scripts/export_lama.py

"""Export LaMa-manga from safetensors → ONNX with native DFT ops.

Torch 2.9+ exports torch.fft.rfft2/irfft2 as ONNX DFT operators,
producing a compact graph with dynamic spatial dimensions.

Requirements:
  pip install torch>=2.9 safetensors huggingface_hub onnx onnxscript

Usage:
  python scripts/export_lama.py -o models/lama-manga.onnx
"""
import argparse
import os
import logging
from collections import Counter, OrderedDict

logging.basicConfig(level=logging.INFO)

import torch
import torch.nn as nn


# ── LaMa architecture (FFCResNetGenerator) ──────────────────────────────────

class FourierUnit(nn.Module):
    """FourierUnit using native torch.fft (exports as ONNX DFT ops)."""
    def __init__(self, in_channels, out_channels, groups=1, **kwargs):
        super().__init__()
        self.groups = groups
        self.conv_layer = nn.Conv2d(in_channels * 2, out_channels * 2,
                                     kernel_size=1, stride=1, padding=0,
                                     groups=groups, bias=False)
        self.bn = nn.BatchNorm2d(out_channels * 2)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, x):
        batch, channels, h, w = x.shape
        spec = torch.fft.rfft2(x, norm='backward')
        real = spec.real   # [B, C, H, W_half]
        imag = spec.imag
        w_half = w // 2 + 1

        # Interleave real/imag as channels without 5D tensors:
        # [B, C, 1, H*W_half] + cat(dim=2) → [B, C, 2, H*W_half] → [B, 2C, H, W_half]
        real_flat = real.reshape(batch, channels, 1, h * w_half)
        imag_flat = imag.reshape(batch, channels, 1, h * w_half)
        ffted = torch.cat([real_flat, imag_flat], dim=2).reshape(batch, channels * 2, h, w_half)
        ffted = self.relu(self.bn(self.conv_layer(ffted)))
        # De-interleave back: [B, 2C, H, W_half] → [B, C, 2, H*W_half] → split
        out_c = ffted.shape[1] // 2
        ffted = ffted.reshape(batch, out_c, 2, h * w_half)
        out_r = ffted[:, :, 0, :].reshape(batch, out_c, h, w_half)
        out_i = ffted[:, :, 1, :].reshape(batch, out_c, h, w_half)
        spec_out = torch.complex(out_r, out_i)
        return torch.fft.irfft2(spec_out, s=(h, w), norm='backward')


class SpectralTransform(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, groups=1,
                 enable_lfu=True, **kwargs):
        super().__init__()
        self.enable_lfu = enable_lfu
        self.downsample = nn.AvgPool2d(2, 2) if stride == 2 else nn.Identity()
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels, out_channels // 2, 1, groups=groups, bias=False),
            nn.BatchNorm2d(out_channels // 2),
            nn.ReLU(inplace=True),
        )
        self.fu = FourierUnit(out_channels // 2, out_channels // 2, groups)
        if enable_lfu:
            self.lfu = FourierUnit(out_channels // 2, out_channels // 2, groups)
        self.conv2 = nn.Conv2d(out_channels // 2, out_channels, 1,
                                groups=groups, bias=False)

    def forward(self, x):
        x = self.downsample(x)
        x = self.conv1(x)
        output = self.fu(x)
        if self.enable_lfu:
            n, c, h, w = x.shape
            split_s = h // 2
            xs = torch.cat(torch.split(x[:, :c // 4], split_s, dim=-2), dim=1).contiguous()
            xs = torch.cat(torch.split(xs, split_s, dim=-1), dim=1).contiguous()
            xs = self.lfu(xs)
            xs = xs.repeat(1, 1, 2, 2).contiguous()
        else:
            xs = 0
        return self.conv2(x + output + xs)


class FFC(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size,
                 ratio_gin, ratio_gout, stride=1, padding=0,
                 dilation=1, groups=1, bias=False, enable_lfu=True,
                 padding_type='reflect', **spectral_kwargs):
        super().__init__()
        self.ratio_gin = ratio_gin
        self.ratio_gout = ratio_gout

        in_cg = int(in_channels * ratio_gin)
        in_cl = in_channels - in_cg
        out_cg = int(out_channels * ratio_gout)
        out_cl = out_channels - out_cg

        module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d
        self.convl2l = module(in_cl, out_cl, kernel_size, stride, padding,
                              dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d
        self.convl2g = module(in_cl, out_cg, kernel_size, stride, padding,
                              dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d
        self.convg2l = module(in_cg, out_cl, kernel_size, stride, padding,
                              dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform
        self.convg2g = module(in_cg, out_cg, stride, 1 if groups == 1 else groups // 2,
                              enable_lfu, **spectral_kwargs)

    def forward(self, x):
        # ratio_gin/ratio_gout are Python floats — tracer follows the correct branch.
        if self.ratio_gin == 0:
            x_l = x if not isinstance(x, tuple) else x[0]
            out_xl = self.convl2l(x_l) if self.ratio_gout != 1 else None
            out_xg = self.convl2g(x_l) if self.ratio_gout != 0 else None
        else:
            x_l, x_g = x
            out_xl = (self.convl2l(x_l) + self.convg2l(x_g)) if self.ratio_gout != 1 else None
            out_xg = (self.convl2g(x_l) + self.convg2g(x_g)) if self.ratio_gout != 0 else None
        return out_xl, out_xg


class FFC_BN_ACT(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size,
                 ratio_gin, ratio_gout, stride=1, padding=0,
                 dilation=1, groups=1, bias=False,
                 norm_layer=nn.BatchNorm2d, activation_layer=nn.Identity,
                 padding_type='reflect', enable_lfu=True, **kwargs):
        super().__init__()
        self.ffc = FFC(in_channels, out_channels, kernel_size,
                       ratio_gin, ratio_gout, stride, padding, dilation,
                       groups, bias, enable_lfu, padding_type, **kwargs)
        global_channels = int(out_channels * ratio_gout)
        self.bn_l = nn.Identity() if ratio_gout == 1 else norm_layer(out_channels - global_channels)
        self.bn_g = nn.Identity() if ratio_gout == 0 else norm_layer(global_channels)
        self.act_l = nn.Identity() if ratio_gout == 1 else activation_layer(inplace=True)
        self.act_g = nn.Identity() if ratio_gout == 0 else activation_layer(inplace=True)

    def forward(self, x):
        x_l, x_g = self.ffc(x)
        out_l = self.act_l(self.bn_l(x_l)) if x_l is not None else None
        out_g = self.act_g(self.bn_g(x_g)) if x_g is not None else None
        return out_l, out_g


class FFCResnetBlock(nn.Module):
    def __init__(self, dim, padding_type, norm_layer, activation_layer=nn.ReLU,
                 dilation=1, inline=False, **conv_kwargs):
        super().__init__()
        self.conv1 = FFC_BN_ACT(dim, dim, 3, padding=dilation, dilation=dilation,
                                norm_layer=norm_layer, activation_layer=activation_layer,
                                padding_type=padding_type, **conv_kwargs)
        self.conv2 = FFC_BN_ACT(dim, dim, 3, padding=dilation, dilation=dilation,
                                norm_layer=norm_layer, activation_layer=activation_layer,
                                padding_type=padding_type, **conv_kwargs)
        self.inline = inline

    def forward(self, x):
        if isinstance(x, tuple):
            x_l, x_g = x
        else:
            x_l, x_g = x, None
        id_l, id_g = x_l, x_g
        x_l, x_g = self.conv1((x_l, x_g))
        x_l, x_g = self.conv2((x_l, x_g))
        out_l = id_l + x_l if x_l is not None else None
        out_g = id_g + x_g if x_g is not None else None
        return out_l, out_g


class ConcatTupleLayer(nn.Module):
    def forward(self, x):
        x_l, x_g = x
        if x_g is None:
            return x_l
        return torch.cat([x_l, x_g], dim=1)


class LaMaModel(nn.Module):
    """LaMa FFCResNetGenerator with separate image/mask inputs."""
    def __init__(self):
        super().__init__()
        ngf = 64
        n_downsampling = 3
        n_blocks = 18
        norm_layer = nn.BatchNorm2d

        resnet_kw = dict(ratio_gin=0.75, ratio_gout=0.75, enable_lfu=False)
        init_kw = dict(ratio_gin=0, ratio_gout=0, enable_lfu=False)
        down_kw = dict(ratio_gin=0, ratio_gout=0, enable_lfu=False)

        layers = [
            nn.ReflectionPad2d(3),
            FFC_BN_ACT(4, ngf, 7, padding=0, norm_layer=norm_layer,
                       activation_layer=nn.ReLU, **init_kw),
        ]

