X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=2edfa14de0107376a431632791564451e729c298;hb=b1d28a1ed672be21947509dac2f90666b65b5034;hp=ff831f462130a3c959699ac544f72b23584ed504;hpb=b87078aec53ead1e0a3ca44d4ac46c319bbcd63e;p=picoclvr.git

diff --git a/main.py b/main.py
index ff831f4..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, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl",
+    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=" ")
@@ -46,7 +46,7 @@ parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
 
 ########################################
 
-parser.add_argument("--nb_epochs", type=int, default=None)
+parser.add_argument("--nb_epochs", type=int, default=25)
 
 parser.add_argument("--batch_size", type=int, default=None)
 
@@ -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
 
@@ -99,6 +106,13 @@ parser.add_argument("--rpl_nb_runs", type=int, default=5)
 
 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
 
+##############################
+# grid options
+
+parser.add_argument("--grid_size", type=int, default=6)
+
+parser.add_argument("--grid_fraction_play", type=float, default=0)
+
 ##############################
 # picoclvr options
 
@@ -155,9 +169,20 @@ parser.add_argument("--expr_result_max", type=int, default=99)
 parser.add_argument("--expr_input_file", type=str, default=None)
 
 ##############################
-# World options
+# Mixing
 
-parser.add_argument("--world_vqae_nb_epochs", type=int, default=25)
+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)
 
 ######################################################################
 
@@ -171,96 +196,113 @@ if args.result_dir is None:
 ######################################################################
 
 default_task_args = {
-    "byheart": {
-        "model": "37M",
-        "nb_epochs": 2,
-        "batch_size": 25,
-        "nb_train_samples": 50000,
-        "nb_test_samples": 10000,
-    },
-    "learnop": {
+    "file": {
         "model": "37M",
-        "nb_epochs": 15,
         "batch_size": 25,
-        "nb_train_samples": 50000,
+        "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
-    "guessop": {
+    "addition": {
         "model": "352M",
-        "nb_epochs": 5,
         "batch_size": 25,
-        "nb_train_samples": 1000000,
+        "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
-    "twotargets": {
+    "byheart": {
         "model": "37M",
-        "nb_epochs": 10,
         "batch_size": 25,
         "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
-    "addition": {
+    "expr": {
         "model": "352M",
-        "nb_epochs": 50,
         "batch_size": 25,
-        "nb_train_samples": 250000,
+        "nb_train_samples": 2500000,
         "nb_test_samples": 10000,
     },
-    "picoclvr": {
+    "grid": {
         "model": "37M",
-        "nb_epochs": 25,
         "batch_size": 25,
         "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
-    "mnist": {
+    "qmlp": {
         "model": "37M",
-        "nb_epochs": 25,
         "batch_size": 10,
-        "nb_train_samples": 60000,
+        "nb_train_samples": 100000,
+        "nb_test_samples": 1000,
+    },
+    "guessop": {
+        "model": "352M",
+        "batch_size": 25,
+        "nb_train_samples": 1000000,
+        "nb_test_samples": 10000,
+    },
+    "learnop": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
     "maze": {
         "model": "37M",
-        "nb_epochs": 25,
         "batch_size": 5,
         "nb_train_samples": 100000,
         "nb_test_samples": 10000,
     },
+    "picoclvr": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 250000,
+        "nb_test_samples": 10000,
+    },
+    "rpl": {
+        "model": "352M",
+        "batch_size": 5,
+        "nb_train_samples": 2500000,
+        "nb_test_samples": 10000,
+    },
     "snake": {
         "model": "37M",
-        "nb_epochs": 5,
         "batch_size": 25,
         "nb_train_samples": 250000,
         "nb_test_samples": 10000,
     },
     "stack": {
         "model": "37M",
-        "nb_epochs": 15,
         "batch_size": 25,
         "nb_train_samples": 100000,
         "nb_test_samples": 1000,
     },
-    "expr": {
-        "model": "352M",
-        "nb_epochs": 25,
+    "twotargets": {
+        "model": "37M",
         "batch_size": 25,
-        "nb_train_samples": 2500000,
+        "nb_train_samples": 50000,
         "nb_test_samples": 10000,
     },
-    "rpl": {
-        "model": "122M",
-        "nb_epochs": 50,
-        "batch_size": 5,
-        "nb_train_samples": 1000000,
+    "memory": {
+        "model": "37M",
+        "batch_size": 100,
+        "nb_train_samples": 25000,
+        "nb_test_samples": 1000,
+    },
+    "mixing": {
+        "model": "37M",
+        "batch_size": 25,
+        "nb_train_samples": 250000,
+        "nb_test_samples": 10000,
+    },
+    "mnist": {
+        "model": "37M",
+        "batch_size": 10,
+        "nb_train_samples": 60000,
         "nb_test_samples": 10000,
     },
-    "world": {
+    "escape": {
         "model": "37M",
-        "nb_epochs": 10,
         "batch_size": 25,
         "nb_train_samples": 25000,
-        "nb_test_samples": 1000,
+        "nb_test_samples": 10000,
     },
 }
 
@@ -279,6 +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,
@@ -342,6 +391,8 @@ def log_string(s):
     sys.stdout.flush()
 
 
+log_string(f"argv {' '.join(sys.argv)}")
+
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
@@ -367,7 +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,
@@ -410,6 +476,28 @@ 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(
+            hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
+        ),
+        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 == "addition":
     task = tasks.SandBox(
         problem=problems.ProblemAddition(),
@@ -505,12 +593,35 @@ elif args.task == "rpl":
         device=device,
     )
 
-elif args.task == "world":
-    task = tasks.World(
+elif args.task == "grid":
+    task = tasks.Grid(
+        nb_train_samples=args.nb_train_samples,
+        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,
+    )
+
+elif args.task == "qmlp":
+    task = tasks.QMLP(
+        nb_train_samples=args.nb_train_samples,
+        nb_test_samples=args.nb_test_samples,
+        batch_size=args.batch_size,
+        result_dir=args.result_dir,
+        logger=log_string,
+        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,
-        vqae_nb_epochs=args.world_vqae_nb_epochs,
+        height=args.escape_height,
+        width=args.escape_width,
+        T=args.escape_T,
         logger=log_string,
         device=device,
     )
@@ -665,6 +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]
 
@@ -723,6 +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(),