From 19ed324e69b4323933630ff608cddd557233cdb4 Mon Sep 17 00:00:00 2001
From: Leonard Hackel <l.hackel@tu-berlin.de>
Date: Mon, 22 May 2023 09:32:31 +0200
Subject: [PATCH] adding precision to matmul op for torch>1.12
---
train_lit4rsvqa.py | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/train_lit4rsvqa.py b/train_lit4rsvqa.py
index bd24a69..6117e57 100644
--- a/train_lit4rsvqa.py
+++ b/train_lit4rsvqa.py
@@ -249,13 +249,15 @@ def main(
data_dir: str = None,
test_run: bool = False,
num_workers_dataloader: int = 4,
- vision_checkpoint: str = None
+ vision_checkpoint: str = None,
+ matmul_precision: str = "medium",
):
if test_run:
max_img_index = 10 * batch_size
epochs = 10
else:
max_img_index = -1
+ torch.set_float32_matmul_precision(matmul_precision)
pl.seed_everything(seed, workers=True)
@@ -340,7 +342,8 @@ def main(
"Seed": seed,
"# Workers": num_workers_dataloader,
"Vision Checkpoint": vision_checkpoint,
- "GPU": torch.cuda.get_device_name()
+ "GPU": torch.cuda.get_device_name(),
+ "MatMul Precision": matmul_precision,
}
)
--
GitLab