Questions about ModDiscrBlock

[xiaoqian-shen]

Hi, Great Work!

I have a question about the discriminator. The current discriminator is conditioned on δ_{i}^x only in DiscrEpilogue, but not the DiscriminatorBlock. Please kindly enlighten me if I misunderstand anything.

[universome]

Hi, thank you!

Yes, the current implementation uses conditioning only in DiscrEpilogue: we decided to remove it from DiscriminatorBlock, since it considerably increases the training time (~30% as far as I remember) and does not help that much (in terms of raw metrics at least). We also updated the paper accordingly a couple of weeks ago.

If you want to see our old "submission-time" version of `networks.py`, then here it is (sorry for the dirty code)

```python # Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved. # # NVIDIA CORPORATION and its licensors retain all intellectual property # and proprietary rights in and to this software, related documentation # and any modifications thereto. Any use, reproduction, disclosure or # distribution of this software and related documentation without an express # license agreement from NVIDIA CORPORATION is strictly prohibited. import numpy as np import torch import torch.nn as nn from torch import Tensor from omegaconf import OmegaConf, DictConfig import torch.nn.functional as F from src.torch_utils import misc from src.torch_utils import persistence from src.torch_utils.ops import conv2d_resample, upfirdn2d, bias_act, fma from training.motion import MotionEncoder from training.layers import ( FullyConnectedLayer, GenInput, CoordFuser, TimeFuser, TemporalDifferenceEncoder, MultiTimeEncoder, JointTimeEncoder, Conv2dLayer, MappingNetwork, remove_diag, get_max_dist, ) #---------------------------------------------------------------------------- @misc.profiled_function def modulated_conv2d( x, # Input tensor of shape [batch_size, in_channels, in_height, in_width]. weight, # Weight tensor of shape [out_channels, in_channels, kernel_height, kernel_width]. styles, # Modulation coefficients of shape [batch_size, in_channels]. noise = None, # Optional noise tensor to add to the output activations. up = 1, # Integer upsampling factor. down = 1, # Integer downsampling factor. padding = 0, # Padding with respect to the upsampled image. resample_filter = None, # Low-pass filter to apply when resampling activations. Must be prepared beforehand by calling upfirdn2d.setup_filter(). demodulate = True, # Apply weight demodulation? flip_weight = True, # False = convolution, True = correlation (matches torch.nn.functional.conv2d). fused_modconv = True, # Perform modulation, convolution, and demodulation as a single fused operation? ): batch_size = x.shape[0] out_channels, in_channels, kh, kw = weight.shape misc.assert_shape(weight, [out_channels, in_channels, kh, kw]) # [OIkk] misc.assert_shape(x, [batch_size, in_channels, None, None]) # [NIHW] misc.assert_shape(styles, [batch_size, in_channels]) # [NI] # Pre-normalize inputs to avoid FP16 overflow. if x.dtype == torch.float16 and demodulate: weight = weight * (1 / np.sqrt(in_channels * kh * kw) / weight.norm(float('inf'), dim=[1,2,3], keepdim=True)) # max_Ikk styles = styles / styles.norm(float('inf'), dim=1, keepdim=True) # max_I # Calculate per-sample weights and demodulation coefficients. w = None dcoefs = None if demodulate or fused_modconv: w = weight.unsqueeze(0) * styles.reshape(batch_size, 1, -1, 1, 1) # [NOIkk] if demodulate: dcoefs = (w.square().sum(dim=[2,3,4]) + 1e-8).rsqrt() # [NO] if demodulate and fused_modconv: w = w * dcoefs.reshape(batch_size, -1, 1, 1, 1) # [NOIkk] # Execute by scaling the activations before and after the convolution. if not fused_modconv: x = x * styles.to(x.dtype).reshape(batch_size, -1, 1, 1) x = conv2d_resample.conv2d_resample(x=x, w=weight.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, flip_weight=flip_weight) if demodulate and noise is not None: x = fma.fma(x, dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1), noise.to(x.dtype)) elif demodulate: x = x * dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1) elif noise is not None: x = x.add_(noise.to(x.dtype)) return x # Execute as one fused op using grouped convolution. with misc.suppress_tracer_warnings(): # this value will be treated as a constant batch_size = int(batch_size) misc.assert_shape(x, [batch_size, in_channels, None, None]) x = x.reshape(1, -1, *x.shape[2:]) w = w.reshape(-1, in_channels, kh, kw) x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, groups=batch_size, flip_weight=flip_weight) x = x.reshape(batch_size, -1, *x.shape[2:]) if noise is not None: x = x.add_(noise) return x #---------------------------------------------------------------------------- @misc.profiled_function def fmm_modulate_linear(x: Tensor, weight: Tensor, styles: Tensor, noise=None, activation: str="demod") -> Tensor: """ x: [batch_size, c_in, height, width] weight: [c_out, c_in, 1, 1] style: [batch_size, num_mod_params] noise: Optional[batch_size, 1, height, width] """ batch_size, c_in, h, w = x.shape c_out, c_in, kh, kw = weight.shape rank = styles.shape[1] // (c_in + c_out) assert kh == 1 and kw == 1 assert styles.shape[1] % (c_in + c_out) == 0 # Now, we need to construct a [c_out, c_in] matrix left_matrix = styles[:, :c_out * rank] # [batch_size, left_matrix_size] right_matrix = styles[:, c_out * rank:] # [batch_size, right_matrix_size] left_matrix = left_matrix.view(batch_size, c_out, rank) # [batch_size, c_out, rank] right_matrix = right_matrix.view(batch_size, rank, c_in) # [batch_size, rank, c_in] # Imagine, that the output of `self.affine` (in SynthesisLayer) is N(0, 1) # Then, std of weights is sqrt(rank). Converting it back to N(0, 1) modulation = left_matrix @ right_matrix / np.sqrt(rank) # [batch_size, c_out, c_in] if activation == "tanh": modulation = modulation.tanh() elif activation == "sigmoid": modulation = modulation.sigmoid() - 0.5 W = weight.squeeze(3).squeeze(2).unsqueeze(0) * (modulation + 1.0) # [batch_size, c_out, c_in] if activation == "demod": W = W / (W.norm(dim=2, keepdim=True) + 1e-8) # [batch_size, c_out, c_in] W = W.to(dtype=x.dtype) # out = torch.einsum('boi,bihw->bohw', W, x) x = x.view(batch_size, c_in, h * w) # [batch_size, c_in, h * w] out = torch.bmm(W, x) # [batch_size, c_out, h * w] out = out.view(batch_size, c_out, h, w) # [batch_size, c_out, h, w] if not noise is None: out = out.add_(noise) return out #---------------------------------------------------------------------------- @misc.profiled_function def maybe_upsample(x, upsampling_mode: str, up: int) -> Tensor: if up == 1: return x if upsampling_mode == 'bilinear': x = F.interpolate(x, mode='bilinear', align_corners=True, scale_factor=up) elif upsampling_mode == 'nearest': x = F.interpolate(x, mode='nearest', scale_factor=up) else: raise NotImplementedError(f"Unknown upsampling mode: {upsampling_mode}") return x #---------------------------------------------------------------------------- @persistence.persistent_class class SynthesisLayer(torch.nn.Module): def __init__(self, in_channels, # Number of input channels. out_channels, # Number of output channels. w_dim, # Intermediate latent (W) dimensionality. resolution, # Resolution of this layer. kernel_size = 3, # Convolution kernel size. up = 1, # Integer upsampling factor. activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc. resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations. conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping. channels_last = False, # Use channels_last format for the weights? instance_norm = False, # Use instance norm? cfg = {}, # Additional config ): super().__init__() self.cfg = cfg self.resolution = resolution self.use_fmm = self.resolution in self.cfg.fmm.get('resolutions', []) self.up = up self.activation = activation self.conv_clamp = conv_clamp self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter)) self.padding = kernel_size // 2 self.act_gain = bias_act.activation_funcs[activation].def_gain if self.use_fmm: self.affine = FullyConnectedLayer(w_dim, (in_channels + out_channels) * self.cfg.fmm.rank, bias_init=0) else: self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1) memory_format = torch.channels_last if channels_last else torch.contiguous_format self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format)) if self.cfg.use_noise: self.register_buffer('noise_const', torch.randn([resolution, resolution])) self.noise_strength = torch.nn.Parameter(torch.zeros([])) self.bias = torch.nn.Parameter(torch.zeros([out_channels])) self.instance_norm = instance_norm def forward(self, x, w, noise_mode='random', fused_modconv=True, gain=1): assert noise_mode in ['random', 'const', 'none'] in_resolution = self.resolution // self.up misc.assert_shape(x, [None, self.weight.shape[1], in_resolution, in_resolution]) styles = self.affine(w) noise = None if self.cfg.use_noise and noise_mode == 'random': noise = torch.randn([x.shape[0], 1, self.resolution, self.resolution], device=x.device) * self.noise_strength if self.cfg.use_noise and noise_mode == 'const': noise = self.noise_const * self.noise_strength flip_weight = (self.up == 1) # slightly faster if self.instance_norm: x = x / (x.std(dim=[2,3], keepdim=True) + 1e-8) # [batch_size, c, h, w] if self.use_fmm: x = fmm_modulate_linear(x=x, weight=self.weight, styles=styles, noise=noise, activation=self.cfg.fmm.activation) else: x = modulated_conv2d(x=x, weight=self.weight, styles=styles, noise=noise, up=self.up, padding=self.padding, resample_filter=self.resample_filter, flip_weight=flip_weight, fused_modconv=fused_modconv) act_gain = self.act_gain * gain act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None x = bias_act.bias_act(x, self.bias.to(x.dtype), act=self.activation, gain=act_gain, clamp=act_clamp) return x #---------------------------------------------------------------------------- @persistence.persistent_class class ToRGBLayer(torch.nn.Module): def __init__(self, in_channels, out_channels, w_dim, kernel_size=1, conv_clamp=None, channels_last=False): super().