utils.py

import torch
import torch.nn as nn


def patch_first_conv(model, in_channels):
    """Change first convolution layer input channels.
    In case:
        in_channels == 1 or in_channels == 2 -> reuse original weights
        in_channels > 3 -> make random kaiming normal initialization
    """

    # get first conv
    for module in model.modules():
        if isinstance(module, nn.Conv2d):
            break

    # change input channels for first conv
    module.in_channels = in_channels
    weight = module.weight.detach()
    reset = False

    if in_channels == 1:
        weight = weight.sum(1, keepdim=True)
    elif in_channels == 2:
        weight = weight[:, :2] * (3.0 / 2.0)
    else:
        reset = True
        weight = torch.Tensor(
            module.out_channels,
            module.in_channels // module.groups,
            *module.kernel_size
        )

    module.weight = nn.parameter.Parameter(weight)
    if reset:
        module.reset_parameters()


def replace_strides_with_dilation(module, dilation_rate):
    """Patch Conv2d modules replacing strides with dilation"""
    for mod in module.modules():
        if isinstance(mod, nn.Conv2d):
            mod.stride = (1, 1)
            mod.dilation = (dilation_rate, dilation_rate)
            kh, kw = mod.kernel_size
            mod.padding = ((kh // 2) * dilation_rate, (kh // 2) * dilation_rate)

            # Kostyl for EfficientNet
            if hasattr(mod, "static_padding"):
                mod.static_padding = nn.Identity()