keras2ncnn

Export Keras model to Tencent/NCNN.

Supported Keras layers and Features

• Functional Model API (the Sequential and Model as Layer features should be transformed in to flat Functional API Model)
• InputLayer
• ReLU/ReLU6/LeakyReLU/Softmax/Sigmoid
• Clip
• Reshape/Flatten(converted with NCNN::Reshape)
• MaxPooling2D/AveragePooling2D/MaxPool2D/AvgPool2D
• GlobalMaxPooling2D/GlobalAveragePooling2D/GlobalMaxPool2D/GlobalAvgPool2D
• ZeroPadding2D/ReflectPadding2D
• BatchNormalization/InstanceNormalization
• Conv2D/DepthwiseConv2D/SeparableConv2D(converted with split into NCNN::ConvolutionDepthWise->NCNN::Convolution)
• Concatenate/Add/Multiply
• UpSampling2D(nearest neighbour/bilinear)
• Conv2DTranspose(only for the even strides)

Supported Keras Layer Optimization

• Conv2D/DepthwiseConv2D/Conv2DTranspose with BatchNormalization
• BatchNormalization with Conv2D/DepthwiseConv2D/Conv2DTranspose(In Progress)
• ZeroPadding2D(In Progress :: Fusion with Convolution/Pooling)

Unit tests is written

• Unit tests with Python NCNN inference - pyncnn installed
• Latest models tested tf_nightly-2.5.0.dev20201130-cp37-cp37m-win_amd64.whl / tf_nightly_gpu-2.5.0.dev20201130-cp37-cp37m-win_amd64.whl
• Latest NCNN revision https://github.com/Tencent/ncnn/commit/25b224479cbe535ce35ca92556c8f17d9b9f1951
• Latest PyNCNN revision https://github.com/caishanli/pyncnn/commit/63c77b8bd75dae8e2601b918e92a5050a3fee8df

Preconverted models

Some 'preconverted' models can be downloaded from DropBox

Requirements installation

The code was tested with python3.7 with TensorFlow 1.x/2.x (CPU). The code should work with python3.x . The behaviour with TensorFlow GPU/TPU.

git clone https://github.com/azeme1/keras2ncnn.git
cd keras2ncnn
pip3 install -r requirements.txt 

Usage

The model zoo folder contains the sample model (CelebA_PrismaNet_256_hair_seg_model_opt_001.hdf5) as well as the result of the conversion (graph: CelebA_PrismaNet_256_hair_seg_model_opt_001.param and weights: CelebA_PrismaNet_256_hair_seg_model_opt_001.bin) Load the model from the '.hdf5' file

python3 keras2ncnn.py --model_path=model_zoo/segmentation/hair/model_000/CelebA_PrismaNet_256_hair_seg_model_opt_001.hdf5

Load the model from the '.json' file (the weights should be located at the same folder in '.hdf5')

python3 keras2ncnn.py --model_path=model_zoo/segmentation/hair/model_000/CelebA_PrismaNet_256_hair_seg_model_opt_001.json

Useful Links

Tencent/NCNN documentation

• [https://github.com/Tencent/ncnn/wiki/how-to-implement-custom-layer-step-by-step]
• [https://github.com/Tencent/ncnn/wiki/param-and-model-file-structure]
• [https://github.com/Tencent/ncnn/wiki/operation-param-weight-table]
• [https://github.com/MarsTechHAN/keras2ncnn]

  Graph visualization

• [https://github.com/lutzroeder/netron]

TODO List

Code

• Add support for the Dense Layer
• Fix layer name length issue
• Export models from Keras applications applications
• Model in Model support
• Sequential API support
• Mixed mode API support

  Upcoming Models

• [https://github.com/thangtran480/hair-segmentation]
• [https://github.com/ItchyHiker/Hair_Segmentation_Keras]
• [https://github.com/gaelkt/HairNets]
• [https://github.com/JungUnYun/Hair_segmentation]
• [https://github.com/YawenSun9/deep-hair-segmentation]
• [https://github.com/mostafa-shalaby84/hair_segmentation_matting]

Thanx

• [https://github.com/cvzakharchenko]
• [https://github.com/nvoronetskiy]

Important note

Sometimes good result can be achieved with Tensorflow conversion approach

import tensorflow as tf
model = tf.keras.models.load_model("model.h5")
model.save("saved_model")

after that convert 'model.ckpt' or 'model.pb' with following scripts

• [https://github.com/Tencent/ncnn/tree/master/tools/mlir]
• [https://github.com/Tencent/ncnn/tree/master/tools/tensorflow]