codeformer

facelib/detection/yolov5face/utils/extract_ckpt.py

+import torch
+import sys
+sys.path.insert(0,'./facelib/detection/yolov5face')
+model = torch.load('facelib/detection/yolov5face/yolov5n-face.pt', map_location='cpu')['model']
+torch.save(model.state_dict(),'weights/facelib/yolov5n-face.pth')
\ No newline at end of file

facelib/detection/yolov5face/utils/general.py

+import math
+import time
+
+import numpy as np
+import torch
+import torchvision
+
+
+def check_img_size(img_size, s=32):
+    # Verify img_size is a multiple of stride s
+    new_size = make_divisible(img_size, int(s))  # ceil gs-multiple
+    # if new_size != img_size:
+    #     print(f"WARNING: --img-size {img_size:g} must be multiple of max stride {s:g}, updating to {new_size:g}")
+    return new_size
+
+
+def make_divisible(x, divisor):
+    # Returns x evenly divisible by divisor
+    return math.ceil(x / divisor) * divisor
+
+
+def xyxy2xywh(x):
+    # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = (x[:, 0] + x[:, 2]) / 2  # x center
+    y[:, 1] = (x[:, 1] + x[:, 3]) / 2  # y center
+    y[:, 2] = x[:, 2] - x[:, 0]  # width
+    y[:, 3] = x[:, 3] - x[:, 1]  # height
+    return y
+
+
+def xywh2xyxy(x):
+    # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+
+def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None):
+    # Rescale coords (xyxy) from img1_shape to img0_shape
+    if ratio_pad is None:  # calculate from img0_shape
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
+        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
+    else:
+        gain = ratio_pad[0][0]
+        pad = ratio_pad[1]
+
+    coords[:, [0, 2]] -= pad[0]  # x padding
+    coords[:, [1, 3]] -= pad[1]  # y padding
+    coords[:, :4] /= gain
+    clip_coords(coords, img0_shape)
+    return coords
+
+
+def clip_coords(boxes, img_shape):
+    # Clip bounding xyxy bounding boxes to image shape (height, width)
+    boxes[:, 0].clamp_(0, img_shape[1])  # x1
+    boxes[:, 1].clamp_(0, img_shape[0])  # y1
+    boxes[:, 2].clamp_(0, img_shape[1])  # x2
+    boxes[:, 3].clamp_(0, img_shape[0])  # y2
+
+
+def box_iou(box1, box2):
+    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+    Arguments:
+        box1 (Tensor[N, 4])
+        box2 (Tensor[M, 4])
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    return inter / (area1[:, None] + area2 - inter)
+
+
+def non_max_suppression_face(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, labels=()):
+    """Performs Non-Maximum Suppression (NMS) on inference results
+    Returns:
+         detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
+    """
+
+    nc = prediction.shape[2] - 15  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Settings
+    # (pixels) maximum box width and height
+    max_wh = 4096
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label = nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0, 16), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            label = labels[xi]
+            v = torch.zeros((len(label), nc + 15), device=x.device)
+            v[:, :4] = label[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(label)), label[:, 0].long() + 15] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 15:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, landmarks, cls)
+        if multi_label:
+            i, j = (x[:, 15:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 15, None], x[:, 5:15], j[:, None].float()), 1)
+        else:  # best class only
+            conf, j = x[:, 15:].max(1, keepdim=True)
+            x = torch.cat((box, conf, x[:, 5:15], j.float()), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # If none remain process next image
+        n = x.shape[0]  # number of boxes
+        if not n:
+            continue
+
+        # Batched NMS
+        c = x[:, 15:16] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+
+        if merge and (1 < n < 3e3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            break  # time limit exceeded
+
+    return output
+
+
+def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, labels=()):
+    """Performs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
+    """
+
+    nc = prediction.shape[2] - 5  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Settings
+    # (pixels) maximum box width and height
+    max_wh = 4096
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label = nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            label_id = labels[xi]
+            v = torch.zeros((len(label_id), nc + 5), device=x.device)
+            v[:, :4] = label_id[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(label_id)), label_id[:, 0].long() + 5] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 5:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, cls)
+        if multi_label:
+            i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
+        else:  # best class only
+            conf, j = x[:, 5:].max(1, keepdim=True)
+            x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # Check shape
+        n = x.shape[0]  # number of boxes
+        if not n:  # no boxes
+            continue
+
+        x = x[x[:, 4].argsort(descending=True)]  # sort by confidence
+
+        # Batched NMS
+        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        if merge and (1 < n < 3e3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            print(f"WARNING: NMS time limit {time_limit}s exceeded")
+            break  # time limit exceeded
+
+    return output
+
+
+def scale_coords_landmarks(img1_shape, coords, img0_shape, ratio_pad=None):
+    # Rescale coords (xyxy) from img1_shape to img0_shape
+    if ratio_pad is None:  # calculate from img0_shape
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
+        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
+    else:
+        gain = ratio_pad[0][0]
+        pad = ratio_pad[1]
+
+    coords[:, [0, 2, 4, 6, 8]] -= pad[0]  # x padding
+    coords[:, [1, 3, 5, 7, 9]] -= pad[1]  # y padding
+    coords[:, :10] /= gain
+    coords[:, 0].clamp_(0, img0_shape[1])  # x1
+    coords[:, 1].clamp_(0, img0_shape[0])  # y1
+    coords[:, 2].clamp_(0, img0_shape[1])  # x2
+    coords[:, 3].clamp_(0, img0_shape[0])  # y2
+    coords[:, 4].clamp_(0, img0_shape[1])  # x3
+    coords[:, 5].clamp_(0, img0_shape[0])  # y3
+    coords[:, 6].clamp_(0, img0_shape[1])  # x4
+    coords[:, 7].clamp_(0, img0_shape[0])  # y4
+    coords[:, 8].clamp_(0, img0_shape[1])  # x5
+    coords[:, 9].clamp_(0, img0_shape[0])  # y5
+    return coords

facelib/detection/yolov5face/utils/torch_utils.py

+import torch
+from torch import nn
+
+
+def fuse_conv_and_bn(conv, bn):
+    # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
+    fusedconv = (
+        nn.Conv2d(
+            conv.in_channels,
+            conv.out_channels,
+            kernel_size=conv.kernel_size,
+            stride=conv.stride,
+            padding=conv.padding,
+            groups=conv.groups,
+            bias=True,
+        )
+        .requires_grad_(False)
+        .to(conv.weight.device)
+    )
+
+    # prepare filters
+    w_conv = conv.weight.clone().view(conv.out_channels, -1)
+    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
+    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size()))
+
+    # prepare spatial bias
+    b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
+    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
+    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
+
+    return fusedconv
+
+
+def copy_attr(a, b, include=(), exclude=()):
+    # Copy attributes from b to a, options to only include [...] and to exclude [...]
+    for k, v in b.__dict__.items():
+        if (include and k not in include) or k.startswith("_") or k in exclude:
+            continue
+
+        setattr(a, k, v)

facelib/parsing/__init__.py

+import torch
+
+from facelib.utils import load_file_from_url
+from .bisenet import BiSeNet
+from .parsenet import ParseNet
+
+
+def init_parsing_model(model_name='bisenet', half=False, device='cuda'):
+    if model_name == 'bisenet':
+        model = BiSeNet(num_class=19)
+        model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/parsing_bisenet.pth'
+    elif model_name == 'parsenet':
+        model = ParseNet(in_size=512, out_size=512, parsing_ch=19)
+        model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/parsing_parsenet.pth'
+    else:
+        raise NotImplementedError(f'{model_name} is not implemented.')
+
+    model_path = load_file_from_url(url=model_url, model_dir='weights/facelib', progress=True, file_name=None)
+    load_net = torch.load(model_path, map_location=lambda storage, loc: storage)
+    model.load_state_dict(load_net, strict=True)
+    model.eval()
+    model = model.to(device)
+    return model

facelib/parsing/bisenet.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .resnet import ResNet18
+
+
+class ConvBNReLU(nn.Module):
+
+    def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1):
+        super(ConvBNReLU, self).__init__()
+        self.conv = nn.Conv2d(in_chan, out_chan, kernel_size=ks, stride=stride, padding=padding, bias=False)
+        self.bn = nn.BatchNorm2d(out_chan)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = F.relu(self.bn(x))
+        return x
+
+
+class BiSeNetOutput(nn.Module):
+
+    def __init__(self, in_chan, mid_chan, num_class):
+        super(BiSeNetOutput, self).__init__()
+        self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1)
+        self.conv_out = nn.Conv2d(mid_chan, num_class, kernel_size=1, bias=False)
+
+    def forward(self, x):
+        feat = self.conv(x)
+        out = self.conv_out(feat)
+        return out, feat
+
+
+class AttentionRefinementModule(nn.Module):
+
+    def __init__(self, in_chan, out_chan):
+        super(AttentionRefinementModule, self).__init__()
+        self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1)
+        self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size=1, bias=False)
+        self.bn_atten = nn.BatchNorm2d(out_chan)
+        self.sigmoid_atten = nn.Sigmoid()
+
+    def forward(self, x):
+        feat = self.conv(x)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv_atten(atten)
+        atten = self.bn_atten(atten)
+        atten = self.sigmoid_atten(atten)
+        out = torch.mul(feat, atten)
+        return out
+
+
+class ContextPath(nn.Module):
+
+    def __init__(self):
+        super(ContextPath, self).__init__()
+        self.resnet = ResNet18()
+        self.arm16 = AttentionRefinementModule(256, 128)
+        self.arm32 = AttentionRefinementModule(512, 128)
+        self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
+        self.conv_avg = ConvBNReLU(512, 128, ks=1, stride=1, padding=0)
+
+    def forward(self, x):
+        feat8, feat16, feat32 = self.resnet(x)
+        h8, w8 = feat8.size()[2:]
+        h16, w16 = feat16.size()[2:]
+        h32, w32 = feat32.size()[2:]
+
+        avg = F.avg_pool2d(feat32, feat32.size()[2:])
+        avg = self.conv_avg(avg)
+        avg_up = F.interpolate(avg, (h32, w32), mode='nearest')
+
+        feat32_arm = self.arm32(feat32)
+        feat32_sum = feat32_arm + avg_up
+        feat32_up = F.interpolate(feat32_sum, (h16, w16), mode='nearest')
+        feat32_up = self.conv_head32(feat32_up)
+
+        feat16_arm = self.arm16(feat16)
+        feat16_sum = feat16_arm + feat32_up
+        feat16_up = F.interpolate(feat16_sum, (h8, w8), mode='nearest')
+        feat16_up = self.conv_head16(feat16_up)
+
+        return feat8, feat16_up, feat32_up  # x8, x8, x16
+
+
+class FeatureFusionModule(nn.Module):
+
+    def __init__(self, in_chan, out_chan):
+        super(FeatureFusionModule, self).__init__()
+        self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0)
+        self.conv1 = nn.Conv2d(out_chan, out_chan // 4, kernel_size=1, stride=1, padding=0, bias=False)
+        self.conv2 = nn.Conv2d(out_chan // 4, out_chan, kernel_size=1, stride=1, padding=0, bias=False)
+        self.relu = nn.ReLU(inplace=True)
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, fsp, fcp):
+        fcat = torch.cat([fsp, fcp], dim=1)
+        feat = self.convblk(fcat)
+        atten = F.avg_pool2d(feat, feat.size()[2:])
+        atten = self.conv1(atten)
+        atten = self.relu(atten)
+        atten = self.conv2(atten)
+        atten = self.sigmoid(atten)
+        feat_atten = torch.mul(feat, atten)
+        feat_out = feat_atten + feat
+        return feat_out
+
+
+class BiSeNet(nn.Module):
+
+    def __init__(self, num_class):
+        super(BiSeNet, self).__init__()
+        self.cp = ContextPath()
+        self.ffm = FeatureFusionModule(256, 256)
+        self.conv_out = BiSeNetOutput(256, 256, num_class)
+        self.conv_out16 = BiSeNetOutput(128, 64, num_class)
+        self.conv_out32 = BiSeNetOutput(128, 64, num_class)
+
+    def forward(self, x, return_feat=False):
+        h, w = x.size()[2:]
+        feat_res8, feat_cp8, feat_cp16 = self.cp(x)  # return res3b1 feature
+        feat_sp = feat_res8  # replace spatial path feature with res3b1 feature
+        feat_fuse = self.ffm(feat_sp, feat_cp8)
+
+        out, feat = self.conv_out(feat_fuse)
+        out16, feat16 = self.conv_out16(feat_cp8)
+        out32, feat32 = self.conv_out32(feat_cp16)
+
+        out = F.interpolate(out, (h, w), mode='bilinear', align_corners=True)
+        out16 = F.interpolate(out16, (h, w), mode='bilinear', align_corners=True)
+        out32 = F.interpolate(out32, (h, w), mode='bilinear', align_corners=True)
+
+        if return_feat:
+            feat = F.interpolate(feat, (h, w), mode='bilinear', align_corners=True)
+            feat16 = F.interpolate(feat16, (h, w), mode='bilinear', align_corners=True)
+            feat32 = F.interpolate(feat32, (h, w), mode='bilinear', align_corners=True)
+            return out, out16, out32, feat, feat16, feat32
+        else:
+            return out, out16, out32

