first_order.py

# code was heavily based on https://github.com/AliaksandrSiarohin/first-order-model
# Users should be careful about adopting these functions in any commercial matters.
# https://github.com/AliaksandrSiarohin/first-order-model/blob/master/LICENSE.md

import paddle
import paddle.nn as nn
import paddle.nn.functional as F


def SyncBatchNorm(*args, **kwargs):
    if paddle.distributed.get_world_size() > 1:
        return nn.SyncBatchNorm(*args, **kwargs)
    else:
        return nn.BatchNorm(*args, **kwargs)


class ImagePyramide(nn.Layer):
    """
    Create image pyramide for computing pyramide perceptual loss. See Sec 3.3
    """
    def __init__(self, scales, num_channels):
        super(ImagePyramide, self).__init__()
        self.downs = paddle.nn.LayerList()
        self.name_list = []
        for scale in scales:
            self.downs.add_sublayer(
                str(scale).replace('.', '-'),
                AntiAliasInterpolation2d(num_channels, scale))
            self.name_list.append(str(scale).replace('.', '-'))

    def forward(self, x):
        out_dict = {}
        for scale, down_module in zip(self.name_list, self.downs):
            out_dict['prediction_' +
                     str(scale).replace('-', '.')] = down_module(x)
        return out_dict


def detach_kp(kp):
    return {key: value.detach() for key, value in kp.items()}


def kp2gaussian(kp, spatial_size, kp_variance):
    """
    Transform a keypoint into gaussian like representation
    """
    mean = kp['value']

    coordinate_grid = make_coordinate_grid(spatial_size, mean.dtype)
    number_of_leading_dimensions = len(mean.shape) - 1
    shape = (1, ) * number_of_leading_dimensions + tuple(coordinate_grid.shape)
    repeats = tuple(mean.shape[:number_of_leading_dimensions]) + (1, 1, 1)
    coordinate_grid = coordinate_grid.reshape(shape)
    coordinate_grid = coordinate_grid.tile(repeats)

    # Preprocess kp shape
    shape = tuple(mean.shape[:number_of_leading_dimensions]) + (1, 1, 2)
    mean = mean.reshape(shape)

    mean_sub = (coordinate_grid - mean)

    out = paddle.exp(-0.5 * (mean_sub**2).sum(-1) / kp_variance)

    return out


def make_coordinate_grid(spatial_size, type='float32'):
    """
    Create a meshgrid [-1,1] x [-1,1] of given spatial_size.
    """
    h, w = spatial_size
    x = paddle.arange(w, dtype=type)  #.type(type)
    y = paddle.arange(h, dtype=type)  #.type(type)

    x = (2 * (x / (w - 1)) - 1)
    y = (2 * (y / (h - 1)) - 1)

    yy = paddle.tile(y.reshape([-1, 1]), [1, w])
    xx = paddle.tile(x.reshape([1, -1]), [h, 1])

    meshed = paddle.concat([xx.unsqueeze(2), yy.unsqueeze(2)], 2)

    return meshed


class ResBlock2d(nn.Layer):
    """
    Res block, preserve spatial resolution.
    """
    def __init__(self, in_features, kernel_size, padding):
        super(ResBlock2d, self).__init__()
        self.conv1 = nn.Conv2D(in_channels=in_features,
                               out_channels=in_features,
                               kernel_size=kernel_size,
                               padding=padding)
        self.conv2 = nn.Conv2D(in_channels=in_features,
                               out_channels=in_features,
                               kernel_size=kernel_size,
                               padding=padding)
        self.norm1 = SyncBatchNorm(in_features)
        self.norm2 = SyncBatchNorm(in_features)

    def forward(self, x):
        out = self.norm1(x)
        out = F.relu(out)
        out = self.conv1(out)
        out = self.norm2(out)
        out = F.relu(out)
        out = self.conv2(out)
        out += x
        return out


class MobileResBlock2d(nn.Layer):
    """
    Res block, preserve spatial resolution.
    """
    def __init__(self, in_features, kernel_size, padding):
        super(MobileResBlock2d, self).__init__()
        out_features = in_features * 2
        self.conv_pw = nn.Conv2D(in_channels=in_features,
                                 out_channels=out_features,
                                 kernel_size=1,
                                 padding=0,
                                 bias_attr=False)
        self.conv_dw = nn.Conv2D(in_channels=out_features,
                                 out_channels=out_features,
                                 kernel_size=kernel_size,
                                 padding=padding,
                                 groups=out_features,
                                 bias_attr=False)
        self.conv_pw_linear = nn.Conv2D(in_channels=out_features,
                                        out_channels=in_features,
                                        kernel_size=1,
                                        padding=0,
                                        bias_attr=False)
        self.norm1 = SyncBatchNorm(in_features)
        self.norm_pw = SyncBatchNorm(out_features)
        self.norm_dw = SyncBatchNorm(out_features)
        self.norm_pw_linear = SyncBatchNorm(in_features)

