keras2ncnn

Now availabel on pypi!

If you failed to convert a model, welcome to open an issue and attach the h5 file.

Usage:

# Install keras2ncnn (only h5py and numpy is required)
python3 -mpip install --upgrade keras2ncnn

# If you only want to convert the model
python3 -m keras2ncnn -i SOME_H5DF_FILE.h5 -o ./  

# You can see the structure of the converted model and the original model(after optimization)
python3 -m keras2ncnn -i SOME_H5DF_FILE.h5 -o DIR_TO_SAVE_NCNN_PARAM --plot_graph/-p

Supported Op

• InputLayer
• Conv2D (Linear, Softmax, ReLU, Sigmoid)
• Conv2DTranspose (Linear, ReLU, Sigmoid)
• DepthwiseConv2D
• SeparableConv2D (Linear, Softmax, ReLU, Sigmoid)
• Add
• Multiply
• Concatenate
• ZeroPadding2D
• ReLU
• LeakyReLU
• Activation (Softmax, ReLU, Sigmoid)
• UpSampling2D
• BilinearUpsampling
• Cropping2D
• GlobalAveragePooling2D
• GlobalMaxPooling2D
• AveragePooling2D
• MaxPooling2D
• BatchNormalization
• Dense (Linear, Softmax, ReLU, Sigmoid)
• Flatten
• Reshape
• Maximum
• TensorFlowOpLayer (Mul with constant)
• Permute (Need more testing)

Ops that will be dropped by converter

• Dropout
• Lambda
• TimeDistributed
• InputLayer with inbound nodes