facelib/parsing/parsenet.py

+"""Modified from https://github.com/chaofengc/PSFRGAN
+"""
+import numpy as np
+import torch.nn as nn
+from torch.nn import functional as F
+
+
+class NormLayer(nn.Module):
+    """Normalization Layers.
+
+    Args:
+        channels: input channels, for batch norm and instance norm.
+        input_size: input shape without batch size, for layer norm.
+    """
+
+    def __init__(self, channels, normalize_shape=None, norm_type='bn'):
+        super(NormLayer, self).__init__()
+        norm_type = norm_type.lower()
+        self.norm_type = norm_type
+        if norm_type == 'bn':
+            self.norm = nn.BatchNorm2d(channels, affine=True)
+        elif norm_type == 'in':
+            self.norm = nn.InstanceNorm2d(channels, affine=False)
+        elif norm_type == 'gn':
+            self.norm = nn.GroupNorm(32, channels, affine=True)
+        elif norm_type == 'pixel':
+            self.norm = lambda x: F.normalize(x, p=2, dim=1)
+        elif norm_type == 'layer':
+            self.norm = nn.LayerNorm(normalize_shape)
+        elif norm_type == 'none':
+            self.norm = lambda x: x * 1.0
+        else:
+            assert 1 == 0, f'Norm type {norm_type} not support.'
+
+    def forward(self, x, ref=None):
+        if self.norm_type == 'spade':
+            return self.norm(x, ref)
+        else:
+            return self.norm(x)
+
+
+class ReluLayer(nn.Module):
+    """Relu Layer.
+
+    Args:
+        relu type: type of relu layer, candidates are
+            - ReLU
+            - LeakyReLU: default relu slope 0.2
+            - PRelu
+            - SELU
+            - none: direct pass
+    """
+
+    def __init__(self, channels, relu_type='relu'):
+        super(ReluLayer, self).__init__()
+        relu_type = relu_type.lower()
+        if relu_type == 'relu':
+            self.func = nn.ReLU(True)
+        elif relu_type == 'leakyrelu':
+            self.func = nn.LeakyReLU(0.2, inplace=True)
+        elif relu_type == 'prelu':
+            self.func = nn.PReLU(channels)
+        elif relu_type == 'selu':
+            self.func = nn.SELU(True)
+        elif relu_type == 'none':
+            self.func = lambda x: x * 1.0
+        else:
+            assert 1 == 0, f'Relu type {relu_type} not support.'
+
+    def forward(self, x):
+        return self.func(x)
+
+
+class ConvLayer(nn.Module):
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 scale='none',
+                 norm_type='none',
+                 relu_type='none',
+                 use_pad=True,
+                 bias=True):
+        super(ConvLayer, self).__init__()
+        self.use_pad = use_pad
+        self.norm_type = norm_type
+        if norm_type in ['bn']:
+            bias = False
+
+        stride = 2 if scale == 'down' else 1
+
+        self.scale_func = lambda x: x
+        if scale == 'up':
+            self.scale_func = lambda x: nn.functional.interpolate(x, scale_factor=2, mode='nearest')
+
+        self.reflection_pad = nn.ReflectionPad2d(int(np.ceil((kernel_size - 1.) / 2)))
+        self.conv2d = nn.Conv2d(in_channels, out_channels, kernel_size, stride, bias=bias)
+
+        self.relu = ReluLayer(out_channels, relu_type)
+        self.norm = NormLayer(out_channels, norm_type=norm_type)
+
+    def forward(self, x):
+        out = self.scale_func(x)
+        if self.use_pad:
+            out = self.reflection_pad(out)
+        out = self.conv2d(out)
+        out = self.norm(out)
+        out = self.relu(out)
+        return out
+
+
+class ResidualBlock(nn.Module):
+    """
+    Residual block recommended in: http://torch.ch/blog/2016/02/04/resnets.html
+    """
+
+    def __init__(self, c_in, c_out, relu_type='prelu', norm_type='bn', scale='none'):
+        super(ResidualBlock, self).__init__()
+
+        if scale == 'none' and c_in == c_out:
+            self.shortcut_func = lambda x: x
+        else:
+            self.shortcut_func = ConvLayer(c_in, c_out, 3, scale)
+
+        scale_config_dict = {'down': ['none', 'down'], 'up': ['up', 'none'], 'none': ['none', 'none']}
+        scale_conf = scale_config_dict[scale]
+
+        self.conv1 = ConvLayer(c_in, c_out, 3, scale_conf[0], norm_type=norm_type, relu_type=relu_type)
+        self.conv2 = ConvLayer(c_out, c_out, 3, scale_conf[1], norm_type=norm_type, relu_type='none')
+
+    def forward(self, x):
+        identity = self.shortcut_func(x)
+
+        res = self.conv1(x)
+        res = self.conv2(res)
+        return identity + res
+
+
+class ParseNet(nn.Module):
+
+    def __init__(self,
+                 in_size=128,
+                 out_size=128,
+                 min_feat_size=32,
+                 base_ch=64,
+                 parsing_ch=19,
+                 res_depth=10,
+                 relu_type='LeakyReLU',
+                 norm_type='bn',
+                 ch_range=[32, 256]):
+        super().__init__()
+        self.res_depth = res_depth
+        act_args = {'norm_type': norm_type, 'relu_type': relu_type}
+        min_ch, max_ch = ch_range
+
+        ch_clip = lambda x: max(min_ch, min(x, max_ch))  # noqa: E731
+        min_feat_size = min(in_size, min_feat_size)
+
+        down_steps = int(np.log2(in_size // min_feat_size))
+        up_steps = int(np.log2(out_size // min_feat_size))
+
+        # =============== define encoder-body-decoder ====================
+        self.encoder = []
+        self.encoder.append(ConvLayer(3, base_ch, 3, 1))
+        head_ch = base_ch
+        for i in range(down_steps):
+            cin, cout = ch_clip(head_ch), ch_clip(head_ch * 2)
+            self.encoder.append(ResidualBlock(cin, cout, scale='down', **act_args))
+            head_ch = head_ch * 2
+
+        self.body = []
+        for i in range(res_depth):
+            self.body.append(ResidualBlock(ch_clip(head_ch), ch_clip(head_ch), **act_args))
+
+        self.decoder = []
+        for i in range(up_steps):
+            cin, cout = ch_clip(head_ch), ch_clip(head_ch // 2)
+            self.decoder.append(ResidualBlock(cin, cout, scale='up', **act_args))
+            head_ch = head_ch // 2
+
+        self.encoder = nn.Sequential(*self.encoder)
+        self.body = nn.Sequential(*self.body)
+        self.decoder = nn.Sequential(*self.decoder)
+        self.out_img_conv = ConvLayer(ch_clip(head_ch), 3)
+        self.out_mask_conv = ConvLayer(ch_clip(head_ch), parsing_ch)
+
+    def forward(self, x):
+        feat = self.encoder(x)
+        x = feat + self.body(feat)
+        x = self.decoder(x)
+        out_img = self.out_img_conv(x)
+        out_mask = self.out_mask_conv(x)
+        return out_mask, out_img

facelib/parsing/resnet.py

+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+
+    def __init__(self, in_chan, out_chan, stride=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(in_chan, out_chan, stride)
+        self.bn1 = nn.BatchNorm2d(out_chan)
+        self.conv2 = conv3x3(out_chan, out_chan)
+        self.bn2 = nn.BatchNorm2d(out_chan)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        if in_chan != out_chan or stride != 1:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_chan, out_chan, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(out_chan),
+            )
+
+    def forward(self, x):
+        residual = self.conv1(x)
+        residual = F.relu(self.bn1(residual))
+        residual = self.conv2(residual)
+        residual = self.bn2(residual)
+
+        shortcut = x
+        if self.downsample is not None:
+            shortcut = self.downsample(x)
+
+        out = shortcut + residual
+        out = self.relu(out)
+        return out
+
+
+def create_layer_basic(in_chan, out_chan, bnum, stride=1):
+    layers = [BasicBlock(in_chan, out_chan, stride=stride)]
+    for i in range(bnum - 1):
+        layers.append(BasicBlock(out_chan, out_chan, stride=1))
+    return nn.Sequential(*layers)
+
+
+class ResNet18(nn.Module):
+
+    def __init__(self):
+        super(ResNet18, self).__init__()
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = create_layer_basic(64, 64, bnum=2, stride=1)
+        self.layer2 = create_layer_basic(64, 128, bnum=2, stride=2)
+        self.layer3 = create_layer_basic(128, 256, bnum=2, stride=2)
+        self.layer4 = create_layer_basic(256, 512, bnum=2, stride=2)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = F.relu(self.bn1(x))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        feat8 = self.layer2(x)  # 1/8
+        feat16 = self.layer3(feat8)  # 1/16
+        feat32 = self.layer4(feat16)  # 1/32
+        return feat8, feat16, feat32

