init

.gitattributes

@@ -29,3 +29,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+checkpoints/stylefacev/latest_net_FE.pth filter=lfs diff=lfs merge=lfs -text
+checkpoints/stylefacev/latest_net_FE_lm.pth filter=lfs diff=lfs merge=lfs -text
+checkpoints/stylefacev/latest_net_FE_pose.pth filter=lfs diff=lfs merge=lfs -text
+pretrained_models/network-snapshot-005000.pkl filter=lfs diff=lfs merge=lfs -text
+pretrained_models/wing.ckpt filter=lfs diff=lfs merge=lfs -text
+pretrained_models/motion_net.pth filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,8 +1,8 @@
 ---
 title: StyleFaceV
-emoji: 👁
-colorFrom: green
-colorTo: red
+emoji: 🏢
+colorFrom: indigo
+colorTo: purple
 sdk: gradio
 sdk_version: 3.1.7
 app_file: app.py

app.py

@@ -0,0 +1,8 @@
+import gradio as gr
+
+def greet(name):
+    return "Hello " + name + "!!"
+
+iface = gr.Interface(fn=greet, inputs="text", outputs="text")
+iface.launch()
+

checkpoints/stylefacev/latest_net_FE.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52d5f2eb71cb79fce9faa1448450033d83037a427f1c69cd3924fcc9771fe1ef
+size 559223985

checkpoints/stylefacev/latest_net_FE_lm.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34827db294ffd611971460fead1445844144311543e2a35fb2ccd1b52ae8d07c
+size 25497505

checkpoints/stylefacev/latest_net_FE_pose.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27a838efca2be63aa71a1b6da020998ac466aaeee2772c9e2631ba65f993174b
+size 6447709

data/__init__.py

@@ -0,0 +1,93 @@
+"""This package includes all the modules related to data loading and preprocessing
+
+ To add a custom dataset class called 'dummy', you need to add a file called 'dummy_dataset.py' and define a subclass 'DummyDataset' inherited from BaseDataset.
+ You need to implement four functions:
+    -- <__init__>:                      initialize the class, first call BaseDataset.__init__(self, opt).
+    -- <__len__>:                       return the size of dataset.
+    -- <__getitem__>:                   get a data point from data loader.
+    -- <modify_commandline_options>:    (optionally) add dataset-specific options and set default options.
+
+Now you can use the dataset class by specifying flag '--dataset_mode dummy'.
+See our template dataset class 'template_dataset.py' for more details.
+"""
+import importlib
+import torch.utils.data
+from data.base_dataset import BaseDataset
+
+
+def find_dataset_using_name(dataset_name):
+    """Import the module "data/[dataset_name]_dataset.py".
+
+    In the file, the class called DatasetNameDataset() will
+    be instantiated. It has to be a subclass of BaseDataset,
+    and it is case-insensitive.
+    """
+    dataset_filename = "data." + dataset_name + "_dataset"
+    datasetlib = importlib.import_module(dataset_filename)
+
+    dataset = None
+    target_dataset_name = dataset_name.replace('_', '') + 'dataset'
+    for name, cls in datasetlib.__dict__.items():
+        if name.lower() == target_dataset_name.lower() \
+           and issubclass(cls, BaseDataset):
+            dataset = cls
+
+    if dataset is None:
+        raise NotImplementedError("In %s.py, there should be a subclass of BaseDataset with class name that matches %s in lowercase." % (dataset_filename, target_dataset_name))
+
+    return dataset
+
+
+def get_option_setter(dataset_name):
+    """Return the static method <modify_commandline_options> of the dataset class."""
+    dataset_class = find_dataset_using_name(dataset_name)
+    return dataset_class.modify_commandline_options
+
+
+def create_dataset(opt):
+    """Create a dataset given the option.
+
+    This function wraps the class CustomDatasetDataLoader.
+        This is the main interface between this package and 'train.py'/'test.py'
+
+    Example:
+        >>> from data import create_dataset
+        >>> dataset = create_dataset(opt)
+    """
+    data_loader = CustomDatasetDataLoader(opt)
+    dataset = data_loader.load_data()
+    return dataset
+
+
+class CustomDatasetDataLoader():
+    """Wrapper class of Dataset class that performs multi-threaded data loading"""
+
+    def __init__(self, opt):
+        """Initialize this class
+
+        Step 1: create a dataset instance given the name [dataset_mode]
+        Step 2: create a multi-threaded data loader.
+        """
+        self.opt = opt
+        dataset_class = find_dataset_using_name(opt.dataset_mode)
+        self.dataset = dataset_class(opt)
+        print("dataset [%s] was created" % type(self.dataset).__name__)
+        self.dataloader = torch.utils.data.DataLoader(
+            self.dataset,
+            batch_size=opt.batch_size,
+            shuffle=not opt.serial_batches,
+            num_workers=int(opt.num_threads))
+
+    def load_data(self):
+        return self
+
+    def __len__(self):
+        """Return the number of data in the dataset"""
+        return min(len(self.dataset), self.opt.max_dataset_size)
+
+    def __iter__(self):
+        """Return a batch of data"""
+        for i, data in enumerate(self.dataloader):
+            if i * self.opt.batch_size >= self.opt.max_dataset_size:
+                break
+            yield data

