X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=3e67a73fab1452f0d1f97aff1f890fe3ab70bcab;hb=2434c00a82ebb0b23f45d891cc9f80324e3200bd;hp=c51035c118f6480f2f33560b8aeea020c4cfbf28;hpb=ffe183868ac8563fd82fc8312fda90f6f8a95833;p=mygptrnn.git

diff --git a/main.py b/main.py
index c51035c..3e67a73 100755
--- a/main.py
+++ b/main.py
@@ -16,14 +16,6 @@ import mygpt, tasks, problems
 
 ######################################################################
 
-if torch.cuda.is_available():
-    device = torch.device("cuda")
-    torch.backends.cuda.matmul.allow_tf32 = True
-else:
-    device = torch.device("cpu")
-
-######################################################################
-
 
 def str2bool(x):
     x = x.lower()
@@ -55,6 +47,8 @@ parser.add_argument("--seed", type=int, default=0)
 
 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
 
+parser.add_argument("--force_cpu", type=str2bool, default=False)
+
 ########################################
 
 parser.add_argument("--nb_epochs", type=int, default=50)
@@ -208,15 +202,25 @@ parser.add_argument("--mixing_deterministic_start", action="store_true", default
 
 ######################################################################
 
-args = parser.parse_args()
+# args = parser.parse_args()
 
-assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
+args, sup_args = parser.parse_known_args()
+
+sup_args = dict([x.removeprefix("--").split("=") for x in sup_args])
 
 if args.result_dir is None:
     args.result_dir = f"results_{args.task}_{args.model}"
 
 ######################################################################
 
+if not args.force_cpu and torch.cuda.is_available():
+    device = torch.device("cuda")
+    torch.backends.cuda.matmul.allow_tf32 = True
+else:
+    device = torch.device("cpu")
+
+######################################################################
+
 default_task_args = {
     "addition": {
         "model": "352M",
@@ -430,6 +434,8 @@ except FileExistsError:
         print(f"result directory {args.result_dir} already exists")
         exit(1)
 
+loss_file = open(os.path.join(args.result_dir, "loss.dat"), "a")
+
 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
 
 if args.seed >= 0:
@@ -466,6 +472,9 @@ log_string(f"argv {' '.join(sys.argv)}")
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
+for k, v in sup_args.items():
+    log_string(f'sup_args["{k}"] "{v}"')
+
 
 ######################################################################
 
@@ -503,6 +512,9 @@ def get_lr(n_epoch, it):
 ######################################################################
 
 
+assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
+
+
 def picoclvr_pruner_horizontal_green(p):
     return not ("green" in p and ("left" in p or "right" in p))
 
@@ -728,6 +740,8 @@ model = mygpt.MyGPT(
     causal=True,
     dropout=args.dropout,
     attention_layer=args.attention,
+    logger=log_string,
+    **sup_args,
 )
 
 model.to(device)
@@ -832,10 +846,12 @@ if nb_epochs_finished >= nb_epochs:
         deterministic_synthesis=args.deterministic_synthesis,
     )
 
-time_pred_result = None
+time_pred_result = datetime.datetime.now()
 
 it = 0
 
+n_batch = 0
+
 for n_epoch in range(nb_epochs_finished, nb_epochs):
     if args.optim == "sgd":
         optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
@@ -877,6 +893,12 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
         total_loss.backward()
         optimizer.step()
 
+        grad_norm = sum([p.grad.pow(2).sum() for p in model.parameters()]).sqrt()
+
+        loss_file.write(f"{n_epoch} {n_batch} {loss.item()} {grad_norm.item()}\n")
+
+        n_batch += 1
+
     with torch.autograd.no_grad():
         model.eval()
 
@@ -910,10 +932,9 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
         )
 
         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)}"
-            )
+        log_string(
+            f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+        )
         time_pred_result = time_current_result
 
     checkpoint = {