Koreksi teks dari pengguna Amabitur
I work on a project with computer vision, we create some object detection tools for inspections.
I have been working on a pretty interesting task this week. It’s called ‘shared backbone’. Usually, every neural network has a feature extractor (backbone) and decoder (head). Neural networks aren't the fastest thing in the world, you know. And what if we have to use several networks at one time? Well, it takes time…
BUT you can use a trick: on the train stage you can freeze backbone layers, so you will train only head layers. And after that you will get several networks with the same backbones and different heads. And if you split your networks, you can be able to separately use backbone and heads.
You will no longer need to run backbone+head several times, you can run backbone one time and extract all necessary features and after that show these features to all heads. This will significantly speed up the inference time.
bahasa: Inggris
Pengetahuan bahasa: Penutur asli, Kecakapan, Maju, Menengah-Atas, Menengah, Dasar
Log in atau
register untuk memperbaiki teks ini!
Anda dapat mengubah teks ini jika Anda dispesifikasikan Inggris di salah satu tingkat berikut:
Penutur asli, Kecakapan, Maju, Menengah-Atas, Menengah, Dasar
I work on a project that uses computer vision to create object detection tools.
I have been working on a pretty interesting task this week. It’s called ‘shared backbone. ' Usually every neural network has a feature extractor ("backbone") and a decoder ("head"). Neural networks are not the fastest thing in the world, especially if we have to use several networks at one time.
There is a trick: on the training stage you can freeze the backbone layers, so you will train only head layers. After that, you will get several networks with the same backbone and different heads. If you split your networks, you separate out how you use the backbone and heads.
You will no longer need to run the backbone + head combination several times because you can now run the backbone one time and extract all necessary features and after that show these features to all the associated heads. This will significantly speed up the inferencing stage.
Silakan perbaiki teks ke Inggris:
Tolong bantu terjemahan: