Skip to content

GitLab

Commit fba8bde8 authored by bailuo's avatar bailuo
Browse files

update

parents
Pipeline #1808 failed with stages
Hide whitespace changes
Inline Side-by-side
Showing with 2614 additions and 0 deletions
stylegan_human/training_scripts/sg3/training/dataset.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
# 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.
"""Streaming images and labels from datasets created with dataset_tool.py."""
import os
import numpy as np
import zipfile
import PIL.Image
import json
import torch
import dnnlib
from petrel_client.client import Client
import cv2
try:
import pyspng
except ImportError:
pyspng = None
#----------------------------------------------------------------------------
class Dataset(torch.utils.data.Dataset):
def __init__(self,
name, # Name of the dataset.
raw_shape, # Shape of the raw image data (NCHW).
max_size = None, # Artificially limit the size of the dataset. None = no limit. Applied before xflip.
use_labels = False, # Enable conditioning labels? False = label dimension is zero.
xflip = False, # Artificially double the size of the dataset via x-flips. Applied after max_size.
random_seed = 0, # Random seed to use when applying max_size.
square = False,
):
print('Inside Dataset')
self._name = name
self._raw_shape = list(raw_shape)
self._use_labels = use_labels
self._raw_labels = None
self._label_shape = None
self._square = square
print("inside dataset, _square: ", self._square)
# Apply max_size.
self._raw_idx = np.arange(self._raw_shape[0], dtype=np.int64)
if (max_size is not None) and (self._raw_idx.size > max_size):
np.random.RandomState(random_seed).shuffle(self._raw_idx)
self._raw_idx = np.sort(self._raw_idx[:max_size])
# Apply xflip.
self._xflip = np.zeros(self._raw_idx.size, dtype=np.uint8)
if xflip:
self._raw_idx = np.tile(self._raw_idx, 2)
self._xflip = np.concatenate([self._xflip, np.ones_like(self._xflip)])
def _get_raw_labels(self):
if self._raw_labels is None:
self._raw_labels = self._load_raw_labels() if self._use_labels else None
if self._raw_labels is None:
self._raw_labels = np.zeros([self._raw_shape[0], 0], dtype=np.float32)
assert isinstance(self._raw_labels, np.ndarray)
assert self._raw_labels.shape[0] == self._raw_shape[0]
assert self._raw_labels.dtype in [np.float32, np.int64]
if self._raw_labels.dtype == np.int64:
assert self._raw_labels.ndim == 1
assert np.all(self._raw_labels >= 0)
return self._raw_labels
def close(self): # to be overridden by subclass
pass
def _load_raw_image(self, raw_idx): # to be overridden by subclass
raise NotImplementedError
def _load_raw_labels(self): # to be overridden by subclass
raise NotImplementedError
def __getstate__(self):
return dict(self.__dict__, _raw_labels=None)
def __del__(self):
try:
self.close()
except:
pass
def __len__(self):
return self._raw_idx.size
def __getitem__(self, idx):
image = self._load_raw_image(self._raw_idx[idx])
assert isinstance(image, np.ndarray)
assert list(image.shape) == self.image_shape
assert image.dtype == np.uint8
if self._xflip[idx]:
assert image.ndim == 3 # CHW
image = image[:, :, ::-1]
return image.copy(), self.get_label(idx)
def get_label(self, idx):
label = self._get_raw_labels()[self._raw_idx[idx]]
if label.dtype == np.int64:
onehot = np.zeros(self.label_shape, dtype=np.float32)
onehot[label] = 1
label = onehot
return label.copy()
def get_details(self, idx):
d = dnnlib.EasyDict()
d.raw_idx = int(self._raw_idx[idx])
d.xflip = (int(self._xflip[idx]) != 0)
d.raw_label = self._get_raw_labels()[d.raw_idx].copy()
return d
@property
def name(self):
return self._name
@property
def image_shape(self):
return list(self._raw_shape[1:])
@property
def num_channels(self):
assert len(self.image_shape) == 3 # CHW
return self.image_shape[0]
@property
def resolution(self):
assert len(self.image_shape) == 3 # CHW
if self._square:
assert self.image_shape[1] == self.image_shape[2]
else:
assert self.image_shape[1] == self.image_shape[2] * 2
return self.image_shape[1]
@property
def label_shape(self):
if self._label_shape is None:
raw_labels = self._get_raw_labels()
if raw_labels.dtype == np.int64:
self._label_shape = [int(np.max(raw_labels)) + 1]
else:
self._label_shape = raw_labels.shape[1:]
return list(self._label_shape)
@property
def label_dim(self):
assert len(self.label_shape) == 1
return self.label_shape[0]
@property
def has_labels(self):
return any(x != 0 for x in self.label_shape)
@property
def has_onehot_labels(self):
return self._get_raw_labels().dtype == np.int64
#----------------------------------------------------------------------------
class ImageFolderDataset(Dataset):
def __init__(self,
path, # Path to directory or zip.
resolution = None, # Ensure specific resolution, None = highest available.
ceph = False,
square = False,
**super_kwargs, # Additional arguments for the Dataset base class.
):
self._path = path
self._zipfile = None
self._square = square
if os.path.isdir(self._path):
self._type = 'dir'
self._all_fnames = {os.path.relpath(os.path.join(root, fname), start=self._path) for root, _dirs, files in os.walk(self._path) for fname in files}
elif self._file_ext(self._path) == '.zip':
self._type = 'zip'
self._all_fnames = set(self._get_zipfile().namelist())
else:
raise IOError('Path must point to a directory or zip')
PIL.Image.init()
self._image_fnames = sorted(fname for fname in self._all_fnames if self._file_ext(fname) in PIL.Image.EXTENSION)
if len(self._image_fnames) == 0:
raise IOError('No image files found in the specified path')
name = os.path.splitext(os.path.basename(self._path))[0]
raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0).shape)
# if resolution is not None and (raw_shape[2] != resolution or raw_shape[3] != resolution):
# raise IOError('Image files do not match the specified resolution')
if resolution is not None:
if self._square:
raw_shape[2] = raw_shape[3] = resolution
else:
raw_shape[2] = resolution
raw_shape[3] = resolution // 2
# print(raw_shape)
super().__init__(name=name, raw_shape=raw_shape,square=square, **super_kwargs)
@staticmethod
def _file_ext(fname):
return os.path.splitext(fname)[1].lower()
def _get_zipfile(self):
assert self._type == 'zip'
if self._zipfile is None:
self._zipfile = zipfile.ZipFile(self._path)
return self._zipfile
def _open_file(self, fname):
if self._type == 'dir':
return open(os.path.join(self._path, fname), 'rb')
if self._type == 'zip':
return self._get_zipfile().open(fname, 'r')
return None
def close(self):
try:
if self._zipfile is not None:
self._zipfile.close()
finally:
self._zipfile = None
def __getstate__(self):
return dict(super().__getstate__(), _zipfile=None)
def _load_raw_image(self, raw_idx):
fname = self._image_fnames[raw_idx]
with self._open_file(fname) as f:
if pyspng is not None and self._file_ext(fname) == '.png':
image = pyspng.load(f.read())
else:
image = np.array(PIL.Image.open(f))
if image.ndim == 2:
image = image[:, :, np.newaxis] # HW => HWC
image = image.transpose(2, 0, 1) # HWC => CHW
return image
def _load_raw_labels(self):
fname = 'dataset.json'
if fname not in self._all_fnames:
return None
with self._open_file(fname) as f:
labels = json.load(f)['labels']
if labels is None:
return None
labels = dict(labels)
labels = [labels[fname.replace('\\', '/')] for fname in self._image_fnames]
labels = np.array(labels)
labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])
return labels
#----------------------------------------------------------------------------
stylegan_human/training_scripts/sg3/training/networks_stylegan2.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. 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.
"""Network architectures from the paper
"Analyzing and Improving the Image Quality of StyleGAN".
Matches the original implementation of configs E-F by Karras et al. at
https://github.com/NVlabs/stylegan2/blob/master/training/networks_stylegan2.py"""
import numpy as np
import torch
from torch_utils import misc
from torch_utils import persistence
from torch_utils.ops import conv2d_resample
from torch_utils.ops import upfirdn2d
from torch_utils.ops import bias_act
from torch_utils.ops import fma
#----------------------------------------------------------------------------
@misc.profiled_function
def normalize_2nd_moment(x, dim=1, eps=1e-8):
return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
#----------------------------------------------------------------------------
@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) # [NOIkk]
w = w * 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
#----------------------------------------------------------------------------
@persistence.persistent_class
class FullyConnectedLayer(torch.nn.Module):
def __init__(self,
in_features, # Number of input features.
out_features, # Number of output features.
bias = True, # Apply additive bias before the activation function?
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
lr_multiplier = 1, # Learning rate multiplier.
bias_init = 0, # Initial value for the additive bias.
):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.activation = activation
self.weight = torch.nn.Parameter(torch.randn([out_features, in_features]) / lr_multiplier)
self.bias = torch.nn.Parameter(torch.full([out_features], np.float32(bias_init))) if bias else None
self.weight_gain = lr_multiplier / np.sqrt(in_features)
self.bias_gain = lr_multiplier
def forward(self, x):
w = self.weight.to(x.dtype) * self.weight_gain
b = self.bias
if b is not None:
b = b.to(x.dtype)
if self.bias_gain != 1:
b = b * self.bias_gain
if self.activation == 'linear' and b is not None:
x = torch.addmm(b.unsqueeze(0), x, w.t())
else:
x = x.matmul(w.t())
x = bias_act.bias_act(x, b, act=self.activation)
return x
def extra_repr(self):
return f'in_features={self.in_features:d}, out_features={self.out_features:d}, activation={self.activation:s}'
#----------------------------------------------------------------------------
@persistence.persistent_class
class Conv2dLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
kernel_size, # Width and height of the convolution kernel.
bias = True, # Apply additive bias before the activation function?
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
up = 1, # Integer upsampling factor.
down = 1, # Integer downsampling factor.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output to +-X, None = disable clamping.
channels_last = False, # Expect the input to have memory_format=channels_last?
trainable = True, # Update the weights of this layer during training?
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.activation = activation
self.up = up
self.down = down
self.conv_clamp = conv_clamp
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
self.act_gain = bias_act.activation_funcs[activation].def_gain
memory_format = torch.channels_last if channels_last else torch.contiguous_format
weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format)
bias = torch.zeros([out_channels]) if bias else None
if trainable:
self.weight = torch.nn.Parameter(weight)
self.bias = torch.nn.Parameter(bias) if bias is not None else None
else:
self.register_buffer('weight', weight)
if bias is not None:
self.register_buffer('bias', bias)
else:
self.bias = None
def forward(self, x, gain=1):
w = self.weight * self.weight_gain
b = self.bias.to(x.dtype) if self.bias is not None else None
flip_weight = (self.up == 1) # slightly faster
x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=self.resample_filter, up=self.up, down=self.down, padding=self.padding, flip_weight=flip_weight)
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, b, act=self.activation, gain=act_gain, clamp=act_clamp)
return x
def extra_repr(self):
return ' '.join([
f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, activation={self.activation:s},',
f'up={self.up}, down={self.down}'])
#----------------------------------------------------------------------------
@persistence.persistent_class
class MappingNetwork(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality, 0 = no latent.
c_dim, # Conditioning label (C) dimensionality, 0 = no label.
w_dim, # Intermediate latent (W) dimensionality.
num_ws, # Number of intermediate latents to output, None = do not broadcast.
num_layers = 8, # Number of mapping layers.
embed_features = None, # Label embedding dimensionality, None = same as w_dim.
layer_features = None, # Number of intermediate features in the mapping layers, None = same as w_dim.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
lr_multiplier = 0.01, # Learning rate multiplier for the mapping layers.
w_avg_beta = 0.998, # Decay for tracking the moving average of W during training, None = do not track.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.num_ws = num_ws
self.num_layers = num_layers
self.w_avg_beta = w_avg_beta
if embed_features is None:
embed_features = w_dim
if c_dim == 0:
embed_features = 0
if layer_features is None:
layer_features = w_dim
features_list = [z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]
if c_dim > 0:
self.embed = FullyConnectedLayer(c_dim, embed_features)
for idx in range(num_layers):
in_features = features_list[idx]
out_features = features_list[idx + 1]
layer = FullyConnectedLayer(in_features, out_features, activation=activation, lr_multiplier=lr_multiplier)
setattr(self, f'fc{idx}', layer)
if num_ws is not None and w_avg_beta is not None:
self.register_buffer('w_avg', torch.zeros([w_dim]))
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False):
# Embed, normalize, and concat inputs.
x = None
with torch.autograd.profiler.record_function('input'):
if self.z_dim > 0:
misc.assert_shape(z, [None, self.z_dim])
x = normalize_2nd_moment(z.to(torch.float32))
if self.c_dim > 0:
misc.assert_shape(c, [None, self.c_dim])
y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
x = torch.cat([x, y], dim=1) if x is not None else y
# Main layers.
for idx in range(self.num_layers):
layer = getattr(self, f'fc{idx}')
x = layer(x)
# Update moving average of W.
if update_emas and self.w_avg_beta is not None:
with torch.autograd.profiler.record_function('update_w_avg'):
self.w_avg.copy_(x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta))
# Broadcast.
if self.num_ws is not None:
with torch.autograd.profiler.record_function('broadcast'):
x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
# Apply truncation.
if truncation_psi != 1:
with torch.autograd.profiler.record_function('truncate'):
assert self.w_avg_beta is not None
if self.num_ws is None or truncation_cutoff is None:
x = self.w_avg.lerp(x, truncation_psi)
else:
x[:, :truncation_cutoff] = self.w_avg.lerp(x[:, :truncation_cutoff], truncation_psi)
return x
def extra_repr(self):
return f'z_dim={self.z_dim:d}, c_dim={self.c_dim:d}, w_dim={self.w_dim:d}, num_ws={self.num_ws:d}'
#----------------------------------------------------------------------------
@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.
use_noise = True, # Enable noise input?
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?
square = False, # default if for rectangle images
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.w_dim = w_dim
self.resolution = resolution
self.up = up
self.use_noise = use_noise
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
self.square=square
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 use_noise:
if self.square:
self.register_buffer('noise_const', torch.randn([resolution, resolution]))
else:
self.register_buffer('noise_const', torch.randn([resolution, resolution // 2]))
self.noise_strength = torch.nn.Parameter(torch.zeros([]))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
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
if self.square:
misc.assert_shape(x, [None, self.weight.shape[1], in_resolution, in_resolution])
else:
misc.assert_shape(x, [None, self.weight.shape[1], in_resolution, in_resolution // 2])
styles = self.affine(w)
noise = None
if self.use_noise and noise_mode == 'random':
if self.square:
noise = torch.randn([x.shape[0], 1, self.resolution, self.resolution], device=x.device) * self.noise_strength
else:
noise = torch.randn([x.shape[0], 1, self.resolution, self.resolution // 2], device=x.device) * self.noise_strength
if self.use_noise and noise_mode == 'const':
noise = self.noise_const * self.noise_strength
flip_weight = (self.up == 1) # slightly faster
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
def extra_repr(self):
return ' '.join([
f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, w_dim={self.w_dim:d},',
f'resolution={self.resolution:d}, up={self.up}, activation={self.activation:s}'])
#----------------------------------------------------------------------------
@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.in_channels = in_channels
self.out_channels = out_channels
self.w_dim = w_dim
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
def extra_repr(self):
return f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}, w_dim={self.w_dim:d}'
#----------------------------------------------------------------------------
@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.
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 = 256, # 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?
square = False, # default is for rectangle images
fused_modconv_default = True, # Default value of fused_modconv. 'inference_only' = True for inference, False for training.
**layer_kwargs, # Arguments for SynthesisLayer.
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.w_dim = w_dim
self.resolution = resolution
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.fused_modconv_default = fused_modconv_default
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.num_conv = 0
self.num_torgb = 0
self.square = square
if in_channels == 0:
if self.square:
self.const = torch.nn.Parameter(torch.randn([out_channels, resolution, resolution]))
else: # rectangle
self.const = torch.nn.Parameter(torch.randn([out_channels, resolution, resolution // 2]))
if in_channels != 0:
self.conv0 = SynthesisLayer(in_channels, out_channels, w_dim=w_dim, resolution=resolution, up=2,
resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last, square=square, **layer_kwargs)
self.num_conv += 1
self.conv1 = SynthesisLayer(out_channels, out_channels, w_dim=w_dim, resolution=resolution,
conv_clamp=conv_clamp, channels_last=self.channels_last, square=square, **layer_kwargs)
self.num_conv += 1
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=2,
resample_filter=resample_filter, channels_last=self.channels_last)
def forward(self, x, img, ws, force_fp32=False, fused_modconv=None, update_emas=False, **layer_kwargs):
_ = update_emas # unused
misc.assert_shape(ws, [None, self.num_conv + self.num_torgb, self.w_dim])
w_iter = iter(ws.unbind(dim=1))
if ws.device.type != 'cuda':
force_fp32 = True
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:
fused_modconv = self.fused_modconv_default
if fused_modconv == 'inference_only':
fused_modconv = (not self.training)
# Input.
if self.in_channels == 0:
x = self.const.to(dtype=dtype, memory_format=memory_format)
x = x.unsqueeze(0).repeat([ws.shape[0], 1, 1, 1])
else:
if self.square:
misc.assert_shape(x, [None, self.in_channels, self.resolution // 2, self.resolution // 2])
else: # rectangle
misc.assert_shape(x, [None, self.in_channels, self.resolution // 2, self.resolution // 4])
x = x.to(dtype=dtype, memory_format=memory_format)
# Main layers.
if self.in_channels == 0:
x = self.conv1(x, next(w_iter), 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:
x = self.conv0(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs)
x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv, **layer_kwargs)
# ToRGB.
if img is not None:
if self.square:
misc.assert_shape(img, [None, self.img_channels, self.resolution // 2, self.resolution // 2])
else:
misc.assert_shape(img, [None, self.img_channels, self.resolution // 2, self.resolution // 4])
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
def extra_repr(self):
return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
#----------------------------------------------------------------------------
@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.
square,
channel_base = 32768, # Overall multiplier for the number of channels.
channel_max = 512, # Maximum number of channels in any layer.
num_fp16_res = 4, # Use FP16 for the N highest resolutions.
**block_kwargs, # Arguments for SynthesisBlock.
):
assert img_resolution >= 4 and img_resolution & (img_resolution - 1) == 0
super().__init__()
self.w_dim = w_dim
self.img_resolution = img_resolution
self.img_resolution_log2 = int(np.log2(img_resolution))
self.img_channels = img_channels
self.square=square
self.num_fp16_res = num_fp16_res
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)
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=w_dim, resolution=res,
img_channels=img_channels, is_last=is_last, use_fp16=use_fp16, square=square,**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, **block_kwargs):
block_ws = []
with torch.autograd.profiler.record_function('split_ws'):
misc.assert_shape(ws, [None, self.num_ws, self.w_dim])
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}')
x, img = block(x, img, cur_ws, **block_kwargs)
return img
def extra_repr(self):
return ' '.join([
f'w_dim={self.w_dim:d}, num_ws={self.num_ws:d},',
f'img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d},',
f'num_fp16_res={self.num_fp16_res:d}'])
#----------------------------------------------------------------------------
@persistence.persistent_class
class Generator(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality.
c_dim, # Conditioning label (C) dimensionality.
w_dim, # Intermediate latent (W) dimensionality.
square,
img_resolution, # Output resolution.
img_channels, # Number of output color channels.
mapping_kwargs = {}, # Arguments for MappingNetwork.
**synthesis_kwargs, # Arguments for SynthesisNetwork.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.square = square
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, square=square, **synthesis_kwargs)
self.num_ws = self.synthesis.num_ws
self.mapping = MappingNetwork(z_dim=z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, **synthesis_kwargs):
ws = self.mapping(z, c, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff, update_emas=update_emas)
img = self.synthesis(ws, update_emas=update_emas, **synthesis_kwargs)
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.
square = False,
):
assert in_channels in [0, tmp_channels]
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
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.square = square
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()
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)
self.conv0 = Conv2dLayer(tmp_channels, tmp_channels, kernel_size=3, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv1 = Conv2dLayer(tmp_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(tmp_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, force_fp32=False):
if (x if x is not None else img).device.type != 'cuda':
force_fp32 = True
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:
if self.square:
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution])
else:
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution // 2])
x = x.to(dtype=dtype, memory_format=memory_format)
# FromRGB.
if self.in_channels == 0 or self.architecture == 'skip':
if self.square:
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
else:
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution // 2])
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
# Main layers.
if self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x)
x = self.conv1(x, gain=np.sqrt(0.5))
x = y.add_(x)
else:
x = self.conv0(x)
x = self.conv1(x)
assert x.dtype == dtype
return x, img
def extra_repr(self):
return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
#----------------------------------------------------------------------------
@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) # [NCHW] Append to input as new channels.
return x
def extra_repr(self):
return f'group_size={self.group_size}, num_channels={self.num_channels: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.
square = False,
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.cmap_dim = cmap_dim
self.resolution = resolution
self.img_channels = img_channels
self.architecture = architecture
self.square = square
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)
if self.square:
self.fc = FullyConnectedLayer(in_channels * (resolution ** 2), in_channels, activation=activation)
else:
self.fc = FullyConnectedLayer(in_channels * (resolution ** 2 // 2), in_channels, activation=activation)
self.out = FullyConnectedLayer(in_channels, 1 if cmap_dim == 0 else cmap_dim)
def forward(self, x, img, cmap, force_fp32=False):
if self.square:
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution])
else:
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution // 2]) # [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':
if self.square:
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
else:
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution // 2])
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)
x = self.fc(x.flatten(1))
x = self.out(x)
# 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))
assert x.dtype == dtype
return x
def extra_repr(self):
return f'resolution={self.resolution:d}, architecture={self.architecture:s}'
#----------------------------------------------------------------------------
@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 = 4, # Use FP16 for the N highest resolutions.
conv_clamp = 256, # Clamp the output of convolution layers to +-X, None = disable clamping.
cmap_dim = None, # Dimensionality of mapped conditioning label, None = default.
square = False, # default for rectangle images
block_kwargs = {}, # Arguments for DiscriminatorBlock.
mapping_kwargs = {}, # Arguments for MappingNetwork.
epilogue_kwargs = {}, # Arguments for DiscriminatorEpilogue.
):
super().__init__()
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.square = square
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 c_dim == 0:
cmap_dim = 0
common_kwargs = dict(img_channels=img_channels, architecture=architecture, conv_clamp=conv_clamp)
cur_layer_idx = 0
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]
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, square=square, **block_kwargs, **common_kwargs)
setattr(self, f'b{res}', block)
cur_layer_idx += block.num_layers
if c_dim > 0:
self.mapping = MappingNetwork(z_dim=0, c_dim=c_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, square=square, **epilogue_kwargs, **common_kwargs)
def forward(self, img, c, update_emas=False, **block_kwargs):
_ = update_emas # unused
x = None
for res in self.block_resolutions:
block = getattr(self, f'b{res}')
x, img = block(x, img, **block_kwargs)
cmap = None
if self.c_dim > 0:
cmap = self.mapping(None, c)
x = self.b4(x, img, cmap)
return x
def extra_repr(self):
return f'c_dim={self.c_dim:d}, img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d}'
#----------------------------------------------------------------------------
stylegan_human/training_scripts/sg3/training/networks_stylegan3.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. 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.
"""Generator architecture from the paper
"Alias-Free Generative Adversarial Networks"."""
import numpy as np
import scipy.signal
import scipy.optimize
import torch
from torch_utils import misc
from torch_utils import persistence
from torch_utils.ops import conv2d_gradfix
from torch_utils.ops import filtered_lrelu
from torch_utils.ops import bias_act
#----------------------------------------------------------------------------
@misc.profiled_function
def modulated_conv2d(
x, # Input tensor: [batch_size, in_channels, in_height, in_width]
w, # Weight tensor: [out_channels, in_channels, kernel_height, kernel_width]
s, # Style tensor: [batch_size, in_channels]
demodulate = True, # Apply weight demodulation?
padding = 0, # Padding: int or [padH, padW]
input_gain = None, # Optional scale factors for the input channels: [], [in_channels], or [batch_size, in_channels]
):
with misc.suppress_tracer_warnings(): # this value will be treated as a constant
batch_size = int(x.shape[0])
out_channels, in_channels, kh, kw = w.shape
misc.assert_shape(w, [out_channels, in_channels, kh, kw]) # [OIkk]
misc.assert_shape(x, [batch_size, in_channels, None, None]) # [NIHW]
misc.assert_shape(s, [batch_size, in_channels]) # [NI]
# Pre-normalize inputs.
if demodulate:
w = w * w.square().mean([1,2,3], keepdim=True).rsqrt()
s = s * s.square().mean().rsqrt()
# Modulate weights.
w = w.unsqueeze(0) # [NOIkk]
w = w * s.unsqueeze(1).unsqueeze(3).unsqueeze(4) # [NOIkk]
# Demodulate weights.
if demodulate:
dcoefs = (w.square().sum(dim=[2,3,4]) + 1e-8).rsqrt() # [NO]
w = w * dcoefs.unsqueeze(2).unsqueeze(3).unsqueeze(4) # [NOIkk]
# Apply input scaling.
if input_gain is not None:
input_gain = input_gain.expand(batch_size, in_channels) # [NI]
w = w * input_gain.unsqueeze(1).unsqueeze(3).unsqueeze(4) # [NOIkk]
# Execute as one fused op using grouped convolution.
x = x.reshape(1, -1, *x.shape[2:])
w = w.reshape(-1, in_channels, kh, kw)
x = conv2d_gradfix.conv2d(input=x, weight=w.to(x.dtype), padding=padding, groups=batch_size)
x = x.reshape(batch_size, -1, *x.shape[2:])
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class FullyConnectedLayer(torch.nn.Module):
def __init__(self,
in_features, # Number of input features.
out_features, # Number of output features.
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
bias = True, # Apply additive bias before the activation function?
lr_multiplier = 1, # Learning rate multiplier.
weight_init = 1, # Initial standard deviation of the weight tensor.
bias_init = 0, # Initial value of the additive bias.
):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.activation = activation
self.weight = torch.nn.Parameter(torch.randn([out_features, in_features]) * (weight_init / lr_multiplier))
bias_init = np.broadcast_to(np.asarray(bias_init, dtype=np.float32), [out_features])
self.bias = torch.nn.Parameter(torch.from_numpy(bias_init / lr_multiplier)) if bias else None
self.weight_gain = lr_multiplier / np.sqrt(in_features)
self.bias_gain = lr_multiplier
def forward(self, x):
w = self.weight.to(x.dtype) * self.weight_gain
b = self.bias
if b is not None:
b = b.to(x.dtype)
if self.bias_gain != 1:
b = b * self.bias_gain
if self.activation == 'linear' and b is not None:
x = torch.addmm(b.unsqueeze(0), x, w.t())
else:
x = x.matmul(w.t())
x = bias_act.bias_act(x, b, act=self.activation)
return x
def extra_repr(self):
return f'in_features={self.in_features:d}, out_features={self.out_features:d}, activation={self.activation:s}'
#----------------------------------------------------------------------------
@persistence.persistent_class
class MappingNetwork(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality.
c_dim, # Conditioning label (C) dimensionality, 0 = no labels.
w_dim, # Intermediate latent (W) dimensionality.
num_ws, # Number of intermediate latents to output.
num_layers = 2, # Number of mapping layers.
lr_multiplier = 0.01, # Learning rate multiplier for the mapping layers.
w_avg_beta = 0.998, # Decay for tracking the moving average of W during training.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.num_ws = num_ws
self.num_layers = num_layers
self.w_avg_beta = w_avg_beta
# Construct layers.
self.embed = FullyConnectedLayer(self.c_dim, self.w_dim) if self.c_dim > 0 else None
features = [self.z_dim + (self.w_dim if self.c_dim > 0 else 0)] + [self.w_dim] * self.num_layers
for idx, in_features, out_features in zip(range(num_layers), features[:-1], features[1:]):
layer = FullyConnectedLayer(in_features, out_features, activation='lrelu', lr_multiplier=lr_multiplier)
setattr(self, f'fc{idx}', layer)
self.register_buffer('w_avg', torch.zeros([w_dim]))
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False):
misc.assert_shape(z, [None, self.z_dim])
if truncation_cutoff is None:
truncation_cutoff = self.num_ws
# Embed, normalize, and concatenate inputs.
x = z.to(torch.float32)
x = x * (x.square().mean(1, keepdim=True) + 1e-8).rsqrt()
if self.c_dim > 0:
misc.assert_shape(c, [None, self.c_dim])
y = self.embed(c.to(torch.float32))
y = y * (y.square().mean(1, keepdim=True) + 1e-8).rsqrt()
x = torch.cat([x, y], dim=1) if x is not None else y
# Execute layers.
for idx in range(self.num_layers):
x = getattr(self, f'fc{idx}')(x)
# Update moving average of W.
if update_emas:
self.w_avg.copy_(x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta))
# Broadcast and apply truncation.
x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
if truncation_psi != 1:
x[:, :truncation_cutoff] = self.w_avg.lerp(x[:, :truncation_cutoff], truncation_psi)
return x
def extra_repr(self):
return f'z_dim={self.z_dim:d}, c_dim={self.c_dim:d}, w_dim={self.w_dim:d}, num_ws={self.num_ws:d}'
#----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisInput(torch.nn.Module):
def __init__(self,
w_dim, # Intermediate latent (W) dimensionality.
channels, # Number of output channels.
size, # Output spatial size: int or [width, height].
sampling_rate, # Output sampling rate.
bandwidth, # Output bandwidth.
square,
):
super().__init__()
self.w_dim = w_dim
self.channels = channels
self.square = square
if self.square:
self.size = np.broadcast_to(np.asarray(size), [2])
else:
self.size = np.array([size // 2, size]) # [width, height]
self.sampling_rate = sampling_rate
self.bandwidth = bandwidth
# Draw random frequencies from uniform 2D disc.
freqs = torch.randn([self.channels, 2])
radii = freqs.square().sum(dim=1, keepdim=True).sqrt()
freqs /= radii * radii.square().exp().pow(0.25)
freqs *= bandwidth
phases = torch.rand([self.channels]) - 0.5
# Setup parameters and buffers.
self.weight = torch.nn.Parameter(torch.randn([self.channels, self.channels]))
self.affine = FullyConnectedLayer(w_dim, 4, weight_init=0, bias_init=[1,0,0,0])
self.register_buffer('transform', torch.eye(3, 3)) # User-specified inverse transform wrt. resulting image.
self.register_buffer('freqs', freqs)
self.register_buffer('phases', phases)
def forward(self, w):
# Introduce batch dimension.
transforms = self.transform.unsqueeze(0) # [batch, row, col]
freqs = self.freqs.unsqueeze(0) # [batch, channel, xy]
phases = self.phases.unsqueeze(0) # [batch, channel]
# Apply learned transformation.
t = self.affine(w) # t = (r_c, r_s, t_x, t_y)
t = t / t[:, :2].norm(dim=1, keepdim=True) # t' = (r'_c, r'_s, t'_x, t'_y)
m_r = torch.eye(3, device=w.device).unsqueeze(0).repeat([w.shape[0], 1, 1]) # Inverse rotation wrt. resulting image.
m_r[:, 0, 0] = t[:, 0] # r'_c
m_r[:, 0, 1] = -t[:, 1] # r'_s
m_r[:, 1, 0] = t[:, 1] # r'_s
m_r[:, 1, 1] = t[:, 0] # r'_c
m_t = torch.eye(3, device=w.device).unsqueeze(0).repeat([w.shape[0], 1, 1]) # Inverse translation wrt. resulting image.
m_t[:, 0, 2] = -t[:, 2] # t'_x
m_t[:, 1, 2] = -t[:, 3] # t'_y
transforms = m_r @ m_t @ transforms # First rotate resulting image, then translate, and finally apply user-specified transform.
# Transform frequencies.
phases = phases + (freqs @ transforms[:, :2, 2:]).squeeze(2)
freqs = freqs @ transforms[:, :2, :2]
# Dampen out-of-band frequencies that may occur due to the user-specified transform.
amplitudes = (1 - (freqs.norm(dim=2) - self.bandwidth) / (self.sampling_rate / 2 - self.bandwidth)).clamp(0, 1)
# Construct sampling grid.
theta = torch.eye(2, 3, device=w.device)
theta[0, 0] = 0.5 * self.size[0] / self.sampling_rate
theta[1, 1] = 0.5 * self.size[1] / self.sampling_rate
grids = torch.nn.functional.affine_grid(theta.unsqueeze(0), [1, 1, self.size[1], self.size[0]], align_corners=False)
# Compute Fourier features.
x = (grids.unsqueeze(3) @ freqs.permute(0, 2, 1).unsqueeze(1).unsqueeze(2)).squeeze(3) # [batch, height, width, channel]
x = x + phases.unsqueeze(1).unsqueeze(2)
x = torch.sin(x * (np.pi * 2))
x = x * amplitudes.unsqueeze(1).unsqueeze(2)
# Apply trainable mapping.
weight = self.weight / np.sqrt(self.channels)
x = x @ weight.t()
# Ensure correct shape.
x = x.permute(0, 3, 1, 2) # [batch, channel, height, width]
misc.assert_shape(x, [w.shape[0], self.channels, int(self.size[1]), int(self.size[0])])
return x
def extra_repr(self):
return '\n'.join([
f'w_dim={self.w_dim:d}, channels={self.channels:d}, size={list(self.size)},',
f'sampling_rate={self.sampling_rate:g}, bandwidth={self.bandwidth:g}'])
#----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisLayer(torch.nn.Module):
def __init__(self,
w_dim, # Intermediate latent (W) dimensionality.
is_torgb, # Is this the final ToRGB layer?
is_critically_sampled, # Does this layer use critical sampling?
use_fp16, # Does this layer use FP16?
# Input & output specifications.
in_channels, # Number of input channels.
out_channels, # Number of output channels.
in_size, # Input spatial size: int or [width, height].
out_size, # Output spatial size: int or [width, height].
in_sampling_rate, # Input sampling rate (s).
out_sampling_rate, # Output sampling rate (s).
in_cutoff, # Input cutoff frequency (f_c).
out_cutoff, # Output cutoff frequency (f_c).
in_half_width, # Input transition band half-width (f_h).
out_half_width, # Output Transition band half-width (f_h).
# Hyperparameters.
conv_kernel = 3, # Convolution kernel size. Ignored for final the ToRGB layer.
filter_size = 6, # Low-pass filter size relative to the lower resolution when up/downsampling.
lrelu_upsampling = 2, # Relative sampling rate for leaky ReLU. Ignored for final the ToRGB layer.
use_radial_filters = False, # Use radially symmetric downsampling filter? Ignored for critically sampled layers.
conv_clamp = 256, # Clamp the output to [-X, +X], None = disable clamping.
magnitude_ema_beta = 0.999, # Decay rate for the moving average of input magnitudes.
square = False, # default if for rectangle images
):
super().__init__()
self.w_dim = w_dim
self.is_torgb = is_torgb
self.is_critically_sampled = is_critically_sampled
self.use_fp16 = use_fp16
self.in_channels = in_channels
self.out_channels = out_channels
self.square = square
if self.square:
self.in_size = np.broadcast_to(np.asarray(in_size), [2])
self.out_size = np.broadcast_to(np.asarray(out_size), [2])
else:
# self.in_size = np.array[in_size, in_size//2]
self.in_size = np.array([in_size // 2, in_size])
# self.out_size = np.array[out_size, out_size//2]
self.out_size = np.array([out_size // 2, out_size])
self.in_sampling_rate = in_sampling_rate
self.out_sampling_rate = out_sampling_rate
self.tmp_sampling_rate = max(in_sampling_rate, out_sampling_rate) * (1 if is_torgb else lrelu_upsampling)
self.in_cutoff = in_cutoff
self.out_cutoff = out_cutoff
self.in_half_width = in_half_width
self.out_half_width = out_half_width
self.conv_kernel = 1 if is_torgb else conv_kernel
self.conv_clamp = conv_clamp
self.magnitude_ema_beta = magnitude_ema_beta
# Setup parameters and buffers.
self.affine = FullyConnectedLayer(self.w_dim, self.in_channels, bias_init=1)
self.weight = torch.nn.Parameter(torch.randn([self.out_channels, self.in_channels, self.conv_kernel, self.conv_kernel]))
self.bias = torch.nn.Parameter(torch.zeros([self.out_channels]))
self.register_buffer('magnitude_ema', torch.ones([]))
# Design upsampling filter.
self.up_factor = int(np.rint(self.tmp_sampling_rate / self.in_sampling_rate))
assert self.in_sampling_rate * self.up_factor == self.tmp_sampling_rate
self.up_taps = filter_size * self.up_factor if self.up_factor > 1 and not self.is_torgb else 1
self.register_buffer('up_filter', self.design_lowpass_filter(
numtaps=self.up_taps, cutoff=self.in_cutoff, width=self.in_half_width*2, fs=self.tmp_sampling_rate))
# Design downsampling filter.
self.down_factor = int(np.rint(self.tmp_sampling_rate / self.out_sampling_rate))
assert self.out_sampling_rate * self.down_factor == self.tmp_sampling_rate
self.down_taps = filter_size * self.down_factor if self.down_factor > 1 and not self.is_torgb else 1
self.down_radial = use_radial_filters and not self.is_critically_sampled
self.register_buffer('down_filter', self.design_lowpass_filter(
numtaps=self.down_taps, cutoff=self.out_cutoff, width=self.out_half_width*2, fs=self.tmp_sampling_rate, radial=self.down_radial))
# Compute padding.
pad_total = (self.out_size - 1) * self.down_factor + 1 # Desired output size before downsampling.
pad_total -= (self.in_size + self.conv_kernel - 1) * self.up_factor # Input size after upsampling.
pad_total += self.up_taps + self.down_taps - 2 # Size reduction caused by the filters.
pad_lo = (pad_total + self.up_factor) // 2 # Shift sample locations according to the symmetric interpretation (Appendix C.3).
