Port LivePortrait base

modules/config/inference_config.py

@@ -0,0 +1,8 @@
+
+import os
+
+current_file_path = os.path.abspath(__file__)
+current_directory = os.path.dirname(current_file_path)
+class InferenceConfig:
+    def __init__(self):
+        self.flag_use_half_precision: bool = False  # whether to use half precision

modules/config/models.yaml

@@ -0,0 +1,43 @@
+model_params:
+  appearance_feature_extractor_params: # the F in the paper
+    image_channel: 3
+    block_expansion: 64
+    num_down_blocks: 2
+    max_features: 512
+    reshape_channel: 32
+    reshape_depth: 16
+    num_resblocks: 6
+  motion_extractor_params: # the M in the paper
+    num_kp: 21
+    backbone: convnextv2_tiny
+  warping_module_params: # the W in the paper
+    num_kp: 21
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    reshape_channel: 32
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 32
+      max_features: 1024
+      num_blocks: 5
+      reshape_depth: 16
+      compress: 4
+  spade_generator_params: # the G in the paper
+    upscale: 2 # represents upsample factor 256x256 -> 512x512
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+  stitching_retargeting_module_params: # the S in the paper
+    stitching:
+      input_size: 126 # (21*3)*2
+      hidden_sizes: [128, 128, 64]
+      output_size: 65 # (21*3)+2(tx,ty)
+    lip:
+      input_size: 65 # (21*3)+2
+      hidden_sizes: [128, 128, 64]
+      output_size: 63 # (21*3)
+    eye:
+      input_size: 66 # (21*3)+3
+      hidden_sizes: [256, 256, 128, 128, 64]
+      output_size: 63 # (21*3)

modules/live_portrait/appearance_feature_extractor.py

@@ -0,0 +1,48 @@
+# coding: utf-8
+
+"""
+Appearance extractor(F) defined in paper, which maps the source image s to a 3D appearance feature volume.
+"""
+
+import torch
+from torch import nn
+from .util import SameBlock2d, DownBlock2d, ResBlock3d
+
+
+class AppearanceFeatureExtractor(nn.Module):
+
+    def __init__(self, image_channel, block_expansion, num_down_blocks, max_features, reshape_channel, reshape_depth, num_resblocks):
+        super(AppearanceFeatureExtractor, self).__init__()
+        self.image_channel = image_channel
+        self.block_expansion = block_expansion
+        self.num_down_blocks = num_down_blocks
+        self.max_features = max_features
+        self.reshape_channel = reshape_channel
+        self.reshape_depth = reshape_depth
+
+        self.first = SameBlock2d(image_channel, block_expansion, kernel_size=(3, 3), padding=(1, 1))
+
+        down_blocks = []
+        for i in range(num_down_blocks):
+            in_features = min(max_features, block_expansion * (2 ** i))
+            out_features = min(max_features, block_expansion * (2 ** (i + 1)))
+            down_blocks.append(DownBlock2d(in_features, out_features, kernel_size=(3, 3), padding=(1, 1)))
+        self.down_blocks = nn.ModuleList(down_blocks)
+
+        self.second = nn.Conv2d(in_channels=out_features, out_channels=max_features, kernel_size=1, stride=1)
+
+        self.resblocks_3d = torch.nn.Sequential()
+        for i in range(num_resblocks):
+            self.resblocks_3d.add_module('3dr' + str(i), ResBlock3d(reshape_channel, kernel_size=3, padding=1))
+
+    def forward(self, source_image):
+        out = self.first(source_image)  # Bx3x256x256 -> Bx64x256x256
+
+        for i in range(len(self.down_blocks)):
+            out = self.down_blocks[i](out)
+        out = self.second(out)
+        bs, c, h, w = out.shape  # ->Bx512x64x64
+
+        f_s = out.view(bs, self.reshape_channel, self.reshape_depth, h, w)  # ->Bx32x16x64x64
+        f_s = self.resblocks_3d(f_s)  # ->Bx32x16x64x64
+        return f_s

modules/live_portrait/convnextv2.py

@@ -0,0 +1,149 @@
+# coding: utf-8
+
+"""
+This moudle is adapted to the ConvNeXtV2 version for the extraction of implicit keypoints, poses, and expression deformation.
+"""
+
+import torch
+import torch.nn as nn
+# from timm.models.layers import trunc_normal_, DropPath
+from .util import LayerNorm, DropPath, trunc_normal_, GRN
+
+__all__ = ['convnextv2_tiny']
+
+
+class Block(nn.Module):
+    """ ConvNeXtV2 Block.
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+    """
+
+    def __init__(self, dim, drop_path=0.):
+        super().__init__()
+        self.dwconv = nn.Conv2d(dim, dim, kernel_size=7, padding=3, groups=dim)  # depthwise conv
+        self.norm = LayerNorm(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(dim, 4 * dim)  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.grn = GRN(4 * dim)
+        self.pwconv2 = nn.Linear(4 * dim, dim)
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+    def forward(self, x):
+        input = x
+        x = self.dwconv(x)
+        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+        x = self.norm(x)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.grn(x)
+        x = self.pwconv2(x)
+        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)
+
+        x = input + self.drop_path(x)
+        return x
+
+
+class ConvNeXtV2(nn.Module):
+    """ ConvNeXt V2
+
+    Args:
+        in_chans (int): Number of input image channels. Default: 3
+        num_classes (int): Number of classes for classification head. Default: 1000
+        depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
+        dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
+        drop_path_rate (float): Stochastic depth rate. Default: 0.
+        head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
+    """
+
+    def __init__(
+        self,
+        in_chans=3,
+        depths=[3, 3, 9, 3],
+        dims=[96, 192, 384, 768],
+        drop_path_rate=0.,
+        **kwargs
+    ):
+        super().__init__()
+        self.depths = depths
+        self.downsample_layers = nn.ModuleList()  # stem and 3 intermediate downsampling conv layers
+        stem = nn.Sequential(
+            nn.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),
+            LayerNorm(dims[0], eps=1e-6, data_format="channels_first")
+        )
+        self.downsample_layers.append(stem)
+        for i in range(3):
+            downsample_layer = nn.Sequential(
+                LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),
+                nn.Conv2d(dims[i], dims[i+1], kernel_size=2, stride=2),
+            )
+            self.downsample_layers.append(downsample_layer)
+
+        self.stages = nn.ModuleList()  # 4 feature resolution stages, each consisting of multiple residual blocks
+        dp_rates = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
+        cur = 0
+        for i in range(4):
+            stage = nn.Sequential(
+                *[Block(dim=dims[i], drop_path=dp_rates[cur + j]) for j in range(depths[i])]
+            )
+            self.stages.append(stage)
+            cur += depths[i]
+
+        self.norm = nn.LayerNorm(dims[-1], eps=1e-6)  # final norm layer
+
+        # NOTE: the output semantic items
+        num_bins = kwargs.get('num_bins', 66)
+        num_kp = kwargs.get('num_kp', 24)  # the number of implicit keypoints
+        self.fc_kp = nn.Linear(dims[-1], 3 * num_kp)  # implicit keypoints
+
+        # print('dims[-1]: ', dims[-1])
+        self.fc_scale = nn.Linear(dims[-1], 1)  # scale
+        self.fc_pitch = nn.Linear(dims[-1], num_bins)  # pitch bins
+        self.fc_yaw = nn.Linear(dims[-1], num_bins)  # yaw bins
+        self.fc_roll = nn.Linear(dims[-1], num_bins)  # roll bins
+        self.fc_t = nn.Linear(dims[-1], 3)  # translation
+        self.fc_exp = nn.Linear(dims[-1], 3 * num_kp)  # expression / delta
+
+    def _init_weights(self, m):
+        if isinstance(m, (nn.Conv2d, nn.Linear)):
+            trunc_normal_(m.weight, std=.02)
+            nn.init.constant_(m.bias, 0)
+
+    def forward_features(self, x):
+        for i in range(4):
+            x = self.downsample_layers[i](x)
+            x = self.stages[i](x)
+        return self.norm(x.mean([-2, -1]))  # global average pooling, (N, C, H, W) -> (N, C)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+
+        # implicit keypoints
+        kp = self.fc_kp(x)
+
+        # pose and expression deformation
+        pitch = self.fc_pitch(x)
+        yaw = self.fc_yaw(x)
+        roll = self.fc_roll(x)
+        t = self.fc_t(x)
+        exp = self.fc_exp(x)
+        scale = self.fc_scale(x)
+
+        ret_dct = {
+            'pitch': pitch,
+            'yaw': yaw,
+            'roll': roll,
+            't': t,
+            'exp': exp,
+            'scale': scale,
+
+            'kp': kp,  # canonical keypoint
+        }
+
+        return ret_dct
+
+
+def convnextv2_tiny(**kwargs):
+    model = ConvNeXtV2(depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], **kwargs)
+    return model

modules/live_portrait/dense_motion.py

@@ -0,0 +1,104 @@
+# coding: utf-8
+
+"""
+The module that predicting a dense motion from sparse motion representation given by kp_source and kp_driving
+"""
+
+from torch import nn
+import torch.nn.functional as F
+import torch
+from .util import Hourglass, make_coordinate_grid, kp2gaussian
+
+
+class DenseMotionNetwork(nn.Module):
+    def __init__(self, block_expansion, num_blocks, max_features, num_kp, feature_channel, reshape_depth, compress, estimate_occlusion_map=True):
+        super(DenseMotionNetwork, self).__init__()
+        self.hourglass = Hourglass(block_expansion=block_expansion, in_features=(num_kp+1)*(compress+1), max_features=max_features, num_blocks=num_blocks)  # ~60+G
+
+        self.mask = nn.Conv3d(self.hourglass.out_filters, num_kp + 1, kernel_size=7, padding=3)  # 65G! NOTE: computation cost is large
+        self.compress = nn.Conv3d(feature_channel, compress, kernel_size=1)  # 0.8G
+        self.norm = nn.BatchNorm3d(compress, affine=True)
+        self.num_kp = num_kp
+        self.flag_estimate_occlusion_map = estimate_occlusion_map
+
+        if self.flag_estimate_occlusion_map:
+            self.occlusion = nn.Conv2d(self.hourglass.out_filters*reshape_depth, 1, kernel_size=7, padding=3)
+        else:
+            self.occlusion = None
+
+    def create_sparse_motions(self, feature, kp_driving, kp_source):
+        bs, _, d, h, w = feature.shape  # (bs, 4, 16, 64, 64)
+        identity_grid = make_coordinate_grid((d, h, w), ref=kp_source)  # (16, 64, 64, 3)
+        identity_grid = identity_grid.view(1, 1, d, h, w, 3)  # (1, 1, d=16, h=64, w=64, 3)
+        coordinate_grid = identity_grid - kp_driving.view(bs, self.num_kp, 1, 1, 1, 3)
+
+        k = coordinate_grid.shape[1]
+
+        # NOTE: there lacks an one-order flow
+        driving_to_source = coordinate_grid + kp_source.view(bs, self.num_kp, 1, 1, 1, 3)    # (bs, num_kp, d, h, w, 3)
+
+        # adding background feature
+        identity_grid = identity_grid.repeat(bs, 1, 1, 1, 1, 1)
+        sparse_motions = torch.cat([identity_grid, driving_to_source], dim=1)  # (bs, 1+num_kp, d, h, w, 3)
+        return sparse_motions
+
+    def create_deformed_feature(self, feature, sparse_motions):
+        bs, _, d, h, w = feature.shape
+        feature_repeat = feature.unsqueeze(1).unsqueeze(1).repeat(1, self.num_kp+1, 1, 1, 1, 1, 1)      # (bs, num_kp+1, 1, c, d, h, w)
+        feature_repeat = feature_repeat.view(bs * (self.num_kp+1), -1, d, h, w)                         # (bs*(num_kp+1), c, d, h, w)
+        sparse_motions = sparse_motions.view((bs * (self.num_kp+1), d, h, w, -1))                       # (bs*(num_kp+1), d, h, w, 3)
+        sparse_deformed = F.grid_sample(feature_repeat, sparse_motions, align_corners=False)
+        sparse_deformed = sparse_deformed.view((bs, self.num_kp+1, -1, d, h, w))                        # (bs, num_kp+1, c, d, h, w)
+
+        return sparse_deformed
+
+    def create_heatmap_representations(self, feature, kp_driving, kp_source):
+        spatial_size = feature.shape[3:]  # (d=16, h=64, w=64)
+        gaussian_driving = kp2gaussian(kp_driving, spatial_size=spatial_size, kp_variance=0.01)  # (bs, num_kp, d, h, w)
+        gaussian_source = kp2gaussian(kp_source, spatial_size=spatial_size, kp_variance=0.01)  # (bs, num_kp, d, h, w)
+        heatmap = gaussian_driving - gaussian_source  # (bs, num_kp, d, h, w)
+
+        # adding background feature
+        zeros = torch.zeros(heatmap.shape[0], 1, spatial_size[0], spatial_size[1], spatial_size[2]).type(heatmap.dtype).to(heatmap.device)
+        heatmap = torch.cat([zeros, heatmap], dim=1)
+        heatmap = heatmap.unsqueeze(2)         # (bs, 1+num_kp, 1, d, h, w)
+        return heatmap
+
+    def forward(self, feature, kp_driving, kp_source):
+        bs, _, d, h, w = feature.shape  # (bs, 32, 16, 64, 64)
+
+        feature = self.compress(feature)  # (bs, 4, 16, 64, 64)
+        feature = self.norm(feature)  # (bs, 4, 16, 64, 64)
+        feature = F.relu(feature)  # (bs, 4, 16, 64, 64)
+
+        out_dict = dict()
+
+        # 1. deform 3d feature
+        sparse_motion = self.create_sparse_motions(feature, kp_driving, kp_source)  # (bs, 1+num_kp, d, h, w, 3)
+        deformed_feature = self.create_deformed_feature(feature, sparse_motion)  # (bs, 1+num_kp, c=4, d=16, h=64, w=64)
+
+        # 2. (bs, 1+num_kp, d, h, w)
+        heatmap = self.create_heatmap_representations(deformed_feature, kp_driving, kp_source)  # (bs, 1+num_kp, 1, d, h, w)
+
+        input = torch.cat([heatmap, deformed_feature], dim=2)  # (bs, 1+num_kp, c=5, d=16, h=64, w=64)
+        input = input.view(bs, -1, d, h, w)  # (bs, (1+num_kp)*c=105, d=16, h=64, w=64)
+
+        prediction = self.hourglass(input)
+
+        mask = self.mask(prediction)
+        mask = F.softmax(mask, dim=1)  # (bs, 1+num_kp, d=16, h=64, w=64)
+        out_dict['mask'] = mask
+        mask = mask.unsqueeze(2)                                   # (bs, num_kp+1, 1, d, h, w)
+        sparse_motion = sparse_motion.permute(0, 1, 5, 2, 3, 4)    # (bs, num_kp+1, 3, d, h, w)
+        deformation = (sparse_motion * mask).sum(dim=1)            # (bs, 3, d, h, w)  mask take effect in this place
+        deformation = deformation.permute(0, 2, 3, 4, 1)           # (bs, d, h, w, 3)
+
+        out_dict['deformation'] = deformation
+
+        if self.flag_estimate_occlusion_map:
+            bs, _, d, h, w = prediction.shape
+            prediction_reshape = prediction.view(bs, -1, h, w)
+            occlusion_map = torch.sigmoid(self.occlusion(prediction_reshape))  # Bx1x64x64
+            out_dict['occlusion_map'] = occlusion_map
+
+        return out_dict

modules/live_portrait/motion_extractor.py

@@ -0,0 +1,35 @@
+# coding: utf-8
+
+"""
+Motion extractor(M), which directly predicts the canonical keypoints, head pose and expression deformation of the input image
+"""
+
+from torch import nn
+import torch
+
+from .convnextv2 import convnextv2_tiny
+from .util import filter_state_dict
+
+model_dict = {
+    'convnextv2_tiny': convnextv2_tiny,
+}
+
+
+class MotionExtractor(nn.Module):
+    def __init__(self, **kwargs):
+        super(MotionExtractor, self).__init__()
+
+        # default is convnextv2_base
+        backbone = kwargs.get('backbone', 'convnextv2_tiny')
+        self.detector = model_dict.get(backbone)(**kwargs)
+
+    def load_pretrained(self, init_path: str):
+        if init_path not in (None, ''):
+            state_dict = torch.load(init_path, map_location=lambda storage, loc: storage)['model']
+            state_dict = filter_state_dict(state_dict, remove_name='head')
+            ret = self.detector.load_state_dict(state_dict, strict=False)
+            print(f'Load pretrained model from {init_path}, ret: {ret}')
+
+    def forward(self, x):
+        out = self.detector(x)
+        return out