data/base_dataset.py

@@ -0,0 +1,157 @@
+"""This module implements an abstract base class (ABC) 'BaseDataset' for datasets.
+
+It also includes common transformation functions (e.g., get_transform, __scale_width), which can be later used in subclasses.
+"""
+import random
+import numpy as np
+import torch.utils.data as data
+from PIL import Image
+import torchvision.transforms as transforms
+from abc import ABC, abstractmethod
+
+
+class BaseDataset(data.Dataset, ABC):
+    """This class is an abstract base class (ABC) for datasets.
+
+    To create a subclass, you need to implement the following four functions:
+    -- <__init__>:                      initialize the class, first call BaseDataset.__init__(self, opt).
+    -- <__len__>:                       return the size of dataset.
+    -- <__getitem__>:                   get a data point.
+    -- <modify_commandline_options>:    (optionally) add dataset-specific options and set default options.
+    """
+
+    def __init__(self, opt):
+        """Initialize the class; save the options in the class
+
+        Parameters:
+            opt (Option class)-- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        self.opt = opt
+        self.root = opt.dataroot
+
+    @staticmethod
+    def modify_commandline_options(parser, is_train):
+        """Add new dataset-specific options, and rewrite default values for existing options.
+
+        Parameters:
+            parser          -- original option parser
+            is_train (bool) -- whether training phase or test phase. You can use this flag to add training-specific or test-specific options.
+
+        Returns:
+            the modified parser.
+        """
+        return parser
+
+    @abstractmethod
+    def __len__(self):
+        """Return the total number of images in the dataset."""
+        return 0
+
+    @abstractmethod
+    def __getitem__(self, index):
+        """Return a data point and its metadata information.
+
+        Parameters:
+            index - - a random integer for data indexing
+
+        Returns:
+            a dictionary of data with their names. It ususally contains the data itself and its metadata information.
+        """
+        pass
+
+
+def get_params(opt, size):
+    w, h = size
+    new_h = h
+    new_w = w
+    if opt.preprocess == 'resize_and_crop':
+        new_h = new_w = opt.load_size
+    elif opt.preprocess == 'scale_width_and_crop':
+        new_w = opt.load_size
+        new_h = opt.load_size * h // w
+
+    x = random.randint(0, np.maximum(0, new_w - opt.crop_size))
+    y = random.randint(0, np.maximum(0, new_h - opt.crop_size))
+
+    flip = random.random() > 0.5
+
+    return {'crop_pos': (x, y), 'flip': flip}
+
+
+def get_transform(opt, params=None, grayscale=False, method=Image.BICUBIC, convert=True):
+    transform_list = []
+    if grayscale:
+        transform_list.append(transforms.Grayscale(1))
+    if 'resize' in opt.preprocess:
+        osize = [opt.load_size, opt.load_size]
+        transform_list.append(transforms.Resize(osize, method))
+    elif 'scale_width' in opt.preprocess:
+        transform_list.append(transforms.Lambda(lambda img: __scale_width(img, opt.load_size, opt.crop_size, method)))
+
+    if 'crop' in opt.preprocess:
+        if params is None:
+            transform_list.append(transforms.RandomCrop(opt.crop_size))
+        else:
+            transform_list.append(transforms.Lambda(lambda img: __crop(img, params['crop_pos'], opt.crop_size)))
+
+    if opt.preprocess == 'none':
+        transform_list.append(transforms.Lambda(lambda img: __make_power_2(img, base=4, method=method)))
+
+    if not opt.no_flip:
+        if params is None:
+            transform_list.append(transforms.RandomHorizontalFlip())
+        elif params['flip']:
+            transform_list.append(transforms.Lambda(lambda img: __flip(img, params['flip'])))
+
+    if convert:
+        transform_list += [transforms.ToTensor()]
+        if grayscale:
+            transform_list += [transforms.Normalize((0.5,), (0.5,))]
+        else:
+            transform_list += [transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
+    return transforms.Compose(transform_list)
+
+
+def __make_power_2(img, base, method=Image.BICUBIC):
+    ow, oh = img.size
+    h = int(round(oh / base) * base)
+    w = int(round(ow / base) * base)
+    if h == oh and w == ow:
+        return img
+
+    __print_size_warning(ow, oh, w, h)
+    return img.resize((w, h), method)
+
+
+def __scale_width(img, target_size, crop_size, method=Image.BICUBIC):
+    ow, oh = img.size
+    if ow == target_size and oh >= crop_size:
+        return img
+    w = target_size
+    h = int(max(target_size * oh / ow, crop_size))
+    return img.resize((w, h), method)
+
+
+def __crop(img, pos, size):
+    ow, oh = img.size
+    x1, y1 = pos
+    tw = th = size
+    if (ow > tw or oh > th):
+        return img.crop((x1, y1, x1 + tw, y1 + th))
+    return img
+
+
+def __flip(img, flip):
+    if flip:
+        return img.transpose(Image.FLIP_LEFT_RIGHT)
+    return img
+
+
+def __print_size_warning(ow, oh, w, h):
+    """Print warning information about image size(only print once)"""
+    if not hasattr(__print_size_warning, 'has_printed'):
+        print("The image size needs to be a multiple of 4. "
+              "The loaded image size was (%d, %d), so it was adjusted to "
+              "(%d, %d). This adjustment will be done to all images "
+              "whose sizes are not multiples of 4" % (ow, oh, w, h))
+        __print_size_warning.has_printed = True

