FoundationVision/LlamaGen

1# Modified from:2#   taming-transformers:  https://github.com/CompVis/taming-transformers3import functools4import torch5import torch.nn as nn6 7 8class NLayerDiscriminator(nn.Module):9    """Defines a PatchGAN discriminator as in Pix2Pix10        --> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py11    """12    def __init__(self, input_nc=3, ndf=64, n_layers=3, use_actnorm=False):13        """Construct a PatchGAN discriminator14        Parameters:15            input_nc (int)  -- the number of channels in input images16            ndf (int)       -- the number of filters in the last conv layer17            n_layers (int)  -- the number of conv layers in the discriminator18            norm_layer      -- normalization layer19        """20        super(NLayerDiscriminator, self).__init__()21        if not use_actnorm:22            norm_layer = nn.BatchNorm2d23        else:24            norm_layer = ActNorm25        if type(norm_layer) == functools.partial:  # no need to use bias as BatchNorm2d has affine parameters26            use_bias = norm_layer.func != nn.BatchNorm2d27        else:28            use_bias = norm_layer != nn.BatchNorm2d29 30        kw = 431        padw = 132        sequence = [nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), nn.LeakyReLU(0.2, True)]33        nf_mult = 134        nf_mult_prev = 135        for n in range(1, n_layers):  # gradually increase the number of filters36            nf_mult_prev = nf_mult37            nf_mult = min(2 ** n, 8)38            sequence += [39                nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=2, padding=padw, bias=use_bias),40                norm_layer(ndf * nf_mult),41                nn.LeakyReLU(0.2, True)42            ]43 44        nf_mult_prev = nf_mult45        nf_mult = min(2 ** n_layers, 8)46        sequence += [47            nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias),48            norm_layer(ndf * nf_mult),49            nn.LeakyReLU(0.2, True)50        ]51 52        sequence += [53            nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)]  # output 1 channel prediction map54        self.main = nn.Sequential(*sequence)55 56        self.apply(self._init_weights)57    58    def _init_weights(self, module):    59        if isinstance(module, nn.Conv2d):60            nn.init.normal_(module.weight.data, 0.0, 0.02)61        elif isinstance(module, nn.BatchNorm2d):62            nn.init.normal_(module.weight.data, 1.0, 0.02)63            nn.init.constant_(module.bias.data, 0)64 65    def forward(self, input):66        """Standard forward."""67        return self.main(input)68 69 70class ActNorm(nn.Module):71    def __init__(self, num_features, logdet=False, affine=True,72                 allow_reverse_init=False):73        assert affine74        super().__init__()75        self.logdet = logdet76        self.loc = nn.Parameter(torch.zeros(1, num_features, 1, 1))77        self.scale = nn.Parameter(torch.ones(1, num_features, 1, 1))78        self.allow_reverse_init = allow_reverse_init79 80        self.register_buffer('initialized', torch.tensor(0, dtype=torch.uint8))81 82    def initialize(self, input):83        with torch.no_grad():84            flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1)85            mean = (86                flatten.mean(1)87                .unsqueeze(1)88                .unsqueeze(2)89                .unsqueeze(3)90                .permute(1, 0, 2, 3)91            )92            std = (93                flatten.std(1)94                .unsqueeze(1)95                .unsqueeze(2)96                .unsqueeze(3)97                .permute(1, 0, 2, 3)98            )99 100            self.loc.data.copy_(-mean)101            self.scale.data.copy_(1 / (std + 1e-6))102 103    def forward(self, input, reverse=False):104        if reverse:105            return self.reverse(input)106        if len(input.shape) == 2:107            input = input[:,:,None,None]108            squeeze = True109        else:110            squeeze = False111 112        _, _, height, width = input.shape113 114        if self.training and self.initialized.item() == 0:115            self.initialize(input)116            self.initialized.fill_(1)117 118        h = self.scale * (input + self.loc)119 120        if squeeze:121            h = h.squeeze(-1).squeeze(-1)122 123        if self.logdet:124            log_abs = torch.log(torch.abs(self.scale))125            logdet = height*width*torch.sum(log_abs)126            logdet = logdet * torch.ones(input.shape[0]).to(input)127            return h, logdet128 129        return h130 131    def reverse(self, output):132        if self.training and self.initialized.item() == 0:133            if not self.allow_reverse_init:134                raise RuntimeError(135                    "Initializing ActNorm in reverse direction is "136                    "disabled by default. Use allow_reverse_init=True to enable."137                )138            else:139                self.initialize(output)140                self.initialized.fill_(1)141 142        if len(output.shape) == 2:143            output = output[:,:,None,None]144            squeeze = True145        else:146            squeeze = False147 148        h = output / self.scale - self.loc149 150        if squeeze:151            h = h.squeeze(-1).squeeze(-1)152        return h