EfficientNet Lite models adapted to Keras functional API.
- Nov 2021:
- Added separate
get_preprocessing_layer
utility function.
- Added separate
This is a package with EfficientNet-Lite model variants adapted to Keras.
EfficientNet-Lite variants are modified versions of EfficientNet models, better suited for mobile and embedded devices.
The model's weights are converted from original repository.
The design was meant to mimic the usage of keras.applications
:
# Install
!pip install git+https://github.com/sebastian-sz/efficientnet-lite-keras@main
# Import package:
from efficientnet_lite import EfficientNetLiteB0
import tensorflow as tf
# Use model directly:
model = EfficientNetLiteB0(weights='imagenet', input_shape=(224, 224, 3))
model.summary()
# Or to extract features / fine tune:
backbone = EfficientNetLiteB0(
weights='imagenet',
input_shape=(224,224, 3),
include_top=False
)
model = tf.keras.Sequential([
backbone,
tf.keras.layers.GlobalAveragePooling2D(),
tf.keras.layers.Dense(10) # 10 = num classes
])
model.compile(...)
model.fit(...)
You can fine tune these models, just like other Keras models.
For end-to-end fine-tuning and conversion examples check out the Colab Notebook.
There are multiple ways to install.
The only requirements are Tensorflow 2.2+ and Python 3.6+.
(Although, Tensorflow at least 2.4 is strongly recommended)
pip install git+https://github.com/sebastian-sz/efficientnet-lite-keras@main
git clone https://github.com/sebastian-sz/efficientnet-lite-keras.git
cd efficientnet_lite_keras
pip install .
If you do not want to install you could just drop the efficientnet_lite/efficientnet_lite.py
file directly into your project.
You can also install this package as an extension to official Tensorflow docker container:
Build: docker build -t efficientnet_lite_keras .
Run: docker run -it --rm efficientnet_lite_keras
For GPU support or different TAG you can (for example) pass
--build-arg IMAGE_TAG=2.5.0-gpu
in build command.
If all goes well you should be able to import:
from efficientnet_lite import *
There are 5 lite model variants you can use (B0-B4).
The imagenet weights are automatically downloaded if you pass weights="imagenet"
option while creating the models.
The models expect image values in range -1:+1
. In more detail the preprocessing
function (for pretrained models) looks as follows:
def preprocess(image): # input image is in range 0-255.
return (image - 127.00) / 128.00
Or you can use Preprocessing Layer:
from efficientnet_lite import get_preprocessing_layer
layer = get_preprocessing_layer()
inputs = layer(image)
The following table shows input shapes for each model variant:
Model variant | Input shape |
---|---|
B0 | 224,224 |
B1 | 240,240 |
B2 | 260,260 |
B3 | 280,280 |
B4 | 300,300 |
For fine-tuning example, check out the Colab Notebook.
The models are TFLite compatible. You can convert them like any other Keras model:
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
with open("efficientnet_lite.tflite", "wb") as file:
file.write(tflite_model)
The models are ONNX compatible. For ONNX Conversion you can use tf2onnx package:
!pip install tf2onnx==1.8.4
# Save the model in TF's Saved Model format:
model.save("my_saved_model/")
# Convert:
!python -m tf2onnx.convert \
--saved-model my_saved_model/ \
--output efficientnet_lite.onnx
Lite model variants were intended for mobile use and embedded systems, so I tested if they work with Tensorflow Model Optimization Toolkit.
For example, preparing the model for pruning should work:
import tensorflow_model_optimization as tfmot
from efficientnet_lite import EfficientNetLiteB0
model = EfficientNetLiteB0()
model = tfmot.sparsity.keras.prune_low_magnitude(model)
The original weights are present in the
original repository
for Efficient Net Lite in the form of Tensorflow's .ckpt
files. Also, on Tensorflow's
GitHub, there is a utility script
for converting EfficientNet weights.
The scripts worked for me, after I modified the model's architecture, to match the description of Lite variants.
The converted weights are on this repository's GitHub. If, for some reason, you wish to download and convert original weights yourself, I prepared the utility scripts:
bash scripts/download_all_weights.sh
bash scripts/convert_all_weights.sh
[1] Original repository
[2] Existing non-lite Keras EfficientNet models
[3] Weight update util
If you found this repo useful, please consider giving it a star!