modules/live_portrait/spade_generator.py

@@ -0,0 +1,59 @@
+# coding: utf-8
+
+"""
+Spade decoder(G) defined in the paper, which input the warped feature to generate the animated image.
+"""
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from .util import SPADEResnetBlock
+
+
+class SPADEDecoder(nn.Module):
+    def __init__(self, upscale=1, max_features=256, block_expansion=64, out_channels=64, num_down_blocks=2):
+        for i in range(num_down_blocks):
+            input_channels = min(max_features, block_expansion * (2 ** (i + 1)))
+        self.upscale = upscale
+        super().__init__()
+        norm_G = 'spadespectralinstance'
+        label_num_channels = input_channels  # 256
+
+        self.fc = nn.Conv2d(input_channels, 2 * input_channels, 3, padding=1)
+        self.G_middle_0 = SPADEResnetBlock(2 * input_channels, 2 * input_channels, norm_G, label_num_channels)
+        self.G_middle_1 = SPADEResnetBlock(2 * input_channels, 2 * input_channels, norm_G, label_num_channels)
+        self.G_middle_2 = SPADEResnetBlock(2 * input_channels, 2 * input_channels, norm_G, label_num_channels)
+        self.G_middle_3 = SPADEResnetBlock(2 * input_channels, 2 * input_channels, norm_G, label_num_channels)
+        self.G_middle_4 = SPADEResnetBlock(2 * input_channels, 2 * input_channels, norm_G, label_num_channels)
+        self.G_middle_5 = SPADEResnetBlock(2 * input_channels, 2 * input_channels, norm_G, label_num_channels)
+        self.up_0 = SPADEResnetBlock(2 * input_channels, input_channels, norm_G, label_num_channels)
+        self.up_1 = SPADEResnetBlock(input_channels, out_channels, norm_G, label_num_channels)
+        self.up = nn.Upsample(scale_factor=2)
+
+        if self.upscale is None or self.upscale <= 1:
+            self.conv_img = nn.Conv2d(out_channels, 3, 3, padding=1)
+        else:
+            self.conv_img = nn.Sequential(
+                nn.Conv2d(out_channels, 3 * (2 * 2), kernel_size=3, padding=1),
+                nn.PixelShuffle(upscale_factor=2)
+            )
+
+    def forward(self, feature):
+        seg = feature  # Bx256x64x64
+        x = self.fc(feature)  # Bx512x64x64
+        x = self.G_middle_0(x, seg)
+        x = self.G_middle_1(x, seg)
+        x = self.G_middle_2(x, seg)
+        x = self.G_middle_3(x, seg)
+        x = self.G_middle_4(x, seg)
+        x = self.G_middle_5(x, seg)
+
+        x = self.up(x)  # Bx512x64x64 -> Bx512x128x128
+        x = self.up_0(x, seg)  # Bx512x128x128 -> Bx256x128x128
+        x = self.up(x)  # Bx256x128x128 -> Bx256x256x256
+        x = self.up_1(x, seg)  # Bx256x256x256 -> Bx64x256x256
+
+        x = self.conv_img(F.leaky_relu(x, 2e-1))  # Bx64x256x256 -> Bx3xHxW
+        x = torch.sigmoid(x)  # Bx3xHxW
+
+        return x

modules/live_portrait/stitching_retargeting_network.py

@@ -0,0 +1,38 @@
+# coding: utf-8
+
+"""
+Stitching module(S) and two retargeting live_portrait(R) defined in the paper.
+
+- The stitching module pastes the animated portrait back into the original image space without pixel misalignment, such as in
+the stitching region.
+
+- The eyes retargeting module is designed to address the issue of incomplete eye closure during cross-id reenactment, especially
+when a person with small eyes drives a person with larger eyes.
+
+- The lip retargeting module is designed similarly to the eye retargeting module, and can also normalize the input by ensuring that
+the lips are in a closed state, which facilitates better animation driving.
+"""
+from torch import nn
+
+
+class StitchingRetargetingNetwork(nn.Module):
+    def __init__(self, input_size, hidden_sizes, output_size):
+        super(StitchingRetargetingNetwork, self).__init__()
+        layers = []
+        for i in range(len(hidden_sizes)):
+            if i == 0:
+                layers.append(nn.Linear(input_size, hidden_sizes[i]))
+            else:
+                layers.append(nn.Linear(hidden_sizes[i - 1], hidden_sizes[i]))
+            layers.append(nn.ReLU(inplace=True))
+        layers.append(nn.Linear(hidden_sizes[-1], output_size))
+        self.mlp = nn.Sequential(*layers)
+
+    def initialize_weights_to_zero(self):
+        for m in self.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.zeros_(m.weight)
+                nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        return self.mlp(x)

