Conv2DTranspose

classin packageklyn.math.nn.layers

public class Conv2DTranspose<T extends AbstractFloat = Float> extends Convolution2DLayer<T>

Conv2DTranspose

Trainable transposed convolution for spatial upsampling, not a mathematical inverse. Supports FP16/FP32/FP64, valid/same padding, independent strides and optional relu, sigmoid or tanh. Only channels-last tensors [N, H, W, C] are accepted. Groups, dilation and explicit output padding are not part of this API.

Example

layer = Conv2DTranspose<Float>(32, 3, padding="same", activation="relu")

Constructors

Modifier and Type	Member	Description
public	Conv2DTranspose `Conv2DTranspose(filters as Long, kernelSize as Long, strides as Long = 1l, padding as String = "valid", activation as String = "linear", useBias as Boolean = true, seed as Long = 1l)`	Configures a layer whose input channel count is inferred on build.
public	Conv2DTranspose `Conv2DTranspose(filters as Long, kernelSize as Long, strides as Tuple, padding as String = "valid", activation as String = "linear", useBias as Boolean = true, seed as Long = 1l)`	Configures a layer whose input channel count is inferred on build.
public	Conv2DTranspose `Conv2DTranspose(filters as Long, kernelSize as Tuple, strides as Long, padding as String = "valid", activation as String = "linear", useBias as Boolean = true, seed as Long = 1l)`	Configures a layer whose input channel count is inferred on build.
public	Conv2DTranspose `Conv2DTranspose(filters as Long, kernelSize as Tuple, strides as Tuple = (1, 1), padding as String = "valid", activation as String = "linear", useBias as Boolean = true, seed as Long = 1l)`	Configures a layer whose input channel count is inferred on build.