X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=2edfa14de0107376a431632791564451e729c298;hb=b1d28a1ed672be21947509dac2f90666b65b5034;hp=17936c34dc7c3ea4a0108a219a3a386151a34755;hpb=6e09c88d26d0bfd675af9afd9cdc32aa3485d1b7;p=picoclvr.git

diff --git a/main.py b/main.py
index 17936c3..2edfa14 100755
--- a/main.py
+++ b/main.py
@@ -5,7 +5,7 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
-import math, sys, argparse, time, tqdm, os
+import math, sys, argparse, time, tqdm, os, datetime
 
 import torch, torchvision
 from torch import nn
@@ -33,7 +33,7 @@ parser.add_argument(
     "--task",
     type=str,
     default="twotargets",
-    help="byheart, learnop, guessop, mixing, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp",
+    help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, escape",
 )
 
 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
@@ -86,6 +86,13 @@ parser.add_argument("--overwrite_results", action="store_true", default=False)
 
 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 
+##############################
+# filetask
+
+parser.add_argument("--filetask_train_file", type=str, default=None)
+
+parser.add_argument("--filetask_test_file", type=str, default=None)
+
 ##############################
 # rpl options
 
@@ -104,6 +111,8 @@ parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 
 parser.add_argument("--grid_size", type=int, default=6)
 
+parser.add_argument("--grid_fraction_play", type=float, default=0)
+
 ##############################
 # picoclvr options
 
@@ -166,6 +175,15 @@ parser.add_argument("--mixing_hard", action="store_true", default=False)
 
 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
 
+##############################
+# escape options
+
+parser.add_argument("--escape_height", type=int, default=4)
+
+parser.add_argument("--escape_width", type=int, default=6)
+
+parser.add_argument("--escape_T", type=int, default=25)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -178,6 +196,12 @@ if args.result_dir is None:
 ######################################################################
 
 default_task_args = {
+    "file": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 250000,
+        "nb_test_samples": 10000,
+    },
     "addition": {
         "model": "352M",
         "batch_size": 25,
@@ -256,6 +280,12 @@ default_task_args = {
         "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
+    "memory": {
+        "model": "37M",
+        "batch_size": 100,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 1000,
+    },
     "mixing": {
         "model": "37M",
         "batch_size": 25,
@@ -268,6 +298,12 @@ default_task_args = {
         "nb_train_samples": 60000,
         "nb_test_samples": 10000,
     },
+    "escape": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 10000,
+    },
 }
 
 if args.task in default_task_args:
@@ -285,6 +321,13 @@ default_model_args = {
         "nb_heads": 2,
         "nb_blocks": 2,
     },
+    "4M": {
+        "dim_model": 256,
+        "dim_keys": 32,
+        "dim_hidden": 1024,
+        "nb_heads": 4,
+        "nb_blocks": 6,
+    },
     "37M": {
         "dim_model": 512,
         "dim_keys": 64,
@@ -375,7 +418,22 @@ picoclvr_pruner_eval = (
 
 ######################################################################
 
-if args.task == "byheart":
+if args.task == "file":
+    assert (
+        args.filetask_train_file is not None and args.filetask_test_file is not None
+    ), "You have to specify the task train and test files"
+    task = tasks.TaskFromFile(
+        args.filetask_train_file,
+        args.filetask_test_file,
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        shuffle=True,
+        device=device,
+    )
+    args.max_percents_of_test_in_train = 0
+
+elif args.task == "byheart":
     task = tasks.SandBox(
         problem=problems.ProblemByHeart(),
         nb_train_samples=args.nb_train_samples,
@@ -418,6 +476,16 @@ elif args.task == "twotargets":
         device=device,
     )
 
+elif args.task == "memory":
+    task = tasks.SandBox(
+        problem=problems.ProblemMemory(),
+        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 == "mixing":
     task = tasks.SandBox(
         problem=problems.ProblemMixing(
@@ -531,6 +599,7 @@ elif args.task == "grid":
         nb_test_samples=args.nb_test_samples,
         batch_size=args.batch_size,
         size=args.grid_size,
+        fraction_play=args.grid_fraction_play,
         logger=log_string,
         device=device,
     )
@@ -545,6 +614,18 @@ elif args.task == "qmlp":
         device=device,
     )
 
+elif args.task == "escape":
+    task = tasks.Escape(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        height=args.escape_height,
+        width=args.escape_width,
+        T=args.escape_T,
+        logger=log_string,
+        device=device,
+    )
+
 else:
     raise ValueError(f"Unknown task {args.task}")
 
@@ -695,6 +776,8 @@ if nb_epochs_finished >= nb_epochs:
         deterministic_synthesis=args.deterministic_synthesis,
     )
 
+time_pred_result = None
+
 for n_epoch in range(nb_epochs_finished, nb_epochs):
     learning_rate = learning_rate_schedule[n_epoch]
 
@@ -753,6 +836,13 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
             deterministic_synthesis=args.deterministic_synthesis,
         )
 
+        time_current_result = datetime.datetime.now()
+        if time_pred_result is not None:
+            log_string(
+                f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+            )
+        time_pred_result = time_current_result
+
     checkpoint = {
         "nb_epochs_finished": n_epoch + 1,
         "model_state": model.state_dict(),