_quizzes = c_quizzes[i]
+        with self.LOCK_C_QUIZZES:
+            if split == "train":
+                w_quizzes = model.train_w_quizzes
+                c_quizzes = self.train_c_quizzes
+            else:
+                w_quizzes = model.test_w_quizzes
+                c_quizzes = self.test_c_quizzes
+
+            if len(c_quizzes) > 0:
+                c_quizzes = torch.cat(c_quizzes, dim=0)
+                if c_quizzes.size(0) > w_quizzes.size(0) // 2:
+                    i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
+                    c_quizzes = c_quizzes[i]
+
+                i = torch.randperm(w_quizzes.size(0))[
+                    : w_quizzes.size(0) - c_quizzes.size(0)
+                ]
+                w_quizzes = w_quizzes[i]
 
-            i = torch.randperm(w_quizzes.size(0))[
-                : w_quizzes.size(0) - c_quizzes.size(0)
-            ]
-            w_quizzes = w_quizzes[i]
+                self.nb_batch_w_quizzes = w_quizzes.size(0)
+                self.nb_batch_c_quizzes = c_quizzes.size(0)
 
-            self.nb_batch_w_quizzes = w_quizzes.size(0)
-            self.nb_batch_c_quizzes = c_quizzes.size(0)
-
-            input = torch.cat([w_quizzes, c_quizzes], dim=0)
-        else:
-            input = w_quizzes
-            self.nb_batch_w_quizzes = w_quizzes.size(0)
-            self.nb_batch_c_quizzes = 0
+                input = torch.cat([w_quizzes, c_quizzes], dim=0)
+            else:
+                input = w_quizzes
+                self.nb_batch_w_quizzes = w_quizzes.size(0)
+                self.nb_batch_c_quizzes = 0
 
         # Shuffle
         input = input[torch.randperm(input.size(0))]
@@ -417,10 +409,11 @@ class QuizMachine:
     ######################################################################
 
     def store_c_quizzes(self, new_c_quizzes, for_train=True):
-        if for_train:
-            self.train_c_quizzes.append(new_c_quizzes)
-        else:
-            self.test_c_quizzes.append(new_c_quizzes)
+        with self.LOCK_C_QUIZZES:
+            if for_train:
+                self.train_c_quizzes.append(new_c_quizzes)
+            else:
+                self.test_c_quizzes.append(new_c_quizzes)
 
     ######################################################################