modules/live_portrait/util.py

@@ -0,0 +1,441 @@
+# coding: utf-8
+
+"""
+This file defines various neural network live_portrait and utility functions, including convolutional and residual blocks,
+normalizations, and functions for spatial transformation and tensor manipulation.
+"""
+
+from torch import nn
+import torch.nn.functional as F
+import torch
+import torch.nn.utils.spectral_norm as spectral_norm
+import math
+import warnings
+
+
+def kp2gaussian(kp, spatial_size, kp_variance):
+    """
+    Transform a keypoint into gaussian like representation
+    """
+    mean = kp
+
+    coordinate_grid = make_coordinate_grid(spatial_size, mean)
+    number_of_leading_dimensions = len(mean.shape) - 1
+    shape = (1,) * number_of_leading_dimensions + coordinate_grid.shape
+    coordinate_grid = coordinate_grid.view(*shape)
+    repeats = mean.shape[:number_of_leading_dimensions] + (1, 1, 1, 1)
+    coordinate_grid = coordinate_grid.repeat(*repeats)
+
+    # Preprocess kp shape
+    shape = mean.shape[:number_of_leading_dimensions] + (1, 1, 1, 3)
+    mean = mean.view(*shape)
+
+    mean_sub = (coordinate_grid - mean)
+
+    out = torch.exp(-0.5 * (mean_sub ** 2).sum(-1) / kp_variance)
+
+    return out
+
+
+def make_coordinate_grid(spatial_size, ref, **kwargs):
+    d, h, w = spatial_size
+    x = torch.arange(w).type(ref.dtype).to(ref.device)
+    y = torch.arange(h).type(ref.dtype).to(ref.device)
+    z = torch.arange(d).type(ref.dtype).to(ref.device)
+
+    # NOTE: must be right-down-in
+    x = (2 * (x / (w - 1)) - 1)  # the x axis faces to the right
+    y = (2 * (y / (h - 1)) - 1)  # the y axis faces to the bottom
+    z = (2 * (z / (d - 1)) - 1)  # the z axis faces to the inner
+
+    yy = y.view(1, -1, 1).repeat(d, 1, w)
+    xx = x.view(1, 1, -1).repeat(d, h, 1)
+    zz = z.view(-1, 1, 1).repeat(1, h, w)
+
+    meshed = torch.cat([xx.unsqueeze_(3), yy.unsqueeze_(3), zz.unsqueeze_(3)], 3)
+
+    return meshed
+
+
+class ConvT2d(nn.Module):
+    """
+    Upsampling block for use in decoder.
+    """
+
+    def __init__(self, in_features, out_features, kernel_size=3, stride=2, padding=1, output_padding=1):
+        super(ConvT2d, self).__init__()
+
+        self.convT = nn.ConvTranspose2d(in_features, out_features, kernel_size=kernel_size, stride=stride,
+                                        padding=padding, output_padding=output_padding)
+        self.norm = nn.InstanceNorm2d(out_features)
+
+    def forward(self, x):
+        out = self.convT(x)
+        out = self.norm(out)
+        out = F.leaky_relu(out)
+        return out
+
+
+class ResBlock3d(nn.Module):
+    """
+    Res block, preserve spatial resolution.
+    """
+
+    def __init__(self, in_features, kernel_size, padding):
+        super(ResBlock3d, self).__init__()
+        self.conv1 = nn.Conv3d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size, padding=padding)
+        self.conv2 = nn.Conv3d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size, padding=padding)
+        self.norm1 = nn.BatchNorm3d(in_features, affine=True)
+        self.norm2 = nn.BatchNorm3d(in_features, affine=True)
+
+    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 UpBlock3d(nn.Module):
+    """
+    Upsampling block for use in decoder.
+    """
+
+    def __init__(self, in_features, out_features, kernel_size=3, padding=1, groups=1):
+        super(UpBlock3d, self).__init__()
+
+        self.conv = nn.Conv3d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size,
+                              padding=padding, groups=groups)
+        self.norm = nn.BatchNorm3d(out_features, affine=True)
+
+    def forward(self, x):
+        out = F.interpolate(x, scale_factor=(1, 2, 2))
+        out = self.conv(out)
+        out = self.norm(out)
+        out = F.relu(out)
+        return out
+
+
+class DownBlock2d(nn.Module):
+    """
+    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 = nn.BatchNorm2d(out_features, affine=True)
+        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 DownBlock3d(nn.Module):
+    """
+    Downsampling block for use in encoder.
+    """
+
+    def __init__(self, in_features, out_features, kernel_size=3, padding=1, groups=1):
+        super(DownBlock3d, self).__init__()
+        '''
+        self.conv = nn.Conv3d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size,
+                                padding=padding, groups=groups, stride=(1, 2, 2))
+        '''
+        self.conv = nn.Conv3d(in_channels=in_features, out_channels=out_features, kernel_size=kernel_size,
+                              padding=padding, groups=groups)
+        self.norm = nn.BatchNorm3d(out_features, affine=True)
+        self.pool = nn.AvgPool3d(kernel_size=(1, 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 SameBlock2d(nn.Module):
+    """
+    Simple block, preserve spatial resolution.
+    """
+
+    def __init__(self, in_features, out_features, groups=1, kernel_size=3, padding=1, lrelu=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)
+        self.norm = nn.BatchNorm2d(out_features, affine=True)
+        if lrelu:
+            self.ac = nn.LeakyReLU()
+        else:
+            self.ac = nn.ReLU()
+
+    def forward(self, x):
+        out = self.conv(x)
+        out = self.norm(out)
+        out = self.ac(out)
+        return out
+
+
+class Encoder(nn.Module):
+    """
+    Hourglass Encoder
+    """
+
+    def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256):
+        super(Encoder, self).__init__()
+
+        down_blocks = []
+        for i in range(num_blocks):
+            down_blocks.append(DownBlock3d(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.ModuleList(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.Module):
+    """
+    Hourglass Decoder
+    """
+
+    def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256):
+        super(Decoder, self).__init__()
+
+        up_blocks = []
+
+        for i in range(num_blocks)[::-1]:
+            in_filters = (1 if i == num_blocks - 1 else 2) * min(max_features, block_expansion * (2 ** (i + 1)))
+            out_filters = min(max_features, block_expansion * (2 ** i))
+            up_blocks.append(UpBlock3d(in_filters, out_filters, kernel_size=3, padding=1))
+
+        self.up_blocks = nn.ModuleList(up_blocks)
+        self.out_filters = block_expansion + in_features
+
+        self.conv = nn.Conv3d(in_channels=self.out_filters, out_channels=self.out_filters, kernel_size=3, padding=1)
+        self.norm = nn.BatchNorm3d(self.out_filters, affine=True)
+
+    def forward(self, x):
+        out = x.pop()
+        for up_block in self.up_blocks:
+            out = up_block(out)
+            skip = x.pop()
+            out = torch.cat([out, skip], dim=1)
+        out = self.conv(out)
+        out = self.norm(out)
+        out = F.relu(out)
+        return out
+
+
+class Hourglass(nn.Module):
+    """
+    Hourglass architecture.
+    """
+
+    def __init__(self, block_expansion, in_features, num_blocks=3, max_features=256):
+        super(Hourglass, self).__init__()
+        self.encoder = Encoder(block_expansion, in_features, num_blocks, max_features)
+        self.decoder = Decoder(block_expansion, in_features, num_blocks, max_features)
+        self.out_filters = self.decoder.out_filters
+
+    def forward(self, x):
+        return self.decoder(self.encoder(x))
+
+
+class SPADE(nn.Module):
+    def __init__(self, norm_nc, label_nc):
+        super().__init__()
+
+        self.param_free_norm = nn.InstanceNorm2d(norm_nc, affine=False)
+        nhidden = 128
+
+        self.mlp_shared = nn.Sequential(
+            nn.Conv2d(label_nc, nhidden, kernel_size=3, padding=1),
+            nn.ReLU())
+        self.mlp_gamma = nn.Conv2d(nhidden, norm_nc, kernel_size=3, padding=1)
+        self.mlp_beta = nn.Conv2d(nhidden, norm_nc, kernel_size=3, padding=1)
+
+    def forward(self, x, segmap):
+        normalized = self.param_free_norm(x)
+        segmap = F.interpolate(segmap, size=x.size()[2:], mode='nearest')
+        actv = self.mlp_shared(segmap)
+        gamma = self.mlp_gamma(actv)
+        beta = self.mlp_beta(actv)
+        out = normalized * (1 + gamma) + beta
+        return out
+
+
+class SPADEResnetBlock(nn.Module):
+    def __init__(self, fin, fout, norm_G, label_nc, use_se=False, dilation=1):
+        super().__init__()
+        # Attributes
+        self.learned_shortcut = (fin != fout)
+        fmiddle = min(fin, fout)
+        self.use_se = use_se
+        # create conv layers
+        self.conv_0 = nn.Conv2d(fin, fmiddle, kernel_size=3, padding=dilation, dilation=dilation)
+        self.conv_1 = nn.Conv2d(fmiddle, fout, kernel_size=3, padding=dilation, dilation=dilation)
+        if self.learned_shortcut:
+            self.conv_s = nn.Conv2d(fin, fout, kernel_size=1, bias=False)
+        # apply spectral norm if specified
+        if 'spectral' in norm_G:
+            self.conv_0 = spectral_norm(self.conv_0)
+            self.conv_1 = spectral_norm(self.conv_1)
+            if self.learned_shortcut:
+                self.conv_s = spectral_norm(self.conv_s)
+        # define normalization layers
+        self.norm_0 = SPADE(fin, label_nc)
+        self.norm_1 = SPADE(fmiddle, label_nc)
+        if self.learned_shortcut:
+            self.norm_s = SPADE(fin, label_nc)
+
+    def forward(self, x, seg1):
+        x_s = self.shortcut(x, seg1)
+        dx = self.conv_0(self.actvn(self.norm_0(x, seg1)))
+        dx = self.conv_1(self.actvn(self.norm_1(dx, seg1)))
+        out = x_s + dx
+        return out
+
+    def shortcut(self, x, seg1):
+        if self.learned_shortcut:
+            x_s = self.conv_s(self.norm_s(x, seg1))
+        else:
+            x_s = x
+        return x_s
+
+    def actvn(self, x):
+        return F.leaky_relu(x, 2e-1)
+
+
+def filter_state_dict(state_dict, remove_name='fc'):
+    new_state_dict = {}
+    for key in state_dict:
+        if remove_name in key:
+            continue
+        new_state_dict[key] = state_dict[key]
+    return new_state_dict
+
+
+class GRN(nn.Module):
+    """ GRN (Global Response Normalization) layer
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim))
+        self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim))
+
+    def forward(self, x):
+        Gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
+        Nx = Gx / (Gx.mean(dim=-1, keepdim=True) + 1e-6)
+        return self.gamma * (x * Nx) + self.beta + x
+
+
+class LayerNorm(nn.Module):
+    r""" LayerNorm that supports two data formats: channels_last (default) or channels_first.
+    The ordering of the dimensions in the inputs. channels_last corresponds to inputs with
+    shape (batch_size, height, width, channels) while channels_first corresponds to inputs
+    with shape (batch_size, channels, height, width).
+    """
+
+    def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(normalized_shape))
+        self.bias = nn.Parameter(torch.zeros(normalized_shape))
+        self.eps = eps
+        self.data_format = data_format
+        if self.data_format not in ["channels_last", "channels_first"]:
+            raise NotImplementedError
+        self.normalized_shape = (normalized_shape, )
+
+    def forward(self, x):
+        if self.data_format == "channels_last":
+            return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        elif self.data_format == "channels_first":
+            u = x.mean(1, keepdim=True)
+            s = (x - u).pow(2).mean(1, keepdim=True)
+            x = (x - u) / torch.sqrt(s + self.eps)
+            x = self.weight[:, None, None] * x + self.bias[:, None, None]
+            return x
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                      "The distribution of values may be incorrect.",
+                      stacklevel=2)
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def drop_path(x, drop_prob=0., training=False, scale_by_keep=True):
+    """ Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+
+    """
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0 and scale_by_keep:
+        random_tensor.div_(keep_prob)
+    return x * random_tensor
+
+
+class DropPath(nn.Module):
+    """ Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob=None, scale_by_keep=True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)

