The following table lists the supported operations and APIs for vai_q_tensorflow2.
| Layer Types | Supported Layers | Description |
|---|---|---|
| Core | tf.keras.layers.InputLayer | |
| Core | tf.keras.layers.Dense | |
| Core | tf.keras.layers.Activation | If activation is relu or
linear, it is quantized. If activation is sigmoid or swish, it is converted to hard-sigmoid or hard-swish and then quantized by default. Otherwise it is not quantized. |
| Convolution | tf.keras.layers.Conv2D | |
| Convolution | tf.keras.layers.DepthwiseConv2D | |
| Convolution | tf.keras.layers.Conv2DTranspose | |
| Convolution | tf.keras.layers.SeparableConv2D | |
| Pooling | tf.keras.layers.AveragePooling2D | |
| Pooling | tf.keras.layers.MaxPooling2D | |
| Pooling | tf.keras.layers.GlobalAveragePooling | |
| Normalization | tf.keras.layers.BatchNormalization | By default, BatchNormalization layers are fused with the previous
convolution layers. If they cannot be combined, they are converted to
depthwise convolutions. In the QAT mode,
BatchNormalization layers are pseudo-fused if
|
| Regularization | tf.keras.layers.Dropout | By default, the dropout layers are removed. In the QAT mode, dropout layers are retained if remove_dropout is set to FALSE. It is removed when the get_deploy_model function is called. |
| Reshaping | tf.keras.layers.Reshape | |
| Reshaping | tf.keras.layers.Flatten | |
| Reshaping | tf.keras.UpSampling2D | |
| Reshaping | tf.keras.ZeroPadding2D | |
| Merging | tf.keras.layers.Concatenate | |
| Merging | tf.keras.layers.Add | |
| Merging | tf.keras.layers.Multiply | |
| Activation | tf.keras.layers.ReLU | |
| Activation | tf.keras.layers.Softmax | The input for the Softmax layer is quantized. It can run on the standalone Softmax IP for acceleration. |
| Activation | tf.keras.layers.LeakyReLU | Only LeakyReLU layer with "alpha=0." One can be quantized and mapped to DPU. Internally, the quantizer automatically converts the alpha value to 26/256 to match the DPU implementation. LeakyReLU layer with other values is not quantized and is mapped to the CPU. |
| Hard_sigmoid | tf.keras.layer.ReLU(6.)(x + 3.) * (1. / 6.) | The supported hard_sigmoid is from Mobilenet_v3. tf.keras.Activation.hard_sigmoid is not supported now and is not quantized. |
| Activation | tf.keras.layers.PReLU |
Note: The DPU might have limitations on these supported
layers. They can be rolled back to the CPU during compilation. For more information, see
Supported Operators and DPU Limitations.