halonet

Keras HaloNet

---

Summary

• Github lucidrains/halonet-pytorch.
• HaloAttention article: PDF 2103.12731 Scaling Local Self-Attention for Parameter Efficient Visual Backbones.
• No pretrained available for H models. Architecture is guessed from article, so it's NOT certain.
• T model weights are reloaded from timm Github rwightman/pytorch-image-models.

Models

Model	Params	FLOPs	Input	Top1 Acc	Download
HaloNextECA26T	10.7M	2.43G	256	79.50	halonext_eca26t_256_imagenet.h5
HaloNet26T	12.5M	3.18G	256	79.13	halonet26t_256_imagenet.h5
HaloNetSE33T	13.7M	3.55G	256	80.99	halonet_se33t_256_imagenet.h5
HaloRegNetZB	11.68M	1.97G	224	81.042	haloregnetz_b_224_imagenet.h5
HaloNet50T	22.7M	5.29G	256	81.70	halonet50t_256_imagenet.h5
HaloBotNet50T	22.6M	5.02G	256	82.0	halobotnet50t_256_imagenet.h5

Comparing HaloNetH7 accuracy by replacing Conv layers with Attention in each stage:

Conv Stages	Attention Stages	Top-1 Acc (%)	Norm. Train Time
-	1, 2, 3, 4	84.9	1.9
1	2, 3, 4	84.6	1.4
1, 2	3, 4	84.7	1.0
1, 2, 3	4	83.8	0.5

Usage

from keras_cv_attention_models import halonet

# Will download and load pretrained imagenet weights.
mm = halonet.HaloNet26T(pretrained="imagenet")

# Run prediction
import tensorflow as tf
from tensorflow import keras
from skimage.data import chelsea
imm = keras.applications.imagenet_utils.preprocess_input(chelsea(), mode='torch') # Chelsea the cat
pred = mm(tf.expand_dims(tf.image.resize(imm, mm.input_shape[1:3]), 0)).numpy()
print(keras.applications.imagenet_utils.decode_predictions(pred)[0])
# [('n02124075', 'Egyptian_cat', 0.8999013),
#  ('n02123159', 'tiger_cat', 0.012704549),
#  ('n02123045', 'tabby', 0.009713952),
#  ('n07760859', 'custard_apple', 0.00056676986),
#  ('n02487347', 'macaque', 0.00050636294)]

---