facelib/utils/__init__.py

+from .face_utils import align_crop_face_landmarks, compute_increased_bbox, get_valid_bboxes, paste_face_back
+from .misc import img2tensor, load_file_from_url, download_pretrained_models, scandir
+
+__all__ = [
+    'align_crop_face_landmarks', 'compute_increased_bbox', 'get_valid_bboxes', 'load_file_from_url', 
+    'download_pretrained_models', 'paste_face_back', 'img2tensor', 'scandir'
+]

facelib/utils/face_restoration_helper.py

+import cv2
+import numpy as np
+import os
+import torch
+from torchvision.transforms.functional import normalize
+
+from facelib.detection import init_detection_model
+from facelib.parsing import init_parsing_model
+from facelib.utils.misc import img2tensor, imwrite, is_gray, bgr2gray, adain_npy
+from basicsr.utils.download_util import load_file_from_url
+from basicsr.utils.misc import get_device
+
+dlib_model_url = {
+    'face_detector': 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/mmod_human_face_detector-4cb19393.dat',
+    'shape_predictor_5': 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/shape_predictor_5_face_landmarks-c4b1e980.dat'
+}
+
+def get_largest_face(det_faces, h, w):
+
+    def get_location(val, length):
+        if val < 0:
+            return 0
+        elif val > length:
+            return length
+        else:
+            return val
+
+    face_areas = []
+    for det_face in det_faces:
+        left = get_location(det_face[0], w)
+        right = get_location(det_face[2], w)
+        top = get_location(det_face[1], h)
+        bottom = get_location(det_face[3], h)
+        face_area = (right - left) * (bottom - top)
+        face_areas.append(face_area)
+    largest_idx = face_areas.index(max(face_areas))
+    return det_faces[largest_idx], largest_idx
+
+
+def get_center_face(det_faces, h=0, w=0, center=None):
+    if center is not None:
+        center = np.array(center)
+    else:
+        center = np.array([w / 2, h / 2])
+    center_dist = []
+    for det_face in det_faces:
+        face_center = np.array([(det_face[0] + det_face[2]) / 2, (det_face[1] + det_face[3]) / 2])
+        dist = np.linalg.norm(face_center - center)
+        center_dist.append(dist)
+    center_idx = center_dist.index(min(center_dist))
+    return det_faces[center_idx], center_idx
+
+
+class FaceRestoreHelper(object):
+    """Helper for the face restoration pipeline (base class)."""
+
+    def __init__(self,
+                 upscale_factor,
+                 face_size=512,
+                 crop_ratio=(1, 1),
+                 det_model='retinaface_resnet50',
+                 save_ext='png',
+                 template_3points=False,
+                 pad_blur=False,
+                 use_parse=False,
+                 device=None):
+        self.template_3points = template_3points  # improve robustness
+        self.upscale_factor = int(upscale_factor)
+        # the cropped face ratio based on the square face
+        self.crop_ratio = crop_ratio  # (h, w)
+        assert (self.crop_ratio[0] >= 1 and self.crop_ratio[1] >= 1), 'crop ration only supports >=1'
+        self.face_size = (int(face_size * self.crop_ratio[1]), int(face_size * self.crop_ratio[0]))
+        self.det_model = det_model
+
+        if self.det_model == 'dlib':
+            # standard 5 landmarks for FFHQ faces with 1024 x 1024
+            self.face_template = np.array([[686.77227723, 488.62376238], [586.77227723, 493.59405941],
+                                        [337.91089109, 488.38613861], [437.95049505, 493.51485149],
+                                        [513.58415842, 678.5049505]])
+            self.face_template = self.face_template / (1024 // face_size)
+        elif self.template_3points:
+            self.face_template = np.array([[192, 240], [319, 240], [257, 371]])
+        else:
+            # standard 5 landmarks for FFHQ faces with 512 x 512 
+            # facexlib
+            self.face_template = np.array([[192.98138, 239.94708], [318.90277, 240.1936], [256.63416, 314.01935],
+                                           [201.26117, 371.41043], [313.08905, 371.15118]])
+
+            # dlib: left_eye: 36:41  right_eye: 42:47  nose: 30,32,33,34  left mouth corner: 48  right mouth corner: 54
+            # self.face_template = np.array([[193.65928, 242.98541], [318.32558, 243.06108], [255.67984, 328.82894],
+            #                                 [198.22603, 372.82502], [313.91018, 372.75659]])
+
+        self.face_template = self.face_template * (face_size / 512.0)
+        if self.crop_ratio[0] > 1:
+            self.face_template[:, 1] += face_size * (self.crop_ratio[0] - 1) / 2
+        if self.crop_ratio[1] > 1:
+            self.face_template[:, 0] += face_size * (self.crop_ratio[1] - 1) / 2
+        self.save_ext = save_ext
+        self.pad_blur = pad_blur
+        if self.pad_blur is True:
+            self.template_3points = False
+
+        self.all_landmarks_5 = []
+        self.det_faces = []
+        self.affine_matrices = []
+        self.inverse_affine_matrices = []
+        self.cropped_faces = []
+        self.restored_faces = []
+        self.pad_input_imgs = []
+
+        if device is None:
+            # self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+            self.device = get_device()
+        else:
+            self.device = device
+
+        # init face detection model
+        if self.det_model == 'dlib':
+            self.face_detector, self.shape_predictor_5 = self.init_dlib(dlib_model_url['face_detector'], dlib_model_url['shape_predictor_5'])
+        else:
+            self.face_detector = init_detection_model(det_model, half=False, device=self.device)
+
+        # init face parsing model
+        self.use_parse = use_parse
+        self.face_parse = init_parsing_model(model_name='parsenet', device=self.device)
+
+    def set_upscale_factor(self, upscale_factor):
+        self.upscale_factor = upscale_factor
+
+    def read_image(self, img):
+        """img can be image path or cv2 loaded image."""
+        # self.input_img is Numpy array, (h, w, c), BGR, uint8, [0, 255]
+        if isinstance(img, str):
+            img = cv2.imread(img)
+
+        if np.max(img) > 256:  # 16-bit image
+            img = img / 65535 * 255
+        if len(img.shape) == 2:  # gray image
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+        elif img.shape[2] == 4:  # BGRA image with alpha channel
+            img = img[:, :, 0:3]
+
+        self.input_img = img
+        self.is_gray = is_gray(img, threshold=10)
+        if self.is_gray:
+            print('Grayscale input: True')
+
+        if min(self.input_img.shape[:2])<512:
+            f = 512.0/min(self.input_img.shape[:2])
+            self.input_img = cv2.resize(self.input_img, (0,0), fx=f, fy=f, interpolation=cv2.INTER_LINEAR)
+
+    def init_dlib(self, detection_path, landmark5_path):
+        """Initialize the dlib detectors and predictors."""
+        try:
+            import dlib
+        except ImportError:
+            print('Please install dlib by running:' 'conda install -c conda-forge dlib')
+        detection_path = load_file_from_url(url=detection_path, model_dir='weights/dlib', progress=True, file_name=None)
+        landmark5_path = load_file_from_url(url=landmark5_path, model_dir='weights/dlib', progress=True, file_name=None)
+        face_detector = dlib.cnn_face_detection_model_v1(detection_path)
+        shape_predictor_5 = dlib.shape_predictor(landmark5_path)
+        return face_detector, shape_predictor_5
+
+    def get_face_landmarks_5_dlib(self,
+                                only_keep_largest=False,
+                                scale=1):
+        det_faces = self.face_detector(self.input_img, scale)
+
+        if len(det_faces) == 0:
+            print('No face detected. Try to increase upsample_num_times.')
+            return 0
+        else:
+            if only_keep_largest:
+                print('Detect several faces and only keep the largest.')
+                face_areas = []
+                for i in range(len(det_faces)):
+                    face_area = (det_faces[i].rect.right() - det_faces[i].rect.left()) * (
+                        det_faces[i].rect.bottom() - det_faces[i].rect.top())
+                    face_areas.append(face_area)
+                largest_idx = face_areas.index(max(face_areas))
+                self.det_faces = [det_faces[largest_idx]]
+            else:
+                self.det_faces = det_faces
+
+        if len(self.det_faces) == 0:
+            return 0
+
+        for face in self.det_faces:
+            shape = self.shape_predictor_5(self.input_img, face.rect)
+            landmark = np.array([[part.x, part.y] for part in shape.parts()])
+            self.all_landmarks_5.append(landmark)
+
+        return len(self.all_landmarks_5)
+
+
+    def get_face_landmarks_5(self,
+                             only_keep_largest=False,
+                             only_center_face=False,
+                             resize=None,
+                             blur_ratio=0.01,
+                             eye_dist_threshold=None):
+        if self.det_model == 'dlib':
+            return self.get_face_landmarks_5_dlib(only_keep_largest)
+
+        if resize is None:
+            scale = 1
+            input_img = self.input_img
+        else:
+            h, w = self.input_img.shape[0:2]
+            scale = resize / min(h, w)
+            scale = max(1, scale) # always scale up
+            h, w = int(h * scale), int(w * scale)
+            interp = cv2.INTER_AREA if scale < 1 else cv2.INTER_LINEAR
+            input_img = cv2.resize(self.input_img, (w, h), interpolation=interp)
+
+        with torch.no_grad():
+            bboxes = self.face_detector.detect_faces(input_img)
+
+        if bboxes is None or bboxes.shape[0] == 0:
+            return 0
+        else:
+            bboxes = bboxes / scale
+
+        for bbox in bboxes:
+            # remove faces with too small eye distance: side faces or too small faces
+            eye_dist = np.linalg.norm([bbox[6] - bbox[8], bbox[7] - bbox[9]])
+            if eye_dist_threshold is not None and (eye_dist < eye_dist_threshold):
+                continue
+
+            if self.template_3points:
+                landmark = np.array([[bbox[i], bbox[i + 1]] for i in range(5, 11, 2)])
+            else:
+                landmark = np.array([[bbox[i], bbox[i + 1]] for i in range(5, 15, 2)])
+            self.all_landmarks_5.append(landmark)
+            self.det_faces.append(bbox[0:5])
+            
+        if len(self.det_faces) == 0:
+            return 0
+        if only_keep_largest:
+            h, w, _ = self.input_img.shape
+            self.det_faces, largest_idx = get_largest_face(self.det_faces, h, w)
+            self.all_landmarks_5 = [self.all_landmarks_5[largest_idx]]
+        elif only_center_face:
+            h, w, _ = self.input_img.shape
+            self.det_faces, center_idx = get_center_face(self.det_faces, h, w)
+            self.all_landmarks_5 = [self.all_landmarks_5[center_idx]]
+
+        # pad blurry images
+        if self.pad_blur:
+            self.pad_input_imgs = []
+            for landmarks in self.all_landmarks_5:
+                # get landmarks
+                eye_left = landmarks[0, :]
+                eye_right = landmarks[1, :]
+                eye_avg = (eye_left + eye_right) * 0.5
+                mouth_avg = (landmarks[3, :] + landmarks[4, :]) * 0.5
+                eye_to_eye = eye_right - eye_left
+                eye_to_mouth = mouth_avg - eye_avg
+
+                # Get the oriented crop rectangle
+                # x: half width of the oriented crop rectangle
+                x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+                #  - np.flipud(eye_to_mouth) * [-1, 1]: rotate 90 clockwise
+                # norm with the hypotenuse: get the direction
+                x /= np.hypot(*x)  # get the hypotenuse of a right triangle
+                rect_scale = 1.5
+                x *= max(np.hypot(*eye_to_eye) * 2.0 * rect_scale, np.hypot(*eye_to_mouth) * 1.8 * rect_scale)
+                # y: half height of the oriented crop rectangle
+                y = np.flipud(x) * [-1, 1]
+
+                # c: center
+                c = eye_avg + eye_to_mouth * 0.1
+                # quad: (left_top, left_bottom, right_bottom, right_top)
+                quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+                # qsize: side length of the square
+                qsize = np.hypot(*x) * 2
+                border = max(int(np.rint(qsize * 0.1)), 3)
+
+                # get pad
+                # pad: (width_left, height_top, width_right, height_bottom)
+                pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+                       int(np.ceil(max(quad[:, 1]))))
+                pad = [
+                    max(-pad[0] + border, 1),
+                    max(-pad[1] + border, 1),
+                    max(pad[2] - self.input_img.shape[0] + border, 1),
