export_onnx.py

import os
import urllib.request
import torch
import torch.nn as nn
from torch.nn import functional as F

# =================================================================
# OFFICIAL REAL-ESRGAN 6B ANIME ARCHITECTURE DEFINITION
# =================================================================
class ResidualDenseBlock_5C(nn.Module):
    def __init__(self, nf=64, gc=32, bias=True):
        super(ResidualDenseBlock_5C, self).__init__()
        self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias)
        self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias)
        self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias)
        self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias)
        self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias)

    def forward(self, x):
        x1 = F.lrelu(self.conv1(x), 0.2, inplace=True)
        x2 = F.lrelu(self.conv2(torch.cat((x, x1), 1)), 0.2, inplace=True)
        x3 = F.lrelu(self.conv3(torch.cat((x, x1, x2), 1)), 0.2, inplace=True)
        x4 = F.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)), 0.2, inplace=True)
        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
        return x5 * 0.2 + x

class RRDB(nn.Module):
    def __init__(self, nf, gc=32):
        super(RRDB, self).__init__()
        self.RDB1 = ResidualDenseBlock_5C(nf, gc)
        self.RDB2 = ResidualDenseBlock_5C(nf, gc)
        self.RDB3 = ResidualDenseBlock_5C(nf, gc)

    def forward(self, x):
        return self.RDB3(self.RDB2(self.RDB1(x))) * 0.2 + x

class RRDBNet(nn.Module):
    def __init__(self, in_nc=3, out_nc=3, nf=64, nb=6, gc=32, scale=4):
        super(RRDBNet, self).__init__()
        self.scale = scale
        self.conv_first = nn.Conv2d(in_nc, nf, 3, 1, 1)
        self.body = nn.Sequential(*[RRDB(nf, gc) for _ in range(nb)])
        self.conv_body = nn.Conv2d(nf, nf, 3, 1, 1)
        self.conv_hr = nn.Conv2d(nf, nf, 3, 1, 1)
        self.conv_last = nn.Conv2d(nf, out_nc, 3, 1, 1)
        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)

    def forward(self, x):
        fea = self.conv_first(x)
        body_fea = self.conv_body(self.body(fea))
        fea = fea + body_fea

        # PixelShuffle Upsampling Layer
        fea = self.lrelu(self.conv_hr(F.interpolate(fea, scale_factor=self.scale, mode='nearest')))
        out = self.conv_last(fea)
        return out

# =================================================================
# DOWNLOAD AND EXPORT EXECUTOR
# =================================================================
if __name__ == "__main__":
    pth_url = "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth"
    pth_path = "RealESRGAN_x4plus_anime_6B.pth"
    onnx_path = "RealESRGAN_x4plus_anime_6B.onnx"

    if not os.path.exists(pth_path):
        print("Downloading official PyTorch .pth weights...")
        opener = urllib.request.build_opener()
        opener.addheaders = [('User-Agent', 'Mozilla/5.0')]
        urllib.request.install_opener(opener)
        urllib.request.urlretrieve(pth_url, pth_path)

    print("Initializing RRDBNet Architecture (6 Blocks)...")
    model = RRDBNet(in_nc=3, out_nc=3, nf=64, nb=6, gc=32, scale=4)
    
    print("Loading weight checkpoints into model...")
    loadnet = torch.load(pth_path, map_location=torch.device('cpu'))
    if 'params_ema' in loadnet:
        keyname = 'params_ema'
    elif 'params' in loadnet:
        keyname = 'params'
    else:
        keyname = next(iter(loadnet))
    model.load_state_dict(loadnet[keyname], strict=True)
    model.eval()

    # Create dummy tensor representing [Batch, Channels, Height, Width]
    dummy_input = torch.randn(1, 3, 64, 64, dtype=torch.float32)

    print("Exporting to ONNX layout with DYNAMIC SHAPES...")
    torch.onnx.export(
        model,
        dummy_input,
        onnx_path,
        export_params=True,
        opset_version=14,
        do_constant_folding=True,
        input_names=['input'],
        output_names=['output'],
        dynamic_axes={
            'input': {2: 'height', 3: 'width'},
            'output': {2: 'height', 3: 'width'}
        }
    )
    print(f"Successfully baked dynamic shape model to: {onnx_path}")