training/ffc.py

# Fast Fourier Convolution NeurIPS 2020
# original implementation https://github.com/pkumivision/FFC/blob/main/model_zoo/ffc.py
# paper https://proceedings.neurips.cc/paper/2020/file/2fd5d41ec6cfab47e32164d5624269b1-Paper.pdf

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch
import torch.nn as nn
import torch.nn.functional as F
from kornia.geometry.transform import rotate
import torch.fft as fft
from icecream import ic
import PIL

def save_image_grid(feats, fname, gridsize):
    gw, gh = gridsize
    idx = gw * gh

    max_num = torch.max(feats[:idx]).item()
    min_num = torch.min(feats[:idx]).item()
    feats = feats[:idx].cpu() * 255 / (max_num - min_num) 
    feats = np.asarray(feats, dtype=np.float32)
    feats = np.rint(feats).clip(0, 255).astype(np.uint8)

    C, H, W = feats.shape

    feats = feats.reshape(gh, gw, 1, H, W)
    feats = feats.transpose(0, 3, 1, 4, 2)
    feats = feats.reshape(gh * H, gw * W, 1)
    feats = np.stack([feats]*3, axis=2).squeeze() * 10
    feats = np.rint(feats).clip(0, 255).astype(np.uint8)
    
    from icecream import ic
    ic(feats.shape)
    
    feats = PIL.Image.fromarray(feats)
    feats.save(fname + '.png')

def _conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
    return F.conv2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)

class LearnableSpatialTransformWrapper(nn.Module):
    def __init__(self, impl, pad_coef=0.5, angle_init_range=80, train_angle=True):
        super().__init__()
        self.impl = impl
        self.angle = torch.rand(1) * angle_init_range
        if train_angle:
            self.angle = nn.Parameter(self.angle, requires_grad=True)
        self.pad_coef = pad_coef

    def forward(self, x):
        if torch.is_tensor(x):
            return self.inverse_transform(self.impl(self.transform(x)), x)
        elif isinstance(x, tuple):
            x_trans = tuple(self.transform(elem) for elem in x)
            y_trans = self.impl(x_trans)
            return tuple(self.inverse_transform(elem, orig_x) for elem, orig_x in zip(y_trans, x))
        else:
            raise ValueError(f'Unexpected input type {type(x)}')

    def transform(self, x):
        height, width = x.shape[2:]
        pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef)
        x_padded = F.pad(x, [pad_w, pad_w, pad_h, pad_h], mode='reflect')
        x_padded_rotated = rotate(x_padded, angle=self.angle.to(x_padded))
        return x_padded_rotated

    def inverse_transform(self, y_padded_rotated, orig_x):
        height, width = orig_x.shape[2:]
        pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef)

        y_padded = rotate(y_padded_rotated, angle=-self.angle.to(y_padded_rotated))
        y_height, y_width = y_padded.shape[2:]
        y = y_padded[:, :, pad_h : y_height - pad_h, pad_w : y_width - pad_w]
        return y


class SELayer(nn.Module):
    def __init__(self, channel, reduction=16):
        super(SELayer, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
            nn.Linear(channel, channel // reduction, bias=False),
            nn.ReLU(inplace=False),
            nn.Linear(channel // reduction, channel, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        res = x * y.expand_as(x)
        return res


class FourierUnit(nn.Module):

    def __init__(self, in_channels, out_channels, groups=1, spatial_scale_factor=None, spatial_scale_mode='bilinear',
                 spectral_pos_encoding=False, use_se=False, se_kwargs=None, ffc3d=False, fft_norm='ortho'):
        # bn_layer not used
        super(FourierUnit, self).__init__()
        self.groups = groups

        self.conv_layer = torch.nn.Conv2d(in_channels=in_channels * 2 + (2 if spectral_pos_encoding else 0),
                                          out_channels=out_channels * 2,
                                          kernel_size=1, stride=1, padding=0, groups=self.groups, bias=False)
        self.relu = torch.nn.ReLU(inplace=False)

        # squeeze and excitation block
        self.use_se = use_se
        if use_se:
            if se_kwargs is None:
                se_kwargs = {}
            self.se = SELayer(self.conv_layer.in_channels, **se_kwargs)

        self.spatial_scale_factor = spatial_scale_factor
        self.spatial_scale_mode = spatial_scale_mode
        self.spectral_pos_encoding = spectral_pos_encoding
        self.ffc3d = ffc3d
        self.fft_norm = fft_norm

    def forward(self, x):
        batch = x.shape[0]

        if self.spatial_scale_factor is not None:
            orig_size = x.shape[-2:]
            x = F.interpolate(x, scale_factor=self.spatial_scale_factor, mode=self.spatial_scale_mode, align_corners=False)

        r_size = x.size()
        # (batch, c, h, w/2+1, 2)
        fft_dim = (-3, -2, -1) if self.ffc3d else (-2, -1)
        ffted = fft.rfftn(x, dim=fft_dim, norm=self.fft_norm)
        ffted = torch.stack((ffted.real, ffted.imag), dim=-1)
        ffted = ffted.permute(0, 1, 4, 2, 3).contiguous()  # (batch, c, 2, h, w/2+1)
        ffted = ffted.view((batch, -1,) + ffted.size()[3:])

