Comments (6)
Thank you for your great work!
This issue can be closed.
from stylegan-v.
Hi! We encode the positional information in 2 ways: with Fourier Features and with learnable embeddings of size 256.
We concatenate the positional encoding vectors from those two modules together.
You might like to wait for the complete source code release (I only need to refactor the CLIP editing scripts), but if you like, here is the positional embedding in D part:
from typing import Tuple
import torch
import torch.nn as nn
import numpy as np
from omegaconf import DictConfig
class TemporalDifferenceEncoder(nn.Module):
def __init__(self, cfg: DictConfig):
super().__init__()
self.cfg = cfg
if self.cfg.sampling.num_frames_per_video > 1:
self.d = 256
self.const_embed = nn.Embedding(self.cfg.sampling.max_num_frames, self.d)
self.time_encoder = ScalarEncoder1d(self.cfg.sampling.max_num_frames)
def get_dim(self) -> int:
if self.cfg.sampling.num_frames_per_video == 1:
return 1
elif self.cfg.sampling.type == 'uniform':
return self.d + self.time_encoder.get_dim()
else:
return (self.d + self.time_encoder.get_dim()) * (self.cfg.sampling.num_frames_per_video - 1)
def forward(self, t: torch.Tensor) -> torch.Tensor:
batch_size = t.shape[0]
if self.cfg.sampling.num_frames_per_video == 1:
out = torch.zeros(len(t), 1, device=t.device)
else:
if self.cfg.sampling.type == 'uniform':
num_diffs_to_use = 1
t_diffs = t[:, 1] - t[:, 0] # [batch_size]
else:
num_diffs_to_use = self.cfg.sampling.num_frames_per_video - 1
t_diffs = (t[:, 1:] - t[:, :-1]).view(-1) # [batch_size * (num_frames - 1)]
# Note: float => round => long is necessary when it's originally long
const_embs = self.const_embed(t_diffs.float().round().long()) # [batch_size * num_diffs_to_use, d]
fourier_embs = self.time_encoder(t_diffs.unsqueeze(1)) # [batch_size * num_diffs_to_use, num_fourier_feats]
out = torch.cat([const_embs, fourier_embs], dim=1) # [batch_size * num_diffs_to_use, d + num_fourier_feats]
out = out.view(batch_size, num_diffs_to_use, -1).view(batch_size, -1) # [batch_size, num_diffs_to_use * (d + num_fourier_feats)]
return out
#----------------------------------------------------------------------------
class ScalarEncoder1d(nn.Module):
def __init__(self, max_num_frames: int):
super().__init__()
assert max_num_frames >= 1, f"Wrong max_num_frames: {max_num_frames}"
fourier_coefs = construct_log_spaced_freqs(max_num_frames)
self.register_buffer('fourier_coefs', fourier_coefs) # [1, num_fourier_feats]
def get_dim(self) -> int:
return self.fourier_coefs.shape[1] * 2
def forward(self, t: torch.Tensor) -> torch.Tensor:
assert t.ndim == 2, f"Wrong shape: {t.shape}"
t = t.view(-1).float() # [batch_size * num_frames]
fourier_raw_embs = self.fourier_coefs * t.unsqueeze(1) # [bf, num_fourier_feats]
fourier_embs = torch.cat([
fourier_raw_embs.sin(),
fourier_raw_embs.cos(),
], dim=1) # [bf, num_fourier_feats * 2]
return fourier_embs
#----------------------------------------------------------------------------
def construct_log_spaced_freqs(max_num_frames: int) -> Tuple[int, torch.Tensor]:
time_resolution = 2 ** np.ceil(np.log2(max_num_frames))
num_fourier_feats = np.ceil(np.log2(time_resolution)).astype(int)
powers = torch.tensor([2]).repeat(num_fourier_feats).pow(torch.arange(num_fourier_feats)) # [num_fourier_feats]
fourier_coefs = powers.unsqueeze(0).float() * np.pi # [1, num_fourier_feats]
return fourier_coefs / time_resolution
#----------------------------------------------------------------------------
Also note that we didn't experiment with this module much. We tried a couple of modifications, but they all performed the same.
from stylegan-v.
Many thanks to your reply!
Currently I am implementing a positional encoding provided in Section 3.1 as p(t) = alpha * sin(omega * t + rho), with alpha, omega, rho learnable.
class PositionalEncoding(Module):
"""
TODO: Other PE type implementation, E.g. [..., cos(2*pi*sigma^(j/m)*v), sin(2*pi*sigma^(j/m)*v), ...]
"""
def __init__(self, dim):
super().__init__()
self.dim = dim
self.omega = nn.Parameter(torch.randn(1, self.dim, requires_grad=True))
self.rho = nn.Parameter(torch.randn(self.dim, requires_grad=True))
self.alpha = nn.Parameter(torch.randn(self.dim, requires_grad=True))
def forward(self, Ts):
"""
Ts: [b, t, 1]
"""
ot = torch.einsum("btc,cd->btd", [Ts, self.omega])
pe = torch.sin(ot + self.rho) * self.alpha
return pe
Is it supposed to make sense?
from stylegan-v.
Hi! Yes, it does make sense, we tried such parametrization, and it produced very repetitive motions.
Is your application also video generation?
from stylegan-v.
Yes!
We are following a generally different idea, but a PE of time stamps is still necessary.
Many thanks for your sharing.
from stylegan-v.
Hi! We've pushed the training/evaluation/sampling code to this repo several days ago.
You can check our discriminator details
from stylegan-v.
Related Issues (20)
- Large GPU memory consumption at the beginning of training HOT 1
- Questions about details of hyperparameters in AlignedTimeEncoder HOT 4
- Notebook release
- FVD calculation is not deterministic HOT 1
- About FaceForensics Dataset HOT 2
- Linearly spaced periods stated in Appendix B.3 is not correct HOT 1
- Low IS on UCF101 HOT 16
- Face Forensics dataset preprocessing HOT 1
- FileNotFoundError: [Errno 2] No such file or directory: 'src\\experiment_config.yaml'
- Run stylegan-v Conditionally
- Question about FVD calculation HOT 2
- Question about the release of Pre-trained checkpoints HOT 2
- I tried to generate fake video samples with your pretrained weights but....
- ImportError: cannot import name 'CoordFuser' from 'src.training.layers' HOT 4
- TypeError: Descriptors cannot be created directly
- Projection of real video with multiple frames
- RainbowJelly dataset
- Problem regarding true data statics computation for FVD HOT 4
- Rainbow jelly isn't available
- MOCOGAN: training isn't runing
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from stylegan-v.