resnet.py

from functools import partial

import torch
import torch.nn as nn
import torch.nn.functional as F


class Upsample2D(nn.Module):
    """
    An upsampling layer with an optional convolution.

    Parameters:
        channels: channels in the inputs and outputs.
        use_conv: a bool determining if a convolution is applied.
        dims: determines if the signal is 1D, 2D, or 3D. If 3D, then upsampling occurs in the inner-two dimensions.
    """

    def __init__(self, channels, use_conv=False, use_conv_transpose=False, out_channels=None, name="conv"):
        super().__init__()
        self.channels = channels
        self.out_channels = out_channels or channels
        self.use_conv = use_conv
        self.use_conv_transpose = use_conv_transpose
        self.name = name

        conv = None
        if use_conv_transpose:
            conv = nn.ConvTranspose2d(channels, self.out_channels, 4, 2, 1)
        elif use_conv:
            conv = nn.Conv2d(self.channels, self.out_channels, 3, padding=1)

        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
        if name == "conv":
            self.conv = conv
        else:
            self.Conv2d_0 = conv

    def forward(self, hidden_states, output_size=None):
        assert hidden_states.shape[1] == self.channels

        if self.use_conv_transpose:
            return self.conv(hidden_states)

        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
        # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
        # https://github.com/pytorch/pytorch/issues/86679
        dtype = hidden_states.dtype
        if dtype == torch.bfloat16:
            hidden_states = hidden_states.to(torch.float32)

        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
        if hidden_states.shape[0] >= 64:
            hidden_states = hidden_states.contiguous()

        # if `output_size` is passed we force the interpolation output
        # size and do not make use of `scale_factor=2`
        if output_size is None:
            hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
        else:
            hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")

        # If the input is bfloat16, we cast back to bfloat16
        if dtype == torch.bfloat16:
            hidden_states = hidden_states.to(dtype)

        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
        if self.use_conv:
            if self.name == "conv":
                hidden_states = self.conv(hidden_states)
            else:
                hidden_states = self.Conv2d_0(hidden_states)

        return hidden_states


class Downsample2D(nn.Module):
    """
    A downsampling layer with an optional convolution.

    Parameters:
        channels: channels in the inputs and outputs.
        use_conv: a bool determining if a convolution is applied.
        dims: determines if the signal is 1D, 2D, or 3D. If 3D, then downsampling occurs in the inner-two dimensions.
    """

    def __init__(self, channels, use_conv=False, out_channels=None, padding=1, name="conv"):
        super().__init__()
        self.channels = channels
        self.out_channels = out_channels or channels
        self.use_conv = use_conv
        self.padding = padding
        stride = 2
        self.name = name

        if use_conv:
            conv = nn.Conv2d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
        else:
            assert self.channels == self.out_channels
            conv = nn.AvgPool2d(kernel_size=stride, stride=stride)

        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
        if name == "conv":
            self.Conv2d_0 = conv
            self.conv = conv
        elif name == "Conv2d_0":
            self.conv = conv
        else:
            self.conv = conv

    def forward(self, hidden_states):
        assert hidden_states.shape[1] == self.channels
        if self.use_conv and self.padding == 0:
            pad = (0, 1, 0, 1)
            hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)

        assert hidden_states.shape[1] == self.channels
        hidden_states = self.conv(hidden_states)

        return hidden_states


class FirUpsample2D(nn.Module):
    def __init__(self, channels=None, out_channels=None, use_conv=False, fir_kernel=(1, 3, 3, 1)):
        super().__init__()
        out_channels = out_channels if out_channels else channels
        if use_conv:
            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
        self.use_conv = use_conv
        self.fir_kernel = fir_kernel
        self.out_channels = out_channels

    def _upsample_2d(self, hidden_states, weight=None, kernel=None, factor=2, gain=1):
        """Fused `upsample_2d()` followed by `Conv2d()`.

        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
        arbitrary order.