data/image_folder.py

@@ -0,0 +1,90 @@
+"""A modified image folder class
+
+We modify the official PyTorch image folder (https://github.com/pytorch/vision/blob/master/torchvision/datasets/folder.py)
+so that this class can load images from both current directory and its subdirectories.
+"""
+
+import torch.utils.data as data
+
+from PIL import Image
+import os
+
+IMG_EXTENSIONS = [
+    '.jpg', '.JPG', '.jpeg', '.JPEG',
+    '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP',
+    '.tif', '.TIF', '.tiff', '.TIFF',
+]
+
+
+def is_image_file(filename):
+    return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
+
+
+def make_id_dataset(dir, max_dataset_size=float("inf")):
+    ids = []
+    images = []
+    assert os.path.isdir(dir), '%s is not a valid directory' % dir
+
+    id_names = sorted(os.listdir(dir))
+    for id_name in id_names:
+        id_path = os.path.join(dir, id_name)
+        fnames = os.listdir(id_path)
+        for fname in fnames:
+            path = os.path.join(dir, id_name, fname)
+            images.append(path)
+            ids.append(id_name)
+    return images[:min(max_dataset_size, len(images))], ids[:min(max_dataset_size, len(ids))]
+
+def make_noid_dataset(dir, max_dataset_size=float("inf")):
+    images = []
+    assert os.path.isdir(dir), '%s is not a valid directory' % dir
+
+    fnames = sorted(os.listdir(dir))
+    for fname in fnames:
+        path = os.path.join(dir, fname)
+        images.append(path)
+    return images[:min(max_dataset_size, len(images))]
+
+def make_dataset(dir, max_dataset_size=float("inf")):
+    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)
+                images.append(path)
+    return images[:min(max_dataset_size, len(images))]
+
+
+def default_loader(path):
+    return Image.open(path).convert('RGB')
+
+
+class ImageFolder(data.Dataset):
+
+    def __init__(self, root, transform=None, return_paths=False,
+                 loader=default_loader):
+        imgs = make_dataset(root)
+        if len(imgs) == 0:
+            raise(RuntimeError("Found 0 images in: " + root + "\n"
+                               "Supported image extensions are: " + ",".join(IMG_EXTENSIONS)))
+
+        self.root = root
+        self.imgs = imgs
+        self.transform = transform
+        self.return_paths = return_paths
+        self.loader = loader
+
+    def __getitem__(self, index):
+        path = self.imgs[index]
+        img = self.loader(path)
+        if self.transform is not None:
+            img = self.transform(img)
+        if self.return_paths:
+            return img, path
+        else:
+            return img
+
+    def __len__(self):
+        return len(self.imgs)

