Source code for kiui.sr

import os
import cv2
import math

import torch
import torch.nn as nn
import torch.nn.init as init
from torch.nn import functional as F
from torch.nn.modules.batchnorm import _BatchNorm

import numpy as np
from PIL import Image

from huggingface_hub import hf_hub_download

from kiui.typing import *

HF_MODELS = {
    2: dict(
        repo_id='ai-forever/Real-ESRGAN',
        filename='RealESRGAN_x2.pth',
    ),
    4: dict(
        repo_id='ai-forever/Real-ESRGAN',
        filename='RealESRGAN_x4.pth',
    ),
    8: dict(
        repo_id='ai-forever/Real-ESRGAN',
        filename='RealESRGAN_x8.pth',
    ),
}

@torch.no_grad()
def default_init_weights(module_list, scale=1, bias_fill=0, **kwargs):
    """Initialize network weights.

    Args:
        module_list (list[nn.Module] | nn.Module): Modules to be initialized.
        scale (float): Scale initialized weights, especially for residual
            blocks. Default: 1.
        bias_fill (float): The value to fill bias. Default: 0
        kwargs (dict): Other arguments for initialization function.
    """
    if not isinstance(module_list, list):
        module_list = [module_list]
    for module in module_list:
        for m in module.modules():
            if isinstance(m, nn.Conv2d):
                init.kaiming_normal_(m.weight, **kwargs)
                m.weight.data *= scale
                if m.bias is not None:
                    m.bias.data.fill_(bias_fill)
            elif isinstance(m, nn.Linear):
                init.kaiming_normal_(m.weight, **kwargs)
                m.weight.data *= scale
                if m.bias is not None:
                    m.bias.data.fill_(bias_fill)
            elif isinstance(m, _BatchNorm):
                init.constant_(m.weight, 1)
                if m.bias is not None:
                    m.bias.data.fill_(bias_fill)


def make_layer(basic_block, num_basic_block, **kwarg):
    """Make layers by stacking the same blocks.

    Args:
        basic_block (nn.module): nn.module class for basic block.
        num_basic_block (int): number of blocks.

    Returns:
        nn.Sequential: Stacked blocks in nn.Sequential.
    """
    layers = []
    for _ in range(num_basic_block):
        layers.append(basic_block(**kwarg))
    return nn.Sequential(*layers)


class ResidualBlockNoBN(nn.Module):
    """Residual block without BN.

    It has a style of:
        ---Conv-ReLU-Conv-+-
         |________________|

    Args:
        num_feat (int): Channel number of intermediate features.
            Default: 64.
        res_scale (float): Residual scale. Default: 1.
        pytorch_init (bool): If set to True, use pytorch default init,
            otherwise, use default_init_weights. Default: False.
    """

    def __init__(self, num_feat=64, res_scale=1, pytorch_init=False):
        super(ResidualBlockNoBN, self).__init__()
        self.res_scale = res_scale
        self.conv1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=True)
        self.conv2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=True)
        self.relu = nn.ReLU(inplace=True)

        if not pytorch_init:
            default_init_weights([self.conv1, self.conv2], 0.1)

    def forward(self, x):
        identity = x
        out = self.conv2(self.relu(self.conv1(x)))
        return identity + out * self.res_scale


class Upsample(nn.Sequential):
    """Upsample module.

    Args:
        scale (int): Scale factor. Supported scales: 2^n and 3.
        num_feat (int): Channel number of intermediate features.
    """

    def __init__(self, scale, num_feat):
        m = []
        if (scale & (scale - 1)) == 0:  # scale = 2^n
            for _ in range(int(math.log(scale, 2))):
                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
                m.append(nn.PixelShuffle(2))
        elif scale == 3:
            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
            m.append(nn.PixelShuffle(3))
        else:
            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
        super(Upsample, self).__init__(*m)


def flow_warp(x, flow, interp_mode='bilinear', padding_mode='zeros', align_corners=True):
    """Warp an image or feature map with optical flow.