+                    max(pad[3] - self.input_img.shape[1] + border, 1)
+                ]
+
+                if max(pad) > 1:
+                    # pad image
+                    pad_img = np.pad(self.input_img, ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
+                    # modify landmark coords
+                    landmarks[:, 0] += pad[0]
+                    landmarks[:, 1] += pad[1]
+                    # blur pad images
+                    h, w, _ = pad_img.shape
+                    y, x, _ = np.ogrid[:h, :w, :1]
+                    mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0],
+                                                       np.float32(w - 1 - x) / pad[2]),
+                                      1.0 - np.minimum(np.float32(y) / pad[1],
+                                                       np.float32(h - 1 - y) / pad[3]))
+                    blur = int(qsize * blur_ratio)
+                    if blur % 2 == 0:
+                        blur += 1
+                    blur_img = cv2.boxFilter(pad_img, 0, ksize=(blur, blur))
+                    # blur_img = cv2.GaussianBlur(pad_img, (blur, blur), 0)
+
+                    pad_img = pad_img.astype('float32')
+                    pad_img += (blur_img - pad_img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+                    pad_img += (np.median(pad_img, axis=(0, 1)) - pad_img) * np.clip(mask, 0.0, 1.0)
+                    pad_img = np.clip(pad_img, 0, 255)  # float32, [0, 255]
+                    self.pad_input_imgs.append(pad_img)
+                else:
+                    self.pad_input_imgs.append(np.copy(self.input_img))
+
+        return len(self.all_landmarks_5)
+
+    def align_warp_face(self, save_cropped_path=None, border_mode='constant'):
+        """Align and warp faces with face template.
+        """
+        if self.pad_blur:
+            assert len(self.pad_input_imgs) == len(
+                self.all_landmarks_5), f'Mismatched samples: {len(self.pad_input_imgs)} and {len(self.all_landmarks_5)}'
+        for idx, landmark in enumerate(self.all_landmarks_5):
+            # use 5 landmarks to get affine matrix
+            # use cv2.LMEDS method for the equivalence to skimage transform
+            # ref: https://blog.csdn.net/yichxi/article/details/115827338
+            affine_matrix = cv2.estimateAffinePartial2D(landmark, self.face_template, method=cv2.LMEDS)[0]
+            self.affine_matrices.append(affine_matrix)
+            # warp and crop faces
+            if border_mode == 'constant':
+                border_mode = cv2.BORDER_CONSTANT
+            elif border_mode == 'reflect101':
+                border_mode = cv2.BORDER_REFLECT101
+            elif border_mode == 'reflect':
+                border_mode = cv2.BORDER_REFLECT
+            if self.pad_blur:
+                input_img = self.pad_input_imgs[idx]
+            else:
+                input_img = self.input_img
+            cropped_face = cv2.warpAffine(
+                input_img, affine_matrix, self.face_size, borderMode=border_mode, borderValue=(135, 133, 132))  # gray
+            self.cropped_faces.append(cropped_face)
+            # save the cropped face
+            if save_cropped_path is not None:
+                path = os.path.splitext(save_cropped_path)[0]
+                save_path = f'{path}_{idx:02d}.{self.save_ext}'
+                imwrite(cropped_face, save_path)
+
+    def get_inverse_affine(self, save_inverse_affine_path=None):
+        """Get inverse affine matrix."""
+        for idx, affine_matrix in enumerate(self.affine_matrices):
+            inverse_affine = cv2.invertAffineTransform(affine_matrix)
+            inverse_affine *= self.upscale_factor
+            self.inverse_affine_matrices.append(inverse_affine)
+            # save inverse affine matrices
+            if save_inverse_affine_path is not None:
+                path, _ = os.path.splitext(save_inverse_affine_path)
+                save_path = f'{path}_{idx:02d}.pth'
+                torch.save(inverse_affine, save_path)
+
+
+    def add_restored_face(self, restored_face, input_face=None):
+        if self.is_gray:
+            restored_face = bgr2gray(restored_face) # convert img into grayscale
+            if input_face is not None:
+                restored_face = adain_npy(restored_face, input_face) # transfer the color
+        self.restored_faces.append(restored_face)
+
+
+    def paste_faces_to_input_image(self, save_path=None, upsample_img=None, draw_box=False, face_upsampler=None):
+        h, w, _ = self.input_img.shape
+        h_up, w_up = int(h * self.upscale_factor), int(w * self.upscale_factor)
+
+        if upsample_img is None:
+            # simply resize the background
+            # upsample_img = cv2.resize(self.input_img, (w_up, h_up), interpolation=cv2.INTER_LANCZOS4)
+            upsample_img = cv2.resize(self.input_img, (w_up, h_up), interpolation=cv2.INTER_LINEAR)
+        else:
+            upsample_img = cv2.resize(upsample_img, (w_up, h_up), interpolation=cv2.INTER_LANCZOS4)
+
+        assert len(self.restored_faces) == len(
+            self.inverse_affine_matrices), ('length of restored_faces and affine_matrices are different.')
+        
+        inv_mask_borders = []
+        for restored_face, inverse_affine in zip(self.restored_faces, self.inverse_affine_matrices):
+            if face_upsampler is not None:
+                restored_face = face_upsampler.enhance(restored_face, outscale=self.upscale_factor)[0]
+                inverse_affine /= self.upscale_factor
+                inverse_affine[:, 2] *= self.upscale_factor
+                face_size = (self.face_size[0]*self.upscale_factor, self.face_size[1]*self.upscale_factor)
+            else:
+                # Add an offset to inverse affine matrix, for more precise back alignment
+                if self.upscale_factor > 1:
+                    extra_offset = 0.5 * self.upscale_factor
+                else:
+                    extra_offset = 0
+                inverse_affine[:, 2] += extra_offset
+                face_size = self.face_size
+            inv_restored = cv2.warpAffine(restored_face, inverse_affine, (w_up, h_up))
+
+            # if draw_box or not self.use_parse:  # use square parse maps
+            #     mask = np.ones(face_size, dtype=np.float32)
+            #     inv_mask = cv2.warpAffine(mask, inverse_affine, (w_up, h_up))
+            #     # remove the black borders
+            #     inv_mask_erosion = cv2.erode(
+            #         inv_mask, np.ones((int(2 * self.upscale_factor), int(2 * self.upscale_factor)), np.uint8))
+            #     pasted_face = inv_mask_erosion[:, :, None] * inv_restored
+            #     total_face_area = np.sum(inv_mask_erosion)  # // 3
+            #     # add border
+            #     if draw_box:
+            #         h, w = face_size
+            #         mask_border = np.ones((h, w, 3), dtype=np.float32)
+            #         border = int(1400/np.sqrt(total_face_area))
+            #         mask_border[border:h-border, border:w-border,:] = 0
+            #         inv_mask_border = cv2.warpAffine(mask_border, inverse_affine, (w_up, h_up))
+            #         inv_mask_borders.append(inv_mask_border)
+            #     if not self.use_parse:
+            #         # compute the fusion edge based on the area of face
+            #         w_edge = int(total_face_area**0.5) // 20
+            #         erosion_radius = w_edge * 2
+            #         inv_mask_center = cv2.erode(inv_mask_erosion, np.ones((erosion_radius, erosion_radius), np.uint8))
+            #         blur_size = w_edge * 2
+            #         inv_soft_mask = cv2.GaussianBlur(inv_mask_center, (blur_size + 1, blur_size + 1), 0)
+            #         if len(upsample_img.shape) == 2:  # upsample_img is gray image
+            #             upsample_img = upsample_img[:, :, None]
+            #         inv_soft_mask = inv_soft_mask[:, :, None]
+
+            # always use square mask
+            mask = np.ones(face_size, dtype=np.float32)
+            inv_mask = cv2.warpAffine(mask, inverse_affine, (w_up, h_up))
+            # remove the black borders
+            inv_mask_erosion = cv2.erode(
+                inv_mask, np.ones((int(2 * self.upscale_factor), int(2 * self.upscale_factor)), np.uint8))
+            pasted_face = inv_mask_erosion[:, :, None] * inv_restored
+            total_face_area = np.sum(inv_mask_erosion)  # // 3
+            # add border
+            if draw_box:
+                h, w = face_size
+                mask_border = np.ones((h, w, 3), dtype=np.float32)
+                border = int(1400/np.sqrt(total_face_area))
+                mask_border[border:h-border, border:w-border,:] = 0
+                inv_mask_border = cv2.warpAffine(mask_border, inverse_affine, (w_up, h_up))
+                inv_mask_borders.append(inv_mask_border)
+            # compute the fusion edge based on the area of face
+            w_edge = int(total_face_area**0.5) // 20
+            erosion_radius = w_edge * 2
+            inv_mask_center = cv2.erode(inv_mask_erosion, np.ones((erosion_radius, erosion_radius), np.uint8))
+            blur_size = w_edge * 2
+            inv_soft_mask = cv2.GaussianBlur(inv_mask_center, (blur_size + 1, blur_size + 1), 0)
+            if len(upsample_img.shape) == 2:  # upsample_img is gray image
+                upsample_img = upsample_img[:, :, None]
+            inv_soft_mask = inv_soft_mask[:, :, None]
+
+            # parse mask
+            if self.use_parse:
+                # inference
+                face_input = cv2.resize(restored_face, (512, 512), interpolation=cv2.INTER_LINEAR)
+                face_input = img2tensor(face_input.astype('float32') / 255., bgr2rgb=True, float32=True)
+                normalize(face_input, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+                face_input = torch.unsqueeze(face_input, 0).to(self.device)
+                with torch.no_grad():
+                    out = self.face_parse(face_input)[0]
+                out = out.argmax(dim=1).squeeze().cpu().numpy()
+
+                parse_mask = np.zeros(out.shape)
+                MASK_COLORMAP = [0, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 0, 255, 0, 0, 0]
+                for idx, color in enumerate(MASK_COLORMAP):
+                    parse_mask[out == idx] = color
+                #  blur the mask
+                parse_mask = cv2.GaussianBlur(parse_mask, (101, 101), 11)
+                parse_mask = cv2.GaussianBlur(parse_mask, (101, 101), 11)
+                # remove the black borders
+                thres = 10
+                parse_mask[:thres, :] = 0
+                parse_mask[-thres:, :] = 0
+                parse_mask[:, :thres] = 0
+                parse_mask[:, -thres:] = 0
+                parse_mask = parse_mask / 255.
+
+                parse_mask = cv2.resize(parse_mask, face_size)
+                parse_mask = cv2.warpAffine(parse_mask, inverse_affine, (w_up, h_up), flags=3)
+                inv_soft_parse_mask = parse_mask[:, :, None]
+                # pasted_face = inv_restored
+                fuse_mask = (inv_soft_parse_mask<inv_soft_mask).astype('int')
+                inv_soft_mask = inv_soft_parse_mask*fuse_mask + inv_soft_mask*(1-fuse_mask)
+
+            if len(upsample_img.shape) == 3 and upsample_img.shape[2] == 4:  # alpha channel
+                alpha = upsample_img[:, :, 3:]
+                upsample_img = inv_soft_mask * pasted_face + (1 - inv_soft_mask) * upsample_img[:, :, 0:3]
+                upsample_img = np.concatenate((upsample_img, alpha), axis=2)
+            else:
+                upsample_img = inv_soft_mask * pasted_face + (1 - inv_soft_mask) * upsample_img
+
+        if np.max(upsample_img) > 256:  # 16-bit image
+            upsample_img = upsample_img.astype(np.uint16)
+        else:
+            upsample_img = upsample_img.astype(np.uint8)
+
+        # draw bounding box
+        if draw_box:
+            # upsample_input_img = cv2.resize(input_img, (w_up, h_up))
+            img_color = np.ones([*upsample_img.shape], dtype=np.float32)
+            img_color[:,:,0] = 0
+            img_color[:,:,1] = 255
+            img_color[:,:,2] = 0
+            for inv_mask_border in inv_mask_borders:
+                upsample_img = inv_mask_border * img_color + (1 - inv_mask_border) * upsample_img
+                # upsample_input_img = inv_mask_border * img_color + (1 - inv_mask_border) * upsample_input_img
+
+        if save_path is not None:
+            path = os.path.splitext(save_path)[0]
+            save_path = f'{path}.{self.save_ext}'
+            imwrite(upsample_img, save_path)
+        return upsample_img
+
+    def clean_all(self):
+        self.all_landmarks_5 = []
+        self.restored_faces = []
+        self.affine_matrices = []
+        self.cropped_faces = []
+        self.inverse_affine_matrices = []
+        self.det_faces = []
+        self.pad_input_imgs = []
\ No newline at end of file