        # Downsample
        for i in range(n_downsampling):
            mult = 2 ** i
            kw = dict(down_kw)
            if i == n_downsampling - 1:
                kw['ratio_gout'] = resnet_kw['ratio_gin']
            layers.append(FFC_BN_ACT(
                min(1024, ngf * mult), min(1024, ngf * mult * 2),
                3, stride=2, padding=1,
                norm_layer=norm_layer, activation_layer=nn.ReLU, **kw))

        # ResNet blocks
        mult = 2 ** n_downsampling
        feats = min(1024, ngf * mult)
        for _ in range(n_blocks):
            layers.append(FFCResnetBlock(feats, padding_type='reflect',
                                          norm_layer=norm_layer,
                                          activation_layer=nn.ReLU,
                                          **resnet_kw))

        layers.append(ConcatTupleLayer())

        # Upsample
        for i in range(n_downsampling):
            mult = 2 ** (n_downsampling - i)
            layers += [
                nn.ConvTranspose2d(min(1024, ngf * mult),
                                    min(1024, ngf * mult // 2),
                                    3, stride=2, padding=1, output_padding=1),
                norm_layer(min(1024, ngf * mult // 2)),
                nn.ReLU(True),
            ]

        layers += [nn.ReflectionPad2d(3), nn.Conv2d(ngf, 3, 7), nn.Sigmoid()]
        self.generator = nn.Sequential(*layers)

    def forward(self, image, mask):
        masked_img = image * (1.0 - mask)
        x = torch.cat([masked_img, mask], dim=1)
        return self.generator(x)


def load_weights(model, sf_path):
    """Load safetensors weights, mapping key prefixes.

    Safetensors keys: model.N.xxx  (nn.Sequential index)
    Our model keys:   generator.N.xxx
    """
    from safetensors.torch import load_file
    state = load_file(sf_path)

    mapped = OrderedDict()
    for k, v in state.items():
        # model.N.xxx → generator.N.xxx
        nk = k.replace("model.", "generator.", 1)
        mapped[nk] = v

    result = model.load_state_dict(mapped, strict=False)
    if result.missing_keys:
        real_missing = [k for k in result.missing_keys if 'num_batches_tracked' not in k]
        if real_missing:
            print(f"WARNING: {len(real_missing)} genuinely missing keys:")
            for k in real_missing[:20]:
                print(f"  {k}")
    if result.unexpected_keys:
        print(f"NOTE: {len(result.unexpected_keys)} unexpected keys (ignored)")
    loaded = len(state) - len(result.unexpected_keys)
    print(f"Loaded {loaded}/{len(state)} weights")
    return model


def resolve_safetensors_path(local_path):
    if local_path:
        return local_path
    from huggingface_hub import hf_hub_download
    print("Downloading safetensors from mayocream/lama-manga...")
    return hf_hub_download("mayocream/lama-manga", "lama-manga.safetensors")


def verify_coreml_session(model_path):
    try:
        import onnxruntime as ort
    except Exception as e:
        print(f"CoreML verify skipped: onnxruntime unavailable ({e})")
        return

    providers = ort.get_available_providers()
    if "CoreMLExecutionProvider" not in providers:
        print(f"CoreML verify skipped: CoreMLExecutionProvider not available ({providers})")
        return

    ml_opts = {
        "ModelFormat": "MLProgram",
        "MLComputeUnits": "CPUAndNeuralEngine",
    }
    sess = ort.InferenceSession(
        model_path,
        providers=[("CoreMLExecutionProvider", ml_opts), "CPUExecutionProvider"],
    )
    got_coreml = "CoreMLExecutionProvider" in sess.get_providers()
    if got_coreml:
        print("CoreML MLProgram session init: OK")
    else:
        raise RuntimeError(
            f"CoreML MLProgram init fallback detected; active providers={sess.get_providers()}"
        )


def main():
    parser = argparse.ArgumentParser(description="Export LaMa-manga ONNX for CoreML/ANE")
    parser.add_argument("-o", "--output", default="models/lama-manga.onnx")
    parser.add_argument("--safetensors", default=None, help="Local safetensors path")
    parser.add_argument("--verify", action="store_true", help="Run test inference")
    parser.add_argument(
        "--verify-coreml",
        action="store_true",
        help="Try creating an ONNX Runtime CoreML MLProgram session after export",
    )

    args = parser.parse_args()
    sf_path = resolve_safetensors_path(args.safetensors)