    Args:
        x (Tensor): Tensor with size (n, c, h, w).
        flow (Tensor): Tensor with size (n, h, w, 2), normal value.
        interp_mode (str): 'nearest' or 'bilinear'. Default: 'bilinear'.
        padding_mode (str): 'zeros' or 'border' or 'reflection'.
            Default: 'zeros'.
        align_corners (bool): Before pytorch 1.3, the default value is
            align_corners=True. After pytorch 1.3, the default value is
            align_corners=False. Here, we use the True as default.

    Returns:
        Tensor: Warped image or feature map.
    """
    assert x.size()[-2:] == flow.size()[1:3]
    _, _, h, w = x.size()
    # create mesh grid
    grid_y, grid_x = torch.meshgrid(torch.arange(0, h).type_as(x), torch.arange(0, w).type_as(x))
    grid = torch.stack((grid_x, grid_y), 2).float()  # W(x), H(y), 2
    grid.requires_grad = False

    vgrid = grid + flow
    # scale grid to [-1,1]
    vgrid_x = 2.0 * vgrid[:, :, :, 0] / max(w - 1, 1) - 1.0
    vgrid_y = 2.0 * vgrid[:, :, :, 1] / max(h - 1, 1) - 1.0
    vgrid_scaled = torch.stack((vgrid_x, vgrid_y), dim=3)
    output = F.grid_sample(x, vgrid_scaled, mode=interp_mode, padding_mode=padding_mode, align_corners=align_corners)

    # TODO, what if align_corners=False
    return output


def resize_flow(flow, size_type, sizes, interp_mode='bilinear', align_corners=False):
    """Resize a flow according to ratio or shape.

    Args:
        flow (Tensor): Precomputed flow. shape [N, 2, H, W].
        size_type (str): 'ratio' or 'shape'.
        sizes (list[int | float]): the ratio for resizing or the final output
            shape.
            1) The order of ratio should be [ratio_h, ratio_w]. For
            downsampling, the ratio should be smaller than 1.0 (i.e., ratio
            < 1.0). For upsampling, the ratio should be larger than 1.0 (i.e.,
            ratio > 1.0).
            2) The order of output_size should be [out_h, out_w].
        interp_mode (str): The mode of interpolation for resizing.
            Default: 'bilinear'.
        align_corners (bool): Whether align corners. Default: False.

    Returns:
        Tensor: Resized flow.
    """
    _, _, flow_h, flow_w = flow.size()
    if size_type == 'ratio':
        output_h, output_w = int(flow_h * sizes[0]), int(flow_w * sizes[1])
    elif size_type == 'shape':
        output_h, output_w = sizes[0], sizes[1]
    else:
        raise ValueError(f'Size type should be ratio or shape, but got type {size_type}.')

    input_flow = flow.clone()
    ratio_h = output_h / flow_h
    ratio_w = output_w / flow_w
    input_flow[:, 0, :, :] *= ratio_w
    input_flow[:, 1, :, :] *= ratio_h
    resized_flow = F.interpolate(
        input=input_flow, size=(output_h, output_w), mode=interp_mode, align_corners=align_corners)
    return resized_flow


def pixel_unshuffle(x, scale):
    """ Pixel unshuffle.

    Args:
        x (Tensor): Input feature with shape (b, c, hh, hw).
        scale (int): Downsample ratio.

    Returns:
        Tensor: the pixel unshuffled feature.
    """
    b, c, hh, hw = x.size()
    out_channel = c * (scale**2)
    assert hh % scale == 0 and hw % scale == 0
    h = hh // scale
    w = hw // scale
    x_view = x.view(b, c, h, scale, w, scale)
    return x_view.permute(0, 1, 3, 5, 2, 4).reshape(b, out_channel, h, w)

def pad_reflect(image, pad_size):
    imsize = image.shape
    height, width = imsize[:2]
    new_img = np.zeros([height+pad_size*2, width+pad_size*2, imsize[2]]).astype(np.uint8)
    new_img[pad_size:-pad_size, pad_size:-pad_size, :] = image
    