modules/live_portrait/warping_network.py

@@ -0,0 +1,77 @@
+# coding: utf-8
+
+"""
+Warping field estimator(W) defined in the paper, which generates a warping field using the implicit
+keypoint representations x_s and x_d, and employs this flow field to warp the source feature volume f_s.
+"""
+
+from torch import nn
+import torch.nn.functional as F
+from .util import SameBlock2d
+from .dense_motion import DenseMotionNetwork
+
+
+class WarpingNetwork(nn.Module):
+    def __init__(
+        self,
+        num_kp,
+        block_expansion,
+        max_features,
+        num_down_blocks,
+        reshape_channel,
+        estimate_occlusion_map=False,
+        dense_motion_params=None,
+        **kwargs
+    ):
+        super(WarpingNetwork, self).__init__()
+
+        self.upscale = kwargs.get('upscale', 1)
+        self.flag_use_occlusion_map = kwargs.get('flag_use_occlusion_map', True)
+
+        if dense_motion_params is not None:
+            self.dense_motion_network = DenseMotionNetwork(
+                num_kp=num_kp,
+                feature_channel=reshape_channel,
+                estimate_occlusion_map=estimate_occlusion_map,
+                **dense_motion_params
+            )
+        else:
+            self.dense_motion_network = None
+
+        self.third = SameBlock2d(max_features, block_expansion * (2 ** num_down_blocks), kernel_size=(3, 3), padding=(1, 1), lrelu=True)
+        self.fourth = nn.Conv2d(in_channels=block_expansion * (2 ** num_down_blocks), out_channels=block_expansion * (2 ** num_down_blocks), kernel_size=1, stride=1)
+
+        self.estimate_occlusion_map = estimate_occlusion_map
+
+    def deform_input(self, inp, deformation):
+        return F.grid_sample(inp, deformation, align_corners=False)
+
+    def forward(self, feature_3d, kp_driving, kp_source):
+        if self.dense_motion_network is not None:
+            # Feature warper, Transforming feature representation according to deformation and occlusion
+            dense_motion = self.dense_motion_network(
+                feature=feature_3d, kp_driving=kp_driving, kp_source=kp_source
+            )
+            if 'occlusion_map' in dense_motion:
+                occlusion_map = dense_motion['occlusion_map']  # Bx1x64x64
+            else:
+                occlusion_map = None
+
+            deformation = dense_motion['deformation']  # Bx16x64x64x3
+            out = self.deform_input(feature_3d, deformation)  # Bx32x16x64x64
+
+            bs, c, d, h, w = out.shape  # Bx32x16x64x64
+            out = out.view(bs, c * d, h, w)  # -> Bx512x64x64
+            out = self.third(out)  # -> Bx256x64x64
+            out = self.fourth(out)  # -> Bx256x64x64
+
+            if self.flag_use_occlusion_map and (occlusion_map is not None):
+                out = out * occlusion_map
+
+        ret_dct = {
+            'occlusion_map': occlusion_map,
+            'deformation': deformation,
+            'out': out,
+        }
+
+        return ret_dct

modules/live_portrait_wrapper.py

@@ -0,0 +1,150 @@
+import numpy as np
+import torch
+
+from .utils.helper import concat_feat
+from .utils.camera import headpose_pred_to_degree, get_rotation_matrix
+from .config.inference_config import InferenceConfig
+
+class LivePortraitWrapper(object):
+
+    def __init__(self, cfg: InferenceConfig, appearance_feature_extractor, motion_extractor,
+                                            warping_module, spade_generator, stitching_retargeting_module):
+
+        self.appearance_feature_extractor = appearance_feature_extractor
+        self.motion_extractor = motion_extractor
+        self.warping_module = warping_module
+        self.spade_generator = spade_generator
+        self.stitching_retargeting_module = stitching_retargeting_module
+
+        self.cfg = cfg
+
+    def extract_feature_3d(self, x: torch.Tensor) -> torch.Tensor:
+        """ get the appearance feature of the image by F
+        x: Bx3xHxW, normalized to 0~1
+        """
+        with torch.no_grad():
+            feature_3d = self.appearance_feature_extractor(x)
+
+        return feature_3d.float()
+
+    def get_kp_info(self, x: torch.Tensor, **kwargs) -> dict:
+        """ get the implicit keypoint information
+        x: Bx3xHxW, normalized to 0~1
+        flag_refine_info: whether to trandform the pose to degrees and the dimention of the reshape
+        return: A dict contains keys: 'pitch', 'yaw', 'roll', 't', 'exp', 'scale', 'kp'
+        """
+        with torch.no_grad():
+            kp_info = self.motion_extractor(x)
+
+            if self.cfg.flag_use_half_precision:
+                # float the dict
+                for k, v in kp_info.items():
+                    if isinstance(v, torch.Tensor):
+                        kp_info[k] = v.float()
+
+        flag_refine_info: bool = kwargs.get('flag_refine_info', True)
+        if flag_refine_info:
+            bs = kp_info['kp'].shape[0]
+            kp_info['pitch'] = headpose_pred_to_degree(kp_info['pitch'])[:, None]  # Bx1
+            kp_info['yaw'] = headpose_pred_to_degree(kp_info['yaw'])[:, None]  # Bx1
+            kp_info['roll'] = headpose_pred_to_degree(kp_info['roll'])[:, None]  # Bx1
+            kp_info['kp'] = kp_info['kp'].reshape(bs, -1, 3)  # BxNx3
+            kp_info['exp'] = kp_info['exp'].reshape(bs, -1, 3)  # BxNx3
+
+        return kp_info
+    def transform_keypoint(self, kp_info: dict):
+        """
+        transform the implicit keypoints with the pose, shift, and expression deformation
+        kp: BxNx3
+        """
+        kp = kp_info['kp']    # (bs, k, 3)
+        pitch, yaw, roll = kp_info['pitch'], kp_info['yaw'], kp_info['roll']
+
+        t, exp = kp_info['t'], kp_info['exp']
+        scale = kp_info['scale']
+
+        pitch = headpose_pred_to_degree(pitch)
+        yaw = headpose_pred_to_degree(yaw)
+        roll = headpose_pred_to_degree(roll)
+
+        bs = kp.shape[0]
+        if kp.ndim == 2:
+            num_kp = kp.shape[1] // 3  # Bx(num_kpx3)
+        else:
+            num_kp = kp.shape[1]  # Bxnum_kpx3
+
+        rot_mat = get_rotation_matrix(pitch, yaw, roll)    # (bs, 3, 3)
+
+        # Eqn.2: s * (R * x_c,s + exp) + t
+        kp_transformed = kp.view(bs, num_kp, 3) @ rot_mat + exp.view(bs, num_kp, 3)
+        kp_transformed *= scale[..., None]  # (bs, k, 3) * (bs, 1, 1) = (bs, k, 3)
+        kp_transformed[:, :, 0:2] += t[:, None, 0:2]  # remove z, only apply tx ty
+
+        return kp_transformed
+
+    def stitch(self, kp_source: torch.Tensor, kp_driving: torch.Tensor) -> torch.Tensor:
+        """
+        kp_source: BxNx3
+        kp_driving: BxNx3
+        Return: Bx(3*num_kp+2)
+        """
+        feat_stiching = concat_feat(kp_source, kp_driving)
+
+        with torch.no_grad():
+            delta = self.stitching_retargeting_module['stitching'](feat_stiching)
+
+        return delta
+
+    def stitching(self, kp_source: torch.Tensor, kp_driving: torch.Tensor) -> torch.Tensor:
+        """ conduct the stitching
+        kp_source: Bxnum_kpx3
+        kp_driving: Bxnum_kpx3
+        """
+
+        if self.stitching_retargeting_module is not None:
+
+            bs, num_kp = kp_source.shape[:2]
+
+            kp_driving_new = kp_driving.clone()
+            delta = self.stitch(kp_source, kp_driving_new)
+
+            delta_exp = delta[..., :3*num_kp].reshape(bs, num_kp, 3)  # 1x20x3
+            delta_tx_ty = delta[..., 3*num_kp:3*num_kp+2].reshape(bs, 1, 2)  # 1x1x2
+
+            kp_driving_new += delta_exp
+            kp_driving_new[..., :2] += delta_tx_ty
+
+            return kp_driving_new
+
+        return kp_driving
+
+    def warp_decode(self, feature_3d: torch.Tensor, kp_source: torch.Tensor, kp_driving: torch.Tensor) -> torch.Tensor:
+        """ get the image after the warping of the implicit keypoints
+        feature_3d: Bx32x16x64x64, feature volume
+        kp_source: BxNx3
+        kp_driving: BxNx3
+        """
+        # The line 18 in Algorithm 1: D(W(f_s; x_s, x′_d,i)）
+        with torch.no_grad():
+            # get decoder input
+            ret_dct = self.warping_module(feature_3d, kp_source=kp_source, kp_driving=kp_driving)
+            # decode
+            ret_dct['out'] = self.spade_generator(feature=ret_dct['out'])
+
+            # float the dict
+            if self.cfg.flag_use_half_precision:
+                for k, v in ret_dct.items():
+                    if isinstance(v, torch.Tensor):
+                        ret_dct[k] = v.float()
+
+        return ret_dct
+
+    def parse_output(self, out: torch.Tensor) -> np.ndarray:
+        """ construct the output as standard
+        return: 1xHxWx3, uint8
+        """
+        out = np.transpose(out.data.cpu().numpy(), [0, 2, 3, 1])  # 1x3xHxW -> 1xHxWx3
+        out = np.clip(out, 0, 1)  # clip to 0~1
+        out = np.clip(out * 255, 0, 255).astype(np.uint8)  # 0~1 -> 0~255
+
+        return out