        if self.spectral_pos_encoding:
            height, width = ffted.shape[-2:]
            coords_vert = torch.linspace(0, 1, height)[None, None, :, None].expand(batch, 1, height, width).to(ffted)
            coords_hor = torch.linspace(0, 1, width)[None, None, None, :].expand(batch, 1, height, width).to(ffted)
            ffted = torch.cat((coords_vert, coords_hor, ffted), dim=1)

        if self.use_se:
            ffted = self.se(ffted)

        ffted = self.conv_layer(ffted)  # (batch, c*2, h, w/2+1)
        ffted = self.relu(ffted)

        ffted = ffted.view((batch, -1, 2,) + ffted.size()[2:]).permute(
            0, 1, 3, 4, 2).contiguous()  # (batch,c, t, h, w/2+1, 2)
        ffted = torch.complex(ffted[..., 0], ffted[..., 1])

        ifft_shape_slice = x.shape[-3:] if self.ffc3d else x.shape[-2:]
        output = torch.fft.irfftn(ffted, s=ifft_shape_slice, dim=fft_dim, norm=self.fft_norm)

        if self.spatial_scale_factor is not None:
            output = F.interpolate(output, size=orig_size, mode=self.spatial_scale_mode, align_corners=False)

        return output


class SpectralTransform(nn.Module):
    
    def __init__(self, in_channels, out_channels, stride=1, groups=1, enable_lfu=True, **fu_kwargs):
        # bn_layer not used
        super(SpectralTransform, self).__init__()
        self.enable_lfu = enable_lfu
        if stride == 2:
            self.downsample = nn.AvgPool2d(kernel_size=(2, 2), stride=2)
        else:
            self.downsample = nn.Identity()

        self.stride = stride
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels, out_channels //
                      2, kernel_size=1, groups=groups, bias=False),
            # nn.BatchNorm2d(out_channels // 2),
            nn.ReLU(inplace=True)
        )
        self.fu = FourierUnit(
            out_channels // 2, out_channels // 2, groups, **fu_kwargs)
        if self.enable_lfu:
            self.lfu = FourierUnit(
                out_channels // 2, out_channels // 2, groups)
        self.conv2 = torch.nn.Conv2d(
            out_channels // 2, out_channels, kernel_size=1, groups=groups, bias=False)

    def forward(self, x):

        x = self.downsample(x)
        x = self.conv1(x)
        output = self.fu(x)

        if self.enable_lfu:
            n, c, h, w = x.shape
            split_no = 2
            split_s = h // split_no
            xs = torch.cat(torch.split(
                x[:, :c // 4], split_s, dim=-2), dim=1).contiguous()
            xs = torch.cat(torch.split(xs, split_s, dim=-1),
                           dim=1).contiguous()
            xs = self.lfu(xs)
            xs = xs.repeat(1, 1, split_no, split_no).contiguous()
        else:
            xs = 0

        output = self.conv2(x + output + xs)

        return output
    
class FFC(nn.Module):
    
    def __init__(self, in_channels, out_channels, kernel_size,
                 ratio_gin, ratio_gout, stride=1, padding=0,
                 dilation=1, groups=1, bias=False, enable_lfu=True,
                 padding_type='reflect', gated=False, **spectral_kwargs):
        super(FFC, self).__init__()

        assert stride == 1 or stride == 2, "Stride should be 1 or 2."
        self.stride = stride

        in_cg = int(in_channels * ratio_gin)
        in_cl = in_channels - in_cg
        out_cg = int(out_channels * ratio_gout)
        out_cl = out_channels - out_cg
        #groups_g = 1 if groups == 1 else int(groups * ratio_gout)
        #groups_l = 1 if groups == 1 else groups - groups_g

        self.ratio_gin = ratio_gin
        self.ratio_gout = ratio_gout
        self.global_in_num = in_cg

        module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d
        self.convl2l = module(in_cl, out_cl, kernel_size,
                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d
        self.convl2g = module(in_cl, out_cg, kernel_size,
                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d
        self.convg2l = module(in_cg, out_cl, kernel_size,
                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform
        self.convg2g = module(
            in_cg, out_cg, stride, 1 if groups == 1 else groups // 2, enable_lfu, **spectral_kwargs)

        self.gated = gated
        module = nn.Identity if in_cg == 0 or out_cl == 0 or not self.gated else nn.Conv2d
        self.gate = module(in_channels, 2, 1)

    def forward(self, x, fname=None):
        x_l, x_g = x if type(x) is tuple else (x, 0)
        out_xl, out_xg = 0, 0

        if self.gated:
            total_input_parts = [x_l]
            if torch.is_tensor(x_g):
                total_input_parts.append(x_g)
            total_input = torch.cat(total_input_parts, dim=1)

            gates = torch.sigmoid(self.gate(total_input))
            g2l_gate, l2g_gate = gates.chunk(2, dim=1)
        else:
            g2l_gate, l2g_gate = 1, 1
            