facelib/utils/face_utils.py

+import cv2
+import numpy as np
+import torch
+
+
+def compute_increased_bbox(bbox, increase_area, preserve_aspect=True):
+    left, top, right, bot = bbox
+    width = right - left
+    height = bot - top
+
+    if preserve_aspect:
+        width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width))
+        height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height))
+    else:
+        width_increase = height_increase = increase_area
+    left = int(left - width_increase * width)
+    top = int(top - height_increase * height)
+    right = int(right + width_increase * width)
+    bot = int(bot + height_increase * height)
+    return (left, top, right, bot)
+
+
+def get_valid_bboxes(bboxes, h, w):
+    left = max(bboxes[0], 0)
+    top = max(bboxes[1], 0)
+    right = min(bboxes[2], w)
+    bottom = min(bboxes[3], h)
+    return (left, top, right, bottom)
+
+
+def align_crop_face_landmarks(img,
+                              landmarks,
+                              output_size,
+                              transform_size=None,
+                              enable_padding=True,
+                              return_inverse_affine=False,
+                              shrink_ratio=(1, 1)):
+    """Align and crop face with landmarks.
+
+    The output_size and transform_size are based on width. The height is
+    adjusted based on shrink_ratio_h/shring_ration_w.
+
+    Modified from:
+    https://github.com/NVlabs/ffhq-dataset/blob/master/download_ffhq.py
+
+    Args:
+        img (Numpy array): Input image.
+        landmarks (Numpy array): 5 or 68 or 98 landmarks.
+        output_size (int): Output face size.
+        transform_size (ing): Transform size. Usually the four time of
+            output_size.
+        enable_padding (float): Default: True.
+        shrink_ratio (float | tuple[float] | list[float]): Shring the whole
+            face for height and width (crop larger area). Default: (1, 1).
+
+    Returns:
+        (Numpy array): Cropped face.
+    """
+    lm_type = 'retinaface_5'  # Options: dlib_5, retinaface_5
+
+    if isinstance(shrink_ratio, (float, int)):
+        shrink_ratio = (shrink_ratio, shrink_ratio)
+    if transform_size is None:
+        transform_size = output_size * 4
+
+    # Parse landmarks
+    lm = np.array(landmarks)
+    if lm.shape[0] == 5 and lm_type == 'retinaface_5':
+        eye_left = lm[0]
+        eye_right = lm[1]
+        mouth_avg = (lm[3] + lm[4]) * 0.5
+    elif lm.shape[0] == 5 and lm_type == 'dlib_5':
+        lm_eye_left = lm[2:4]
+        lm_eye_right = lm[0:2]
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        mouth_avg = lm[4]
+    elif lm.shape[0] == 68:
+        lm_eye_left = lm[36:42]
+        lm_eye_right = lm[42:48]
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        mouth_avg = (lm[48] + lm[54]) * 0.5
+    elif lm.shape[0] == 98:
+        lm_eye_left = lm[60:68]
+        lm_eye_right = lm[68:76]
+        eye_left = np.mean(lm_eye_left, axis=0)
+        eye_right = np.mean(lm_eye_right, axis=0)
+        mouth_avg = (lm[76] + lm[82]) * 0.5
+
+    eye_avg = (eye_left + eye_right) * 0.5
+    eye_to_eye = eye_right - eye_left
+    eye_to_mouth = mouth_avg - eye_avg
+
+    # Get the oriented crop rectangle
+    # x: half width of the oriented crop rectangle
+    x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+    #  - np.flipud(eye_to_mouth) * [-1, 1]: rotate 90 clockwise
+    # norm with the hypotenuse: get the direction
+    x /= np.hypot(*x)  # get the hypotenuse of a right triangle
+    rect_scale = 1  # TODO: you can edit it to get larger rect
+    x *= max(np.hypot(*eye_to_eye) * 2.0 * rect_scale, np.hypot(*eye_to_mouth) * 1.8 * rect_scale)
+    # y: half height of the oriented crop rectangle
+    y = np.flipud(x) * [-1, 1]
+
+    x *= shrink_ratio[1]  # width
+    y *= shrink_ratio[0]  # height
+
+    # c: center
+    c = eye_avg + eye_to_mouth * 0.1
+    # quad: (left_top, left_bottom, right_bottom, right_top)
+    quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+    # qsize: side length of the square
+    qsize = np.hypot(*x) * 2
+
+    quad_ori = np.copy(quad)
+    # Shrink, for large face
+    # TODO: do we really need shrink
+    shrink = int(np.floor(qsize / output_size * 0.5))
+    if shrink > 1:
+        h, w = img.shape[0:2]
+        rsize = (int(np.rint(float(w) / shrink)), int(np.rint(float(h) / shrink)))
+        img = cv2.resize(img, rsize, interpolation=cv2.INTER_AREA)
+        quad /= shrink
+        qsize /= shrink
+
+    # Crop
+    h, w = img.shape[0:2]
+    border = max(int(np.rint(qsize * 0.1)), 3)
+    crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+            int(np.ceil(max(quad[:, 1]))))
+    crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, w), min(crop[3] + border, h))
+    if crop[2] - crop[0] < w or crop[3] - crop[1] < h:
+        img = img[crop[1]:crop[3], crop[0]:crop[2], :]
+        quad -= crop[0:2]
+
+    # Pad
+    # pad: (width_left, height_top, width_right, height_bottom)
+    h, w = img.shape[0:2]
+    pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+           int(np.ceil(max(quad[:, 1]))))
+    pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - w + border, 0), max(pad[3] - h + border, 0))
+    if enable_padding and max(pad) > border - 4:
+        pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
+        img = np.pad(img, ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
+        h, w = img.shape[0:2]
+        y, x, _ = np.ogrid[:h, :w, :1]
+        mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0],
+                                           np.float32(w - 1 - x) / pad[2]),
+                          1.0 - np.minimum(np.float32(y) / pad[1],
+                                           np.float32(h - 1 - y) / pad[3]))
+        blur = int(qsize * 0.02)
+        if blur % 2 == 0:
+            blur += 1
+        blur_img = cv2.boxFilter(img, 0, ksize=(blur, blur))
+
+        img = img.astype('float32')
+        img += (blur_img - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+        img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
+        img = np.clip(img, 0, 255)  # float32, [0, 255]
+        quad += pad[:2]
+
+    # Transform use cv2
+    h_ratio = shrink_ratio[0] / shrink_ratio[1]
+    dst_h, dst_w = int(transform_size * h_ratio), transform_size
+    template = np.array([[0, 0], [0, dst_h], [dst_w, dst_h], [dst_w, 0]])
+    # use cv2.LMEDS method for the equivalence to skimage transform
+    # ref: https://blog.csdn.net/yichxi/article/details/115827338
+    affine_matrix = cv2.estimateAffinePartial2D(quad, template, method=cv2.LMEDS)[0]
+    cropped_face = cv2.warpAffine(
+        img, affine_matrix, (dst_w, dst_h), borderMode=cv2.BORDER_CONSTANT, borderValue=(135, 133, 132))  # gray
+
+    if output_size < transform_size:
+        cropped_face = cv2.resize(
+            cropped_face, (output_size, int(output_size * h_ratio)), interpolation=cv2.INTER_LINEAR)
+
+    if return_inverse_affine:
+        dst_h, dst_w = int(output_size * h_ratio), output_size
+        template = np.array([[0, 0], [0, dst_h], [dst_w, dst_h], [dst_w, 0]])
+        # use cv2.LMEDS method for the equivalence to skimage transform
+        # ref: https://blog.csdn.net/yichxi/article/details/115827338
+        affine_matrix = cv2.estimateAffinePartial2D(
+            quad_ori, np.array([[0, 0], [0, output_size], [dst_w, dst_h], [dst_w, 0]]), method=cv2.LMEDS)[0]
+        inverse_affine = cv2.invertAffineTransform(affine_matrix)
+    else:
+        inverse_affine = None
+    return cropped_face, inverse_affine
+
+
+def paste_face_back(img, face, inverse_affine):
+    h, w = img.shape[0:2]
+    face_h, face_w = face.shape[0:2]
+    inv_restored = cv2.warpAffine(face, inverse_affine, (w, h))
+    mask = np.ones((face_h, face_w, 3), dtype=np.float32)
+    inv_mask = cv2.warpAffine(mask, inverse_affine, (w, h))
+    # remove the black borders
+    inv_mask_erosion = cv2.erode(inv_mask, np.ones((2, 2), np.uint8))
+    inv_restored_remove_border = inv_mask_erosion * inv_restored
+    total_face_area = np.sum(inv_mask_erosion) // 3
+    # compute the fusion edge based on the area of face
+    w_edge = int(total_face_area**0.5) // 20
+    erosion_radius = w_edge * 2
+    inv_mask_center = cv2.erode(inv_mask_erosion, np.ones((erosion_radius, erosion_radius), np.uint8))
+    blur_size = w_edge * 2
+    inv_soft_mask = cv2.GaussianBlur(inv_mask_center, (blur_size + 1, blur_size + 1), 0)
+    img = inv_soft_mask * inv_restored_remove_border + (1 - inv_soft_mask) * img
+    # float32, [0, 255]
+    return img
+
+
+if __name__ == '__main__':
+    import os
+
+    from facelib.detection import init_detection_model
+    from facelib.utils.face_restoration_helper import get_largest_face
+
+    img_path = '/home/wxt/datasets/ffhq/ffhq_wild/00009.png'
+    img_name = os.splitext(os.path.basename(img_path))[0]
+
+    # initialize model
+    det_net = init_detection_model('retinaface_resnet50', half=False)
+    img_ori = cv2.imread(img_path)
+    h, w = img_ori.shape[0:2]
+    # if larger than 800, scale it
+    scale = max(h / 800, w / 800)
+    if scale > 1:
+        img = cv2.resize(img_ori, (int(w / scale), int(h / scale)), interpolation=cv2.INTER_LINEAR)
+
+    with torch.no_grad():
+        bboxes = det_net.detect_faces(img, 0.97)
+    if scale > 1:
+        bboxes *= scale  # the score is incorrect
+    bboxes = get_largest_face(bboxes, h, w)[0]
+
+    landmarks = np.array([[bboxes[i], bboxes[i + 1]] for i in range(5, 15, 2)])
+
+    cropped_face, inverse_affine = align_crop_face_landmarks(
+        img_ori,
+        landmarks,
+        output_size=512,
+        transform_size=None,
+        enable_padding=True,
+        return_inverse_affine=True,
+        shrink_ratio=(1, 1))
+
+    cv2.imwrite(f'tmp/{img_name}_cropeed_face.png', cropped_face)
+    img = paste_face_back(img_ori, cropped_face, inverse_affine)
+    cv2.imwrite(f'tmp/{img_name}_back.png', img)