    # 1. Build model
    print("Building LaMa model (rank-4-safe DFT)...")
    model = LaMaModel()

    # 2. Load weights
    print(f"Loading weights from {sf_path}")
    model = load_weights(model, sf_path)
    model.eval()

    # 3. Optional: verify inference
    if args.verify:
        with torch.no_grad():
            img = torch.randn(1, 3, 512, 512)
            mask = torch.zeros(1, 1, 512, 512)
            mask[:, :, 100:200, 100:200] = 1.0
            out = model(img, mask)
            print(f"Test output: shape={out.shape}, range=[{out.min():.3f}, {out.max():.3f}]")

    # 4. Export ONNX
    print("Exporting ONNX...")
    dummy_img = torch.randn(1, 3, 512, 512)
    dummy_mask = torch.zeros(1, 1, 512, 512)

    # LaMa has 3× stride-2 downsampling; input must be multiple of 8.
    # Rust inpaint.rs already pads to multiple of 8.
    h_base = torch.export.Dim("h_blocks", min=8, max=512)
    w_base = torch.export.Dim("w_blocks", min=8, max=512)
    height = h_base * 8
    width = w_base * 8
    dynamic_shapes = {
        "image": {2: height, 3: width},
        "mask": {2: height, 3: width},
    }

    with torch.no_grad():
        torch.onnx.export(
            model,
            (dummy_img, dummy_mask),
            args.output,
            input_names=["image", "mask"],
            output_names=["output"],
            dynamic_shapes=dynamic_shapes,
        )

    # 5. Rewrite DFT(inverse=1, onesided=1) → hermitian pad + DFT(inverse=1, onesided=0).
    #    ORT does not support inverse+onesided combo. The decomposition is:
    #    1. Take half-spectrum input [B,C,H,W//2+1,2]
    #    2. Mirror bins [1:-1] with conjugate to reconstruct full spectrum
    #    3. Run DFT(inverse=1, onesided=0) on the full spectrum
    #    This is the standard irfft decomposition, just done at graph level.
    import onnx
    from onnx import helper, TensorProto

    model_onnx = onnx.load(args.output, load_external_data=True)
    graph = model_onnx.graph

    fixed = 0
    nodes_to_add = []
    nodes_to_remove = []

    for node in list(graph.node):
        if node.op_type != "DFT":
            continue
        attrs = {a.name: a.i for a in node.attribute}
        if not (attrs.get("inverse", 0) == 1 and attrs.get("onesided", 0) == 1):
            continue

        inp = node.input[0]   # half-spectrum [..., W//2+1, 2]
        out = node.output[0]
        uid = f"_irfft_{fixed}"

        # Slice middle bins [1:-1] along axis -2
        starts = helper.make_tensor(f"starts{uid}", TensorProto.INT64, [1], [1])
        ends = helper.make_tensor(f"ends{uid}", TensorProto.INT64, [1], [-1])
        axes = helper.make_tensor(f"axes{uid}", TensorProto.INT64, [1], [-2])
        graph.initializer.extend([starts, ends, axes])
        mid_name = f"mid{uid}"
        nodes_to_add.append(helper.make_node(
            "Slice", [inp, starts.name, ends.name, axes.name], [mid_name]))

        # Flip middle bins along axis -2 via Slice with step=-1
        flip_starts = helper.make_tensor(f"flip_s{uid}", TensorProto.INT64, [1], [-1])
        flip_ends = helper.make_tensor(f"flip_e{uid}", TensorProto.INT64, [1], [-2147483648])  # INT_MIN
        flip_axes = helper.make_tensor(f"flip_ax{uid}", TensorProto.INT64, [1], [-2])
        flip_steps = helper.make_tensor(f"flip_st{uid}", TensorProto.INT64, [1], [-1])
        graph.initializer.extend([flip_starts, flip_ends, flip_axes, flip_steps])
        flip_name = f"flip{uid}"
        nodes_to_add.append(helper.make_node(
            "Slice", [mid_name, flip_starts.name, flip_ends.name, flip_axes.name, flip_steps.name],
            [flip_name]))