data/noiseshufflevideo_dataset.py

@@ -0,0 +1,71 @@
+from data.base_dataset import BaseDataset, get_transform, get_params
+from data.image_folder import make_id_dataset
+from PIL import Image
+import random
+import os
+import numpy as np
+import torch
+
+
+class NoiseShuffleVideoDataset(BaseDataset):
+    """This dataset class can load a set of images specified by the path --dataroot /path/to/data.
+
+    It can be used for generating CycleGAN results only for one side with the model option '-model test'.
+    """
+
+    def __init__(self, opt):
+        """Initialize this dataset class.
+
+        Parameters:
+            opt (Option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        BaseDataset.__init__(self, opt)
+        self.opt = opt
+        self.A_paths, self.A_ids = make_id_dataset(opt.dataroot, opt.max_dataset_size)
+
+        self.input_nc = self.opt.output_nc if self.opt.direction == 'BtoA' else self.opt.input_nc
+
+    def __getitem__(self, index):
+        """Return a data point and its metadata information.
+
+        Parameters:
+            index - - a random integer for data indexing
+
+        Returns a dictionary that contains A and A_paths
+            A(tensor) - - an image in one domain
+            A_paths(str) - - the path of the image
+        """
+        # A_id = self.A_ids[index]
+        A_list = []
+        random.seed(index)
+        A_index = int(random.random() * (len(self.A_paths) - 1))
+        A_video = self.A_paths[A_index]
+        A_frames = sorted(os.listdir(A_video))
+        max_frames = len(A_frames)
+        while max_frames < 60:
+            A_index = (A_index + 1) % len(self.A_paths)
+            A_video = self.A_paths[A_index]
+            A_frames = sorted(os.listdir(A_video))
+            max_frames = len(A_frames)
+
+        for i in range(max_frames):
+            A_frame = A_frames[i]
+            # print(A_frame)
+            A_path = os.path.join(A_video, A_frame)
+            A_img = Image.open(A_path).convert('RGB')
+
+            if i == 0:
+                transform_params = get_params(self.opt, A_img.size)
+                self.transform = get_transform(self.opt, transform_params, grayscale=(self.input_nc == 1))
+
+            A = self.transform(A_img)
+            A_list.append(A.unsqueeze(0))
+
+        A = torch.cat(A_list, 0)
+        B = torch.from_numpy(np.random.RandomState(index).randn(512))
+
+        return {'A': A, 'A_paths': A_path, 'B': B}
+
+    def __len__(self):
+        """Return the total number of images in the dataset."""
+        return len(self.A_paths)

dnnlib/__init__.py

@@ -0,0 +1,9 @@
+# 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.
+
+from .util import EasyDict, make_cache_dir_path