facelib/utils/misc.py

+import cv2
+import os
+import os.path as osp
+import numpy as np
+from PIL import Image
+import torch
+from torch.hub import download_url_to_file, get_dir
+from urllib.parse import urlparse
+# from basicsr.utils.download_util import download_file_from_google_drive
+
+ROOT_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+
+def download_pretrained_models(file_ids, save_path_root):
+    import gdown
+    
+    os.makedirs(save_path_root, exist_ok=True)
+
+    for file_name, file_id in file_ids.items():
+        file_url = 'https://drive.google.com/uc?id='+file_id
+        save_path = osp.abspath(osp.join(save_path_root, file_name))
+        if osp.exists(save_path):
+            user_response = input(f'{file_name} already exist. Do you want to cover it? Y/N\n')
+            if user_response.lower() == 'y':
+                print(f'Covering {file_name} to {save_path}')
+                gdown.download(file_url, save_path, quiet=False)
+                # download_file_from_google_drive(file_id, save_path)
+            elif user_response.lower() == 'n':
+                print(f'Skipping {file_name}')
+            else:
+                raise ValueError('Wrong input. Only accepts Y/N.')
+        else:
+            print(f'Downloading {file_name} to {save_path}')
+            gdown.download(file_url, save_path, quiet=False)
+            # download_file_from_google_drive(file_id, save_path)
+
+
+def imwrite(img, file_path, params=None, auto_mkdir=True):
+    """Write image to file.
+
+    Args:
+        img (ndarray): Image array to be written.
+        file_path (str): Image file path.
+        params (None or list): Same as opencv's :func:`imwrite` interface.
+        auto_mkdir (bool): If the parent folder of `file_path` does not exist,
+            whether to create it automatically.
+
+    Returns:
+        bool: Successful or not.
+    """
+    if auto_mkdir:
+        dir_name = os.path.abspath(os.path.dirname(file_path))
+        os.makedirs(dir_name, exist_ok=True)
+    return cv2.imwrite(file_path, img, params)
+
+
+def img2tensor(imgs, bgr2rgb=True, float32=True):
+    """Numpy array to tensor.
+
+    Args:
+        imgs (list[ndarray] | ndarray): Input images.
+        bgr2rgb (bool): Whether to change bgr to rgb.
+        float32 (bool): Whether to change to float32.
+
+    Returns:
+        list[tensor] | tensor: Tensor images. If returned results only have
+            one element, just return tensor.
+    """
+
+    def _totensor(img, bgr2rgb, float32):
+        if img.shape[2] == 3 and bgr2rgb:
+            if img.dtype == 'float64':
+                img = img.astype('float32')
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+        img = torch.from_numpy(img.transpose(2, 0, 1))
+        if float32:
+            img = img.float()
+        return img
+
+    if isinstance(imgs, list):
+        return [_totensor(img, bgr2rgb, float32) for img in imgs]
+    else:
+        return _totensor(imgs, bgr2rgb, float32)
+
+
+def load_file_from_url(url, model_dir=None, progress=True, file_name=None):
+    """Ref:https://github.com/1adrianb/face-alignment/blob/master/face_alignment/utils.py
+    """
+    if model_dir is None:
+        hub_dir = get_dir()
+        model_dir = os.path.join(hub_dir, 'checkpoints')
+
+    os.makedirs(os.path.join(ROOT_DIR, model_dir), exist_ok=True)
+
+    parts = urlparse(url)
+    filename = os.path.basename(parts.path)
+    if file_name is not None:
+        filename = file_name
+    cached_file = os.path.abspath(os.path.join(ROOT_DIR, model_dir, filename))
+    if not os.path.exists(cached_file):
+        print(f'Downloading: "{url}" to {cached_file}\n')
+        download_url_to_file(url, cached_file, hash_prefix=None, progress=progress)
+    return cached_file
+
+
+def scandir(dir_path, suffix=None, recursive=False, full_path=False):
+    """Scan a directory to find the interested files.
+    Args:
+        dir_path (str): Path of the directory.
+        suffix (str | tuple(str), optional): File suffix that we are
+            interested in. Default: None.
+        recursive (bool, optional): If set to True, recursively scan the
+            directory. Default: False.
+        full_path (bool, optional): If set to True, include the dir_path.
+            Default: False.
+    Returns:
+        A generator for all the interested files with relative paths.
+    """
+
+    if (suffix is not None) and not isinstance(suffix, (str, tuple)):
+        raise TypeError('"suffix" must be a string or tuple of strings')
+
+    root = dir_path
+
+    def _scandir(dir_path, suffix, recursive):
+        for entry in os.scandir(dir_path):
+            if not entry.name.startswith('.') and entry.is_file():
+                if full_path:
+                    return_path = entry.path
+                else:
+                    return_path = osp.relpath(entry.path, root)
+
+                if suffix is None:
+                    yield return_path
+                elif return_path.endswith(suffix):
+                    yield return_path
+            else:
+                if recursive:
+                    yield from _scandir(entry.path, suffix=suffix, recursive=recursive)
+                else:
+                    continue
+
+    return _scandir(dir_path, suffix=suffix, recursive=recursive)
+
+
+def is_gray(img, threshold=10):
+    img = Image.fromarray(img)
+    if len(img.getbands()) == 1:
+        return True
+    img1 = np.asarray(img.getchannel(channel=0), dtype=np.int16)
+    img2 = np.asarray(img.getchannel(channel=1), dtype=np.int16)
+    img3 = np.asarray(img.getchannel(channel=2), dtype=np.int16)
+    diff1 = (img1 - img2).var()
+    diff2 = (img2 - img3).var()
+    diff3 = (img3 - img1).var()
+    diff_sum = (diff1 + diff2 + diff3) / 3.0
+    if diff_sum <= threshold:
+        return True
+    else:
+        return False
+
+def rgb2gray(img, out_channel=3):
+    r, g, b = img[:,:,0], img[:,:,1], img[:,:,2]
+    gray = 0.2989 * r + 0.5870 * g + 0.1140 * b
+    if out_channel == 3:
+        gray = gray[:,:,np.newaxis].repeat(3, axis=2)
+    return gray
+
+def bgr2gray(img, out_channel=3):
+    b, g, r = img[:,:,0], img[:,:,1], img[:,:,2]
+    gray = 0.2989 * r + 0.5870 * g + 0.1140 * b
+    if out_channel == 3:
+        gray = gray[:,:,np.newaxis].repeat(3, axis=2)
+    return gray
+
+
+def calc_mean_std(feat, eps=1e-5):
+    """
+    Args:
+        feat (numpy): 3D [w h c]s
+    """
+    size = feat.shape
+    assert len(size) == 3, 'The input feature should be 3D tensor.'
+    c = size[2]
+    feat_var = feat.reshape(-1, c).var(axis=0) + eps
+    feat_std = np.sqrt(feat_var).reshape(1, 1, c)
+    feat_mean = feat.reshape(-1, c).mean(axis=0).reshape(1, 1, c)
+    return feat_mean, feat_std
+
+
+def adain_npy(content_feat, style_feat):
+    """Adaptive instance normalization for numpy.
+
+    Args:
+        content_feat (numpy): The input feature.
+        style_feat (numpy): The reference feature.
+    """
+    size = content_feat.shape
+    style_mean, style_std = calc_mean_std(style_feat)
+    content_mean, content_std = calc_mean_std(content_feat)
+    normalized_feat = (content_feat - np.broadcast_to(content_mean, size)) / np.broadcast_to(content_std, size)
+    return normalized_feat * np.broadcast_to(style_std, size) + np.broadcast_to(style_mean, size)
\ No newline at end of file