    def forward(self, x):
        out = self.norm1(x)
        out = F.relu(out)
        out = self.conv_pw(out)
        out = self.norm_pw(out)

        out = self.conv_dw(out)
        out = self.norm_dw(out)
        out = F.relu(out)

        out = self.conv_pw_linear(out)
        out = self.norm_pw_linear(out)
        out += x
        return out


class UpBlock2d(nn.Layer):
    """
    Upsampling block for use in decoder.
    """
    def __init__(self,
                 in_features,
                 out_features,
                 kernel_size=3,
                 padding=1,
                 groups=1):
        super(UpBlock2d, self).__init__()

        self.conv = nn.Conv2D(in_channels=in_features,
                              out_channels=out_features,
                              kernel_size=kernel_size,
                              padding=padding,
                              groups=groups)
        self.norm = SyncBatchNorm(out_features)

    def forward(self, x):
        out = F.interpolate(x, scale_factor=2)
        out = self.conv(out)
        out = self.norm(out)
        out = F.relu(out)
        return out


class MobileUpBlock2d(nn.Layer):
    """
    Upsampling block for use in decoder.
    """
    def __init__(self,
                 in_features,
                 out_features,
                 kernel_size=3,
                 padding=1,
                 groups=1):
        super(MobileUpBlock2d, self).__init__()

        self.conv = nn.Conv2D(in_channels=in_features,
                              out_channels=in_features,
                              kernel_size=kernel_size,
                              padding=padding,
                              groups=in_features,
                              bias_attr=False)
        self.conv1 = nn.Conv2D(in_channels=in_features,
                               out_channels=out_features,
                               kernel_size=1,
                               padding=0,
                               bias_attr=False)
        self.norm = SyncBatchNorm(in_features)
        self.norm1 = SyncBatchNorm(out_features)

    def forward(self, x):
        out = F.interpolate(x, scale_factor=2)
        out = self.conv(out)
        out = self.norm(out)
        out = F.relu(out)
        out = self.conv1(out)
        out = self.norm1(out)
        out = F.relu(out)
        return out


class DownBlock2d(nn.Layer):
    """
    Downsampling block for use in encoder.
    """
    def __init__(self,
                 in_features,
                 out_features,
                 kernel_size=3,
                 padding=1,
                 groups=1):
        super(DownBlock2d, self).__init__()
        self.conv = nn.Conv2D(in_channels=in_features,
                              out_channels=out_features,
                              kernel_size=kernel_size,
                              padding=padding,
                              groups=groups)
        self.norm = SyncBatchNorm(out_features)
        self.pool = nn.AvgPool2D(kernel_size=(2, 2))

    def forward(self, x):
        out = self.conv(x)
        out = self.norm(out)
        out = F.relu(out)
        out = self.pool(out)
        return out


class MobileDownBlock2d(nn.Layer):
    """
    Downsampling block for use in encoder.
    """
    def __init__(self,
                 in_features,
                 out_features,
                 kernel_size=3,
                 padding=1,
                 groups=1):
        super(MobileDownBlock2d, self).__init__()
        self.conv = nn.Conv2D(in_channels=in_features,
                              out_channels=in_features,
                              kernel_size=kernel_size,
                              padding=padding,
                              groups=in_features,
                              bias_attr=False)
        self.norm = SyncBatchNorm(in_features)
        self.pool = nn.AvgPool2D(kernel_size=(2, 2))

        self.conv1 = nn.Conv2D(in_features,
                               out_features,
                               kernel_size=1,
                               padding=0,
                               stride=1,
                               bias_attr=False)
        self.norm1 = SyncBatchNorm(out_features)

    def forward(self, x):
        out = self.conv(x)
        out = self.norm(out)
        out = F.relu(out)
        out = self.conv1(out)
        out = self.norm1(out)
        out = F.relu(out)
        out = self.pool(out)
        return out


class SameBlock2d(nn.Layer):
    """
    Simple block, preserve spatial resolution.
    """
    def __init__(self,
                 in_features,
                 out_features,
                 groups=1,
                 kernel_size=3,
                 padding=1,
                 mobile_net=False):
        super(SameBlock2d, self).__init__()
        self.conv = nn.Conv2D(in_channels=in_features,
                              out_channels=out_features,
                              kernel_size=kernel_size,
                              padding=padding,
                              groups=groups,
                              bias_attr=(mobile_net == False),
                              weight_attr=nn.initializer.KaimingUniform())
        self.norm = SyncBatchNorm(out_features)

    def forward(self, x):
        out = self.conv(x)
        out = self.norm(out)
        out = F.relu(out)
        return out


class Encoder(nn.Layer):
    """
    Hourglass Encoder
    """
    def __init__(self,
                 block_expansion,
                 in_features,
                 num_blocks=3,
                 max_features=256,
                 mobile_net=False):
        super(Encoder, self).__init__()

