Update.

[picoclvr.git] / main.py

diff --git a/main.py b/main.py
index 19f918c..1b0d39a 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -5,9 +5,6 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
-# torch.backends.cuda.matmul.allow_tf23
-# torch.autocast(torch.bfloat16)
-
 import math, sys, argparse, time, tqdm, os
 
 import torch, torchvision
@@ -36,7 +33,7 @@ parser.add_argument(
     "--task",
     type=str,
     default="sandbox",
-    help="sandbox, picoclvr, mnist, maze, snake, stack, expr, world",
+    help="sandbox, picoclvr, mnist, maze, snake, stack, expr, rpl, world",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -59,15 +56,17 @@ parser.add_argument("--learning_rate", type=float, default=1e-4)
 
 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
 
-parser.add_argument("--dim_model", type=int, default=512)
+parser.add_argument("--model", type=str, default="37M")
+
+parser.add_argument("--dim_model", type=int, default=None)
 
-parser.add_argument("--dim_keys", type=int, default=64)
+parser.add_argument("--dim_keys", type=int, default=None)
 
-parser.add_argument("--dim_hidden", type=int, default=2048)
+parser.add_argument("--dim_hidden", type=int, default=None)
 
-parser.add_argument("--nb_heads", type=int, default=8)
+parser.add_argument("--nb_heads", type=int, default=None)
 
-parser.add_argument("--nb_blocks", type=int, default=12)
+parser.add_argument("--nb_blocks", type=int, default=None)
 
 parser.add_argument("--dropout", type=float, default=0.1)
 
@@ -80,7 +79,20 @@ parser.add_argument("--overwrite_results", action="store_true", default=False)
 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 
 ##############################
-# picoclvr options
+# rpl options
+
+parser.add_argument("--rpl_nb_starting_values", type=int, default=5)
+
+parser.add_argument("--rpl_max_input", type=int, default=9)
+
+parser.add_argument("--rpl_prog_len", type=int, default=10)
+
+parser.add_argument("--rpl_nb_runs", type=int, default=8)
+
+parser.add_argument("--rpl_no_prog", action="store_true", default=False)
+
+##############################
+# sandbox options
 
 parser.add_argument("--sandbox_level", type=int, default=0)
 
@@ -161,7 +173,7 @@ if args.result_dir is None:
 
 ######################################################################
 
-default_args = {
+default_task_args = {
     "sandbox": {
         "nb_epochs": 50,
         "batch_size": 25,
@@ -204,6 +216,12 @@ default_args = {
         "nb_train_samples": 1000000,
         "nb_test_samples": 10000,
     },
+    "rpl": {
+        "nb_epochs": 40,
+        "batch_size": 25,
+        "nb_train_samples": 100000,
+        "nb_test_samples": 10000,
+    },
     "world": {
         "nb_epochs": 10,
         "batch_size": 25,
@@ -212,10 +230,50 @@ default_args = {
     },
 }
 
-if args.task in default_args:
-    for k, v in default_args[args.task].items():
+if args.task in default_task_args:
+    for k, v in default_task_args[args.task].items():
+        if getattr(args, k) is None:
+            setattr(args, k, v)
+
+######################################################################
+
+default_model_args = {
+    "17K": {
+        "dim_model": 32,
+        "dim_keys": 32,
+        "dim_hidden": 32,
+        "nb_heads": 2,
+        "nb_blocks": 2,
+    },
+    "37M": {
+        "dim_model": 512,
+        "dim_keys": 64,
+        "dim_hidden": 2048,
+        "nb_heads": 8,
+        "nb_blocks": 12,
+    },
+    "122M": {
+        "dim_model": 768,
+        "dim_keys": 64,
+        "dim_hidden": 2048,
+        "nb_heads": 8,
+        "nb_blocks": 24,
+    },
+    "352M": {
+        "dim_model": 1024,
+        "dim_keys": 64,
+        "dim_hidden": 2048,
+        "nb_heads": 8,
+        "nb_blocks": 48,
+    },
+}
+
+if args.model in default_model_args:
+    for k, v in default_model_args[args.model].items():
         if getattr(args, k) is None:
             setattr(args, k, v)
+else:
+    raise ValueError(f"Unknown model {args.model}")
 
 ######################################################################
 
@@ -377,6 +435,20 @@ elif args.task == "expr":
         device=device,
     )
 
+elif args.task == "rpl":
+    task = tasks.RPL(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        nb_starting_values=args.rpl_nb_starting_values,
+        max_input=args.rpl_max_input,
+        prog_len=args.rpl_prog_len,
+        nb_runs=args.rpl_nb_runs,
+        no_prog=args.rpl_no_prog,
+        logger=log_string,
+        device=device,
+    )
+
 elif args.task == "world":
     task = tasks.World(
         nb_train_samples=args.nb_train_samples,
@@ -446,12 +518,12 @@ 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,
+        n_epoch=nb_epochs_finished,
+        model=model,
+        result_dir=args.result_dir,
+        logger=log_string,
+        deterministic_synthesis=args.deterministic_synthesis,
+        input_file=args.expr_input_file,
     )
 
     exit(0)
@@ -527,11 +599,11 @@ nb_samples_seen = 0
 
 if nb_epochs_finished >= nb_epochs:
     task.produce_results(
-        nb_epochs_finished,
-        model,
-        args.result_dir,
-        log_string,
-        args.deterministic_synthesis,
+        n_epoch=nb_epochs_finished,
+        model=model,
+        result_dir=args.result_dir,
+        logger=log_string,
+        deterministic_synthesis=args.deterministic_synthesis,
     )
 
 for n_epoch in range(nb_epochs_finished, nb_epochs):
@@ -585,7 +657,11 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
         )
 
         task.produce_results(
-            n_epoch, model, args.result_dir, log_string, args.deterministic_synthesis
+            n_epoch=n_epoch,
+            model=model,
+            result_dir=args.result_dir,
+            logger=log_string,
+            deterministic_synthesis=args.deterministic_synthesis,
         )
 
     checkpoint = {