        Args:
            hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
            weight: Weight tensor of the shape `[filterH, filterW, inChannels,
                outChannels]`. Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
            kernel: FIR filter of the shape `[firH, firW]` or `[firN]`
                (separable). The default is `[1] * factor`, which corresponds to nearest-neighbor upsampling.
            factor: Integer upsampling factor (default: 2).
            gain: Scaling factor for signal magnitude (default: 1.0).

        Returns:
            output: Tensor of the shape `[N, C, H * factor, W * factor]` or `[N, H * factor, W * factor, C]`, and same
            datatype as `hidden_states`.
        """

        assert isinstance(factor, int) and factor >= 1

        # Setup filter kernel.
        if kernel is None:
            kernel = [1] * factor

        # setup kernel
        kernel = torch.tensor(kernel, dtype=torch.float32)
        if kernel.ndim == 1:
            kernel = torch.outer(kernel, kernel)
        kernel /= torch.sum(kernel)

        kernel = kernel * (gain * (factor**2))

        if self.use_conv:
            convH = weight.shape[2]
            convW = weight.shape[3]
            inC = weight.shape[1]

            pad_value = (kernel.shape[0] - factor) - (convW - 1)

            stride = (factor, factor)
            # Determine data dimensions.
            output_shape = (
                (hidden_states.shape[2] - 1) * factor + convH,
                (hidden_states.shape[3] - 1) * factor + convW,
            )
            output_padding = (
                output_shape[0] - (hidden_states.shape[2] - 1) * stride[0] - convH,
                output_shape[1] - (hidden_states.shape[3] - 1) * stride[1] - convW,
            )
            assert output_padding[0] >= 0 and output_padding[1] >= 0
            num_groups = hidden_states.shape[1] // inC

            # Transpose weights.
            weight = torch.reshape(weight, (num_groups, -1, inC, convH, convW))
            weight = torch.flip(weight, dims=[3, 4]).permute(0, 2, 1, 3, 4)
            weight = torch.reshape(weight, (num_groups * inC, -1, convH, convW))

            inverse_conv = F.conv_transpose2d(
                hidden_states, weight, stride=stride, output_padding=output_padding, padding=0
            )

            output = upfirdn2d_native(
                inverse_conv,
                torch.tensor(kernel, device=inverse_conv.device),
                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2 + 1),
            )
        else:
            pad_value = kernel.shape[0] - factor
            output = upfirdn2d_native(
                hidden_states,
                torch.tensor(kernel, device=hidden_states.device),
                up=factor,
                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
            )

        return output

    def forward(self, hidden_states):
        if self.use_conv:
            height = self._upsample_2d(hidden_states, self.Conv2d_0.weight, kernel=self.fir_kernel)
            height = height + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
        else:
            height = self._upsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)

        return height


class FirDownsample2D(nn.Module):
    def __init__(self, channels=None, out_channels=None, use_conv=False, fir_kernel=(1, 3, 3, 1)):
        super().__init__()
        out_channels = out_channels if out_channels else channels
        if use_conv:
            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
        self.fir_kernel = fir_kernel
        self.use_conv = use_conv
        self.out_channels = out_channels

    def _downsample_2d(self, hidden_states, weight=None, kernel=None, factor=2, gain=1):
        """Fused `Conv2d()` followed by `downsample_2d()`.
        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
        arbitrary order.

        Args:
            hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
            weight:
                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
                performed by `inChannels = x.shape[0] // numGroups`.
            kernel: FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] *
            factor`, which corresponds to average pooling.
            factor: Integer downsampling factor (default: 2).
            gain: Scaling factor for signal magnitude (default: 1.0).