        # Conjugate = negate imag part: [..., 2] → flip sign of channel 1
        conj_scale = helper.make_tensor(f"conj{uid}", TensorProto.FLOAT, [2], [1.0, -1.0])
        graph.initializer.append(conj_scale)
        conj_name = f"conj_out{uid}"
        nodes_to_add.append(helper.make_node(
            "Mul", [flip_name, conj_scale.name], [conj_name]))

        # Concat: [half, conj_mirror] along axis -2
        full_name = f"full{uid}"
        nodes_to_add.append(helper.make_node(
            "Concat", [inp, conj_name], [full_name], axis=-2))

        # DFT(inverse=1, onesided=0) on full spectrum
        new_dft = helper.make_node(
            "DFT", [full_name], [out],
            inverse=1, onesided=0, name=f"idft{uid}")
        # Copy dft_length input if present
        if len(node.input) > 1 and node.input[1]:
            new_dft.input.append(node.input[1])
        nodes_to_add.append(new_dft)

        nodes_to_remove.append(node)
        fixed += 1

    for n in nodes_to_remove:
        graph.node.remove(n)
    graph.node.extend(nodes_to_add)

    print(f"Rewrote {fixed} DFT(inverse+onesided) → hermitian pad + DFT(inverse)")

    # 6. Rewrite Shape(start=..., end=...) to Shape + Slice.
    #    CoreML EP's MLProgram path currently rejects Shape with end == rank
    #    (e.g. start=3,end=4 on rank-4 tensors) with:
    #    "axis 4 not in valid range [-4,3]".
    shape_rewrites = 0
    rewritten_nodes = []
    for node in list(graph.node):
        if node.op_type != "Shape":
            rewritten_nodes.append(node)
            continue

        attrs = {a.name: a.i for a in node.attribute}
        if "start" not in attrs and "end" not in attrs:
            rewritten_nodes.append(node)
            continue

        start = attrs.get("start", 0)
        end = attrs.get("end", 9223372036854775807)
        uid = f"_shapefix_{shape_rewrites}"

        full_shape_out = f"{node.output[0]}{uid}_all"
        shape_name = f"{node.name}_all" if node.name else f"shape_all{uid}"
        slice_name = f"{node.name}_slice" if node.name else f"shape_slice{uid}"

        rewritten_nodes.append(
            helper.make_node("Shape", [node.input[0]], [full_shape_out], name=shape_name)
        )

        starts = helper.make_tensor(f"starts{uid}", TensorProto.INT64, [1], [start])
        ends = helper.make_tensor(f"ends{uid}", TensorProto.INT64, [1], [end])
        axes = helper.make_tensor(f"axes{uid}", TensorProto.INT64, [1], [0])
        graph.initializer.extend([starts, ends, axes])

        rewritten_nodes.append(
            helper.make_node(
                "Slice",
                [full_shape_out, starts.name, ends.name, axes.name],
                list(node.output),
                name=slice_name,
            )
        )
        shape_rewrites += 1

    if shape_rewrites:
        del graph.node[:]
        graph.node.extend(rewritten_nodes)
    print(f"Rewrote {shape_rewrites} Shape(start/end) nodes → Shape+Slice")

    # Re-embed weights into single file
    for tensor in graph.initializer:
        tensor.ClearField("data_location")
    onnx.save(model_onnx, args.output, save_as_external_data=False)

    ops = Counter(n.op_type for n in model_onnx.graph.node)
    print(f"\nFinal: {sum(ops.values())} nodes, {len(ops)} unique ops")
    for op, count in ops.most_common():
        print(f"  {op}: {count}")

    for t in list(model_onnx.graph.input) + list(model_onnx.graph.output):
        shape = t.type.tensor_type.shape
        dims = [d.dim_param or str(d.dim_value) for d in shape.dim]
        print(f"  {t.name}: [{', '.join(dims)}]")

    size_mb = os.path.getsize(args.output) / (1024 * 1024)
    print(f"\nSaved: {args.output} ({size_mb:.1f} MB)")

    if args.verify_coreml:
        verify_coreml_session(args.output)


if __name__ == "__main__":
    main()