        down_blocks = []
        for i in range(num_blocks):
            if mobile_net:
                down_blocks.append(
                    MobileDownBlock2d(in_features if i == 0 else min(
                        max_features, block_expansion * (2**i)),
                                      min(max_features,
                                          block_expansion * (2**(i + 1))),
                                      kernel_size=3,
                                      padding=1))
            else:
                down_blocks.append(
                    DownBlock2d(in_features if i == 0 else min(
                        max_features, block_expansion * (2**i)),
                                min(max_features,
                                    block_expansion * (2**(i + 1))),
                                kernel_size=3,
                                padding=1))
        self.down_blocks = nn.LayerList(down_blocks)

    def forward(self, x):
        outs = [x]
        for down_block in self.down_blocks:
            outs.append(down_block(outs[-1]))
        return outs


class Decoder(nn.Layer):
    """
    Hourglass Decoder
    """
    def __init__(self,
                 block_expansion,
                 in_features,
                 num_blocks=3,
                 max_features=256,
                 mobile_net=False):
        super(Decoder, self).__init__()

        up_blocks = []

        for i in range(num_blocks)[::-1]:
            out_filters = min(max_features, block_expansion * (2**i))
            if mobile_net:
                in_filters = (1 if i == num_blocks - 1 else 2) * min(
                    max_features, block_expansion * (2**(i + 1)))
                up_blocks.append(
                    MobileUpBlock2d(in_filters,
                                    out_filters,
                                    kernel_size=3,
                                    padding=1))
            else:
                in_filters = (1 if i == num_blocks - 1 else 2) * min(
                    max_features, block_expansion * (2**(i + 1)))
                up_blocks.append(
                    UpBlock2d(in_filters, out_filters, kernel_size=3,
                              padding=1))

        self.up_blocks = nn.LayerList(up_blocks)
        self.out_filters = block_expansion + in_features

    def forward(self, x):
        out = x.pop()
        for up_block in self.up_blocks:
            out = up_block(out)
            skip = x.pop()
            out = paddle.concat([out, skip], axis=1)
        return out


class Hourglass(nn.Layer):
    """
    Hourglass architecture.
    """
    def __init__(self,
                 block_expansion,
                 in_features,
                 num_blocks=3,
                 max_features=256,
                 mobile_net=False):
        super(Hourglass, self).__init__()
        self.encoder = Encoder(block_expansion,
                               in_features,
                               num_blocks,
                               max_features,
                               mobile_net=mobile_net)
        self.decoder = Decoder(block_expansion,
                               in_features,
                               num_blocks,
                               max_features,
                               mobile_net=mobile_net)
        self.out_filters = self.decoder.out_filters

    def forward(self, x):
        return self.decoder(self.encoder(x))


class AntiAliasInterpolation2d(nn.Layer):
    """
    Band-limited downsampling, for better preservation of the input signal.
    """
    def __init__(self, channels, scale, mobile_net=False):
        super(AntiAliasInterpolation2d, self).__init__()
        if mobile_net:
            sigma = 0.25
            kernel_size = 3
        else:
            sigma = (1 / scale - 1) / 2
            kernel_size = 2 * round(sigma * 4) + 1
        self.ka = kernel_size // 2
        self.kb = self.ka - 1 if kernel_size % 2 == 0 else self.ka

        kernel_size = [kernel_size, kernel_size]
        sigma = [sigma, sigma]
        # The gaussian kernel is the product of the
        # gaussian function of each dimension.
        kernel = 1
        meshgrids = paddle.meshgrid(
            [paddle.arange(size, dtype='float32') for size in kernel_size])
        for size, std, mgrid in zip(kernel_size, sigma, meshgrids):
            mean = (size - 1) / 2
            kernel *= paddle.exp(-(mgrid - mean)**2 / (2 * std**2 + 1e-9))

        # Make sure sum of values in gaussian kernel equals 1.
        kernel = kernel / paddle.sum(kernel)
        # Reshape to depthwise convolutional weight
        kernel = kernel.reshape([1, 1, *kernel.shape])
        kernel = paddle.tile(kernel, [channels, *[1] * (kernel.dim() - 1)])

        self.register_buffer('weight', kernel)
        self.groups = channels
        self.scale = scale

    def forward(self, input):
        if self.scale == 1.0:
            return input

        out = F.pad(input, [self.ka, self.kb, self.ka, self.kb])
        out = F.conv2d(out, weight=self.weight, groups=self.groups)
        out.stop_gradient = False
        inv_scale = 1 / self.scale
        int_inv_scale = int(inv_scale)
        assert (inv_scale == int_inv_scale)
        # lite: fluid resize_nearest
        # out = paddle.fluid.layers.resize_nearest(out, scale=self.scale)
        out = out[:, :, ::int_inv_scale, ::int_inv_scale]
        # patch end
        return out