X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=71026c59640be3752f3a311dc6f58cbcf49b11d1;hb=5d46a9bd7d032d90ef4c4b38ac3c9b5b66526527;hp=213524e753f4f6732bbfcb9cbb9929bdf82ad19a;hpb=ead4b8e4edd29578c01501d168e416b47fa4047b;p=picoclvr.git

diff --git a/main.py b/main.py
index 213524e..71026c5 100755
--- 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)
 
@@ -82,6 +81,17 @@ parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 ##############################
 # picoclvr options
 
+parser.add_argument("--sandbox_level", type=int, default=0)
+
+parser.add_argument("--sandbox_levels_nb_items", type=int, default=25)
+
+parser.add_argument("--sandbox_levels_len_source", type=int, default=6)
+
+parser.add_argument("--sandbox_levels_len_result", type=int, default=8)
+
+##############################
+# picoclvr options
+
 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
 
 parser.add_argument("--picoclvr_height", type=int, default=12)
@@ -150,11 +160,11 @@ if args.result_dir is None:
 
 ######################################################################
 
-default_args = {
+default_task_args = {
     "sandbox": {
-        "nb_epochs": 10,
+        "nb_epochs": 50,
         "batch_size": 25,
-        "nb_train_samples": 25000,
+        "nb_train_samples": 100000,
         "nb_test_samples": 10000,
     },
     "picoclvr": {
@@ -193,6 +203,12 @@ 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,
@@ -201,13 +217,53 @@ 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}")
+
+######################################################################
+
 try:
     os.mkdir(args.result_dir)
 except FileExistsError:
@@ -265,8 +321,28 @@ picoclvr_pruner_eval = (
 ######################################################################
 
 if args.task == "sandbox":
+    if args.sandbox_level == 0:
+        problem = tasks.ProblemLevel0(
+            nb_sentences=args.sandbox_levels_nb_items,
+            len_prompt=args.sandbox_levels_len_source,
+            len_result=args.sandbox_levels_len_result,
+        )
+    elif args.sandbox_level == 1:
+        problem = tasks.ProblemLevel1(
+            nb_operators=args.sandbox_levels_nb_items,
+            len_source=args.sandbox_levels_len_source,
+            len_result=args.sandbox_levels_len_result,
+        )
+    elif args.sandbox_level == 2:
+        problem = tasks.ProblemLevel2(
+            len_source=args.sandbox_levels_len_source,
+            len_result=args.sandbox_levels_len_result,
+        )
+    else:
+        raise ValueError(f"Unknown sandbox level {args.sandbox_level}")
+
     task = tasks.SandBox(
-        tasks.ProblemLevel1(),
+        problem,
         # tasks.ProblemAddition(zero_padded=False, inverted_result=False),
         nb_train_samples=args.nb_train_samples,
         nb_test_samples=args.nb_test_samples,
@@ -346,6 +422,15 @@ 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,
+        logger=log_string,
+        device=device,
+    )
+
 elif args.task == "world":
     task = tasks.World(
         nb_train_samples=args.nb_train_samples,