        # for i in range(x_g.shape[0]):
        #     c, h, w = x_g[i].shape
        #     gh = 3
        #     gw = 3
        #     save_image_grid(x_g[i].detach(), f'vis/{fname}_xg_{h}', (gh, gw))
        
        # for i in range(x_l.shape[0]):
        #     c, h, w = x_l[i].shape
        #     gh = 3
        #     gw = 3
        #     save_image_grid(x_l[i].detach(), f'vis/{fname}_xl_{h}', (gh, gw))
            
        spec_x = self.convg2g(x_g)
        
        # for i in range(spec_x.shape[0]):
        #     c, h, w = spec_x[i].shape
        #     gh = 3
        #     gw = 3
        #     save_image_grid(spec_x[i].detach(), f'vis/{fname}_spec_x_{h}', (gh, gw))

        if self.ratio_gout != 1:
            out_xl = self.convl2l(x_l) + self.convg2l(x_g) * g2l_gate
        if self.ratio_gout != 0:
            out_xg = self.convl2g(x_l) * l2g_gate + spec_x
        
        # for i in range(out_xg.shape[0]):
        #     c, h, w = out_xg[i].shape
        #     gh = 3
        #     gw = 3
        #     save_image_grid(out_xg[i].detach(), f'vis/{fname}_outg_{h}', (gh, gw))
        
        # for i in range(out_xl.shape[0]):
        #     c, h, w = out_xl[i].shape
        #     gh = 3
        #     gw = 3
        #     save_image_grid(out_xl[i].detach(), f'vis/{fname}_outl_{h}', (gh, gw))

        return out_xl, out_xg

class FFC_BN_ACT(nn.Module):

    def __init__(self, in_channels, out_channels,
                 kernel_size, ratio_gin, ratio_gout,
                 stride=1, padding=0, dilation=1, groups=1, bias=False,
                 norm_layer=nn.SyncBatchNorm, activation_layer=nn.Identity,
                 padding_type='reflect',
                 enable_lfu=True, **kwargs):
        super(FFC_BN_ACT, self).__init__()
        self.ffc = FFC(in_channels, out_channels, kernel_size,
                       ratio_gin, ratio_gout, stride, padding, dilation,
                       groups, bias, enable_lfu, padding_type=padding_type, **kwargs)
        lnorm = nn.Identity if ratio_gout == 1 else norm_layer
        gnorm = nn.Identity if ratio_gout == 0 else norm_layer
        global_channels = int(out_channels * ratio_gout)
        # self.bn_l = lnorm(out_channels - global_channels)
        # self.bn_g = gnorm(global_channels)

        lact = nn.Identity if ratio_gout == 1 else activation_layer
        gact = nn.Identity if ratio_gout == 0 else activation_layer
        self.act_l = lact(inplace=True)
        self.act_g = gact(inplace=True)

    def forward(self, x, fname=None):
        x_l, x_g = self.ffc(x, fname=fname,)
        x_l = self.act_l(x_l)
        x_g = self.act_g(x_g)
        return x_l, x_g


class FFCResnetBlock(nn.Module):
    def __init__(self, dim, padding_type, norm_layer, activation_layer=nn.ReLU, dilation=1,
                 spatial_transform_kwargs=None, inline=False, ratio_gin=0.75, ratio_gout=0.75):
        super().__init__()
        self.conv1 = FFC_BN_ACT(dim, dim, kernel_size=3, padding=dilation, dilation=dilation,
                                norm_layer=norm_layer,
                                activation_layer=activation_layer,
                                padding_type=padding_type,
                                ratio_gin=ratio_gin, ratio_gout=ratio_gout)
        self.conv2 = FFC_BN_ACT(dim, dim, kernel_size=3, padding=dilation, dilation=dilation,
                                norm_layer=norm_layer,
                                activation_layer=activation_layer,
                                padding_type=padding_type,
                                ratio_gin=ratio_gin, ratio_gout=ratio_gout)
        if spatial_transform_kwargs is not None:
            self.conv1 = LearnableSpatialTransformWrapper(self.conv1, **spatial_transform_kwargs)
            self.conv2 = LearnableSpatialTransformWrapper(self.conv2, **spatial_transform_kwargs)
        self.inline = inline

    def forward(self, x, fname=None):
        if self.inline:
            x_l, x_g = x[:, :-self.conv1.ffc.global_in_num], x[:, -self.conv1.ffc.global_in_num:]
        else:
            x_l, x_g = x if type(x) is tuple else (x, 0)

        id_l, id_g = x_l, x_g

        x_l, x_g = self.conv1((x_l, x_g), fname=fname)
        x_l, x_g = self.conv2((x_l, x_g), fname=fname)

        x_l, x_g = id_l + x_l, id_g + x_g
        out = x_l, x_g
        if self.inline:
            out = torch.cat(out, dim=1)
        return out

class ConcatTupleLayer(nn.Module):
    def forward(self, x):
        assert isinstance(x, tuple)
        x_l, x_g = x
        assert torch.is_tensor(x_l) or torch.is_tensor(x_g)
        if not torch.is_tensor(x_g):
            return x_l
        return torch.cat(x, dim=1)