deepblink.networks.unet module

UNet architecture.

deepblink.networks.unet.unet(dropout: float = 0.2, cell_size: int = 4, filters: int = 5, ndown: int = 2, l2: float = 1e-06, block: str = 'convolutional') → keras.engine.training.Model

Unet model with second, cell size dependent encoder.

Note that “convolution” is the currently best block.

Parameters:

• dropout – Percentage of dropout before each MaxPooling step.
• cell_size – Size of one cell in the prediction matrix.
• filters – Log_2 number of filters in the first inception block.
• ndown – Downsampling steps in the first encoder / decoder.
• l2 – L2 value for kernel and bias regularization.
• block – Type of block in each layer. [options: convolutional, inception, residual]