i wonder how kernel get initialized in transpose convolution, such as ConvTranspose1d, ConvTranspose2d, and ConvTranspose3d. is bilinear interpolation applied to the initialization of the kernel?
in addition, how does the kernel(filter) in the transpose convolution become learnable?
how can the kernel learn? does the kernel learn through back-propagation, like the weights in layers in transpose convolution(ConvTranspose1d, ConvTranspose2d, and ConvTranspose3d) ?
Thank you very much for the information!
regarding kernel, I have additional questions.
you mentioned that “the weight parameter of the conv layers are the kernels”.
in the case, will bias will be included in the weights(kernel) when bias is set to “True” in Convolution/ Transpose Convolution ? or
will the bias be added to output value when bias is set to “True”
About the inltialization of weights, i saw that nn.init.uniform_ is used in one case and nn.init.xavier_uniform_ is used in other case.
is there any particular standard for the initialization?
in the linear layer, how can i adjust the size of weights, including bias?
The bias will be added after a filters were applied to the input.
The default initializations are mentioned in the docs for the corresponding layer and are defined in the reset_parameters() method. Depending on your model, use case etc., you might want to change these init methods.
nn.Linear expects the in_features and out_features arguments, which define the shape of the weight and bias parameters.