        Returns:
            output: Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and
            same datatype as `x`.
        """

        assert isinstance(factor, int) and factor >= 1
        if kernel is None:
            kernel = [1] * factor

        # setup kernel
        kernel = torch.tensor(kernel, dtype=torch.float32)
        if kernel.ndim == 1:
            kernel = torch.outer(kernel, kernel)
        kernel /= torch.sum(kernel)

        kernel = kernel * gain

        if self.use_conv:
            _, _, convH, convW = weight.shape
            pad_value = (kernel.shape[0] - factor) + (convW - 1)
            stride_value = [factor, factor]
            upfirdn_input = upfirdn2d_native(
                hidden_states,
                torch.tensor(kernel, device=hidden_states.device),
                pad=((pad_value + 1) // 2, pad_value // 2),
            )
            output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0)
        else:
            pad_value = kernel.shape[0] - factor
            output = upfirdn2d_native(
                hidden_states,
                torch.tensor(kernel, device=hidden_states.device),
                down=factor,
                pad=((pad_value + 1) // 2, pad_value // 2),
            )

        return output

    def forward(self, hidden_states):
        if self.use_conv:
            downsample_input = self._downsample_2d(hidden_states, weight=self.Conv2d_0.weight, kernel=self.fir_kernel)
            hidden_states = downsample_input + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
        else:
            hidden_states = self._downsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)

        return hidden_states


class ResnetBlock2D(nn.Module):
    def __init__(
        self,
        *,
        in_channels,
        out_channels=None,
        conv_shortcut=False,
        dropout=0.0,
        temb_channels=512,
        groups=32,
        groups_out=None,
        pre_norm=True,
        eps=1e-6,
        non_linearity="swish",
        time_embedding_norm="default",
        kernel=None,
        output_scale_factor=1.0,
        use_in_shortcut=None,
        up=False,
        down=False,
    ):
        super().__init__()
        self.pre_norm = pre_norm
        self.pre_norm = True
        self.in_channels = in_channels
        out_channels = in_channels if out_channels is None else out_channels
        self.out_channels = out_channels
        self.use_conv_shortcut = conv_shortcut
        self.time_embedding_norm = time_embedding_norm
        self.up = up
        self.down = down
        self.output_scale_factor = output_scale_factor

        if groups_out is None:
            groups_out = groups

        self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)

        self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)

        if temb_channels is not None:
            self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels)
        else:
            self.time_emb_proj = None

        self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
        self.dropout = torch.nn.Dropout(dropout)
        self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)

        if non_linearity == "swish":
            self.nonlinearity = lambda x: F.silu(x)
        elif non_linearity == "mish":
            self.nonlinearity = Mish()
        elif non_linearity == "silu":
            self.nonlinearity = nn.SiLU()

        self.upsample = self.downsample = None
        if self.up:
            if kernel == "fir":
                fir_kernel = (1, 3, 3, 1)
                self.upsample = lambda x: upsample_2d(x, kernel=fir_kernel)
            elif kernel == "sde_vp":
                self.upsample = partial(F.interpolate, scale_factor=2.0, mode="nearest")
            else:
                self.upsample = Upsample2D(in_channels, use_conv=False)
        elif self.down:
            if kernel == "fir":
                fir_kernel = (1, 3, 3, 1)
                self.downsample = lambda x: downsample_2d(x, kernel=fir_kernel)
            elif kernel == "sde_vp":
                self.downsample = partial(F.avg_pool2d, kernel_size=2, stride=2)
            else:
                self.downsample = Downsample2D(in_channels, use_conv=False, padding=1, name="op")

        self.use_in_shortcut = self.in_channels != self.out_channels if use_in_shortcut is None else use_in_shortcut

        self.conv_shortcut = None
        if self.use_in_shortcut:
            self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)

    def forward(self, input_tensor, temb):
        hidden_states = input_tensor

        hidden_states = self.norm1(hidden_states)
        hidden_states = self.nonlinearity(hidden_states)

        if self.upsample is not None:
            # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
            if hidden_states.shape[0] >= 64:
                input_tensor = input_tensor.contiguous()
                hidden_states = hidden_states.contiguous()
            input_tensor = self.upsample(input_tensor)
            hidden_states = self.upsample(hidden_states)
        elif self.downsample is not None:
            input_tensor = self.downsample(input_tensor)
            hidden_states = self.downsample(hidden_states)

        hidden_states = self.conv1(hidden_states)

        if temb is not None:
            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
            hidden_states = hidden_states + temb

        hidden_states = self.norm2(hidden_states)
        hidden_states = self.nonlinearity(hidden_states)

        hidden_states = self.dropout(hidden_states)
        hidden_states = self.conv2(hidden_states)

        if self.conv_shortcut is not None:
            input_tensor = self.conv_shortcut(input_tensor)