pad_hi = pad_total - pad_lo
self.padding = [int(pad_lo[0]), int(pad_hi[0]), int(pad_lo[1]), int(pad_hi[1])]
def forward(self, x, w, noise_mode='random', force_fp32=False, update_emas=False):
assert noise_mode in ['random', 'const', 'none'] # unused
misc.assert_shape(x, [None, self.in_channels, int(self.in_size[1]), int(self.in_size[0])])
misc.assert_shape(w, [x.shape[0], self.w_dim])
# Track input magnitude.
if update_emas:
with torch.autograd.profiler.record_function('update_magnitude_ema'):
magnitude_cur = x.detach().to(torch.float32).square().mean()
self.magnitude_ema.copy_(magnitude_cur.lerp(self.magnitude_ema, self.magnitude_ema_beta))
input_gain = self.magnitude_ema.rsqrt()
# Execute affine layer.
styles = self.affine(w)
if self.is_torgb:
weight_gain = 1 / np.sqrt(self.in_channels * (self.conv_kernel ** 2))
styles = styles * weight_gain
# Execute modulated conv2d.
dtype = torch.float16 if (self.use_fp16 and not force_fp32 and x.device.type == 'cuda') else torch.float32
x = modulated_conv2d(x=x.to(dtype), w=self.weight, s=styles,
padding=self.conv_kernel-1, demodulate=(not self.is_torgb), input_gain=input_gain)
# Execute bias, filtered leaky ReLU, and clamping.
gain = 1 if self.is_torgb else np.sqrt(2)
slope = 1 if self.is_torgb else 0.2
x = filtered_lrelu.filtered_lrelu(x=x, fu=self.up_filter, fd=self.down_filter, b=self.bias.to(x.dtype),
up=self.up_factor, down=self.down_factor, padding=self.padding, gain=gain, slope=slope, clamp=self.conv_clamp)
# Ensure correct shape and dtype.
misc.assert_shape(x, [None, self.out_channels, int(self.out_size[1]), int(self.out_size[0])])
assert x.dtype == dtype
return x
@staticmethod
def design_lowpass_filter(numtaps, cutoff, width, fs, radial=False):
assert numtaps >= 1
# Identity filter.
if numtaps == 1:
return None
# Separable Kaiser low-pass filter.
if not radial:
f = scipy.signal.firwin(numtaps=numtaps, cutoff=cutoff, width=width, fs=fs)
return torch.as_tensor(f, dtype=torch.float32)
# Radially symmetric jinc-based filter.
x = (np.arange(numtaps) - (numtaps - 1) / 2) / fs
r = np.hypot(*np.meshgrid(x, x))
f = scipy.special.j1(2 * cutoff * (np.pi * r)) / (np.pi * r)
beta = scipy.signal.kaiser_beta(scipy.signal.kaiser_atten(numtaps, width / (fs / 2)))
w = np.kaiser(numtaps, beta)
f *= np.outer(w, w)
f /= np.sum(f)
return torch.as_tensor(f, dtype=torch.float32)
def extra_repr(self):
return '\n'.join([
f'w_dim={self.w_dim:d}, is_torgb={self.is_torgb},',
f'is_critically_sampled={self.is_critically_sampled}, use_fp16={self.use_fp16},',
f'in_sampling_rate={self.in_sampling_rate:g}, out_sampling_rate={self.out_sampling_rate:g},',
f'in_cutoff={self.in_cutoff:g}, out_cutoff={self.out_cutoff:g},',
f'in_half_width={self.in_half_width:g}, out_half_width={self.out_half_width:g},',
f'in_size={list(self.in_size)}, out_size={list(self.out_size)},',
f'in_channels={self.in_channels:d}, out_channels={self.out_channels:d}'])
#----------------------------------------------------------------------------
@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.
square,
channel_base = 32768, # Overall multiplier for the number of channels.
channel_max = 512, # Maximum number of channels in any layer.
num_layers = 14, # Total number of layers, excluding Fourier features and ToRGB.
num_critical = 2, # Number of critically sampled layers at the end.
first_cutoff = 2, # Cutoff frequency of the first layer (f_{c,0}).
first_stopband = 2**2.1, # Minimum stopband of the first layer (f_{t,0}).
last_stopband_rel = 2**0.3, # Minimum stopband of the last layer, expressed relative to the cutoff.
margin_size = 10, # Number of additional pixels outside the image.
output_scale = 0.25, # Scale factor for the output image.
num_fp16_res = 4, # Use FP16 for the N highest resolutions.
**layer_kwargs, # Arguments for SynthesisLayer.
):
super().__init__()
self.w_dim = w_dim
self.num_ws = num_layers + 2
self.img_resolution = img_resolution
self.img_channels = img_channels
self.num_layers = num_layers
self.num_critical = num_critical
self.margin_size = margin_size
self.output_scale = output_scale
self.num_fp16_res = num_fp16_res
self.square = square
# Geometric progression of layer cutoffs and min. stopbands.
last_cutoff = self.img_resolution / 2 # f_{c,N}
last_stopband = last_cutoff * last_stopband_rel # f_{t,N}
exponents = np.minimum(np.arange(self.num_layers + 1) / (self.num_layers - self.num_critical), 1)
cutoffs = first_cutoff * (last_cutoff / first_cutoff) ** exponents # f_c[i]
stopbands = first_stopband * (last_stopband / first_stopband) ** exponents # f_t[i]
# Compute remaining layer parameters.
sampling_rates = np.exp2(np.ceil(np.log2(np.minimum(stopbands * 2, self.img_resolution)))) # s[i]
half_widths = np.maximum(stopbands, sampling_rates / 2) - cutoffs # f_h[i]
sizes = sampling_rates + self.margin_size * 2
sizes[-2:] = self.img_resolution
channels = np.rint(np.minimum((channel_base / 2) / cutoffs, channel_max))
channels[-1] = self.img_channels
# Construct layers.
self.input = SynthesisInput(
w_dim=self.w_dim, channels=int(channels[0]), size=int(sizes[0]),
sampling_rate=sampling_rates[0], bandwidth=cutoffs[0], square=self.square)
self.layer_names = []
for idx in range(self.num_layers + 1):
prev = max(idx - 1, 0)
is_torgb = (idx == self.num_layers)
is_critically_sampled = (idx >= self.num_layers - self.num_critical)
use_fp16 = (sampling_rates[idx] * (2 ** self.num_fp16_res) > self.img_resolution)
layer = SynthesisLayer(
w_dim=self.w_dim, is_torgb=is_torgb, is_critically_sampled=is_critically_sampled, use_fp16=use_fp16,
in_channels=int(channels[prev]), out_channels= int(channels[idx]),
in_size=int(sizes[prev]), out_size=int(sizes[idx]),
in_sampling_rate=int(sampling_rates[prev]), out_sampling_rate=int(sampling_rates[idx]),
in_cutoff=cutoffs[prev], out_cutoff=cutoffs[idx],
in_half_width=half_widths[prev], out_half_width=half_widths[idx],
square=self.square,
**layer_kwargs)
name = f'L{idx}_{layer.out_size[0]}_{layer.out_channels}'
setattr(self, name, layer)
self.layer_names.append(name)
def forward(self, ws, **layer_kwargs):
misc.assert_shape(ws, [None, self.num_ws, self.w_dim])
ws = ws.to(torch.float32).unbind(dim=1)
# Execute layers.
x = self.input(ws[0])
for name, w in zip(self.layer_names, ws[1:]):
x = getattr(self, name)(x, w, **layer_kwargs)
if self.output_scale != 1:
x = x * self.output_scale
# Ensure correct shape and dtype.
if self.square:
misc.assert_shape(x, [None, self.img_channels, self.img_resolution, self.img_resolution])
else:
misc.assert_shape(x, [None, self.img_channels, self.img_resolution, self.img_resolution // 2])
x = x.to(torch.float32)
return x
def extra_repr(self):
return '\n'.join([
f'w_dim={self.w_dim:d}, num_ws={self.num_ws:d},',
f'img_resolution={self.img_resolution:d}, img_channels={self.img_channels:d},',
f'num_layers={self.num_layers:d}, num_critical={self.num_critical:d},',
f'margin_size={self.margin_size:d}, num_fp16_res={self.num_fp16_res:d}'])
#----------------------------------------------------------------------------
@persistence.persistent_class
class Generator(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality.
c_dim, # Conditioning label (C) dimensionality.
w_dim, # Intermediate latent (W) dimensionality.
img_resolution, # Output resolution.
square,
img_channels, # Number of output color channels.
mapping_kwargs = {}, # Arguments for MappingNetwork.