inference_codeformer.py

+import os
+import cv2
+import argparse
+import glob
+import torch
+from torchvision.transforms.functional import normalize
+from basicsr.utils import imwrite, img2tensor, tensor2img
+from basicsr.utils.download_util import load_file_from_url
+from basicsr.utils.misc import gpu_is_available, get_device
+from facelib.utils.face_restoration_helper import FaceRestoreHelper
+from facelib.utils.misc import is_gray
+
+from basicsr.utils.registry import ARCH_REGISTRY
+
+pretrain_model_url = {
+    'restoration': 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth',
+}
+
+def set_realesrgan():
+    from basicsr.archs.rrdbnet_arch import RRDBNet
+    from basicsr.utils.realesrgan_utils import RealESRGANer
+
+    use_half = False
+    if torch.cuda.is_available(): # set False in CPU/MPS mode
+        no_half_gpu_list = ['1650', '1660'] # set False for GPUs that don't support f16
+        if not True in [gpu in torch.cuda.get_device_name(0) for gpu in no_half_gpu_list]:
+            use_half = True
+
+    model = RRDBNet(
+        num_in_ch=3,
+        num_out_ch=3,
+        num_feat=64,
+        num_block=23,
+        num_grow_ch=32,
+        scale=2,
+    )
+    upsampler = RealESRGANer(
+        scale=2,
+        model_path="https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/RealESRGAN_x2plus.pth",
+        model=model,
+        tile=args.bg_tile,
+        tile_pad=40,
+        pre_pad=0,
+        half=use_half
+    )
+
+    if not gpu_is_available():  # CPU
+        import warnings
+        warnings.warn('Running on CPU now! Make sure your PyTorch version matches your CUDA.'
+                        'The unoptimized RealESRGAN is slow on CPU. '
+                        'If you want to disable it, please remove `--bg_upsampler` and `--face_upsample` in command.',
+                        category=RuntimeWarning)
+    return upsampler
+
+if __name__ == '__main__':
+    # device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    device = get_device()
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('-i', '--input_path', type=str, default='./inputs/whole_imgs', 
+            help='Input image, video or folder. Default: inputs/whole_imgs')
+    parser.add_argument('-o', '--output_path', type=str, default=None, 
+            help='Output folder. Default: results/<input_name>_<w>')
+    parser.add_argument('-w', '--fidelity_weight', type=float, default=0.5, 
+            help='Balance the quality and fidelity. Default: 0.5')
+    parser.add_argument('-s', '--upscale', type=int, default=2, 
+            help='The final upsampling scale of the image. Default: 2')
+    parser.add_argument('--has_aligned', action='store_true', help='Input are cropped and aligned faces. Default: False')
+    parser.add_argument('--only_center_face', action='store_true', help='Only restore the center face. Default: False')
+    parser.add_argument('--draw_box', action='store_true', help='Draw the bounding box for the detected faces. Default: False')
+    # large det_model: 'YOLOv5l', 'retinaface_resnet50'
+    # small det_model: 'YOLOv5n', 'retinaface_mobile0.25'
+    parser.add_argument('--detection_model', type=str, default='retinaface_resnet50', 
+            help='Face detector. Optional: retinaface_resnet50, retinaface_mobile0.25, YOLOv5l, YOLOv5n, dlib. \
+                Default: retinaface_resnet50')
+    parser.add_argument('--bg_upsampler', type=str, default='None', help='Background upsampler. Optional: realesrgan')
+    parser.add_argument('--face_upsample', action='store_true', help='Face upsampler after enhancement. Default: False')
+    parser.add_argument('--bg_tile', type=int, default=400, help='Tile size for background sampler. Default: 400')
+    parser.add_argument('--suffix', type=str, default=None, help='Suffix of the restored faces. Default: None')
+    parser.add_argument('--save_video_fps', type=float, default=None, help='Frame rate for saving video. Default: None')
+
+    args = parser.parse_args()
+
+    # ------------------------ input & output ------------------------
+    w = args.fidelity_weight
+    input_video = False
+    if args.input_path.endswith(('jpg', 'jpeg', 'png', 'JPG', 'JPEG', 'PNG')): # input single img path
+        input_img_list = [args.input_path]
+        result_root = f'results/test_img_{w}'
+    elif args.input_path.endswith(('mp4', 'mov', 'avi', 'MP4', 'MOV', 'AVI')): # input video path
+        from basicsr.utils.video_util import VideoReader, VideoWriter
+        input_img_list = []
+        vidreader = VideoReader(args.input_path)
+        image = vidreader.get_frame()
+        while image is not None:
+            input_img_list.append(image)
+            image = vidreader.get_frame()
+        audio = vidreader.get_audio()
+        fps = vidreader.get_fps() if args.save_video_fps is None else args.save_video_fps   
+        video_name = os.path.basename(args.input_path)[:-4]
+        result_root = f'results/{video_name}_{w}'
+        input_video = True
+        vidreader.close()
+    else: # input img folder
+        if args.input_path.endswith('/'):  # solve when path ends with /
+            args.input_path = args.input_path[:-1]
+        # scan all the jpg and png images
+        input_img_list = sorted(glob.glob(os.path.join(args.input_path, '*.[jpJP][pnPN]*[gG]')))
+        result_root = f'results/{os.path.basename(args.input_path)}_{w}'
+
+    if not args.output_path is None: # set output path
+        result_root = args.output_path
+
+    test_img_num = len(input_img_list)
+    if test_img_num == 0:
+        raise FileNotFoundError('No input image/video is found...\n' 
+            '\tNote that --input_path for video should end with .mp4|.mov|.avi')
+
+    # ------------------ set up background upsampler ------------------
+    if args.bg_upsampler == 'realesrgan':
+        bg_upsampler = set_realesrgan()
+    else:
+        bg_upsampler = None
+
+    # ------------------ set up face upsampler ------------------
+    if args.face_upsample:
+        if bg_upsampler is not None:
+            face_upsampler = bg_upsampler
+        else:
+            face_upsampler = set_realesrgan()
+    else:
+        face_upsampler = None
+
+    # ------------------ set up CodeFormer restorer -------------------
+    net = ARCH_REGISTRY.get('CodeFormer')(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, 
+                                            connect_list=['32', '64', '128', '256']).to(device)
+    
+    # ckpt_path = 'experiments/20240106_012817_CodeFormer_stage3/models/net_g_150000.pth'
+    
+    ckpt_path = load_file_from_url(url=pretrain_model_url['restoration'], 
+                                    model_dir='weights/CodeFormer', progress=True, file_name=None)
+    checkpoint = torch.load(ckpt_path)['params_ema']
+    net.load_state_dict(checkpoint)
+    net.eval()
+
+    # ------------------ set up FaceRestoreHelper -------------------
+    # large det_model: 'YOLOv5l', 'retinaface_resnet50'
+    # small det_model: 'YOLOv5n', 'retinaface_mobile0.25'
+    if not args.has_aligned: 
+        print(f'Face detection model: {args.detection_model}')
+    if bg_upsampler is not None: 
+        print(f'Background upsampling: True, Face upsampling: {args.face_upsample}')
+    else:
+        print(f'Background upsampling: False, Face upsampling: {args.face_upsample}')
+
+    face_helper = FaceRestoreHelper(
+        args.upscale,
+        face_size=512,
+        crop_ratio=(1, 1),
+        det_model = args.detection_model,
+        save_ext='png',
+        use_parse=True,
+        device=device)
+
+    # -------------------- start to processing ---------------------
+    for i, img_path in enumerate(input_img_list):
+        # clean all the intermediate results to process the next image
+        face_helper.clean_all()
+        
+        if isinstance(img_path, str):
+            img_name = os.path.basename(img_path)
+            basename, ext = os.path.splitext(img_name)
+            print(f'[{i+1}/{test_img_num}] Processing: {img_name}')
+            img = cv2.imread(img_path, cv2.IMREAD_COLOR)
+        else: # for video processing
+            basename = str(i).zfill(6)
+            img_name = f'{video_name}_{basename}' if input_video else basename
+            print(f'[{i+1}/{test_img_num}] Processing: {img_name}')
+            img = img_path
+
+        if args.has_aligned: 
+            # the input faces are already cropped and aligned
+            img = cv2.resize(img, (512, 512), interpolation=cv2.INTER_LINEAR)
+            face_helper.is_gray = is_gray(img, threshold=10)
+            if face_helper.is_gray:
+                print('Grayscale input: True')
+            face_helper.cropped_faces = [img]
+        else:
+            face_helper.read_image(img)
+            # get face landmarks for each face
+            num_det_faces = face_helper.get_face_landmarks_5(
+                only_center_face=args.only_center_face, resize=640, eye_dist_threshold=5)
+            print(f'\tdetect {num_det_faces} faces')
+            # align and warp each face
+            face_helper.align_warp_face()
+
+        # face restoration for each cropped face
+        for idx, cropped_face in enumerate(face_helper.cropped_faces):
+            # prepare data
+            cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
+            normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+            cropped_face_t = cropped_face_t.unsqueeze(0).to(device)
+
+            try:
+                with torch.no_grad():
+                    output = net(cropped_face_t, w=w, adain=True)[0]
+                    restored_face = tensor2img(output, rgb2bgr=True, min_max=(-1, 1))
+                del output
+                torch.cuda.empty_cache()
+            except Exception as error:
+                print(f'\tFailed inference for CodeFormer: {error}')
+                restored_face = tensor2img(cropped_face_t, rgb2bgr=True, min_max=(-1, 1))
+
+            restored_face = restored_face.astype('uint8')
+            face_helper.add_restored_face(restored_face, cropped_face)
+
+        # paste_back
+        if not args.has_aligned:
+            # upsample the background
+            if bg_upsampler is not None:
+                # Now only support RealESRGAN for upsampling background
+                bg_img = bg_upsampler.enhance(img, outscale=args.upscale)[0]
+            else:
+                bg_img = None
+            face_helper.get_inverse_affine(None)
+            # paste each restored face to the input image
+            if args.face_upsample and face_upsampler is not None: 
+                restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img, draw_box=args.draw_box, face_upsampler=face_upsampler)
+            else:
+                restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img, draw_box=args.draw_box)
+
+        # save faces
+        for idx, (cropped_face, restored_face) in enumerate(zip(face_helper.cropped_faces, face_helper.restored_faces)):
+            # save cropped face
+            if not args.has_aligned: 
+                save_crop_path = os.path.join(result_root, 'cropped_faces', f'{basename}_{idx:02d}.png')
+                imwrite(cropped_face, save_crop_path)
+            # save restored face
+            if args.has_aligned:
+                save_face_name = f'{basename}.png'
+            else:
+                save_face_name = f'{basename}_{idx:02d}.png'
+            if args.suffix is not None:
+                save_face_name = f'{save_face_name[:-4]}_{args.suffix}.png'
+            save_restore_path = os.path.join(result_root, 'restored_faces', save_face_name)
+            imwrite(restored_face, save_restore_path)
+
+        # save restored img
+        if not args.has_aligned and restored_img is not None:
+            if args.suffix is not None:
+                basename = f'{basename}_{args.suffix}'
+            save_restore_path = os.path.join(result_root, 'final_results', f'{basename}.png')
+            imwrite(restored_img, save_restore_path)
+
+    # save enhanced video
+    if input_video:
+        print('Video Saving...')
+        # load images
+        video_frames = []
+        img_list = sorted(glob.glob(os.path.join(result_root, 'final_results', '*.[jp][pn]g')))
+        for img_path in img_list:
+            img = cv2.imread(img_path)
+            video_frames.append(img)
+        # write images to video
+        height, width = video_frames[0].shape[:2]
+        if args.suffix is not None:
+            video_name = f'{video_name}_{args.suffix}.png'
+        save_restore_path = os.path.join(result_root, f'{video_name}.mp4')
+        vidwriter = VideoWriter(save_restore_path, height, width, fps, audio)
+         
+        for f in video_frames:
+            vidwriter.write_frame(f)
+        vidwriter.close()
+
+    print(f'\nAll results are saved in {result_root}')