modules/utils/camera.py

@@ -0,0 +1,75 @@
+# coding: utf-8
+
+"""
+functions for processing and transforming 3D facial keypoints
+"""
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+PI = np.pi
+
+
+def headpose_pred_to_degree(pred):
+    """
+    pred: (bs, 66) or (bs, 1) or others
+    """
+    if pred.ndim > 1 and pred.shape[1] == 66:
+        # NOTE: note that the average is modified to 97.5
+        device = pred.device
+        idx_tensor = [idx for idx in range(0, 66)]
+        idx_tensor = torch.FloatTensor(idx_tensor).to(device)
+        pred = F.softmax(pred, dim=1)
+        degree = torch.sum(pred*idx_tensor, axis=1) * 3 - 97.5
+
+        return degree
+
+    return pred
+
+
+def get_rotation_matrix(pitch_, yaw_, roll_):
+    """ the input is in degree
+    """
+    # calculate the rotation matrix: vps @ rot
+
+    # transform to radian
+    pitch = pitch_ / 180 * PI
+    yaw = yaw_ / 180 * PI
+    roll = roll_ / 180 * PI
+
+    device = pitch.device
+
+    if pitch.ndim == 1:
+        pitch = pitch.unsqueeze(1)
+    if yaw.ndim == 1:
+        yaw = yaw.unsqueeze(1)
+    if roll.ndim == 1:
+        roll = roll.unsqueeze(1)
+
+    # calculate the euler matrix
+    bs = pitch.shape[0]
+    ones = torch.ones([bs, 1]).to(device)
+    zeros = torch.zeros([bs, 1]).to(device)
+    x, y, z = pitch, yaw, roll
+
+    rot_x = torch.cat([
+        ones, zeros, zeros,
+        zeros, torch.cos(x), -torch.sin(x),
+        zeros, torch.sin(x), torch.cos(x)
+    ], dim=1).reshape([bs, 3, 3])
+
+    rot_y = torch.cat([
+        torch.cos(y), zeros, torch.sin(y),
+        zeros, ones, zeros,
+        -torch.sin(y), zeros, torch.cos(y)
+    ], dim=1).reshape([bs, 3, 3])
+
+    rot_z = torch.cat([
+        torch.cos(z), -torch.sin(z), zeros,
+        torch.sin(z), torch.cos(z), zeros,
+        zeros, zeros, ones
+    ], dim=1).reshape([bs, 3, 3])
+
+    rot = rot_z @ rot_y @ rot_x
+    return rot.permute(0, 2, 1)  # transpose

modules/utils/face_analysis_diy.py

@@ -0,0 +1,78 @@
+# coding: utf-8
+
+"""
+face detectoin and alignment using InsightFace
+"""
+
+import numpy as np
+from .rprint import rlog as log
+from insightface.app import FaceAnalysis
+from insightface.app.common import Face
+from .timer import Timer
+
+
+def sort_by_direction(faces, direction: str = 'large-small', face_center=None):
+    if len(faces) <= 0:
+        return faces
+    if direction == 'left-right':
+        return sorted(faces, key=lambda face: face['bbox'][0])
+    if direction == 'right-left':
+        return sorted(faces, key=lambda face: face['bbox'][0], reverse=True)
+    if direction == 'top-bottom':
+        return sorted(faces, key=lambda face: face['bbox'][1])
+    if direction == 'bottom-top':
+        return sorted(faces, key=lambda face: face['bbox'][1], reverse=True)
+    if direction == 'small-large':
+        return sorted(faces, key=lambda face: (face['bbox'][2] - face['bbox'][0]) * (face['bbox'][3] - face['bbox'][1]))
+    if direction == 'large-small':
+        return sorted(faces, key=lambda face: (face['bbox'][2] - face['bbox'][0]) * (face['bbox'][3] - face['bbox'][1]), reverse=True)
+    if direction == 'distance-from-retarget-face':
+        return sorted(faces, key=lambda face: (((face['bbox'][2]+face['bbox'][0])/2-face_center[0])**2+((face['bbox'][3]+face['bbox'][1])/2-face_center[1])**2)**0.5)
+    return faces
+
+
+class FaceAnalysisDIY(FaceAnalysis):
+    def __init__(self, name='buffalo_l', root='~/.insightface', allowed_modules=None, **kwargs):
+        super().__init__(name=name, root=root, allowed_modules=allowed_modules, **kwargs)
+
+        self.timer = Timer()
+
+    def get(self, img_bgr, **kwargs):
+        max_num = kwargs.get('max_num', 0)  # the number of the detected faces, 0 means no limit
+        flag_do_landmark_2d_106 = kwargs.get('flag_do_landmark_2d_106', True)  # whether to do 106-point detection
+        direction = kwargs.get('direction', 'large-small')  # sorting direction
+        face_center = None
+
+        bboxes, kpss = self.det_model.detect(img_bgr, max_num=max_num, metric='default')
+        if bboxes.shape[0] == 0:
+            return []
+        ret = []
+        for i in range(bboxes.shape[0]):
+            bbox = bboxes[i, 0:4]
+            det_score = bboxes[i, 4]
+            kps = None
+            if kpss is not None:
+                kps = kpss[i]
+            face = Face(bbox=bbox, kps=kps, det_score=det_score)
+            for taskname, model in self.models.items():
+                if taskname == 'detection':
+                    continue
+
+                if (not flag_do_landmark_2d_106) and taskname == 'landmark_2d_106':
+                    continue
+
+                # print(f'taskname: {taskname}')
+                model.get(img_bgr, face)
+            ret.append(face)
+
+        ret = sort_by_direction(ret, direction, face_center)
+        return ret
+
+    def warmup(self):
+        self.timer.tic()
+
+        img_bgr = np.zeros((512, 512, 3), dtype=np.uint8)
+        self.get(img_bgr)
+
+        elapse = self.timer.toc()
+        log(f'FaceAnalysisDIY warmup time: {elapse:.3f}s')