dnnlib/util.py

@@ -0,0 +1,491 @@
+# 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.
+
+"""Miscellaneous utility classes and functions."""
+
+import ctypes
+import fnmatch
+import importlib
+import inspect
+import numpy as np
+import os
+import shutil
+import sys
+import types
+import io
+import pickle
+import re
+import requests
+import html
+import hashlib
+import glob
+import tempfile
+import urllib
+import urllib.request
+import uuid
+
+from distutils.util import strtobool
+from typing import Any, List, Tuple, Union
+
+
+# Util classes
+# ------------------------------------------------------------------------------------------
+
+
+class EasyDict(dict):
+    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self[name]
+        except KeyError:
+            raise AttributeError(name)
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        self[name] = value
+
+    def __delattr__(self, name: str) -> None:
+        del self[name]
+
+
+class Logger(object):
+    """Redirect stderr to stdout, optionally print stdout to a file, and optionally force flushing on both stdout and the file."""
+
+    def __init__(self, file_name: str = None, file_mode: str = "w", should_flush: bool = True):
+        self.file = None
+
+        if file_name is not None:
+            self.file = open(file_name, file_mode)
+
+        self.should_flush = should_flush
+        self.stdout = sys.stdout
+        self.stderr = sys.stderr
+
+        sys.stdout = self
+        sys.stderr = self
+
+    def __enter__(self) -> "Logger":
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.close()
+
+    def write(self, text: Union[str, bytes]) -> None:
+        """Write text to stdout (and a file) and optionally flush."""
+        if isinstance(text, bytes):
+            text = text.decode()
+        if len(text) == 0: # workaround for a bug in VSCode debugger: sys.stdout.write(''); sys.stdout.flush() => crash
+            return
+
+        if self.file is not None:
+            self.file.write(text)
+
+        self.stdout.write(text)
+
+        if self.should_flush:
+            self.flush()
+
+    def flush(self) -> None:
+        """Flush written text to both stdout and a file, if open."""
+        if self.file is not None:
+            self.file.flush()
+
+        self.stdout.flush()
+
+    def close(self) -> None:
+        """Flush, close possible files, and remove stdout/stderr mirroring."""
+        self.flush()
+
+        # if using multiple loggers, prevent closing in wrong order
+        if sys.stdout is self:
+            sys.stdout = self.stdout
+        if sys.stderr is self:
+            sys.stderr = self.stderr
+
+        if self.file is not None:
+            self.file.close()
+            self.file = None
+
+
+# Cache directories
+# ------------------------------------------------------------------------------------------
+
+_dnnlib_cache_dir = None
+
+def set_cache_dir(path: str) -> None:
+    global _dnnlib_cache_dir
+    _dnnlib_cache_dir = path
+
+def make_cache_dir_path(*paths: str) -> str:
+    if _dnnlib_cache_dir is not None:
+        return os.path.join(_dnnlib_cache_dir, *paths)
+    if 'DNNLIB_CACHE_DIR' in os.environ:
+        return os.path.join(os.environ['DNNLIB_CACHE_DIR'], *paths)
+    if 'HOME' in os.environ:
+        return os.path.join(os.environ['HOME'], '.cache', 'dnnlib', *paths)
+    if 'USERPROFILE' in os.environ:
+        return os.path.join(os.environ['USERPROFILE'], '.cache', 'dnnlib', *paths)
+    return os.path.join(tempfile.gettempdir(), '.cache', 'dnnlib', *paths)
+
+# Small util functions
+# ------------------------------------------------------------------------------------------
+
+
+def format_time(seconds: Union[int, float]) -> str:
+    """Convert the seconds to human readable string with days, hours, minutes and seconds."""
+    s = int(np.rint(seconds))
+
+    if s < 60:
+        return "{0}s".format(s)
+    elif s < 60 * 60:
+        return "{0}m {1:02}s".format(s // 60, s % 60)
+    elif s < 24 * 60 * 60:
+        return "{0}h {1:02}m {2:02}s".format(s // (60 * 60), (s // 60) % 60, s % 60)
+    else:
+        return "{0}d {1:02}h {2:02}m".format(s // (24 * 60 * 60), (s // (60 * 60)) % 24, (s // 60) % 60)
+
+
+def format_time_brief(seconds: Union[int, float]) -> str:
+    """Convert the seconds to human readable string with days, hours, minutes and seconds."""
+    s = int(np.rint(seconds))
+
+    if s < 60:
+        return "{0}s".format(s)
+    elif s < 60 * 60:
+        return "{0}m {1:02}s".format(s // 60, s % 60)
+    elif s < 24 * 60 * 60:
+        return "{0}h {1:02}m".format(s // (60 * 60), (s // 60) % 60)
+    else:
+        return "{0}d {1:02}h".format(s // (24 * 60 * 60), (s // (60 * 60)) % 24)
+
+
+def ask_yes_no(question: str) -> bool:
+    """Ask the user the question until the user inputs a valid answer."""
+    while True:
+        try:
+            print("{0} [y/n]".format(question))
+            return strtobool(input().lower())
+        except ValueError:
+            pass
+
+
+def tuple_product(t: Tuple) -> Any:
+    """Calculate the product of the tuple elements."""
+    result = 1
+
+    for v in t:
+        result *= v
+
+    return result
+
+
+_str_to_ctype = {
+    "uint8": ctypes.c_ubyte,
+    "uint16": ctypes.c_uint16,
+    "uint32": ctypes.c_uint32,
+    "uint64": ctypes.c_uint64,
+    "int8": ctypes.c_byte,
+    "int16": ctypes.c_int16,
+    "int32": ctypes.c_int32,
+    "int64": ctypes.c_int64,
+    "float32": ctypes.c_float,
+    "float64": ctypes.c_double
+}
+
+
+def get_dtype_and_ctype(type_obj: Any) -> Tuple[np.dtype, Any]:
+    """Given a type name string (or an object having a __name__ attribute), return matching Numpy and ctypes types that have the same size in bytes."""
+    type_str = None
+
+    if isinstance(type_obj, str):
+        type_str = type_obj
+    elif hasattr(type_obj, "__name__"):
+        type_str = type_obj.__name__
+    elif hasattr(type_obj, "name"):
+        type_str = type_obj.name
+    else:
+        raise RuntimeError("Cannot infer type name from input")
+
+    assert type_str in _str_to_ctype.keys()
+
+    my_dtype = np.dtype(type_str)
+    my_ctype = _str_to_ctype[type_str]
+
+    assert my_dtype.itemsize == ctypes.sizeof(my_ctype)
+
+    return my_dtype, my_ctype
+
+
+def is_pickleable(obj: Any) -> bool:
+    try:
+        with io.BytesIO() as stream:
+            pickle.dump(obj, stream)
+        return True
+    except:
+        return False
+
+
+# Functionality to import modules/objects by name, and call functions by name
+# ------------------------------------------------------------------------------------------
+
+def get_module_from_obj_name(obj_name: str) -> Tuple[types.ModuleType, str]:
+    """Searches for the underlying module behind the name to some python object.
+    Returns the module and the object name (original name with module part removed)."""
+
+    # allow convenience shorthands, substitute them by full names
+    obj_name = re.sub("^np.", "numpy.", obj_name)
+    obj_name = re.sub("^tf.", "tensorflow.", obj_name)
+
+    # list alternatives for (module_name, local_obj_name)
+    parts = obj_name.split(".")
+    name_pairs = [(".".join(parts[:i]), ".".join(parts[i:])) for i in range(len(parts), 0, -1)]
+
+    # try each alternative in turn
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+            return module, local_obj_name
+        except:
+            pass
+
+    # maybe some of the modules themselves contain errors?
+    for module_name, _local_obj_name in name_pairs:
+        try:
+            importlib.import_module(module_name) # may raise ImportError
+        except ImportError:
+            if not str(sys.exc_info()[1]).startswith("No module named '" + module_name + "'"):
+                raise
+
+    # maybe the requested attribute is missing?
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+        except ImportError:
+            pass
+
+    # we are out of luck, but we have no idea why
+    raise ImportError(obj_name)
+
+
+def get_obj_from_module(module: types.ModuleType, obj_name: str) -> Any:
+    """Traverses the object name and returns the last (rightmost) python object."""
+    if obj_name == '':
+        return module
+    obj = module
+    for part in obj_name.split("."):
+        obj = getattr(obj, part)
+    return obj
+
+
+def get_obj_by_name(name: str) -> Any:
+    """Finds the python object with the given name."""
+    module, obj_name = get_module_from_obj_name(name)
+    return get_obj_from_module(module, obj_name)