inference_colorization.py

+import os
+import cv2
+import argparse
+import glob
+import torch
+from torchvision.transforms.functional import normalize
+from basicsr.utils import imwrite, img2tensor, tensor2img
+from basicsr.utils.download_util import load_file_from_url
+from basicsr.utils.misc import get_device
+from basicsr.utils.registry import ARCH_REGISTRY
+
+pretrain_model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer_colorization.pth'
+
+if __name__ == '__main__':
+    # device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    device = get_device()
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('-i', '--input_path', type=str, default='./inputs/gray_faces', 
+                    help='Input image or folder. Default: inputs/gray_faces')
+    parser.add_argument('-o', '--output_path', type=str, default=None, 
+                    help='Output folder. Default: results/<input_name>')
+    parser.add_argument('--suffix', type=str, default=None, 
+                    help='Suffix of the restored faces. Default: None')
+    args = parser.parse_args()
+
+    # ------------------------ input & output ------------------------
+    print('[NOTE] The input face images should be aligned and cropped to a resolution of 512x512.')
+    if args.input_path.endswith(('jpg', 'jpeg', 'png', 'JPG', 'JPEG', 'PNG')): # input single img path
+        input_img_list = [args.input_path]
+        result_root = f'results/test_colorization_img'
+    else: # input img folder
+        if args.input_path.endswith('/'):  # solve when path ends with /
+            args.input_path = args.input_path[:-1]
+        # scan all the jpg and png images
+        input_img_list = sorted(glob.glob(os.path.join(args.input_path, '*.[jpJP][pnPN]*[gG]')))
+        result_root = f'results/{os.path.basename(args.input_path)}'
+
+    if not args.output_path is None: # set output path
+        result_root = args.output_path
+
+    test_img_num = len(input_img_list)
+
+    # ------------------ set up CodeFormer restorer -------------------
+    net = ARCH_REGISTRY.get('CodeFormer')(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, 
+                                            connect_list=['32', '64', '128']).to(device)
+    
+    # ckpt_path = 'weights/CodeFormer/codeformer.pth'
+    ckpt_path = load_file_from_url(url=pretrain_model_url, 
+                                    model_dir='weights/CodeFormer', progress=True, file_name=None)
+    checkpoint = torch.load(ckpt_path)['params_ema']
+    net.load_state_dict(checkpoint)
+    net.eval()
+
+    # -------------------- start to processing ---------------------
+    for i, img_path in enumerate(input_img_list):
+        img_name = os.path.basename(img_path)
+        basename, ext = os.path.splitext(img_name)
+        print(f'[{i+1}/{test_img_num}] Processing: {img_name}')
+        input_face = cv2.imread(img_path)
+        assert input_face.shape[:2] == (512, 512), 'Input resolution must be 512x512 for colorization.'
+        # input_face = cv2.resize(input_face, (512, 512), interpolation=cv2.INTER_LINEAR)
+        input_face = img2tensor(input_face / 255., bgr2rgb=True, float32=True)
+        normalize(input_face, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+        input_face = input_face.unsqueeze(0).to(device)
+        try:
+            with torch.no_grad():
+                # w is fixed to 0 since we didn't train the Stage III for colorization
+                output_face = net(input_face, w=0, adain=True)[0] 
+                save_face = tensor2img(output_face, rgb2bgr=True, min_max=(-1, 1))
+            del output_face
+            torch.cuda.empty_cache()
+        except Exception as error:
+            print(f'\tFailed inference for CodeFormer: {error}')
+            save_face = tensor2img(input_face, rgb2bgr=True, min_max=(-1, 1))
+
+        save_face = save_face.astype('uint8')
+
+        # save face
+        if args.suffix is not None:
+            basename = f'{basename}_{args.suffix}'
+        save_restore_path = os.path.join(result_root, f'{basename}.png')
+        imwrite(save_face, save_restore_path)
+
+    print(f'\nAll results are saved in {result_root}')
+

inference_inpainting.py

+import os
+import cv2
+import argparse
+import glob
+import torch
+from torchvision.transforms.functional import normalize
+from basicsr.utils import imwrite, img2tensor, tensor2img
+from basicsr.utils.download_util import load_file_from_url
+from basicsr.utils.misc import get_device
+from basicsr.utils.registry import ARCH_REGISTRY
+
+pretrain_model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer_inpainting.pth'
+
+if __name__ == '__main__':
+    # device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    device = get_device()
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('-i', '--input_path', type=str, default='./inputs/masked_faces', 
+                    help='Input image or folder. Default: inputs/masked_faces')
+    parser.add_argument('-o', '--output_path', type=str, default=None, 
+                    help='Output folder. Default: results/<input_name>')
+    parser.add_argument('--suffix', type=str, default=None, 
+                    help='Suffix of the restored faces. Default: None')
+    args = parser.parse_args()
+
+    # ------------------------ input & output ------------------------
+    print('[NOTE] The input face images should be aligned and cropped to a resolution of 512x512.')
+    if args.input_path.endswith(('jpg', 'jpeg', 'png', 'JPG', 'JPEG', 'PNG')): # input single img path
+        input_img_list = [args.input_path]
+        result_root = f'results/test_inpainting_img'
+    else: # input img folder
+        if args.input_path.endswith('/'):  # solve when path ends with /
+            args.input_path = args.input_path[:-1]
+        # scan all the jpg and png images
+        input_img_list = sorted(glob.glob(os.path.join(args.input_path, '*.[jpJP][pnPN]*[gG]')))
+        result_root = f'results/{os.path.basename(args.input_path)}'
+
+    if not args.output_path is None: # set output path
+        result_root = args.output_path
+
+    test_img_num = len(input_img_list)
+
+    # ------------------ set up CodeFormer restorer -------------------
+    net = ARCH_REGISTRY.get('CodeFormer')(dim_embd=512, codebook_size=512, n_head=8, n_layers=9, 
+                                            connect_list=['32', '64', '128']).to(device)
+    
+    # ckpt_path = 'weights/CodeFormer/codeformer.pth'
+    ckpt_path = load_file_from_url(url=pretrain_model_url, 
+                                    model_dir='weights/CodeFormer', progress=True, file_name=None)
+    checkpoint = torch.load(ckpt_path)['params_ema']
+    net.load_state_dict(checkpoint)
+    net.eval()
+
+    # -------------------- start to processing ---------------------
+    for i, img_path in enumerate(input_img_list):
+        img_name = os.path.basename(img_path)
+        basename, ext = os.path.splitext(img_name)
+        print(f'[{i+1}/{test_img_num}] Processing: {img_name}')
+        input_face = cv2.imread(img_path)
+        assert input_face.shape[:2] == (512, 512), 'Input resolution must be 512x512 for inpainting.'
+        # input_face = cv2.resize(input_face, (512, 512), interpolation=cv2.INTER_LINEAR)
+        input_face = img2tensor(input_face / 255., bgr2rgb=True, float32=True)
+        normalize(input_face, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+        input_face = input_face.unsqueeze(0).to(device)
+        try:
+            with torch.no_grad():
+                mask = torch.zeros(512, 512)
+                m_ind = torch.sum(input_face[0], dim=0)
+                mask[m_ind==3] = 1.0
+                mask = mask.view(1, 1, 512, 512).to(device)
+                # w is fixed to 1, adain=False for inpainting
+                output_face = net(input_face, w=1, adain=False)[0]
+                output_face = (1-mask)*input_face + mask*output_face
+                save_face = tensor2img(output_face, rgb2bgr=True, min_max=(-1, 1))
+            del output_face
+            torch.cuda.empty_cache()
+        except Exception as error:
+            print(f'\tFailed inference for CodeFormer: {error}')
+            save_face = tensor2img(input_face, rgb2bgr=True, min_max=(-1, 1))
+
+        save_face = save_face.astype('uint8')
+
+        # save face
+        if args.suffix is not None:
+            basename = f'{basename}_{args.suffix}'
+        save_restore_path = os.path.join(result_root, f'{basename}.png')
+        imwrite(save_face, save_restore_path)
+
+    print(f'\nAll results are saved in {result_root}')
+