**synthesis_kwargs, # Arguments for SynthesisNetwork.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.img_resolution = img_resolution
self.img_channels = img_channels
self.square = square
self.synthesis = SynthesisNetwork(w_dim=w_dim, img_resolution=img_resolution, img_channels=img_channels, square=self.square, **synthesis_kwargs)
self.num_ws = self.synthesis.num_ws
self.mapping = MappingNetwork(z_dim=z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, update_emas=False, **synthesis_kwargs):
ws = self.mapping(z, c, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff, update_emas=update_emas)
img = self.synthesis(ws, update_emas=update_emas, **synthesis_kwargs)
return img
#----------------------------------------------------------------------------
stylegan_human/utils/ImagesDataset.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
import os
from torch.utils.data import Dataset
from PIL import Image
from utils.data_utils import make_dataset
class ImagesDataset(Dataset):
def __init__(self, source_root, source_transform=None):
self.source_paths = sorted(make_dataset(source_root))
self.source_transform = source_transform
def __len__(self):
return len(self.source_paths)
def __getitem__(self, index):
fname, from_path = self.source_paths[index]
from_im = Image.open(from_path).convert('RGB')
if self.source_transform:
from_im = self.source_transform(from_im)
return fname, from_im
stylegan_human/utils/data_utils.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
import os
from PIL import Image
IMG_EXTENSIONS = [
'.jpg', '.JPG', '.jpeg', '.JPEG',
'.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tiff'
]
def is_image_file(filename):
return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
def tensor2im(var):
# var shape: (3, H, W)
var = var.cpu().detach().transpose(0, 2).transpose(0, 1).numpy()
var = ((var + 1) / 2)
var[var < 0] = 0
var[var > 1] = 1
var = var * 255
return Image.fromarray(var.astype('uint8'))
def make_dataset(dir):
images = []
assert os.path.isdir(dir), '%s is not a valid directory' % dir
for root, _, fnames in sorted(os.walk(dir)):
for fname in fnames:
if is_image_file(fname):
path = os.path.join(root, fname)
fname = fname.split('.')[0]
images.append((fname, path))
return images
stylegan_human/utils/face_alignment.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
import numpy as np
import PIL
import PIL.Image
import scipy
import scipy.ndimage
import dlib
import copy
from PIL import Image
def get_landmark(img, detector, predictor):
"""get landmark with dlib
:return: np.array shape=(68, 2)
"""
# detector = dlib.get_frontal_face_detector()
# dets, _, _ = detector.run(img, 1, -1)
dets = detector(img, 1)
for k, d in enumerate(dets):
shape = predictor(img, d.rect)
t = list(shape.parts())
a = []
for tt in t:
a.append([tt.x, tt.y])
lm = np.array(a)
# face rect
face_rect = [dets[0].rect.left(), dets[0].rect.top(), dets[0].rect.right(), dets[0].rect.bottom()]
return lm, face_rect
def align_face_for_insetgan(img, detector, predictor, output_size=256):
"""
:param img: numpy array rgb
:return: PIL Image
"""
img_cp = copy.deepcopy(img)
lm, face_rect = get_landmark(img, detector, predictor)
lm_chin = lm[0: 17] # left-right
lm_eyebrow_left = lm[17: 22] # left-right
lm_eyebrow_right = lm[22: 27] # left-right
lm_nose = lm[27: 31] # top-down
lm_nostrils = lm[31: 36] # top-down
lm_eye_left = lm[36: 42] # left-clockwise
lm_eye_right = lm[42: 48] # left-clockwise
lm_mouth_outer = lm[48: 60] # left-clockwise
lm_mouth_inner = lm[60: 68] # left-clockwise
# Calculate auxiliary vectors.
eye_left = np.mean(lm_eye_left, axis=0)
eye_right = np.mean(lm_eye_right, axis=0)
eye_avg = (eye_left + eye_right) * 0.5
eye_to_eye = eye_right - eye_left
mouth_left = lm_mouth_outer[0]
mouth_right = lm_mouth_outer[6]
mouth_avg = (mouth_left + mouth_right) * 0.5
eye_to_mouth = mouth_avg - eye_avg
# Choose oriented crop rectangle.
x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
x /= np.hypot(*x)
x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
y = np.flipud(x) * [-1, 1]
c = eye_avg + eye_to_mouth * 0.1
quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
qsize = np.hypot(*x) * 2
# read image
# opencv to PIL
img = PIL.Image.fromarray(img_cp)
# img = PIL.Image.open(filepath)
transform_size = output_size
enable_padding = False
# Shrink.
# shrink = int(np.floor(qsize / output_size * 0.5))
# if shrink > 1:
# rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
# img = img.resize(rsize, PIL.Image.ANTIALIAS)
# quad /= shrink
# qsize /= shrink
# Crop.
border = max(int(np.rint(qsize * 0.1)), 3)
crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
int(np.ceil(max(quad[:, 1]))))
# crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
# min(crop[3] + border, img.size[1]))
# img.save("debug/raw.jpg")
if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
img = img.crop(crop)
quad -= crop[0:2]
# img.save("debug/crop.jpg")
# Pad.
# pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
# int(np.ceil(max(quad[:, 1]))))
# pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
# max(pad[3] - img.size[1] + border, 0))
# if enable_padding and max(pad) > border - 4:
# pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
# img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
# h, w, _ = img.shape
# y, x, _ = np.ogrid[:h, :w, :1]
# mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
# 1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
# blur = qsize * 0.02
# img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
# img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
# img = PIL.Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
# quad += pad[:2]
# Transform.
# crop shape to transform shape
# nw =
# print(img.size, quad+0.5, np.bound((quad+0.5).flatten()))
# assert False
# img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(), PIL.Image.BILINEAR)
# img.save("debug/transform.jpg")
# if output_size < transform_size:
img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
# img.save("debug/resize.jpg")
# print((quad+crop[0:2]).flatten())
# assert False
# Return aligned image.
return img, crop, face_rect
def align_face_for_projector(img, detector, predictor, output_size):
"""
:param filepath: str
:return: PIL Image
"""
img_cp = copy.deepcopy(img)
lm, face_rect = get_landmark(img, detector, predictor)
lm_chin = lm[0: 17] # left-right
lm_eyebrow_left = lm[17: 22] # left-right
lm_eyebrow_right = lm[22: 27] # left-right
lm_nose = lm[27: 31] # top-down
lm_nostrils = lm[31: 36] # top-down
lm_eye_left = lm[36: 42] # left-clockwise
lm_eye_right = lm[42: 48] # left-clockwise
lm_mouth_outer = lm[48: 60] # left-clockwise
lm_mouth_inner = lm[60: 68] # left-clockwise
# Calculate auxiliary vectors.
eye_left = np.mean(lm_eye_left, axis=0)
eye_right = np.mean(lm_eye_right, axis=0)
eye_avg = (eye_left + eye_right) * 0.5
eye_to_eye = eye_right - eye_left
mouth_left = lm_mouth_outer[0]
mouth_right = lm_mouth_outer[6]
mouth_avg = (mouth_left + mouth_right) * 0.5
eye_to_mouth = mouth_avg - eye_avg
# Choose oriented crop rectangle.
x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
x /= np.hypot(*x)
x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
y = np.flipud(x) * [-1, 1]
c = eye_avg + eye_to_mouth * 0.1
quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
qsize = np.hypot(*x) * 2
# read image
img = PIL.Image.fromarray(img_cp)
transform_size = output_size
enable_padding = True
# Shrink.
shrink = int(np.floor(qsize / output_size * 0.5))
if shrink > 1:
rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
img = img.resize(rsize, PIL.Image.ANTIALIAS)
quad /= shrink
qsize /= shrink
# Crop.
border = max(int(np.rint(qsize * 0.1)), 3)
crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
int(np.ceil(max(quad[:, 1]))))
crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
min(crop[3] + border, img.size[1]))
if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
img = img.crop(crop)
quad -= crop[0:2]
# Pad.
pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
int(np.ceil(max(quad[:, 1]))))
pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
max(pad[3] - img.size[1] + border, 0))
if enable_padding and max(pad) > border - 4:
pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
h, w, _ = img.shape
y, x, _ = np.ogrid[:h, :w, :1]
mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
blur = qsize * 0.02
img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
img = PIL.Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
quad += pad[:2]
# Transform.
img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(), PIL.Image.BILINEAR)
if output_size < transform_size:
img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
# Return aligned image.
return img
def reverse_quad_transform(image, quad_to_map_to, alpha):
# forward mapping, for simplicity
result = Image.new("RGBA",image.size)
result_pixels = result.load()
width, height = result.size
for y in range(height):
for x in range(width):
result_pixels[x,y] = (0,0,0,0)
p1 = (quad_to_map_to[0],quad_to_map_to[1])
p2 = (quad_to_map_to[2],quad_to_map_to[3])
p3 = (quad_to_map_to[4],quad_to_map_to[5])
p4 = (quad_to_map_to[6],quad_to_map_to[7])
p1_p2_vec = (p2[0] - p1[0],p2[1] - p1[1])
p4_p3_vec = (p3[0] - p4[0],p3[1] - p4[1])
for y in range(height):
for x in range(width):
pixel = image.getpixel((x,y))
y_percentage = y / float(height)
x_percentage = x / float(width)
# interpolate vertically
pa = (p1[0] + p1_p2_vec[0] * y_percentage, p1[1] + p1_p2_vec[1] * y_percentage)
pb = (p4[0] + p4_p3_vec[0] * y_percentage, p4[1] + p4_p3_vec[1] * y_percentage)
pa_to_pb_vec = (pb[0] - pa[0],pb[1] - pa[1])
# interpolate horizontally
p = (pa[0] + pa_to_pb_vec[0] * x_percentage, pa[1] + pa_to_pb_vec[1] * x_percentage)
try:
result_pixels[p[0],p[1]] = (pixel[0],pixel[1],pixel[2],min(int(alpha * 255),pixel[3]))
except Exception:
pass
return result
\ No newline at end of file
stylegan_human/utils/log_utils.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
import numpy as np
from PIL import Image
import wandb
from pti.pti_configs import global_config
import torch
import matplotlib.pyplot as plt
def log_image_from_w(w, G, name):
img = get_image_from_w(w, G)
pillow_image = Image.fromarray(img)
wandb.log(
{f"{name}": [
wandb.Image(pillow_image, caption=f"current inversion {name}")]},
step=global_config.training_step)
def log_images_from_w(ws, G, names):
for name, w in zip(names, ws):
w = w.to(global_config.device)
log_image_from_w(w, G, name)
def plot_image_from_w(w, G):
img = get_image_from_w(w, G)
pillow_image = Image.fromarray(img)
plt.imshow(pillow_image)
plt.show()
def plot_image(img):
img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8).detach().cpu().numpy()
pillow_image = Image.fromarray(img[0])
plt.imshow(pillow_image)
plt.show()
def save_image(name, method_type, results_dir, image, run_id):
image.save(f'{results_dir}/{method_type}_{name}_{run_id}.jpg')
def save_w(w, G, name, method_type, results_dir):
im = get_image_from_w(w, G)
im = Image.fromarray(im, mode='RGB')
save_image(name, method_type, results_dir, im)
def save_concat_image(base_dir, image_latents, new_inv_image_latent, new_G,
old_G,
file_name,
extra_image=None):
images_to_save = []
if extra_image is not None:
images_to_save.append(extra_image)
for latent in image_latents:
images_to_save.append(get_image_from_w(latent, old_G))
images_to_save.append(get_image_from_w(new_inv_image_latent, new_G))
result_image = create_alongside_images(images_to_save)
result_image.save(f'{base_dir}/{file_name}.jpg')
def save_single_image(base_dir, image_latent, G, file_name):
image_to_save = get_image_from_w(image_latent, G)
image_to_save = Image.fromarray(image_to_save, mode='RGB')
image_to_save.save(f'{base_dir}/{file_name}.jpg')
def create_alongside_images(images):
res = np.concatenate([np.array(image) for image in images], axis=1)
return Image.fromarray(res, mode='RGB')
def get_image_from_w(w, G):
if len(w.size()) <= 2:
w = w.unsqueeze(0)
with torch.no_grad():
img = G.synthesis(w, noise_mode='const')
img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8).detach().cpu().numpy()
return img[0]
stylegan_human/utils/models_utils.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
import pickle
import functools
import torch
from pti.pti_configs import paths_config, global_config
def toogle_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
def load_tuned_G(run_id, type):
new_G_path = f'{paths_config.checkpoints_dir}/model_{run_id}_{type}.pt'
with open(new_G_path, 'rb') as f:
new_G = torch.load(f).to(global_config.device).eval()
new_G = new_G.float()
toogle_grad(new_G, False)
return new_G
def load_old_G():
with open(paths_config.stylegan2_ada_shhq, 'rb') as f:
old_G = pickle.load(f)['G_ema'].to(global_config.device).eval()
old_G = old_G.float()