+
+
+def call_func_by_name(*args, func_name: str = None, **kwargs) -> Any:
+    """Finds the python object with the given name and calls it as a function."""
+    assert func_name is not None
+    func_obj = get_obj_by_name(func_name)
+    assert callable(func_obj)
+    return func_obj(*args, **kwargs)
+
+
+def construct_class_by_name(*args, class_name: str = None, **kwargs) -> Any:
+    """Finds the python class with the given name and constructs it with the given arguments."""
+    return call_func_by_name(*args, func_name=class_name, **kwargs)
+
+
+def get_module_dir_by_obj_name(obj_name: str) -> str:
+    """Get the directory path of the module containing the given object name."""
+    module, _ = get_module_from_obj_name(obj_name)
+    return os.path.dirname(inspect.getfile(module))
+
+
+def is_top_level_function(obj: Any) -> bool:
+    """Determine whether the given object is a top-level function, i.e., defined at module scope using 'def'."""
+    return callable(obj) and obj.__name__ in sys.modules[obj.__module__].__dict__
+
+
+def get_top_level_function_name(obj: Any) -> str:
+    """Return the fully-qualified name of a top-level function."""
+    assert is_top_level_function(obj)
+    module = obj.__module__
+    if module == '__main__':
+        module = os.path.splitext(os.path.basename(sys.modules[module].__file__))[0]
+    return module + "." + obj.__name__
+
+
+# File system helpers
+# ------------------------------------------------------------------------------------------
+
+def list_dir_recursively_with_ignore(dir_path: str, ignores: List[str] = None, add_base_to_relative: bool = False) -> List[Tuple[str, str]]:
+    """List all files recursively in a given directory while ignoring given file and directory names.
+    Returns list of tuples containing both absolute and relative paths."""
+    assert os.path.isdir(dir_path)
+    base_name = os.path.basename(os.path.normpath(dir_path))
+
+    if ignores is None:
+        ignores = []
+
+    result = []
+
+    for root, dirs, files in os.walk(dir_path, topdown=True):
+        for ignore_ in ignores:
+            dirs_to_remove = [d for d in dirs if fnmatch.fnmatch(d, ignore_)]
+
+            # dirs need to be edited in-place
+            for d in dirs_to_remove:
+                dirs.remove(d)
+
+            files = [f for f in files if not fnmatch.fnmatch(f, ignore_)]
+
+        absolute_paths = [os.path.join(root, f) for f in files]
+        relative_paths = [os.path.relpath(p, dir_path) for p in absolute_paths]
+
+        if add_base_to_relative:
+            relative_paths = [os.path.join(base_name, p) for p in relative_paths]
+
+        assert len(absolute_paths) == len(relative_paths)
+        result += zip(absolute_paths, relative_paths)
+
+    return result
+
+
+def copy_files_and_create_dirs(files: List[Tuple[str, str]]) -> None:
+    """Takes in a list of tuples of (src, dst) paths and copies files.
+    Will create all necessary directories."""
+    for file in files:
+        target_dir_name = os.path.dirname(file[1])
+
+        # will create all intermediate-level directories
+        if not os.path.exists(target_dir_name):
+            os.makedirs(target_dir_name)
+
+        shutil.copyfile(file[0], file[1])
+
+
+# URL helpers
+# ------------------------------------------------------------------------------------------
+
+def is_url(obj: Any, allow_file_urls: bool = False) -> bool:
+    """Determine whether the given object is a valid URL string."""
+    if not isinstance(obj, str) or not "://" in obj:
+        return False
+    if allow_file_urls and obj.startswith('file://'):
+        return True
+    try:
+        res = requests.compat.urlparse(obj)
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+        res = requests.compat.urlparse(requests.compat.urljoin(obj, "/"))
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+    except:
+        return False
+    return True
+
+
+def open_url(url: str, cache_dir: str = None, num_attempts: int = 10, verbose: bool = True, return_filename: bool = False, cache: bool = True) -> Any:
+    """Download the given URL and return a binary-mode file object to access the data."""
+    assert num_attempts >= 1
+    assert not (return_filename and (not cache))