        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor

        return output_tensor


class Mish(torch.nn.Module):
    def forward(self, hidden_states):
        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))


def upsample_2d(hidden_states, kernel=None, factor=2, gain=1):
    r"""Upsample2D a batch of 2D images with the given filter.
    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and upsamples each image with the given
    filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified
    `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is
    a: multiple of the upsampling factor.

    Args:
        hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
        kernel: FIR filter of the shape `[firH, firW]` or `[firN]`
          (separable). The default is `[1] * factor`, which corresponds to nearest-neighbor upsampling.
        factor: Integer upsampling factor (default: 2).
        gain: Scaling factor for signal magnitude (default: 1.0).

    Returns:
        output: Tensor of the shape `[N, C, H * factor, W * factor]`
    """
    assert isinstance(factor, int) and factor >= 1
    if kernel is None:
        kernel = [1] * factor

    kernel = torch.tensor(kernel, dtype=torch.float32)
    if kernel.ndim == 1:
        kernel = torch.outer(kernel, kernel)
    kernel /= torch.sum(kernel)

    kernel = kernel * (gain * (factor**2))
    pad_value = kernel.shape[0] - factor
    output = upfirdn2d_native(
        hidden_states,
        kernel.to(device=hidden_states.device),
        up=factor,
        pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
    )
    return output


def downsample_2d(hidden_states, kernel=None, factor=2, gain=1):
    r"""Downsample2D a batch of 2D images with the given filter.
    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
    given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
    specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
    shape is a multiple of the downsampling factor.

    Args:
        hidden_states: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
        kernel: FIR filter of the shape `[firH, firW]` or `[firN]`
          (separable). The default is `[1] * factor`, which corresponds to average pooling.
        factor: Integer downsampling factor (default: 2).
        gain: Scaling factor for signal magnitude (default: 1.0).

    Returns:
        output: Tensor of the shape `[N, C, H // factor, W // factor]`
    """

    assert isinstance(factor, int) and factor >= 1
    if kernel is None:
        kernel = [1] * factor

    kernel = torch.tensor(kernel, dtype=torch.float32)
    if kernel.ndim == 1:
        kernel = torch.outer(kernel, kernel)
    kernel /= torch.sum(kernel)

    kernel = kernel * gain
    pad_value = kernel.shape[0] - factor
    output = upfirdn2d_native(
        hidden_states, kernel.to(device=hidden_states.device), down=factor, pad=((pad_value + 1) // 2, pad_value // 2)
    )
    return output


def upfirdn2d_native(tensor, kernel, up=1, down=1, pad=(0, 0)):
    up_x = up_y = up
    down_x = down_y = down
    pad_x0 = pad_y0 = pad[0]
    pad_x1 = pad_y1 = pad[1]

    _, channel, in_h, in_w = tensor.shape
    tensor = tensor.reshape(-1, in_h, in_w, 1)

    _, in_h, in_w, minor = tensor.shape
    kernel_h, kernel_w = kernel.shape

    out = tensor.view(-1, in_h, 1, in_w, 1, minor)
    out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
    out = out.view(-1, in_h * up_y, in_w * up_x, minor)

    out = F.pad(out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
    out = out.to(tensor.device)  # Move back to mps if necessary
    out = out[
        :,
        max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
        max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
        :,
    ]

    out = out.permute(0, 3, 1, 2)
    out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
    out = F.conv2d(out, w)
    out = out.reshape(
        -1,
        minor,
        in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
        in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
    )
    out = out.permute(0, 2, 3, 1)
    out = out[:, ::down_y, ::down_x, :]

    out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
    out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1

    return out.view(-1, channel, out_h, out_w)