Supported layers and functions
The Edge optimization is compatible with Tensorflow/Keras .h5 models generated with a Tensorflow backend of version 2.x. Models containing unsupported layers will fail to convert.
Supported TensorFlow layers
- Activation (TensorFlow Class Activation)
- Average Pooling 1D (TensorFlow Class AveragePool1D)
- Average Pooling 2D (TensorFlow Class AveragePool2D)
- Batch Normalization (TensorFlow Class BatchNormalization)
- Convolution 2D (TensorFlow Class Conv2D)
- Convolution 1D (TensorFlow Class Conv1D)
- Dense (TensorFlow Class Dense)
- Dropout (TensorFlow Class Dropout)
- Flatten (TensorFlow Class Flatten)
- Gated Recurrent Unit (TensorFlow Class GRU)
- Global Max Pooling 1D (TensorFlow Class GlobalMaxPool1D)
- Global Average Pooling 1D (TensorFlow Class GlobalAveragePool1D)
- Long Short-Term Memory (TensorFlow Class LSTM)
- Max Pooling 2D (TensorFlow Class MaxPool2D)
- Max Pooling 1D (TensorFlow Class MaxPool1D)
- Reshape (TensorFlow Class Reshape)
Supported TensorFlow activation functions
- Hard Sigmoid
- LeakyReLU (TensorFlow Class LeakyReLU)
- ReLU (TensorFlow Class ReLU)
- Sigmoid
- Softmax
- Tanh
Supported Imagimob layers and functions
- Assert Array
- Clear
- Copy
- Discrete Cosine Transform
- Addition
- Argmax
- Division
- Dot product transpose
- Hann smoothing
- Multiplication
- Frobenius norm
- Product
- Real Discrete Fourier Transform
- Average Subtract
- Sum
- Power to Decibel
- Average
- Subtraction
- Add an extra dimension of size 1
- Complex Discrete Fourier Transform
- Max
- FFT Shift
- Subtraction Immediate Reverse (A-x)
- Logarithm
- Mel Filterbank
- Power
- Sliding Window