    new_img[0:pad_size, pad_size:-pad_size, :] = np.flip(image[0:pad_size, :, :], axis=0) #top
    new_img[-pad_size:, pad_size:-pad_size, :] = np.flip(image[-pad_size:, :, :], axis=0) #bottom
    new_img[:, 0:pad_size, :] = np.flip(new_img[:, pad_size:pad_size*2, :], axis=1) #left
    new_img[:, -pad_size:, :] = np.flip(new_img[:, -pad_size*2:-pad_size, :], axis=1) #right
    
    return new_img

def unpad_image(image, pad_size):
    return image[pad_size:-pad_size, pad_size:-pad_size, :]


def process_array(image_array, expand=True):
    """ Process a 3-dimensional array into a scaled, 4 dimensional batch of size 1. """
    
    image_batch = image_array / 255.0
    if expand:
        image_batch = np.expand_dims(image_batch, axis=0)
    return image_batch


def process_output(output_tensor):
    """ Transforms the 4-dimensional output tensor into a suitable image format. """
    
    sr_img = output_tensor.clip(0, 1) * 255
    sr_img = np.uint8(sr_img)
    return sr_img


def pad_patch(image_patch, padding_size, channel_last=True):
    """ Pads image_patch with with padding_size edge values. """
    
    if channel_last:
        return np.pad(
            image_patch,
            ((padding_size, padding_size), (padding_size, padding_size), (0, 0)),
            'edge',
        )
    else:
        return np.pad(
            image_patch,
            ((0, 0), (padding_size, padding_size), (padding_size, padding_size)),
            'edge',
        )


def unpad_patches(image_patches, padding_size):
    return image_patches[:, padding_size:-padding_size, padding_size:-padding_size, :]


def split_image_into_overlapping_patches(image_array, patch_size, padding_size=2):
    """ Splits the image into partially overlapping patches.
    The patches overlap by padding_size pixels.
    Pads the image twice:
        - first to have a size multiple of the patch size,
        - then to have equal padding at the borders.
    Args:
        image_array: numpy array of the input image.
        patch_size: size of the patches from the original image (without padding).
        padding_size: size of the overlapping area.
    """
    
    xmax, ymax, _ = image_array.shape
    x_remainder = xmax % patch_size
    y_remainder = ymax % patch_size
    
    # modulo here is to avoid extending of patch_size instead of 0
    x_extend = (patch_size - x_remainder) % patch_size
    y_extend = (patch_size - y_remainder) % patch_size
    
    # make sure the image is divisible into regular patches
    extended_image = np.pad(image_array, ((0, x_extend), (0, y_extend), (0, 0)), 'edge')
    
    # add padding around the image to simplify computations
    padded_image = pad_patch(extended_image, padding_size, channel_last=True)
    
    xmax, ymax, _ = padded_image.shape
    patches = []
    
    x_lefts = range(padding_size, xmax - padding_size, patch_size)
    y_tops = range(padding_size, ymax - padding_size, patch_size)
    
    for x in x_lefts:
        for y in y_tops:
            x_left = x - padding_size
            y_top = y - padding_size
            x_right = x + patch_size + padding_size
            y_bottom = y + patch_size + padding_size
            patch = padded_image[x_left:x_right, y_top:y_bottom, :]
            patches.append(patch)
    
    return np.array(patches), padded_image.shape


def stich_together(patches, padded_image_shape, target_shape, padding_size=4):
    """ Reconstruct the image from overlapping patches.
    After scaling, shapes and padding should be scaled too.
    Args:
        patches: patches obtained with split_image_into_overlapping_patches
        padded_image_shape: shape of the padded image contructed in split_image_into_overlapping_patches
        target_shape: shape of the final image
        padding_size: size of the overlapping area.
    """
    
    xmax, ymax, _ = padded_image_shape
    patches = unpad_patches(patches, padding_size)
    patch_size = patches.shape[1]
    n_patches_per_row = ymax // patch_size
    
    complete_image = np.zeros((xmax, ymax, 3))
    
    row = -1
    col = 0
    for i in range(len(patches)):
        if i % n_patches_per_row == 0:
            row += 1
            col = 0
        complete_image[
        row * patch_size: (row + 1) * patch_size, col * patch_size: (col + 1) * patch_size,:
        ] = patches[i]
        col += 1
    return complete_image[0: target_shape[0], 0: target_shape[1], :]


class ResidualDenseBlock(nn.Module):
    """Residual Dense Block.

