model.py

import numpy as np
import paddle
import paddle.nn as nn
import paddle.nn.functional as F

from resnet_backbone import resnet34, resnet101
from hook import hook_outputs, model_sizes, dummy_eval
# from weight_norm import weight_norm
from spectral_norm import Spectralnorm
from conv import Conv1D
from paddle import fluid


class SequentialEx(nn.Layer):
    "Like `nn.Sequential`, but with ModuleList semantics, and can access module input"
    def __init__(self, *layers):
        super().__init__()
        self.layers = nn.LayerList(layers)

    def forward(self, x):
        res = x
        for l in self.layers:
            if isinstance(l, MergeLayer):
                l.orig = x
            nres = l(res)
            # We have to remove res.orig to avoid hanging refs and therefore memory leaks
            # l.orig = None
            res = nres
        return res

    def __getitem__(self,i): return self.layers[i]
    def append(self,l): return self.layers.append(l)
    def extend(self,l): return self.layers.extend(l)
    def insert(self,i,l): return self.layers.insert(i,l)


class Deoldify(SequentialEx):
    def __init__(self, encoder, n_classes, blur=False, blur_final=True, self_attention=False, y_range=None, last_cross=True, bottle=False, norm_type='Batch', nf_factor=1, **kwargs):

        imsize = (256, 256)
        sfs_szs = model_sizes(encoder, size=imsize)
        sfs_idxs = list(reversed(_get_sfs_idxs(sfs_szs)))
        self.sfs = hook_outputs([encoder[i] for i in sfs_idxs], detach=False)
        x = dummy_eval(encoder, imsize).detach()

        nf = 512 * nf_factor
        extra_bn = norm_type == 'Spectral'
        ni = sfs_szs[-1][1]
        middle_conv = nn.Sequential(
            custom_conv_layer(
                ni, ni * 2, norm_type=norm_type, extra_bn=extra_bn
            ),
            custom_conv_layer(
                ni * 2, ni, norm_type=norm_type, extra_bn=extra_bn
            ),
        )

        layers = [encoder, nn.BatchNorm(ni), nn.ReLU(), middle_conv]

        for i, idx in enumerate(sfs_idxs):
            not_final = i != len(sfs_idxs) - 1
            up_in_c, x_in_c = int(x.shape[1]), int(sfs_szs[idx][1])
            do_blur = blur and (not_final or blur_final)
            sa = self_attention and (i == len(sfs_idxs) - 3)

            n_out = nf if not_final else nf // 2

            unet_block = UnetBlockWide(
                up_in_c,
                x_in_c,
                n_out,
                self.sfs[i],
                final_div=not_final,
                blur=blur,
                self_attention=sa,
                norm_type=norm_type,
                extra_bn=extra_bn,
                **kwargs
            )
            unet_block.eval()
            layers.append(unet_block)
            x = unet_block(x)

        ni = x.shape[1]
        if imsize != sfs_szs[0][-2:]:
            layers.append(PixelShuffle_ICNR(ni, **kwargs))
        if last_cross:
            layers.append(MergeLayer(dense=True))
            ni += 3
            layers.append(res_block(ni, bottle=bottle, norm_type=norm_type, **kwargs))
        layers += [
            custom_conv_layer(ni, n_classes, ks=1, use_activ=False, norm_type=norm_type)
        ]
        if y_range is not None:
            layers.append(SigmoidRange(*y_range))
        super().__init__(*layers)



def custom_conv_layer(
    ni: int,
    nf: int,
    ks: int = 3,
    stride: int = 1,
    padding: int = None,
    bias: bool = None,
    is_1d: bool = False,
    norm_type='Batch',
    use_activ: bool = True,
    leaky: float = None,
    transpose: bool = False,
    self_attention: bool = False,
    extra_bn: bool = False,
    **kwargs
):
    "Create a sequence of convolutional (`ni` to `nf`), ReLU (if `use_activ`) and batchnorm (if `bn`) layers."
    if padding is None:
        padding = (ks - 1) // 2 if not transpose else 0
    bn = norm_type in ('Batch', 'Batchzero') or extra_bn == True
    if bias is None:
        bias = not bn
    conv_func = nn.ConvTranspose2d if transpose else nn.Conv1d if is_1d else nn.Conv2d

    conv = conv_func(ni, nf, kernel_size=ks, bias_attr=bias, stride=stride, padding=padding)
    if norm_type == 'Weight':
        print('use weight norm')
        conv = nn.utils.weight_norm(conv)
    elif norm_type == 'Spectral':
        # pass
        conv = Spectralnorm(conv)
    layers = [conv]
    if use_activ:
        layers.append(relu(True, leaky=leaky))
    if bn:
        layers.append((nn.BatchNorm if is_1d else nn.BatchNorm)(nf))
    if self_attention:
        layers.append(SelfAttention(nf))
    
    return nn.Sequential(*layers)


def relu(inplace:bool=False, leaky:float=None):
    "Return a relu activation, maybe `leaky` and `inplace`."
    return nn.LeakyReLU(leaky) if leaky is not None else nn.ReLU()


class UnetBlockWide(nn.Layer):
    "A quasi-UNet block, using `PixelShuffle_ICNR upsampling`."

    def __init__(
        self,
        up_in_c: int,
        x_in_c: int,
        n_out: int,
        hook,
        final_div: bool = True,
        blur: bool = False,
        leaky: float = None,
        self_attention: bool = False,
        **kwargs
    ):
        super().__init__()
        self.hook = hook
        up_out = x_out = n_out // 2
        self.shuf = CustomPixelShuffle_ICNR(
            up_in_c, up_out, blur=blur, leaky=leaky, **kwargs
        )
        self.bn = nn.BatchNorm(x_in_c)
        ni = up_out + x_in_c
        self.conv = custom_conv_layer(
            ni, x_out, leaky=leaky, self_attention=self_attention, **kwargs
        )
        self.relu = relu(leaky=leaky)

    def forward(self, up_in):
        s = self.hook.stored
        up_out = self.shuf(up_in)
        ssh = s.shape[-2:]
        if ssh != up_out.shape[-2:]:
            up_out = F.interpolate(up_out, s.shape[-2:], mode='nearest')
        cat_x = self.relu(paddle.concat([up_out, self.bn(s)], axis=1))
        return self.conv(cat_x)


class UnetBlockDeep(paddle.fluid.Layer):
    "A quasi-UNet block, using `PixelShuffle_ICNR upsampling`."

    def __init__(
        self,
        up_in_c: int,
        x_in_c: int,
        # hook: Hook,
        final_div: bool = True,
        blur: bool = False,
        leaky: float = None,
        self_attention: bool = False,
        nf_factor: float = 1.0,
        **kwargs
    ):
        super().__init__()
        
        self.shuf = CustomPixelShuffle_ICNR(
            up_in_c, up_in_c // 2, blur=blur, leaky=leaky, **kwargs
        )
        self.bn = nn.BatchNorm(x_in_c)
        ni = up_in_c // 2 + x_in_c
        nf = int((ni if final_div else ni // 2) * nf_factor)
        self.conv1 = custom_conv_layer(ni, nf, leaky=leaky, **kwargs)
        self.conv2 = custom_conv_layer(
            nf, nf, leaky=leaky, self_attention=self_attention, **kwargs
        )
        self.relu = relu(leaky=leaky)

    def forward(self, up_in):
        s = self.hook.stored
        up_out = self.shuf(up_in)
        ssh = s.shape[-2:]
        if ssh != up_out.shape[-2:]:
            up_out = F.interpolate(up_out, s.shape[-2:], mode='nearest')
        cat_x = self.relu(paddle.concat([up_out, self.bn(s)], axis=1))
        return self.conv2(self.conv1(cat_x))


def ifnone(a, b):
    "`a` if `a` is not None, otherwise `b`."
    return b if a is None else a


class PixelShuffle_ICNR(nn.Layer):
    "Upsample by `scale` from `ni` filters to `nf` (default `ni`), using `nn.PixelShuffle`, `icnr` init, and `weight_norm`."
    def __init__(self, ni:int, nf:int=None, scale:int=2, blur:bool=False, norm_type='Weight', leaky:float=None):
        super().__init__()
        nf = ifnone(nf, ni)
        self.conv = conv_layer(ni, nf*(scale**2), ks=1, norm_type=norm_type, use_activ=False)
        
        self.shuf = PixelShuffle(scale)
        
        self.pad = ReplicationPad2d((1,0,1,0))
        self.blur = nn.Pool2D(2, pool_stride=1, pool_type='avg')
        self.relu = relu(True, leaky=leaky)

    def forward(self,x):
        x = self.shuf(self.relu(self.conv(x)))
        return self.blur(self.pad(x)) if self.blur else x

def conv_layer(ni:int, nf:int, ks:int=3, stride:int=1, padding:int=None, bias:bool=None, is_1d:bool=False,
               norm_type='Batch',  use_activ:bool=True, leaky:float=None,
               transpose:bool=False, init=None, self_attention:bool=False):
    "Create a sequence of convolutional (`ni` to `nf`), ReLU (if `use_activ`) and batchnorm (if `bn`) layers."
    if padding is None: padding = (ks-1)//2 if not transpose else 0
    bn = norm_type in ('Batch', 'BatchZero')
    if bias is None: bias = not bn
    conv_func = nn.ConvTranspose2d if transpose else nn.Conv1d if is_1d else nn.Conv2d
    
    conv = conv_func(ni, nf, kernel_size=ks, bias_attr=bias, stride=stride, padding=padding)
    if norm_type=='Weight':
        conv = nn.utils.weight_norm(conv)
    elif norm_type=='Spectral': 
        conv = Spectralnorm(conv)

    layers = [conv]
    if use_activ: layers.append(relu(True, leaky=leaky))
    if bn: layers.append((nn.BatchNorm if is_1d else nn.BatchNorm)(nf))
    if self_attention: layers.append(SelfAttention(nf))
    return nn.Sequential(*layers)


class CustomPixelShuffle_ICNR(paddle.fluid.Layer):
    "Upsample by `scale` from `ni` filters to `nf` (default `ni`), using `nn.PixelShuffle`, `icnr` init, and `weight_norm`."

    def __init__(
        self,
        ni: int,
        nf: int = None,
        scale: int = 2,
        blur: bool = False,
        leaky: float = None,
        **kwargs
    ):
        super().__init__()
        nf = ifnone(nf, ni)
        self.conv = custom_conv_layer(
            ni, nf * (scale ** 2), ks=1, use_activ=False, **kwargs
        )
        
        self.shuf = PixelShuffle(scale)
        
        self.pad = ReplicationPad2d((1, 0, 1, 0))
        self.blur = nn.Pool2D(2, pool_stride=1, pool_type='avg')
        self.relu = nn.LeakyReLU(leaky) if leaky is not None else nn.ReLU()#relu(True, leaky=leaky)

    def forward(self, x):
        x = self.shuf(self.relu(self.conv(x)))
        return self.blur(self.pad(x)) if self.blur else x


class MergeLayer(paddle.fluid.Layer):
    "Merge a shortcut with the result of the module by adding them or concatenating thme if `dense=True`."
    def __init__(self, dense:bool=False):
        super().__init__()
        self.dense=dense
        self.orig = None

    def forward(self, x): 
        out = paddle.concat([x,self.orig], axis=1) if self.dense else (x+self.orig)
        self.orig = None
        return out


def res_block(nf, dense:bool=False, norm_type='Batch', bottle:bool=False, **conv_kwargs):
    "Resnet block of `nf` features. `conv_kwargs` are passed to `conv_layer`."
    norm2 = norm_type
    if not dense and (norm_type=='Batch'): norm2 = 'BatchZero'
    nf_inner = nf//2 if bottle else nf
    return SequentialEx(conv_layer(nf, nf_inner, norm_type=norm_type, **conv_kwargs),
                      conv_layer(nf_inner, nf, norm_type=norm2, **conv_kwargs),
                      MergeLayer(dense))


class SigmoidRange(paddle.fluid.Layer):
    "Sigmoid module with range `(low,x_max)`"
    def __init__(self, low, high):
        super().__init__()
        self.low,self.high = low,high

    def forward(self, x): return sigmoid_range(x, self.low, self.high)


def sigmoid_range(x, low, high):
    "Sigmoid function with range `(low, high)`"
    return F.sigmoid(x) * (high - low) + low



class PixelShuffle(paddle.fluid.Layer):
    def __init__(self, upscale_factor):
        super(PixelShuffle, self).__init__()
        self.upscale_factor = upscale_factor

    def forward(self, x):
        return paddle.fluid.layers.pixel_shuffle(x, self.upscale_factor)


class ReplicationPad2d(nn.Layer):
    def __init__(self, size):
        super(ReplicationPad2d, self).__init__()
        self.size = size

    def forward(self, x):
        return paddle.fluid.layers.pad2d(x, self.size, mode="edge")

def conv1d(ni:int, no:int, ks:int=1, stride:int=1, padding:int=0, bias:bool=False):
    "Create and initialize a `nn.Conv1d` layer with spectral normalization."
    conv = nn.Conv1d(ni, no, ks, stride=stride, padding=padding, bias_attr=bias)
    return Spectralnorm(conv)


class SelfAttention(nn.Layer):
    "Self attention layer for nd."
    def __init__(self, n_channels):
        super().__init__()
        self.query = conv1d(n_channels, n_channels//8)
        self.key   = conv1d(n_channels, n_channels//8)
        self.value = conv1d(n_channels, n_channels)
        self.gamma = self.create_parameter(shape=[1],
                            default_initializer=paddle.fluid.initializer.Constant(0.0))#nn.Parameter(tensor([0.]))

    def forward(self, x):
        #Notation from https://arxiv.org/pdf/1805.08318.pdf
        size = x.shape
        x = paddle.reshape(x,  list(size[:2]) + [-1])
        f,g,h = self.query(x),self.key(x),self.value(x)
        
        beta = paddle.nn.functional.softmax(paddle.bmm(paddle.transpose(f, [0, 2, 1]), g), axis=1)
        o = self.gamma * paddle.bmm(h, beta) + x
        return paddle.reshape(o, size)

def _get_sfs_idxs(sizes):
    "Get the indexes of the layers where the size of the activation changes."
    feature_szs = [size[-1] for size in sizes]
    sfs_idxs = list(
        np.where(np.array(feature_szs[:-1]) != np.array(feature_szs[1:]))[0]
    )
    if feature_szs[0] != feature_szs[1]:
        sfs_idxs = [0] + sfs_idxs
    return sfs_idxs


def build_model():
    backbone = resnet101()
    cut = -2
    encoder = nn.Sequential(*list(backbone.children())[:cut])

    model = Deoldify(encoder, 3, blur=True, y_range=(-3, 3), norm_type='Spectral', self_attention=True, nf_factor=2)
    return model