modules/utils/helper.py

@@ -0,0 +1,131 @@
+# coding: utf-8
+
+"""
+utility functions and classes to handle feature extraction and model loading
+"""
+
+import os
+import os.path as osp
+import cv2
+import torch
+import yaml
+from rich.console import Console
+from collections import OrderedDict
+
+from ..live_portrait.spade_generator import SPADEDecoder
+from ..live_portrait.warping_network import WarpingNetwork
+from ..live_portrait.motion_extractor import MotionExtractor
+from ..live_portrait.appearance_feature_extractor import AppearanceFeatureExtractor
+from ..live_portrait.stitching_retargeting_network import StitchingRetargetingNetwork
+from .rprint import rlog as log
+
+
+def suffix(filename):
+    """a.jpg -> jpg"""
+    pos = filename.rfind(".")
+    if pos == -1:
+        return ""
+    return filename[pos + 1:]
+
+
+def prefix(filename):
+    """a.jpg -> a"""
+    pos = filename.rfind(".")
+    if pos == -1:
+        return filename
+    return filename[:pos]
+
+
+def basename(filename):
+    """a/b/c.jpg -> c"""
+    return prefix(osp.basename(filename))
+
+
+def is_video(file_path):
+    if file_path.lower().endswith((".mp4", ".mov", ".avi", ".webm")) or osp.isdir(file_path):
+        return True
+    return False
+
+def is_template(file_path):
+    if file_path.endswith(".pkl"):
+        return True
+    return False
+
+
+def mkdir(d, log=False):
+    # return self-assined `d`, for one line code
+    if not osp.exists(d):
+        os.makedirs(d, exist_ok=True)
+        if log:
+            print(f"Make dir: {d}")
+    return d
+
+
+def squeeze_tensor_to_numpy(tensor):
+    out = tensor.data.squeeze(0).cpu().numpy()
+    return out
+
+
+def dct2cuda(dct: dict, device_id: int):
+    for key in dct:
+        dct[key] = torch.tensor(dct[key]).cuda(device_id)
+    return dct
+
+
+def concat_feat(kp_source: torch.Tensor, kp_driving: torch.Tensor) -> torch.Tensor:
+    """
+    kp_source: (bs, k, 3)
+    kp_driving: (bs, k, 3)
+    Return: (bs, 2k*3)
+    """
+    bs_src = kp_source.shape[0]
+    bs_dri = kp_driving.shape[0]
+    assert bs_src == bs_dri, 'batch size must be equal'
+
+    feat = torch.cat([kp_source.view(bs_src, -1), kp_driving.view(bs_dri, -1)], dim=1)
+    return feat
+
+
+# get coefficients of Eqn. 7
+def calculate_transformation(config, s_kp_info, t_0_kp_info, t_i_kp_info, R_s, R_t_0, R_t_i):
+    if config.relative:
+        new_rotation = (R_t_i @ R_t_0.permute(0, 2, 1)) @ R_s
+        new_expression = s_kp_info['exp'] + (t_i_kp_info['exp'] - t_0_kp_info['exp'])
+    else:
+        new_rotation = R_t_i
+        new_expression = t_i_kp_info['exp']
+    new_translation = s_kp_info['t'] + (t_i_kp_info['t'] - t_0_kp_info['t'])
+    new_translation[..., 2].fill_(0)  # Keep the z-axis unchanged
+    new_scale = s_kp_info['scale'] * (t_i_kp_info['scale'] / t_0_kp_info['scale'])
+    return new_rotation, new_expression, new_translation, new_scale
+
+def load_description(fp):
+    with open(fp, 'r', encoding='utf-8') as f:
+        content = f.read()
+    return content
+
+
+def resize_to_limit(img, max_dim=1280, n=2):
+    h, w = img.shape[:2]
+    if max_dim > 0 and max(h, w) > max_dim:
+        if h > w:
+            new_h = max_dim
+            new_w = int(w * (max_dim / h))
+        else:
+            new_w = max_dim
+            new_h = int(h * (max_dim / w))
+        img = cv2.resize(img, (new_w, new_h))
+    n = max(n, 1)
+    new_h = img.shape[0] - (img.shape[0] % n)
+    new_w = img.shape[1] - (img.shape[1] % n)
+    if new_h == 0 or new_w == 0:
+        return img
+    if new_h != img.shape[0] or new_w != img.shape[1]:
+        img = img[:new_h, :new_w]
+    return img
+
+
+def load_yaml(file_path):
+    with open(file_path, 'r') as file:
+        data = yaml.safe_load(file)
+    return data

modules/utils/image_helper.py

@@ -0,0 +1,53 @@
+from PIL import Image
+import torch
+import numpy as np
+
+
+class PreparedSrcImg:
+    def __init__(self, src_rgb, crop_trans_m, x_s_info, f_s_user, x_s_user, mask_ori):
+        self.src_rgb = src_rgb
+        self.crop_trans_m = crop_trans_m
+        self.x_s_info = x_s_info
+        self.f_s_user = f_s_user
+        self.x_s_user = x_s_user
+        self.mask_ori = mask_ori
+
+
+def tensor2pil(image):
+    return Image.fromarray(np.clip(255. * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8))
+
+
+def pil2tensor(image):
+    return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)
+
+
+def rgb_crop(rgb, region):
+    return rgb[region[1]:region[3], region[0]:region[2]]
+
+
+def rgb_crop_batch(rgbs, region):
+    return rgbs[:, region[1]:region[3], region[0]:region[2]]
+
+
+def get_rgb_size(rgb):
+    return rgb.shape[1], rgb.shape[0]
+
+
+def create_transform_matrix(x, y, s_x, s_y):
+    return np.float32([[s_x, 0, x], [0, s_y, y]])
+
+
+def calc_crop_limit(center, img_size, crop_size):
+    pos = center - crop_size / 2
+    if pos < 0:
+        crop_size += pos * 2
+        pos = 0
+
+    pos2 = pos + crop_size
+
+    if img_size < pos2:
+        crop_size -= (pos2 - img_size) * 2
+        pos2 = img_size
+        pos = pos2 - crop_size
+
+    return pos, pos2, crop_size

modules/utils/io.py

@@ -0,0 +1,97 @@
+# coding: utf-8
+
+import os
+from glob import glob
+import os.path as osp
+import imageio
+import numpy as np
+import cv2; cv2.setNumThreads(0); cv2.ocl.setUseOpenCL(False)
+
+
+def load_image_rgb(image_path: str):
+    if not osp.exists(image_path):
+        raise FileNotFoundError(f"Image not found: {image_path}")
+    img = cv2.imread(image_path, cv2.IMREAD_COLOR)
+    return cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+
+
+def load_driving_info(driving_info):
+    driving_video_ori = []
+
+    def load_images_from_directory(directory):
+        image_paths = sorted(glob(osp.join(directory, '*.png')) + glob(osp.join(directory, '*.jpg')))
+        return [load_image_rgb(im_path) for im_path in image_paths]
+
+    def load_images_from_video(file_path):
+        reader = imageio.get_reader(file_path)
+        return [image for idx, image in enumerate(reader)]
+
+    if osp.isdir(driving_info):
+        driving_video_ori = load_images_from_directory(driving_info)
+    elif osp.isfile(driving_info):
+        driving_video_ori = load_images_from_video(driving_info)
+
+    return driving_video_ori
+
+
+def contiguous(obj):
+    if not obj.flags.c_contiguous:
+        obj = obj.copy(order="C")
+    return obj
+
+
+def _resize_to_limit(img: np.ndarray, max_dim=1920, n=2):
+    """
+    ajust the size of the image so that the maximum dimension does not exceed max_dim, and the width and the height of the image are multiples of n.
+    :param img: the image to be processed.
+    :param max_dim: the maximum dimension constraint.
+    :param n: the number that needs to be multiples of.
+    :return: the adjusted image.
+    """
+    h, w = img.shape[:2]
+
+    # ajust the size of the image according to the maximum dimension
+    if max_dim > 0 and max(h, w) > max_dim:
+        if h > w:
+            new_h = max_dim
+            new_w = int(w * (max_dim / h))
+        else:
+            new_w = max_dim
+            new_h = int(h * (max_dim / w))
+        img = cv2.resize(img, (new_w, new_h))
+
+    # ensure that the image dimensions are multiples of n
+    n = max(n, 1)
+    new_h = img.shape[0] - (img.shape[0] % n)
+    new_w = img.shape[1] - (img.shape[1] % n)
+
+    if new_h == 0 or new_w == 0:
+        # when the width or height is less than n, no need to process
+        return img
+
+    if new_h != img.shape[0] or new_w != img.shape[1]:
+        img = img[:new_h, :new_w]
+
+    return img
+
+
+def load_img_online(obj, mode="bgr", **kwargs):
+    max_dim = kwargs.get("max_dim", 1920)
+    n = kwargs.get("n", 2)
+    if isinstance(obj, str):
+        if mode.lower() == "gray":
+            img = cv2.imread(obj, cv2.IMREAD_GRAYSCALE)
+        else:
+            img = cv2.imread(obj, cv2.IMREAD_COLOR)
+    else:
+        img = obj
+
+    # Resize image to satisfy constraints
+    img = _resize_to_limit(img, max_dim=max_dim, n=n)
+
+    if mode.lower() == "bgr":
+        return contiguous(img)
+    elif mode.lower() == "rgb":
+        return contiguous(img[..., ::-1])
+    else:
+        raise Exception(f"Unknown mode {mode}")

modules/utils/rprint.py

@@ -0,0 +1,16 @@
+# coding: utf-8
+
+"""
+custom print and log functions 
+"""
+
+__all__ = ['rprint', 'rlog']
+
+try:
+    from rich.console import Console
+    console = Console()
+    rprint = console.print
+    rlog = console.log
+except:
+    rprint = print
+    rlog = print

modules/utils/timer.py

@@ -0,0 +1,29 @@
+# coding: utf-8
+
+"""
+tools to measure elapsed time
+"""
+
+import time
+
+class Timer(object):
+    """A simple timer."""
+
+    def __init__(self):
+        self.total_time = 0.
+        self.calls = 0
+        self.start_time = 0.
+        self.diff = 0.
+
+    def tic(self):
+        # using time.time instead of time.clock because time time.clock
+        # does not normalize for multithreading
+        self.start_time = time.time()
+
+    def toc(self, average=True):
+        self.diff = time.time() - self.start_time
+        return self.diff
+
+    def clear(self):
+        self.start_time = 0.
+        self.diff = 0.

modules/utils/video.py

@@ -0,0 +1,142 @@
+# coding: utf-8
+
+"""
+functions for processing video
+"""
+
+import os.path as osp
+import numpy as np
+import subprocess
+import imageio
+import cv2
+
+# from rich.progress import track
+from .helper import prefix
+from .rprint import rprint as print
+
+
+def exec_cmd(cmd):
+    subprocess.run(cmd, shell=True, check=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
+
+
+def images2video(images, wfp, **kwargs):
+    fps = kwargs.get('fps', 30)
+    video_format = kwargs.get('format', 'mp4')  # default is mp4 format
+    codec = kwargs.get('codec', 'libx264')  # default is libx264 encoding
+    quality = kwargs.get('quality')  # video quality
+    pixelformat = kwargs.get('pixelformat', 'yuv420p')  # video pixel format
+    image_mode = kwargs.get('image_mode', 'rgb')
+    macro_block_size = kwargs.get('macro_block_size', 2)
+    ffmpeg_params = ['-crf', str(kwargs.get('crf', 18))]
+
+    writer = imageio.get_writer(
+        wfp, fps=fps, format=video_format,
+        codec=codec, quality=quality, ffmpeg_params=ffmpeg_params, pixelformat=pixelformat, macro_block_size=macro_block_size
+    )
+
+    n = len(images)
+    print('writing',n)
+    for i in range(n):
+        if image_mode.lower() == 'bgr':
+            writer.append_data(images[i][..., ::-1])
+        else:
+            writer.append_data(images[i])
+
+    writer.close()
+
+    # print(f':smiley: Dump to {wfp}\n', style="bold green")
+    print(f'Dump to {wfp}\n')
+
+
+def video2gif(video_fp, fps=30, size=256):
+    if osp.exists(video_fp):
+        d = osp.split(video_fp)[0]
+        fn = prefix(osp.basename(video_fp))
+        palette_wfp = osp.join(d, 'palette.png')
+        gif_wfp = osp.join(d, f'{fn}.gif')
+        # generate the palette
+        cmd = f'ffmpeg -i {video_fp} -vf "fps={fps},scale={size}:-1:flags=lanczos,palettegen" {palette_wfp} -y'
+        exec_cmd(cmd)
+        # use the palette to generate the gif
+        cmd = f'ffmpeg -i {video_fp} -i {palette_wfp} -filter_complex "fps={fps},scale={size}:-1:flags=lanczos[x];[x][1:v]paletteuse" {gif_wfp} -y'
+        exec_cmd(cmd)
+    else:
+        print(f'video_fp: {video_fp} not exists!')
+
+
+def merge_audio_video(video_fp, audio_fp, wfp):
+    if osp.exists(video_fp) and osp.exists(audio_fp):
+        cmd = f'ffmpeg -i {video_fp} -i {audio_fp} -c:v copy -c:a aac {wfp} -y'
+        exec_cmd(cmd)
+        print(f'merge {video_fp} and {audio_fp} to {wfp}')
+    else:
+        print(f'video_fp: {video_fp} or audio_fp: {audio_fp} not exists!')
+
+
+def blend(img: np.ndarray, mask: np.ndarray, background_color=(255, 255, 255)):
+    mask_float = mask.astype(np.float32) / 255.
+    background_color = np.array(background_color).reshape([1, 1, 3])
+    bg = np.ones_like(img) * background_color
+    img = np.clip(mask_float * img + (1 - mask_float) * bg, 0, 255).astype(np.uint8)
+    return img
+
+
+def concat_frames(I_p_lst, driving_rgb_lst, img_rgb):
+    # TODO: add more concat style, e.g., left-down corner driving
+    out_lst = []
+    print('Concatenating result...',len(I_p_lst))
+    for idx, _ in enumerate(I_p_lst):
+    # track(enumerate(I_p_lst), total=len(I_p_lst), description='Concatenating result...'):
+        source_image_drived = I_p_lst[idx]
+        image_drive = driving_rgb_lst[idx]
+
+        # resize images to match source_image_drived shape
+        h, w, _ = source_image_drived.shape
+        image_drive_resized = cv2.resize(image_drive, (w, h))
+        img_rgb_resized = cv2.resize(img_rgb, (w, h))
+
+        # concatenate images horizontally
+        frame = np.concatenate((image_drive_resized, img_rgb_resized, source_image_drived), axis=1)
+        out_lst.append(frame)
+    return out_lst
+
+
+class VideoWriter:
+    def __init__(self, **kwargs):
+        self.fps = kwargs.get('fps', 30)
+        self.wfp = kwargs.get('wfp', 'video.mp4')
+        self.video_format = kwargs.get('format', 'mp4')
+        self.codec = kwargs.get('codec', 'libx264')
+        self.quality = kwargs.get('quality')
+        self.pixelformat = kwargs.get('pixelformat', 'yuv420p')
+        self.image_mode = kwargs.get('image_mode', 'rgb')
+        self.ffmpeg_params = kwargs.get('ffmpeg_params')
+
+        self.writer = imageio.get_writer(
+            self.wfp, fps=self.fps, format=self.video_format,
+            codec=self.codec, quality=self.quality,
+            ffmpeg_params=self.ffmpeg_params, pixelformat=self.pixelformat
+        )
+
+    def write(self, image):
+        if self.image_mode.lower() == 'bgr':
+            self.writer.append_data(image[..., ::-1])
+        else:
+            self.writer.append_data(image)
+
+    def close(self):
+        if self.writer is not None:
+            self.writer.close()
+
+
+def change_video_fps(input_file, output_file, fps=20, codec='libx264', crf=5):
+    cmd = f"ffmpeg -i {input_file} -c:v {codec} -crf {crf} -r {fps} {output_file} -y"
+    exec_cmd(cmd)
+
+
+def get_fps(filepath):
+    import ffmpeg
+    probe = ffmpeg.probe(filepath)
+    video_stream = next((stream for stream in probe['streams'] if stream['codec_type'] == 'video'), None)
+    fps = eval(video_stream['avg_frame_rate'])
+    return fps