return old_G
stylegan_human/utils/util.py 0 → 100644
View file @ fba8bde8
# Copyright (c) SenseTime Research. All rights reserved.
import torch
import cv2
from torchvision import transforms
import numpy as np
import math
def visual(output, out_path):
output = (output + 1)/2
output = torch.clamp(output, 0, 1)
if output.shape[1] == 1:
output = torch.cat([output, output, output], 1)
output = output[0].detach().cpu().permute(1,2,0).numpy()
output = (output*255).astype(np.uint8)
output = output[:,:,::-1]
cv2.imwrite(out_path, output)
def get_lr(t, initial_lr, rampdown=0.25, rampup=0.05):
lr_ramp = min(1, (1 - t) / rampdown)
lr_ramp = 0.5 - 0.5 * math.cos(lr_ramp * math.pi)
lr_ramp = lr_ramp * min(1, t / rampup)
return initial_lr * lr_ramp
def latent_noise(latent, strength):
noise = torch.randn_like(latent) * strength
return latent + noise
def noise_regularize_(noises):
loss = 0
for noise in noises:
size = noise.shape[2]
while True:
loss = (
loss
+ (noise * torch.roll(noise, shifts=1, dims=3)).mean().pow(2)
+ (noise * torch.roll(noise, shifts=1, dims=2)).mean().pow(2)
)
if size <= 8:
break
noise = noise.reshape([-1, 1, size // 2, 2, size // 2, 2])
noise = noise.mean([3, 5])
size //= 2
return loss
def noise_normalize_(noises):
for noise in noises:
mean = noise.mean()
std = noise.std()
noise.data.add_(-mean).div_(std)
def tensor_to_numpy(x):
x = x[0].permute(1, 2, 0)
x = torch.clamp(x, -1 ,1)
x = (x+1) * 127.5
x = x.cpu().detach().numpy().astype(np.uint8)
return x
def numpy_to_tensor(x):
x = (x / 255 - 0.5) * 2
x = torch.from_numpy(x).unsqueeze(0).permute(0, 3, 1, 2)
x = x.cuda().float()
return x
def tensor_to_pil(x):
x = torch.clamp(x, -1 ,1)
x = (x+1) * 127.5
return transforms.ToPILImage()(x.squeeze_(0))
torch_utils/__init__.py 0 → 100644
View file @ fba8bde8
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. 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.
# empty
torch_utils/custom_ops.py 0 → 100644
View file @ fba8bde8
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. 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 glob
import hashlib
import importlib
import os
import re
import shutil
import uuid
import torch
import torch.utils.cpp_extension
from torch.utils.file_baton import FileBaton
#----------------------------------------------------------------------------
# Global options.
verbosity = 'brief' # Verbosity level: 'none', 'brief', 'full'
#----------------------------------------------------------------------------
# Internal helper funcs.
def _find_compiler_bindir():
patterns = [
'C:/Program Files*/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64',
'C:/Program Files*/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64',
'C:/Program Files*/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64',
'C:/Program Files*/Microsoft Visual Studio */vc/bin',
]
for pattern in patterns:
matches = sorted(glob.glob(pattern))
if len(matches):
return matches[-1]
return None
#----------------------------------------------------------------------------
def _get_mangled_gpu_name():
name = torch.cuda.get_device_name().lower()
out = []
for c in name:
if re.match('[a-z0-9_-]+', c):
out.append(c)
else:
out.append('-')
return ''.join(out)
#----------------------------------------------------------------------------
# Main entry point for compiling and loading C++/CUDA plugins.
_cached_plugins = dict()
def get_plugin(module_name, sources, headers=None, source_dir=None, **build_kwargs):
assert verbosity in ['none', 'brief', 'full']
if headers is None:
headers = []
if source_dir is not None:
sources = [os.path.join(source_dir, fname) for fname in sources]
headers = [os.path.join(source_dir, fname) for fname in headers]
# Already cached?
if module_name in _cached_plugins:
return _cached_plugins[module_name]
# Print status.
if verbosity == 'full':
print(f'Setting up PyTorch plugin "{module_name}"...')
elif verbosity == 'brief':
print(f'Setting up PyTorch plugin "{module_name}"... ', end='', flush=True)
verbose_build = (verbosity == 'full')
# Compile and load.
try: # pylint: disable=too-many-nested-blocks
# Make sure we can find the necessary compiler binaries.
if os.name == 'nt' and os.system("where cl.exe >nul 2>nul") != 0:
compiler_bindir = _find_compiler_bindir()
if compiler_bindir is None:
raise RuntimeError(f'Could not find MSVC/GCC/CLANG installation on this computer. Check _find_compiler_bindir() in "{__file__}".')
os.environ['PATH'] += ';' + compiler_bindir
# Some containers set TORCH_CUDA_ARCH_LIST to a list that can either
# break the build or unnecessarily restrict what's available to nvcc.
# Unset it to let nvcc decide based on what's available on the
# machine.
os.environ['TORCH_CUDA_ARCH_LIST'] = ''
# Incremental build md5sum trickery. Copies all the input source files
# into a cached build directory under a combined md5 digest of the input
# source files. Copying is done only if the combined digest has changed.
# This keeps input file timestamps and filenames the same as in previous
# extension builds, allowing for fast incremental rebuilds.
#
# This optimization is done only in case all the source files reside in
# a single directory (just for simplicity) and if the TORCH_EXTENSIONS_DIR
# environment variable is set (we take this as a signal that the user
# actually cares about this.)
#
# EDIT: We now do it regardless of TORCH_EXTENSIOS_DIR, in order to work
# around the *.cu dependency bug in ninja config.
#
all_source_files = sorted(sources + headers)
all_source_dirs = set(os.path.dirname(fname) for fname in all_source_files)
if len(all_source_dirs) == 1: # and ('TORCH_EXTENSIONS_DIR' in os.environ):
# Compute combined hash digest for all source files.
hash_md5 = hashlib.md5()
for src in all_source_files:
with open(src, 'rb') as f:
hash_md5.update(f.read())
# Select cached build directory name.
source_digest = hash_md5.hexdigest()
build_top_dir = torch.utils.cpp_extension._get_build_directory(module_name, verbose=verbose_build) # pylint: disable=protected-access
cached_build_dir = os.path.join(build_top_dir, f'{source_digest}-{_get_mangled_gpu_name()}')
if not os.path.isdir(cached_build_dir):
tmpdir = f'{build_top_dir}/srctmp-{uuid.uuid4().hex}'
os.makedirs(tmpdir)
for src in all_source_files:
shutil.copyfile(src, os.path.join(tmpdir, os.path.basename(src)))
try:
os.replace(tmpdir, cached_build_dir) # atomic
except OSError:
# source directory already exists, delete tmpdir and its contents.
shutil.rmtree(tmpdir)
if not os.path.isdir(cached_build_dir): raise
# Compile.
cached_sources = [os.path.join(cached_build_dir, os.path.basename(fname)) for fname in sources]
torch.utils.cpp_extension.load(name=module_name, build_directory=cached_build_dir,
verbose=verbose_build, sources=cached_sources, **build_kwargs)
else:
torch.utils.cpp_extension.load(name=module_name, verbose=verbose_build, sources=sources, **build_kwargs)
# Load.
module = importlib.import_module(module_name)
except:
if verbosity == 'brief':
print('Failed!')
raise
# Print status and add to cache dict.
if verbosity == 'full':
print(f'Done setting up PyTorch plugin "{module_name}".')
elif verbosity == 'brief':
print('Done.')
