tfreplknet

Keras (TensorFlow v2) reimplementation of Re-parameterized Large Kernel Network (RepLKNet) model.

Based on Official Pytorch implementation.

Supports variable-shape inference.

Installation

pip install tfreplknet

Available models and weights

Examples

Default usage (without preprocessing):

from tfreplknet import RepLKNetB224In1k  # + 4 other variants and input preprocessing

model = RepLKNetB224In1k()  # by default will download imagenet{1k, 21k}-pretrained weights
model.compile(...)
model.fit(...)

Custom classification (with preprocessing):

from keras import layers, models
from tfreplknet import RepLKNetB224In1k, preprocess_input_bl

inputs = layers.Input(shape=(224, 224, 3), dtype='uint8')
outputs = layers.Lambda(preprocess_input_bl)(inputs)
outputs = RepLKNetB224In1k(include_top=False)(outputs)
outputs = layers.Dense(100, activation='softmax')(outputs)

model = models.Model(inputs=inputs, outputs=outputs)
model.compile(...)
model.fit(...)

Evaluation

For correctness, RepLKNetB224In1k and RepLKNetB384In1k models (original and ported) tested with ImageNet-v2 test set.

import tensorflow as tf
import tensorflow_datasets as tfds
from tfreplknet import RepLKNetB224In1k, RepLKNetB384In1k, preprocess_input_bl

def _prepare(example):
    # For RepLKNetB224In1k
    image = tf.image.resize(example['image'], (256, 256), method=tf.image.ResizeMethod.BICUBIC)
    image = tf.image.central_crop(image, 0.875)
    
    # For RepLKNetB384In1k
    # image = tf.image.resize(example['image'], (438, 438), method=tf.image.ResizeMethod.BICUBIC)
    # image = tf.image.central_crop(image, 0.877)
    
    image = preprocess_input_bl(image)
    
    return image, example['label']
    
imagenet2 = tfds.load('imagenet_v2', split='test', shuffle_files=True)
imagenet2 = imagenet2.map(_prepare, num_parallel_calls=tf.data.AUTOTUNE)
imagenet2 = imagenet2.batch(8)

model = RepLKNetB224In1k()
model.compile('sgd', 'sparse_categorical_crossentropy', ['accuracy', 'sparse_top_k_categorical_accuracy'])
history = model.evaluate(imagenet2)

print(history)

Citation

@article{2022arXiv220306717D,
  title={Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs},
  author={{Ding}, Xiaohan and {Zhang}, Xiangyu and {Zhou}, Yizhuang and {Han}, Jungong and {Ding}, Guiguang and {Sun}, Jian},
  journal={arXiv preprint arXiv:2203.06717},
  year={2022}
}