| import torch |
| import torch.nn as nn |
| import torch.nn.functional as F |
|
|
| from functools import reduce |
|
|
| class BaseNetwork(nn.Module): |
| def __init__(self): |
| super(BaseNetwork, self).__init__() |
|
|
| def print_network(self): |
| if isinstance(self, list): |
| self = self[0] |
| num_params = 0 |
| for param in self.parameters(): |
| num_params += param.numel() |
| print( |
| 'Network [%s] was created. Total number of parameters: %.1f million. ' |
| 'To see the architecture, do print(network).' % |
| (type(self).__name__, num_params / 1000000)) |
|
|
| def init_weights(self, init_type='normal', gain=0.02): |
| ''' |
| initialize network's weights |
| init_type: normal | xavier | kaiming | orthogonal |
| https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/9451e70673400885567d08a9e97ade2524c700d0/models/networks.py#L39 |
| ''' |
| def init_func(m): |
| classname = m.__class__.__name__ |
| if classname.find('InstanceNorm2d') != -1: |
| if hasattr(m, 'weight') and m.weight is not None: |
| nn.init.constant_(m.weight.data, 1.0) |
| if hasattr(m, 'bias') and m.bias is not None: |
| nn.init.constant_(m.bias.data, 0.0) |
| elif hasattr(m, 'weight') and (classname.find('Conv') != -1 |
| or classname.find('Linear') != -1): |
| if init_type == 'normal': |
| nn.init.normal_(m.weight.data, 0.0, gain) |
| elif init_type == 'xavier': |
| nn.init.xavier_normal_(m.weight.data, gain=gain) |
| elif init_type == 'xavier_uniform': |
| nn.init.xavier_uniform_(m.weight.data, gain=1.0) |
| elif init_type == 'kaiming': |
| nn.init.kaiming_normal_(m.weight.data, a=0, mode='fan_in') |
| elif init_type == 'orthogonal': |
| nn.init.orthogonal_(m.weight.data, gain=gain) |
| elif init_type == 'none': |
| m.reset_parameters() |
| else: |
| raise NotImplementedError( |
| 'initialization method [%s] is not implemented' % |
| init_type) |
| if hasattr(m, 'bias') and m.bias is not None: |
| nn.init.constant_(m.bias.data, 0.0) |
|
|
| self.apply(init_func) |
|
|
| |
| for m in self.children(): |
| if hasattr(m, 'init_weights'): |
| m.init_weights(init_type, gain) |
|
|
|
|
| class Vec2Feat(nn.Module): |
| def __init__(self, channel, hidden, kernel_size, stride, padding): |
| super(Vec2Feat, self).__init__() |
| self.relu = nn.LeakyReLU(0.2, inplace=True) |
| c_out = reduce((lambda x, y: x * y), kernel_size) * channel |
| self.embedding = nn.Linear(hidden, c_out) |
| self.kernel_size = kernel_size |
| self.stride = stride |
| self.padding = padding |
| self.bias_conv = nn.Conv2d(channel, |
| channel, |
| kernel_size=3, |
| stride=1, |
| padding=1) |
|
|
| def forward(self, x, t, output_size): |
| b_, _, _, _, c_ = x.shape |
| x = x.view(b_, -1, c_) |
| feat = self.embedding(x) |
| b, _, c = feat.size() |
| feat = feat.view(b * t, -1, c).permute(0, 2, 1) |
| feat = F.fold(feat, |
| output_size=output_size, |
| kernel_size=self.kernel_size, |
| stride=self.stride, |
| padding=self.padding) |
| feat = self.bias_conv(feat) |
| return feat |
|
|
|
|
| class FusionFeedForward(nn.Module): |
| def __init__(self, dim, hidden_dim=1960, t2t_params=None): |
| super(FusionFeedForward, self).__init__() |
| |
| self.fc1 = nn.Sequential(nn.Linear(dim, hidden_dim)) |
| self.fc2 = nn.Sequential(nn.GELU(), nn.Linear(hidden_dim, dim)) |
| assert t2t_params is not None |
| self.t2t_params = t2t_params |
| self.kernel_shape = reduce((lambda x, y: x * y), t2t_params['kernel_size']) |
|
|
| def forward(self, x, output_size): |
| n_vecs = 1 |
| for i, d in enumerate(self.t2t_params['kernel_size']): |
| n_vecs *= int((output_size[i] + 2 * self.t2t_params['padding'][i] - |
| (d - 1) - 1) / self.t2t_params['stride'][i] + 1) |
|
|
| x = self.fc1(x) |
| b, n, c = x.size() |
| normalizer = x.new_ones(b, n, self.kernel_shape).view(-1, n_vecs, self.kernel_shape).permute(0, 2, 1) |
| normalizer = F.fold(normalizer, |
| output_size=output_size, |
| kernel_size=self.t2t_params['kernel_size'], |
| padding=self.t2t_params['padding'], |
| stride=self.t2t_params['stride']) |
|
|
| x = F.fold(x.view(-1, n_vecs, c).permute(0, 2, 1), |
| output_size=output_size, |
| kernel_size=self.t2t_params['kernel_size'], |
| padding=self.t2t_params['padding'], |
| stride=self.t2t_params['stride']) |
|
|
| x = F.unfold(x / normalizer, |
| kernel_size=self.t2t_params['kernel_size'], |
| padding=self.t2t_params['padding'], |
| stride=self.t2t_params['stride']).permute( |
| 0, 2, 1).contiguous().view(b, n, c) |
| x = self.fc2(x) |
| return x |
|
|