_cached_plugins[module_name] = module
return module
#----------------------------------------------------------------------------
torch_utils/misc.py 0 → 100644
View file @ fba8bde8
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. 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 re
import contextlib
import numpy as np
import torch
import warnings
import dnnlib
#----------------------------------------------------------------------------
# Cached construction of constant tensors. Avoids CPU=>GPU copy when the
# same constant is used multiple times.
_constant_cache = dict()
def constant(value, shape=None, dtype=None, device=None, memory_format=None):
value = np.asarray(value)
if shape is not None:
shape = tuple(shape)
if dtype is None:
dtype = torch.get_default_dtype()
if device is None:
device = torch.device('cpu')
if memory_format is None:
memory_format = torch.contiguous_format
key = (value.shape, value.dtype, value.tobytes(), shape, dtype, device, memory_format)
tensor = _constant_cache.get(key, None)
if tensor is None:
tensor = torch.as_tensor(value.copy(), dtype=dtype, device=device)
if shape is not None:
tensor, _ = torch.broadcast_tensors(tensor, torch.empty(shape))
tensor = tensor.contiguous(memory_format=memory_format)
_constant_cache[key] = tensor
return tensor
#----------------------------------------------------------------------------
# Replace NaN/Inf with specified numerical values.
try:
nan_to_num = torch.nan_to_num # 1.8.0a0
except AttributeError:
def nan_to_num(input, nan=0.0, posinf=None, neginf=None, *, out=None): # pylint: disable=redefined-builtin
assert isinstance(input, torch.Tensor)
if posinf is None:
posinf = torch.finfo(input.dtype).max
if neginf is None:
neginf = torch.finfo(input.dtype).min
assert nan == 0
return torch.clamp(input.unsqueeze(0).nansum(0), min=neginf, max=posinf, out=out)
#----------------------------------------------------------------------------
# Symbolic assert.
try:
symbolic_assert = torch._assert # 1.8.0a0 # pylint: disable=protected-access
except AttributeError:
symbolic_assert = torch.Assert # 1.7.0
#----------------------------------------------------------------------------
# Context manager to temporarily suppress known warnings in torch.jit.trace().
# Note: Cannot use catch_warnings because of https://bugs.python.org/issue29672
@contextlib.contextmanager
def suppress_tracer_warnings():
flt = ('ignore', None, torch.jit.TracerWarning, None, 0)
warnings.filters.insert(0, flt)
yield
warnings.filters.remove(flt)
#----------------------------------------------------------------------------
# Assert that the shape of a tensor matches the given list of integers.
# None indicates that the size of a dimension is allowed to vary.
# Performs symbolic assertion when used in torch.jit.trace().
def assert_shape(tensor, ref_shape):
if tensor.ndim != len(ref_shape):
raise AssertionError(f'Wrong number of dimensions: got {tensor.ndim}, expected {len(ref_shape)}')
for idx, (size, ref_size) in enumerate(zip(tensor.shape, ref_shape)):
if ref_size is None:
pass
elif isinstance(ref_size, torch.Tensor):
with suppress_tracer_warnings(): # as_tensor results are registered as constants
symbolic_assert(torch.equal(torch.as_tensor(size), ref_size), f'Wrong size for dimension {idx}')
elif isinstance(size, torch.Tensor):
with suppress_tracer_warnings(): # as_tensor results are registered as constants
symbolic_assert(torch.equal(size, torch.as_tensor(ref_size)), f'Wrong size for dimension {idx}: expected {ref_size}')
elif size != ref_size:
raise AssertionError(f'Wrong size for dimension {idx}: got {size}, expected {ref_size}')
#----------------------------------------------------------------------------
# Function decorator that calls torch.autograd.profiler.record_function().
def profiled_function(fn):
def decorator(*args, **kwargs):
with torch.autograd.profiler.record_function(fn.__name__):
return fn(*args, **kwargs)
decorator.__name__ = fn.__name__
return decorator
#----------------------------------------------------------------------------
# Sampler for torch.utils.data.DataLoader that loops over the dataset
# indefinitely, shuffling items as it goes.
class InfiniteSampler(torch.utils.data.Sampler):
def __init__(self, dataset, rank=0, num_replicas=1, shuffle=True, seed=0, window_size=0.5):
assert len(dataset) > 0
assert num_replicas > 0
assert 0 <= rank < num_replicas
assert 0 <= window_size <= 1
super().__init__(dataset)
self.dataset = dataset
self.rank = rank
self.num_replicas = num_replicas
self.shuffle = shuffle
self.seed = seed
self.window_size = window_size
def __iter__(self):
order = np.arange(len(self.dataset))
rnd = None
window = 0
if self.shuffle:
rnd = np.random.RandomState(self.seed)
rnd.shuffle(order)
window = int(np.rint(order.size * self.window_size))
idx = 0
while True:
i = idx % order.size
if idx % self.num_replicas == self.rank:
yield order[i]
if window >= 2:
j = (i - rnd.randint(window)) % order.size
order[i], order[j] = order[j], order[i]
idx += 1
#----------------------------------------------------------------------------
# Utilities for operating with torch.nn.Module parameters and buffers.
def params_and_buffers(module):
assert isinstance(module, torch.nn.Module)
return list(module.parameters()) + list(module.buffers())
def named_params_and_buffers(module):
assert isinstance(module, torch.nn.Module)
return list(module.named_parameters()) + list(module.named_buffers())
def copy_params_and_buffers(src_module, dst_module, require_all=False):
assert isinstance(src_module, torch.nn.Module)
assert isinstance(dst_module, torch.nn.Module)
src_tensors = dict(named_params_and_buffers(src_module))
for name, tensor in named_params_and_buffers(dst_module):
assert (name in src_tensors) or (not require_all)
if name in src_tensors:
tensor.copy_(src_tensors[name].detach()).requires_grad_(tensor.requires_grad)
#----------------------------------------------------------------------------
# Context manager for easily enabling/disabling DistributedDataParallel
# synchronization.
@contextlib.contextmanager
def ddp_sync(module, sync):
assert isinstance(module, torch.nn.Module)
if sync or not isinstance(module, torch.nn.parallel.DistributedDataParallel):
yield
else:
with module.no_sync():
yield
#----------------------------------------------------------------------------
# Check DistributedDataParallel consistency across processes.
def check_ddp_consistency(module, ignore_regex=None):
assert isinstance(module, torch.nn.Module)
for name, tensor in named_params_and_buffers(module):
fullname = type(module).__name__ + '.' + name
if ignore_regex is not None and re.fullmatch(ignore_regex, fullname):
continue
tensor = tensor.detach()
if tensor.is_floating_point():
tensor = nan_to_num(tensor)
other = tensor.clone()
torch.distributed.broadcast(tensor=other, src=0)
assert (tensor == other).all(), fullname
#----------------------------------------------------------------------------
# Print summary table of module hierarchy.
def print_module_summary(module, inputs, max_nesting=3, skip_redundant=True):
assert isinstance(module, torch.nn.Module)
assert not isinstance(module, torch.jit.ScriptModule)
assert isinstance(inputs, (tuple, list))
# Register hooks.
entries = []
nesting = [0]
def pre_hook(_mod, _inputs):
nesting[0] += 1
def post_hook(mod, _inputs, outputs):
nesting[0] -= 1
if nesting[0] <= max_nesting:
outputs = list(outputs) if isinstance(outputs, (tuple, list)) else [outputs]
outputs = [t for t in outputs if isinstance(t, torch.Tensor)]
entries.append(dnnlib.EasyDict(mod=mod, outputs=outputs))
hooks = [mod.register_forward_pre_hook(pre_hook) for mod in module.modules()]
hooks += [mod.register_forward_hook(post_hook) for mod in module.modules()]
# Run module.
outputs = module(*inputs)
for hook in hooks:
hook.remove()
# Identify unique outputs, parameters, and buffers.
tensors_seen = set()
for e in entries:
e.unique_params = [t for t in e.mod.parameters() if id(t) not in tensors_seen]
e.unique_buffers = [t for t in e.mod.buffers() if id(t) not in tensors_seen]
e.unique_outputs = [t for t in e.outputs if id(t) not in tensors_seen]
tensors_seen |= {id(t) for t in e.unique_params + e.unique_buffers + e.unique_outputs}
# Filter out redundant entries.
if skip_redundant:
entries = [e for e in entries if len(e.unique_params) or len(e.unique_buffers) or len(e.unique_outputs)]
# Construct table.
rows = [[type(module).__name__, 'Parameters', 'Buffers', 'Output shape', 'Datatype']]
rows += [['---'] * len(rows[0])]
param_total = 0
buffer_total = 0
submodule_names = {mod: name for name, mod in module.named_modules()}
for e in entries:
name = '<top-level>' if e.mod is module else submodule_names[e.mod]
param_size = sum(t.numel() for t in e.unique_params)
buffer_size = sum(t.numel() for t in e.unique_buffers)
output_shapes = [str(list(t.shape)) for t in e.outputs]
output_dtypes = [str(t.dtype).split('.')[-1] for t in e.outputs]
rows += [[
name + (':0' if len(e.outputs) >= 2 else ''),
str(param_size) if param_size else '-',
str(buffer_size) if buffer_size else '-',
(output_shapes + ['-'])[0],
(output_dtypes + ['-'])[0],
]]
for idx in range(1, len(e.outputs)):
rows += [[name + f':{idx}', '-', '-', output_shapes[idx], output_dtypes[idx]]]
param_total += param_size
buffer_total += buffer_size
rows += [['---'] * len(rows[0])]
rows += [['Total', str(param_total), str(buffer_total), '-', '-']]
# Print table.
widths = [max(len(cell) for cell in column) for column in zip(*rows)]
print()
for row in rows:
print(' '.join(cell + ' ' * (width - len(cell)) for cell, width in zip(row, widths)))
print()
return outputs
#----------------------------------------------------------------------------
torch_utils/ops/__init__.py 0 → 100644
View file @ fba8bde8
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. 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.
# empty
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment