feature_extractor_models/encoders/_utils.py

import torch
import torch.nn as nn


def patch_first_conv(model, new_in_channels, default_in_channels=3, pretrained=True):
    """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) and module.in_channels == default_in_channels:
            break

    weight = module.weight.detach()
    module.in_channels = new_in_channels

    if not pretrained:
        module.weight = nn.parameter.Parameter(
            torch.Tensor(
                module.out_channels,
                new_in_channels // module.groups,
                *module.kernel_size,
            )
        )
        module.reset_parameters()

    elif new_in_channels == 1:
        new_weight = weight.sum(1, keepdim=True)
        module.weight = nn.parameter.Parameter(new_weight)

    else:
        new_weight = torch.Tensor(
            module.out_channels, new_in_channels // module.groups, *module.kernel_size
        )

        for i in range(new_in_channels):
            new_weight[:, i] = weight[:, i % default_in_channels]

        new_weight = new_weight * (default_in_channels / new_in_channels)
        module.weight = nn.parameter.Parameter(new_weight)


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()