biyoml / TF2-Pretrained-RegNets

RegNet (Designing Network Design Spaces) implementation in TensorFlow-2 with pretrained weights.

TensorFlow-2 Pretrained RegNets

RegNet (Designing Network Design Spaces) implementation in TensorFlow-2 with pretrained weights.

Dependencies

• Python ≥ 3.7
• TensorFlow ≥ 2.5

Models

RegNetX

Model name	Model name in paper	Pretrained weights	Top-1 (%)
RegNetX_400MF	REGNETX-400MF	RegNetX_400MF.h5	72.87
RegNetX_800MF	REGNETX-800MF	RegNetX_800MF.h5	75.21
RegNetX_1_6GF	REGNETX-1.6GF	RegNetX_1_6GF.h5	77.11
RegNetX_3_2GF	REGNETX-3.2GF	RegNetX_3_2GF.h5	78.33
RegNetX_8GF	REGNETX-8.0GF	RegNetX_8GF.h5	79.36
RegNetX_16GF	REGNETX-16GF	RegNetX_16GF.h5	79.98
RegNetX_32GF	REGNETX-32GF	RegNetX_32GF.h5	80.58

RegNetY

Model name	Model name in paper	Pretrained weights	Top-1 (%)
RegNetY_400MF	REGNETY-400MF	RegNetY_400MF.h5	74.02
RegNetY_800MF	REGNETY-800MF	RegNetY_800MF.h5	76.44
RegNetY_1_6GF	REGNETY-1.6GF	RegNetY_1_6GF.h5	77.98
RegNetY_3_2GF	REGNETY-3.2GF	RegNetY_3_2GF.h5	78.94
RegNetY_8GF	REGNETY-8.0GF	RegNetY_8GF.h5	80.05
RegNetY_16GF	REGNETY-16GF	RegNetY_16GF.h5	80.43
RegNetY_32GF	REGNETY-32GF	RegNetY_32GF.h5	80.84

• Pretrained weights are converted from TorchVision model zoo, and we only provide models of certain flop regimes that are available in the model zoo. Script for conversion is data/scripts/convert.py.
• Top-1: 224x224 single-crop, top-1 accuracy using converted weights. Reproduce by:

# You need to register on http://www.image-net.org/download-images to get the link to
# download ILSVRC2012_img_val.tar.
mkdir ILSVRC2012_img_val/
tar xvf ILSVRC2012_img_val.tar -C ILSVRC2012_img_val/

python data/scripts/eval.py --h5 path/to/pretrained.h5 --data_dir ILSVRC2012_img_val/ --batch_size 32

Usage

import regnet

# Specify include_top=False if you want to remove the classification layer at the top
model = regnet.RegNetX_1_6GF(input_shape=(224, 224, 3),
                             weights="path/to/RegNetX_1_6GF.h5",
                             include_top=True)

model.compile(...)
model.fit(...)

Note: Input images should be loaded in to a range of [0, 1] and then normalized using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].

License

MIT