    Used in RRDB block in ESRGAN.

    Args:
        num_feat (int): Channel number of intermediate features.
        num_grow_ch (int): Channels for each growth.
    """

    def __init__(self, num_feat=64, num_grow_ch=32):
        super(ResidualDenseBlock, self).__init__()
        self.conv1 = nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
        self.conv2 = nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
        self.conv3 = nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
        self.conv4 = nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
        self.conv5 = nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)

        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)

        # initialization
        default_init_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)

    def forward(self, x):
        x1 = self.lrelu(self.conv1(x))
        x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
        # Emperically, we use 0.2 to scale the residual for better performance
        return x5 * 0.2 + x


class RRDB(nn.Module):
    """Residual in Residual Dense Block.

    Used in RRDB-Net in ESRGAN.

    Args:
        num_feat (int): Channel number of intermediate features.
        num_grow_ch (int): Channels for each growth.
    """

    def __init__(self, num_feat, num_grow_ch=32):
        super(RRDB, self).__init__()
        self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
        self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
        self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)

    def forward(self, x):
        out = self.rdb1(x)
        out = self.rdb2(out)
        out = self.rdb3(out)
        # Emperically, we use 0.2 to scale the residual for better performance
        return out * 0.2 + x


class RRDBNet(nn.Module):
    """Networks consisting of Residual in Residual Dense Block, which is used
    in ESRGAN.

    ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks.

    We extend ESRGAN for scale x2 and scale x1.
    Note: This is one option for scale 1, scale 2 in RRDBNet.
    We first employ the pixel-unshuffle (an inverse operation of pixelshuffle to reduce the spatial size
    and enlarge the channel size before feeding inputs into the main ESRGAN architecture.

    Args:
        num_in_ch (int): Channel number of inputs.
        num_out_ch (int): Channel number of outputs.
        num_feat (int): Channel number of intermediate features.
            Default: 64
        num_block (int): Block number in the trunk network. Defaults: 23
        num_grow_ch (int): Channels for each growth. Default: 32.
    """

    def __init__(self, num_in_ch, num_out_ch, scale=4, num_feat=64, num_block=23, num_grow_ch=32):
        super(RRDBNet, self).__init__()
        self.scale = scale
        if scale == 2:
            num_in_ch = num_in_ch * 4
        elif scale == 1:
            num_in_ch = num_in_ch * 16
        self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
        self.body = make_layer(RRDB, num_block, num_feat=num_feat, num_grow_ch=num_grow_ch)
        self.conv_body = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
        # upsample
        self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
        self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
        if scale == 8:
            self.conv_up3 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
        self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
        self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)

        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)

    def forward(self, x):
        if self.scale == 2:
            feat = pixel_unshuffle(x, scale=2)
        elif self.scale == 1:
            feat = pixel_unshuffle(x, scale=4)
        else:
            feat = x
        feat = self.conv_first(feat)
        body_feat = self.conv_body(self.body(feat))
        feat = feat + body_feat
        # upsample
        feat = self.lrelu(self.conv_up1(F.interpolate(feat, scale_factor=2, mode='nearest')))
        feat = self.lrelu(self.conv_up2(F.interpolate(feat, scale_factor=2, mode='nearest')))
        if self.scale == 8:
            feat = self.lrelu(self.conv_up3(F.interpolate(feat, scale_factor=2, mode='nearest')))
        out = self.conv_last(self.lrelu(self.conv_hr(feat)))
        return out


class RealESRGAN:
    def __init__(self, device, scale=4):
        print(f'[INFO] init RealESRGAN_{scale}x: {device}')
        self.device = device
        self.scale = scale
        self.model = RRDBNet(
            num_in_ch=3, num_out_ch=3, num_feat=64, 
            num_block=23, num_grow_ch=32, scale=scale
        )
        self.load_weights()
        
    def load_weights(self):
        model_path = hf_hub_download(repo_id=HF_MODELS[self.scale]['repo_id'], filename=HF_MODELS[self.scale]['filename'])
        checkpoint = torch.load(model_path)
        if 'params' in checkpoint:
            self.model.load_state_dict(checkpoint['params'], strict=True)
        elif 'params_ema' in checkpoint:
            self.model.load_state_dict(checkpoint['params_ema'], strict=True)
        else:
            self.model.load_state_dict(checkpoint, strict=True)
        self.model.eval()
        self.model.to(self.device)
        
    @torch.cuda.amp.autocast()
    def predict(self, lr_image, batch_size=4, patches_size=192, padding=24, pad_size=15):
        # lr_image: np.ndarray, [h, w, 3], RGB uint8
        # return: np.ndarray, [H, W, 3], RGB uint8

        return_tensor = False
        if torch.is_tensor(lr_image):
            # or Tensor, [1, 3, H, W], RGB float32
            lr_image = (lr_image.detach().permute(0,2,3,1)[0].cpu().numpy() * 255).astype(np.uint8)
            return_tensor = True
        
        lr_image = pad_reflect(lr_image, pad_size)

        patches, p_shape = split_image_into_overlapping_patches(lr_image, patch_size=patches_size, padding_size=padding)
        img = torch.from_numpy(patches.astype(np.float32) / 255).permute((0,3,1,2)).to(self.device).detach()

        with torch.no_grad():
            res = self.model(img[0:batch_size])
            for i in range(batch_size, img.shape[0], batch_size):
                res = torch.cat((res, self.model(img[i:i+batch_size])), 0)

        sr_image = res.permute((0,2,3,1)).clamp_(0, 1).cpu()
        np_sr_image = sr_image.numpy()

        padded_size_scaled = tuple(np.multiply(p_shape[0:2], self.scale)) + (3,)
        scaled_image_shape = tuple(np.multiply(lr_image.shape[0:2], self.scale)) + (3,)

        np_sr_image = stich_together(
            np_sr_image, padded_image_shape=padded_size_scaled, 
            target_shape=scaled_image_shape, padding_size=padding * self.scale
        )

        sr_img = (np_sr_image * 255).astype(np.uint8)
        sr_img = unpad_image(sr_img, pad_size * self.scale)

        if return_tensor:
            sr_img = torch.from_numpy(sr_img.astype(np.float32) / 255).permute((2,0,1)).unsqueeze(0).to(self.device)
        
        return sr_img


MODELS = {}


def sr(image: ndarray, scale: Literal[2, 4, 8] = 2, device=None):
    """ lazy load functional super-resolution API for convenience.

    Args:
        image (ndarray): input image, uint8/float32 [H, W, 3]
        scale (Literal[2, 4, 8], optional): upscale factor. Defaults to 2.
        device (torch.device, optional): device to put SR models, if not provided, will try to use 'cuda'. Defaults to None.

    Returns:
        ndarray: super-resolutioned image, uint8/float32 [H * scale, W * scale, 3]
    """
    global MODELS
    if scale not in MODELS:
        if device is None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        MODELS[scale] = RealESRGAN(device, scale=scale)
    
    return_float = False
    if image.dtype == np.float32:
        return_float = True
        image = (image * 255).astype(np.uint8)

    sr_image = MODELS[scale].predict(image)

    if return_float:
        sr_image = sr_image.astype(np.float32) / 255.0

    return sr_image


def main():
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument('input', type=str)
    parser.add_argument('--output', type=str, default=None)
    parser.add_argument('--scale', type=int, default=4)
    args = parser.parse_args()

    model = RealESRGAN('cuda', scale=4)

    if args.output is None:
        args.output = os.path.splitext(args.input)[0] + f'_{args.scale}x.jpg'

    image = cv2.imread(args.input)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    sr_image = model.predict(image)
    sr_image = cv2.cvtColor(sr_image, cv2.COLOR_RGB2BGR)

    cv2.imwrite(args.output, sr_image)


if __name__ == '__main__':
    main()