__init__() self.conv_clamp = conv_clamp self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1) memory_format = torch.channels_last if channels_last else torch.contiguous_format self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format)) self.bias = torch.nn.Parameter(torch.zeros([out_channels])) self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2)) def forward(self, x, w, fused_modconv=True): styles = self.affine(w) * self.weight_gain x = modulated_conv2d(x=x, weight=self.weight, styles=styles, demodulate=False, fused_modconv=fused_modconv) x = bias_act.bias_act(x, self.bias.to(x.dtype), clamp=self.conv_clamp) return x #---------------------------------------------------------------------------- @persistence.persistent_class class SynthesisBlock(torch.nn.Module): def __init__(self, in_channels, # Number of input channels, 0 = first block. out_channels, # Number of output channels. w_dim, # Intermediate latent (W) dimensionality. motion_w_dim, # Motion code size resolution, # Resolution of this block. img_channels, # Number of output color channels. is_last, # Is this the last block? architecture = 'skip', # Architecture: 'orig', 'skip', 'resnet'. resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations. conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping. use_fp16 = False, # Use FP16 for this block? fp16_channels_last = False, # Use channels-last memory format with FP16? cfg = {}, # Additional config **layer_kwargs, # Arguments for SynthesisLayer. ): assert architecture in ['orig', 'skip', 'resnet'] super().__init__() self.cfg = cfg self.in_channels = in_channels self.w_dim = w_dim if resolution <= self.cfg.input.resolution: self.resolution = self.cfg.input.resolution self.up = 1 self.input_resolution = self.cfg.input.resolution else: self.resolution = resolution self.up = 2 self.input_resolution = resolution // 2 self.img_channels = img_channels self.is_last = is_last self.architecture = architecture self.use_fp16 = use_fp16 self.channels_last = (use_fp16 and fp16_channels_last) self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter)) self.num_conv = 0 self.num_torgb = 0 self.use_fmm = self.resolution in self.cfg.fmm.get('resolutions', []) self.kernel_size = 1 if self.use_fmm else 3 self.use_instance_norm = self.use_fmm and in_channels > 0 and cfg.get('fmm', {}).get('instance_norm', False) if self.cfg.time_enc.per_resolution: assert self.architecture != 'resnet' self.time_fuser = TimeFuser(self.cfg, resolution=self.resolution, motion_w_dim=motion_w_dim) self.time_emb_dim = self.time_fuser.get_total_dim() else: self.time_fuser = None self.time_emb_dim = 0 if in_channels == 0: self.input = GenInput(self.cfg, out_channels, w_dim, motion_w_dim=motion_w_dim) conv1_in_channels = self.input.total_dim + self.time_emb_dim else: up_for_conv0 = 1 if self.use_fmm else self.up # For FMM, we'll upsample manually if self.cfg.coords.enabled and (not self.cfg.coords.per_resolution or self.resolution > self.input_resolution): assert self.architecture != 'resnet' self.coord_fuser = CoordFuser( cfg=self.cfg.coords, w_dim=self.w_dim, resolution=self.resolution // up_for_conv0, t_resolution=self.cfg.max_num_frames) conv0_in_channels = in_channels + self.coord_fuser.total_dim + self.time_emb_dim else: self.coord_fuser = None conv0_in_channels = in_channels + self.time_emb_dim # We are not using instance norm in conv0, because we concatenate coords to it (sometimes) # and some coords can be all-zero self.conv0 = SynthesisLayer(conv0_in_channels, out_channels, w_dim=w_dim, resolution=self.resolution, up=up_for_conv0, resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last, kernel_size=self.kernel_size, cfg=cfg, instance_norm=False, **layer_kwargs) self.num_conv += 1 conv1_in_channels = out_channels self.conv1 = SynthesisLayer(conv1_in_channels, out_channels, w_dim=w_dim, resolution=self.resolution, conv_clamp=conv_clamp, channels_last=self.channels_last, kernel_size=self.kernel_size, cfg=cfg, instance_norm=self.use_instance_norm, **layer_kwargs) self.num_conv += 1 if self.cfg.get('num_extra_convs', {}).get(str(self.resolution), 0) > 0: assert self.architecture != 'resnet', "Not implemented for resnet" self.extra_convs = nn.ModuleList([ SynthesisLayer(out_channels, out_channels, w_dim=w_dim, resolution=self.resolution, conv_clamp=conv_clamp, channels_last=self.channels_last, kernel_size=self.kernel_size, instance_norm=self.use_instance_norm, cfg=cfg, **layer_kwargs) for _ in range(self.cfg.num_extra_convs[str(self.resolution)])]) self.num_conv += len(self.extra_convs) else: self.extra_convs = None if is_last or architecture == 'skip': self.torgb = ToRGBLayer(out_channels, img_channels, w_dim=w_dim, conv_clamp=conv_clamp, channels_last=self.channels_last) self.num_torgb += 1 if in_channels != 0 and architecture == 'resnet': self.skip = Conv2dLayer(in_channels, out_channels, kernel_size=1, bias=False, up=self.up, resample_filter=resample_filter, channels_last=self.channels_last) def forward(self, x, img, ws, t=None, motion_w=None, force_fp32=False, fused_modconv=None, **layer_kwargs): misc.assert_shape(ws, [None, self.num_conv + self.num_torgb, self.w_dim]) w_iter = iter(ws.unbind(dim=1)) dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32 memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format if fused_modconv is None: with misc.suppress_tracer_warnings(): # this value will be treated as a constant fused_modconv = (not self.training) and (dtype == torch.float32 or (isinstance(x, Tensor) and int(x.shape[0]) == 1)) # Input. if self.in_channels == 0: conv1_w = next(w_iter) x = self.input(ws.shape[0], conv1_w, t=t, motion_w=motion_w, device=ws.device, dtype=dtype, memory_format=memory_format) else: misc.assert_shape(x, [None, self.in_channels, self.input_resolution, self.input_resolution]) x = x.to(dtype=dtype, memory_format=memory_format) x = maybe_upsample(x, self.cfg.fmm_upsampling_mode, self.up) if self.use_fmm else x # [batch_size, c, h, w] # Main layers. if self.in_channels == 0: x = x if self.time_fuser is None else self.time_fuser(x, t=t, motion_w=motion_w) # [batch_size, c + time_emb_dim, h, w] x = self.conv1(x, conv1_w, fused_modconv=fused_modconv, **layer_kwargs) elif self.architecture == 'resnet': y = self.skip(x, gain=np.sqrt(0.5)) x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs) x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, gain=np.sqrt(0.5), **layer_kwargs) x = y.add_(x) else: conv0_w = next(w_iter) if self.coord_fuser is not None: x = self.coord_fuser(x, conv0_w, t=t, dtype=dtype, memory_format=memory_format) if self.time_fuser is not None: x = self.time_fuser(x, t=t, motion_w=motion_w) # [b, c + coord_dim + time_dim, h, w] x = self.conv0(x, conv0_w, fused_modconv=fused_modconv, **layer_kwargs) x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs) if not self.extra_convs is None: for conv, w in zip(self.extra_convs, w_iter): x = conv(x, w, fused_modconv=fused_modconv, **layer_kwargs) # ToRGB. if img is not None: misc.assert_shape(img, [None, self.img_channels, self.input_resolution, self.input_resolution]) if self.up == 2: if self.use_fmm: img = maybe_upsample(img, self.cfg.fmm_upsampling_mode, 2) else: img = upfirdn2d.upsample2d(img, self.resample_filter) if self.is_last or self.architecture == 'skip': y = self.torgb(x, next(w_iter), fused_modconv=fused_modconv) y = y.to(dtype=torch.float32, memory_format=torch.contiguous_format) img = img.add_(y) if img is not None else y assert x.dtype == dtype assert img is None or img.dtype == torch.float32 return x, img #---------------------------------------------------------------------------- @persistence.persistent_class class SynthesisNetwork(torch.nn.Module): def __init__(self, w_dim, # Intermediate latent (W) dimensionality. img_resolution, # Output image resolution. img_channels, # Number of color channels. channel_base = 32768, # Overall multiplier for the number of channels. channel_max = 512, # Maximum number of channels in any layer. num_fp16_res = 0, # Use FP16 for the N highest resolutions. cfg = {}, # Additional config **block_kwargs, # Arguments for SynthesisBlock. ): assert img_resolution >= 4 and img_resolution & (img_resolution - 1) == 0 super().__init__() self.w_dim = w_dim self.cfg = cfg self.img_resolution = img_resolution self.img_resolution_log2 = int(np.log2(img_resolution)) self.img_channels = img_channels self.block_resolutions = [2 ** i for i in range(2, self.img_resolution_log2 + 1)] channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions} fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8) if self.cfg.motion.w_dim > 0: self.motion_encoder = MotionEncoder(self.cfg, resolutions=self.block_resolutions) self.motion_w_dim = self.motion_encoder.get_output_dim() else: self.motion_encoder = None self.motion_w_dim = 0 self.num_ws = 0 for res in self.block_resolutions: in_channels = channels_dict[res // 2] if res > 4 else 0 out_channels = channels_dict[res] use_fp16 = (res >= fp16_resolution) is_last = (res == self.img_resolution) block = SynthesisBlock( in_channels, out_channels, w_dim=self.w_dim + (self.motion_w_dim if self.cfg.time_enc.cond_type == 'concat_w' else 0), motion_w_dim=self.motion_w_dim, resolution=res, img_channels=img_channels, is_last=is_last, use_fp16=use_fp16, cfg=cfg, **block_kwargs) self.num_ws += block.num_conv if is_last: self.num_ws += block.num_torgb setattr(self, f'b{res}', block) def forward(self, ws, t=None, c=None, l=None, motion_noise=None, motion_w=None, **block_kwargs): assert len(ws) == len(c) == len(t), f"Wrong shape: {ws.shape}, {c.shape}, {t.shape}" assert t.ndim == 2, f"Wrong shape: {t.shape}" misc.assert_shape(ws, [None, self.num_ws, self.w_dim]) block_ws = [] if self.motion_encoder is None: ws = ws.repeat_interleave(t.shape[1], dim=0) # [batch_size * num_frames, num_ws, w_dim] motion_w = None else: if motion_w is None: motion_info = self.motion_encoder(c, t, l=l, w=ws[:, 0], motion_noise=motion_noise) # [batch_size * num_frames, motion_w_dim] motion_w = motion_info['motion_w'] # [batch_size * num_frames, motion_w_dim] if not self.cfg.time_enc.per_resolution and self.cfg.time_enc.cond_type in ['concat_w', 'sum_w']: misc.assert_shape(motion_w, [t.shape[0] * t.shape[1], self.motion_w_dim]) if self.cfg.time_enc.cond_type == 'concat_w': motion_ws = motion_w.unsqueeze(1).repeat(1, self.num_ws, 1) # [batch_size * num_frames, num_ws, motion_w_dim] ws = torch.cat([ws.repeat_interleave(t.shape[1], dim=0), motion_ws], dim=2) # [batch_size * num_frames, num_ws, w_dim + motion_w_dim] elif self.cfg.time_enc.cond_type == 'sum_w': ws = ws.repeat_interleave(t.shape[1], dim=0) + motion_w.unsqueeze(1) # [batch_size * num_frames, num_ws, w_dim + motion_w_dim] else: ws = ws.repeat_interleave(t.shape[1], dim=0) # [batch_size * num_frames, num_ws, w_dim] with torch.autograd.profiler.record_function('split_ws'): ws = ws.to(torch.float32) w_idx = 0 for res in self.block_resolutions: block = getattr(self, f'b{res}') block_ws.append(ws.narrow(1, w_idx, block.num_conv + block.num_torgb)) w_idx += block.num_conv x = img = None for res, cur_ws in zip(self.block_resolutions, block_ws): block = getattr(self, f'b{res}') if self.cfg.time_enc.per_resolution: motion_w = motion_info['motion_w'][res] # [batch_size * num_frames, motion_w_dim] if self.cfg.time_enc.cond_type == "concat_w": cur_ws = torch.cat([cur_ws, motion_w.unsqueeze(1).repeat(1, cur_ws.shape[1], 1)], dim=2) # [batch_size * num_frames, num_cur_ws, w_dim + motion_w_dim] elif self.cfg.time_enc.cond_type == "sum_w": cur_ws = cur_ws + motion_w.unsqueeze(1) # [batch_size * num_frames, num_cur_ws, w_dim] elif self.cfg.time_enc.cond_type == "concat_act": pass else: raise NotImplementedError(f"Unkown agg op: {self.cfg.motion.agg}") if self.cfg.time_enc.cond_type != 'concat_act': motion_w = None # To make sure that we do not leak. x, img = block(x, img, cur_ws, t=t, motion_w=motion_w, **block_kwargs) return img #---------------------------------------------------------------------------- @persistence.persistent_class class Generator(torch.nn.Module): def __init__(self, c_dim, # Conditioning label (C) dimensionality. w_dim, # Intermediate latent (W) dimensionality. img_resolution, # Output resolution. img_channels, # Number of output color channels. mapping_kwargs = {}, # Arguments for MappingNetwork. synthesis_kwargs = {}, # Arguments for SynthesisNetwork. cfg = {}, # Config ): super().__init__() self.cfg = cfg self.sampling_dict = OmegaConf.to_container(OmegaConf.create({**self.cfg.sampling})) self.z_dim = self.cfg.z_dim self.c_dim = c_dim self.w_dim = w_dim self.img_resolution = img_resolution self.img_channels = img_channels self.synthesis = SynthesisNetwork(w_dim=w_dim, img_resolution=img_resolution, img_channels=img_channels, cfg=cfg, **synthesis_kwargs) self.num_ws = self.synthesis.num_ws self.mapping = MappingNetwork(z_dim=self.z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs) def forward(self, z, c, t, l, truncation_psi=1, truncation_cutoff=None, **synthesis_kwargs): assert len(z) == len(c) == len(t), f"Wrong shape: {z.shape}, {c.shape}, {t.shape}" assert t.ndim == 2, f"Wrong shape: {t.shape}" ws = self.mapping(z, c, l=l, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff) # [batch_size, num_ws, w_dim] img = self.synthesis(ws, t=t, c=c, l=l, **synthesis_kwargs) # [batch_size * num_frames, c, h, w] return img #---------------------------------------------------------------------------- @persistence.persistent_class class DiscriminatorBlock(torch.nn.Module): def __init__(self, in_channels, # Number of input channels, 0 = first block. tmp_channels, # Number of intermediate channels. out_channels, # Number of output channels. resolution, # Resolution of this block. img_channels, # Number of input color channels. first_layer_idx, # Index of the first layer. architecture = 'resnet', # Architecture: 'orig', 'skip', 'resnet'. activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc. resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations. conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping. use_fp16 = False, # Use FP16 for this block? fp16_channels_last = False, # Use channels-last memory format with FP16? freeze_layers = 0, # Freeze-D: Number of layers to freeze. c_dim = 0, # Embedding size for t. cfg = {}, # Main config. ): assert architecture in ['orig', 'skip', 'resnet'] super().__init__() self.cfg = cfg self.in_channels = in_channels self.resolution = resolution self.img_channels = img_channels self.first_layer_idx = first_layer_idx self.architecture = architecture self.use_fp16 = use_fp16 self.channels_last = (use_fp16 and fp16_channels_last) self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter)) self.num_layers = 0 def trainable_gen(): while True: layer_idx = self.first_layer_idx + self.num_layers trainable = (layer_idx >= freeze_layers) self.num_layers += 1 yield trainable trainable_iter = trainable_gen() conv0_in_channels = in_channels if in_channels > 0 else tmp_channels if in_channels == 0 or architecture == 'skip': self.fromrgb = Conv2dLayer(img_channels, tmp_channels, kernel_size=1, activation=activation, trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last) if self.cfg.hyper_type in ['hyper', 'dummy_hyper']: assert next(trainable_iter) self.conv0 = SynthesisLayer( conv0_in_channels, tmp_channels, w_dim=c_dim, resolution=self.resolution, kernel_size=3, activation=activation, conv_clamp=conv_clamp, channels_last=self.channels_last, cfg=self.cfg.dummy_synth_cfg) elif self.cfg.hyper_type == 'no_hyper': self.conv0 = Conv2dLayer(conv0_in_channels, tmp_channels, kernel_size=3, activation=activation, trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last) else: raise NotImplementedError("Unknown hyper type:", self.cfg.hyper_type) if int(self.resolution) in [int(r) for r in self.cfg.contr.get('resolutions', [])] and self.cfg.num_frames_per_sample > 1: assert self.cfg.agg.type != "concat" or self.cfg.agg.concat_res < self.resolution, \ f"Cant compute similarities after concatenation: {self.cfg.agg.concat_res} > {self.resolution}" self.contr = GroupwiseContrastiveLayer( cfg=self.cfg, in_channels=conv0_in_channels, c_dim=c_dim, resolution=self.resolution, conv_clamp=conv_clamp, channels_last=self.channels_last) conv1_in_channels = self.contr.get_output_dim() else: self.contr = None conv1_in_channels = tmp_channels self.conv1 = Conv2dLayer(conv1_in_channels, out_channels, kernel_size=3, activation=activation, down=2, trainable=next(trainable_iter), resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last) if architecture == 'resnet': self.skip = Conv2dLayer(conv0_in_channels, out_channels, kernel_size=1, bias=False, down=2, trainable=next(trainable_iter), resample_filter=resample_filter, channels_last=self.channels_last) def forward(self, x, img, c, force_fp32=False): dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32 memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format # Input. if x is not None: misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution]) x = x.to(dtype=dtype, memory_format=memory_format) # FromRGB. if self.in_channels == 0 or self.architecture == 'skip': misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution]) img = img.to(dtype=dtype, memory_format=memory_format) y = self.fromrgb(img) x = x + y if x is not None else y img = upfirdn2d.downsample2d(img, self.resample_filter) if self.architecture == 'skip' else None c_weight = 1.0 if self.cfg.get('is_hyper', True) else 0.0 if self.cfg.hyper_type == 'hyper': cond_kwargs = {'w': c} elif self.cfg.hyper_type == 'dummy_hyper': cond_kwargs = {'w': c * 0.0} elif self.cfg.hyper_type == 'no_hyper': cond_kwargs = {} else: raise NotImplementedError("Unknwon hyper type", self.cfg.hyper_type) # Main layers. if self.architecture == 'resnet': y = self.skip(x, gain=np.sqrt(0.5)) x = self.conv0(x, **cond_kwargs) x = x if self.contr is None else self.contr(x, c) x = self.conv1(x, gain=np.sqrt(0.5)) x = y.add_(x) else: x = self.conv0(x, **cond_kwargs) x = x if self.contr is None else self.contr(x, c) x = self.conv1(x) assert x.dtype == dtype return x, img #---------------------------------------------------------------------------- @persistence.persistent_class class MinibatchStdLayer(torch.nn.Module): def __init__(self, group_size, num_channels=1): super().__init__() self.group_size = group_size self.num_channels = num_channels def forward(self, x): N, C, H, W = x.shape with misc.suppress_tracer_warnings(): # as_tensor results are registered as constants G = torch.min(torch.as_tensor(self.group_size), torch.as_tensor(N)) if self.group_size is not None else N F = self.num_channels c = C // F y = x.reshape(G, -1, F, c, H, W) # [GnFcHW] Split minibatch N into n groups of size G, and channels C into F groups of size c. y = y - y.mean(dim=0) # [GnFcHW] Subtract mean over group. y = y.square().mean(dim=0) # [nFcHW] Calc variance over group. y = (y + 1e-8).sqrt() # [nFcHW] Calc stddev over group. y = y.mean(dim=[2,3,4]) # [nF] Take average over channels and pixels. y = y.reshape(-1, F, 1, 1) # [nF11] Add missing dimensions. y = y.repeat(G, 1, H, W) # [NFHW] Replicate over group and pixels. x = torch.cat([x, y], dim=1) # [N(C+1)HW] Append to input as new channels. return x #---------------------------------------------------------------------------- @persistence.persistent_class class GroupwiseContrastiveLayer(torch.nn.Module): """ This layer compares images of the same video with one another and concatenates the similarity scores back to their original activations """ def __init__(self, cfg, in_channels: int, resolution: int, c_dim: int, conv_clamp: int=None, channels_last: bool=False): super().__init__() self.cfg = cfg self.in_channels = in_channels self.group_size = self.cfg.num_frames_per_sample self.dim = self.cfg.contr.dim self.scale = 1 if self.dim <= 3 else ((self.dim - 2) ** 2 / self.dim) ** 0.5 self.diff_based = self.cfg.contr.get('diff_based', False) self.transform = SynthesisLayer( in_channels=in_channels, out_channels=self.cfg.contr.dim, w_dim=c_dim, resolution=resolution, kernel_size=self.cfg.contr.kernel_size, activation='lrelu', conv_clamp=conv_clamp, channels_last=channels_last, cfg=self.cfg.dummy_synth_cfg, ) self.agg = self.cfg.contr.agg def get_output_dim(self) -> int: if self.agg in ["mean", "min", "max", "fmin_rmax"]: return self.in_channels + 1 elif self.agg == "none": if self.diff_based: return self.in_channels + (self.group_size - 1) * (self.group_size - 2) else: return self.in_channels + self.group_size - 1 else: raise NotImplementedError def forward(self, x: Tensor, c: Tensor) -> Tensor: bn, c_in, h, w = x.shape num_groups = bn // self.group_size y = self.transform(x, c) # [bn, dim, h, w] y = y.reshape(num_groups, self.group_size, self.dim, h, w) # [num_groups, group_size, dim, h, w] if self.diff_based: # We first compute differences between frames features # and then we compute similarities between those vectors # This should show D how the pixels moved d = y[:, 1:] - y[:, :-1] # [num_groups, group_size - 1, dim, h, w] # Unfortunately, we cannot normalize the diffs because R1 penalty falls with NaNs... # So, normalize the scale at least somehow... d_sims = (1.0 / np.sqrt(d.shape[2])) * d.permute(0, 3, 4, 1, 2) @ d.permute(0, 3, 4, 2, 1) # [num_groups, h, w, group_size - 1, group_size - 1] assert not torch.isnan(d_sims).any(), "There are NaNs in the diffs tensor" d_sims = remove_diag(d_sims) # [num_groups, h, w, group_size - 1, group_size - 2] y = d_sims.unsqueeze(1).repeat(1, self.group_size, 1, 1, 1, 1) # [num_groups, group_size, h, w, group_size - 1, group_size - 2] y = y.view(bn, h, w, (self.group_size - 1) * (self.group_size - 2)).permute(0, 3, 1, 2) # [bn, (group_size - 1) * (group_size - 2), h, w] else: y = F.normalize(y, dim=2) # [num_groups, group_size, dim, h, w] y = y.permute(0, 3, 4, 1, 2) @ y.permute(0, 3, 4, 2, 1) # [num_groups, h, w, group_size, group_size] y = y * self.scale # [num_groups, h, w, group_size, group_size] y = remove_diag(y) # [num_groups, h, w, group_size, group_size - 1] y = y.permute(0, 3, 4, 1, 2) # [num_groups, group_size, group_size - 1, h, w] y = y.reshape(bn, self.group_size - 1, h, w) # [bn, group_size - 1, h, w] if self.agg == "mean": y = y.mean(dim=1, keepdim=True) # [bn, 1, h, w] elif self.agg == "max": y = y.max(dim=1, keepdim=True)[0] # [bn, 1, h, w] elif self.agg == "min": y = y.min(dim=1, keepdim=True)[0] # [bn, 1, h, w] elif self.agg == "none": y = y # [bn, group_size - 1, h, w] else: raise NotImplementedError y = torch.cat([x, y], dim=1) # [bn, in_channel + d, h, w] return y #---------------------------------------------------------------------------- @persistence.persistent_class class FeatDiffLayer(torch.nn.Module): """ Computes differences between consecutive frames features """ def __init__(self, cfg, in_channels: int, dim: int, resolution: int, c_dim: int, conv_clamp: int=None, channels_last: bool=False): super().__init__() self.cfg = cfg self.in_channels = in_channels self.group_size = self.cfg.num_frames_per_sample self.dim = dim self.transform = SynthesisLayer( in_channels=in_channels, out_channels=self.dim, w_dim=c_dim, resolution=resolution, kernel_size=3, activation='lrelu', conv_clamp=conv_clamp, channels_last=channels_last, cfg=self.cfg.dummy_synth_cfg, ) def get_output_dim(self) -> int: return (self.group_size - 1) * self.dim def forward(self, x: Tensor, c: Tensor) -> Tensor: bn, c_in, h, w = x.shape num_groups = bn // self.group_size y = self.transform(x, c) # [bn, dim, h, w] y = y.reshape(num_groups, self.group_size, self.dim, h, w) # [num_groups, group_size, dim, h, w] d = y[:, 1:] - y[:, :-1] # [num_groups, group_size - 1, dim, h, w] d = d.view(num_groups, (self.group_size - 1) * self.dim, h, w) # [num_groups, (group_size - 1) * c, h, w] return d #---------------------------------------------------------------------------- @persistence.persistent_class class DiscriminatorEpilogue(torch.nn.Module): def __init__(self, in_channels, # Number of input channels. cmap_dim, # Dimensionality of mapped conditioning label, 0 = no label. resolution, # Resolution of this block. img_channels, # Number of input color channels. architecture = 'resnet', # Architecture: 'orig', 'skip', 'resnet'. mbstd_group_size = 4, # Group size for the minibatch standard deviation layer, None = entire minibatch. mbstd_num_channels = 1, # Number of features for the minibatch standard deviation layer, 0 = disable. activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc. conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping. cfg = {}, # Architecture config. ): assert architecture in ['orig', 'skip', 'resnet'] super().__init__() self.cfg = cfg self.in_channels = in_channels self.cmap_dim = cmap_dim self.resolution = resolution self.img_channels = img_channels self.architecture = architecture if architecture == 'skip': self.fromrgb = Conv2dLayer(img_channels, in_channels, kernel_size=1, activation=activation) self.mbstd = MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels) if mbstd_num_channels > 0 else None self.conv = Conv2dLayer(in_channels + mbstd_num_channels, in_channels, kernel_size=3, activation=activation, conv_clamp=conv_clamp) self.fc = FullyConnectedLayer(in_channels * (resolution ** 2), in_channels, activation=activation) self.out = FullyConnectedLayer(in_channels, 1 if cmap_dim == 0 else cmap_dim) if self.cfg.predict_dists_weight > 0.0: self.dist_predictor = nn.Sequential( FullyConnectedLayer(in_channels * (resolution ** 2), in_channels, activation=activation), torch.nn.Flatten(), FullyConnectedLayer(in_channels, get_max_dist(self.cfg.sampling), activation='linear'), ) else: self.dist_predictor = None def forward(self, x, img, cmap, force_fp32=False): misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution]) # [NCHW] _ = force_fp32 # unused dtype = torch.float32 memory_format = torch.contiguous_format # FromRGB. x = x.to(dtype=dtype, memory_format=memory_format) if self.architecture == 'skip': misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution]) img = img.to(dtype=dtype, memory_format=memory_format) x = x + self.fromrgb(img) # Main layers. if self.mbstd is not None: x = self.mbstd(x) x = self.conv(x) dist_preds = None if self.dist_predictor is None else self.dist_predictor(x.flatten(1)) # [batch_size] x = self.fc(x.flatten(1)) x = self.out(x) # [batch_size, out_dim] if not self.dist_predictor is None: # If one uncomments this, then we'll encounter a DDP consistency error for some reason x = x + dist_preds.sum() * 0.0 # Conditioning. if self.cmap_dim > 0: misc.assert_shape(cmap, [None, self.cmap_dim]) x = (x * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim)) # [batch_size, 1] assert x.dtype == dtype return x, dist_preds #---------------------------------------------------------------------------- @persistence.persistent_class class Discriminator(torch.nn.Module): def __init__(self, c_dim, # Conditioning label (C) dimensionality. img_resolution, # Input resolution. img_channels, # Number of input color channels. architecture = 'resnet', # Architecture: 'orig', 'skip', 'resnet'. channel_base = 32768, # Overall multiplier for the number of channels. channel_max = 512, # Maximum number of channels in any layer. num_fp16_res = 0, # Use FP16 for the N highest resolutions. conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping. cmap_dim = None, # Dimensionality of mapped conditioning label, None = default. block_kwargs = {}, # Arguments for DiscriminatorBlock. mapping_kwargs = {}, # Arguments for MappingNetwork. epilogue_kwargs = {}, # Arguments for DiscriminatorEpilogue. cfg = {}, # Additional config. ): super().__init__() self.cfg = cfg self.c_dim = c_dim self.img_resolution = img_resolution self.img_resolution_log2 = int(np.log2(img_resolution)) self.img_channels = img_channels self.block_resolutions = [2 ** i for i in range(self.img_resolution_log2, 2, -1)] channels_dict = {res: min(channel_base // res, channel_max) for res in self.block_resolutions + [4]} fp16_resolution = max(2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8) if cmap_dim is None: cmap_dim = channels_dict[4] if self.cfg.num_frames_per_sample > 1: if self.cfg.time_enc_type == 'diff': self.time_encoder = TemporalDifferenceEncoder(self.cfg) elif self.cfg.time_enc_type == 'multi': self.time_encoder = MultiTimeEncoder(self.cfg) elif self.cfg.time_enc_type == 'joint': self.time_encoder = JointTimeEncoder(self.cfg) else: raise NotImplementedError(f"Unknown time encoder in D: {self.cfg.time_enc_type}") assert self.time_encoder.get_total_dim() > 0 else: self.time_encoder = None if self.c_dim == 0 and self.time_encoder is None: cmap_dim = 0 common_kwargs = dict(img_channels=img_channels, architecture=architecture, conv_clamp=conv_clamp) conditioning_dim = c_dim + (0 if self.time_encoder is None else self.time_encoder.get_total_dim()) cur_layer_idx = 0 if self.cfg.agg.type == "concat" and self.cfg.agg.get('concat_diff_dim', 0) > 0: self.diff_transform = FeatDiffLayer( cfg=self.cfg, in_channels=channels_dict[self.cfg.agg.concat_res] // self.cfg.num_frames_div_factor, dim=self.cfg.agg.concat_diff_dim, resolution=self.cfg.agg.concat_res, c_dim=conditioning_dim, conv_clamp=conv_clamp) else: self.diff_transform = None for res in self.block_resolutions: in_channels = channels_dict[res] if res < img_resolution else 0 tmp_channels = channels_dict[res] out_channels = channels_dict[res // 2] if self.cfg.agg.type == "concat": # Adjust numbers of channels if res // 2 == self.cfg.agg.concat_res: out_channels = out_channels // self.cfg.num_frames_div_factor if res == self.cfg.agg.concat_res: in_channels = (in_channels // self.cfg.num_frames_div_factor) * self.cfg.num_frames_per_sample in_channels += (0 if self.diff_transform is None else self.diff_transform.get_output_dim()) use_fp16 = (res >= fp16_resolution) block = DiscriminatorBlock(in_channels, tmp_channels, out_channels, resolution=res, first_layer_idx=cur_layer_idx, use_fp16=use_fp16, cfg=self.cfg, c_dim=conditioning_dim, **block_kwargs, **common_kwargs) setattr(self, f'b{res}', block) cur_layer_idx += block.num_layers if self.c_dim > 0 or not self.time_encoder is None: self.mapping = MappingNetwork(z_dim=0, c_dim=conditioning_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None, **mapping_kwargs) self.b4 = DiscriminatorEpilogue(channels_dict[4], cmap_dim=cmap_dim, resolution=4, cfg=self.cfg, **epilogue_kwargs, **common_kwargs) def forward(self, img, c, t, **block_kwargs): # TODO: pass img in [b, c, t, h, w] shape instead of [b * t, c, h, w] assert len(img) == t.shape[0] * t.shape[1], f"Wrong shape: {img.shape}, {t.shape}" assert t.ndim == 2, f"Wrong shape: {t.shape}" if not self.time_encoder is None: t_embs = self.time_encoder(t.view(-1, self.cfg.num_frames_per_sample)) # [batch_size, t_dim] c_orig = torch.cat([c, t_embs], dim=1) # [batch_size, c_dim + t_dim] c = c_orig.repeat_interleave(t.shape[1], dim=0) # [batch_size * num_frames, c_dim + t_dim] if self.cfg.dummy_c: c = c * 0.0 c_orig = c_orig * 0.0 x = None for res in self.block_resolutions: block = getattr(self, f'b{res}') if self.cfg.agg.type == "concat" and res == self.cfg.agg.concat_res: d = None if self.diff_transform is None else self.diff_transform(x, c) # [batch_size, num_frames - 1, diff_c, h, w] x = x.view(-1, self.cfg.num_frames_per_sample, *x.shape[1:]) # [batch_size, num_frames, c, h, w] x = x.view(x.shape[0], -1, *x.shape[3:]) # [batch_size, num_frames * c, h, w] x = x if self.diff_transform is None else torch.cat([x, d], dim=1) # [batch_size, num_frames * c + (num_frames - 1) * d_dim, h, w] c = c_orig x, img = block(x, img, c, **block_kwargs) cmap = None if self.c_dim > 0 or not self.time_encoder is None: assert c.shape[1] > 0 if c.shape[1] > 0: cmap = self.mapping(None, c) x, dist_preds = self.b4(x, img, cmap) x = x.squeeze(1) # [batch_size] return {'image_logits': x, 'dist_preds': dist_preds} #---------------------------------------------------------------------------- ```

[xiaoqian-shen]

Thanks for your reply!