Update.

[picoclvr.git] / main.py

diff --git a/main.py b/main.py
index df3f154..213524e 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -14,6 +14,7 @@ import torch, torchvision
 from torch import nn
 from torch.nn import functional as F
 
+import ffutils
 import mygpt, tasks
 
 ######################################################################
@@ -34,8 +35,8 @@ parser = argparse.ArgumentParser(
 parser.add_argument(
     "--task",
     type=str,
-    default="picoclvr",
-    help="picoclvr, mnist, maze, snake, stack, expr",
+    default="sandbox",
+    help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -110,7 +111,7 @@ parser.add_argument("--snake_nb_colors", type=int, default=5)
 parser.add_argument("--snake_length", type=int, default=200)
 
 ##############################
-# Snake options
+# Stack options
 
 parser.add_argument("--stack_nb_steps", type=int, default=100)
 
@@ -125,7 +126,18 @@ parser.add_argument("--stack_fraction_values_for_train", type=float, default=0.7
 
 parser.add_argument("--expr_nb_variables", type=int, default=5)
 
-parser.add_argument("--expr_sequence_length", type=int, default=30)
+parser.add_argument("--expr_sequence_length", type=int, default=40)
+
+parser.add_argument("--expr_operand_max", type=int, default=9)
+
+parser.add_argument("--expr_result_max", type=int, default=99)
+
+parser.add_argument("--expr_input_file", type=str, default=None)
+
+##############################
+# World options
+
+parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
 
 ######################################################################
 
@@ -139,6 +151,12 @@ if args.result_dir is None:
 ######################################################################
 
 default_args = {
+    "sandbox": {
+        "nb_epochs": 10,
+        "batch_size": 25,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 10000,
+    },
     "picoclvr": {
         "nb_epochs": 25,
         "batch_size": 25,
@@ -170,11 +188,17 @@ default_args = {
         "nb_test_samples": 1000,
     },
     "expr": {
-        "nb_epochs": 50,
+        "nb_epochs": 40,
         "batch_size": 25,
-        "nb_train_samples": 250000,
+        "nb_train_samples": 1000000,
         "nb_test_samples": 10000,
     },
+    "world": {
+        "nb_epochs": 10,
+        "batch_size": 25,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 1000,
+    },
 }
 
 if args.task in default_args:
@@ -240,7 +264,18 @@ picoclvr_pruner_eval = (
 
 ######################################################################
 
-if args.task == "picoclvr":
+if args.task == "sandbox":
+    task = tasks.SandBox(
+        tasks.ProblemLevel1(),
+        # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
+        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 == "picoclvr":
     task = tasks.PicoCLVR(
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
@@ -305,10 +340,22 @@ elif args.task == "expr":
         nb_test_samples=args.nb_test_samples,
         nb_variables=args.expr_nb_variables,
         sequence_length=args.expr_sequence_length,
+        operand_max=args.expr_operand_max,
+        result_max=args.expr_result_max,
         batch_size=args.batch_size,
         device=device,
     )
 
+elif args.task == "world":
+    task = tasks.World(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        vqae_nb_epochs=args.world_vqae_nb_epochs,
+        logger=log_string,
+        device=device,
+    )
+
 else:
     raise ValueError(f"Unknown task {args.task}")
 
@@ -366,6 +413,20 @@ else:
 
 ######################################################################
 
+if args.task == "expr" and args.expr_input_file is not None:
+    task.produce_results(
+        nb_epochs_finished,
+        model,
+        args.result_dir,
+        log_string,
+        args.deterministic_synthesis,
+        args.expr_input_file,
+    )
+
+    exit(0)
+
+######################################################################
+
 nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
 
 # Compute the entropy of the training tokens
@@ -383,20 +444,28 @@ train_set_perplexity = math.exp(entropy)
 
 train_examples = {}
 
+
 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:
-            assert x.size() != y.size() or (x - y).abs().sum() > 0
+            if x.size() == y.size() and (x - y).abs().sum() == 0:
+                nb_collisions += 1
 
 del train_examples
 
+log_string(
+    f"data_check {nb_collisions*100/nb_total:.02f}% ({nb_collisions}/{nb_total}) of test samples are in the train set"
+)
+
 ##############################
 
 if args.learning_rate_schedule == "cos":