+
+    # Doesn't look like an URL scheme so interpret it as a local filename.
+    if not re.match('^[a-z]+://', url):
+        return url if return_filename else open(url, "rb")
+
+    # Handle file URLs.  This code handles unusual file:// patterns that
+    # arise on Windows:
+    #
+    # file:///c:/foo.txt
+    #
+    # which would translate to a local '/c:/foo.txt' filename that's
+    # invalid.  Drop the forward slash for such pathnames.
+    #
+    # If you touch this code path, you should test it on both Linux and
+    # Windows.
+    #
+    # Some internet resources suggest using urllib.request.url2pathname() but
+    # but that converts forward slashes to backslashes and this causes
+    # its own set of problems.
+    if url.startswith('file://'):
+        filename = urllib.parse.urlparse(url).path
+        if re.match(r'^/[a-zA-Z]:', filename):
+            filename = filename[1:]
+        return filename if return_filename else open(filename, "rb")
+
+    assert is_url(url)
+
+    # Lookup from cache.
+    if cache_dir is None:
+        cache_dir = make_cache_dir_path('downloads')
+
+    url_md5 = hashlib.md5(url.encode("utf-8")).hexdigest()
+    if cache:
+        cache_files = glob.glob(os.path.join(cache_dir, url_md5 + "_*"))
+        if len(cache_files) == 1:
+            filename = cache_files[0]
+            return filename if return_filename else open(filename, "rb")
+
+    # Download.
+    url_name = None
+    url_data = None
+    with requests.Session() as session:
+        if verbose:
+            print("Downloading %s ..." % url, end="", flush=True)
+        for attempts_left in reversed(range(num_attempts)):
+            try:
+                with session.get(url) as res:
+                    res.raise_for_status()
+                    if len(res.content) == 0:
+                        raise IOError("No data received")
+
+                    if len(res.content) < 8192:
+                        content_str = res.content.decode("utf-8")
+                        if "download_warning" in res.headers.get("Set-Cookie", ""):
+                            links = [html.unescape(link) for link in content_str.split('"') if "export=download" in link]
+                            if len(links) == 1:
+                                url = requests.compat.urljoin(url, links[0])
+                                raise IOError("Google Drive virus checker nag")
+                        if "Google Drive - Quota exceeded" in content_str:
+                            raise IOError("Google Drive download quota exceeded -- please try again later")
+
+                    match = re.search(r'filename="([^"]*)"', res.headers.get("Content-Disposition", ""))
+                    url_name = match[1] if match else url
+                    url_data = res.content
+                    if verbose:
+                        print(" done")
+                    break
+            except KeyboardInterrupt:
+                raise
+            except:
+                if not attempts_left:
+                    if verbose:
+                        print(" failed")
+                    raise
+                if verbose:
+                    print(".", end="", flush=True)
+
+    # Save to cache.
+    if cache:
+        safe_name = re.sub(r"[^0-9a-zA-Z-._]", "_", url_name)
+        cache_file = os.path.join(cache_dir, url_md5 + "_" + safe_name)
+        temp_file = os.path.join(cache_dir, "tmp_" + uuid.uuid4().hex + "_" + url_md5 + "_" + safe_name)
+        os.makedirs(cache_dir, exist_ok=True)
+        with open(temp_file, "wb") as f:
+            f.write(url_data)
+        os.replace(temp_file, cache_file) # atomic
+        if return_filename:
+            return cache_file
+
+    # Return data as file object.
+    assert not return_filename
+    return io.BytesIO(url_data)

legacy.py

@@ -0,0 +1,323 @@
+# 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.
+
+"""Converting legacy network pickle into the new format."""
+
+import click
+import pickle
+import re
+import copy
+import numpy as np
+import torch
+import dnnlib
+from torch_utils import misc
+
+#----------------------------------------------------------------------------
+
+def load_network_pkl(f, force_fp16=False):
+    data = _LegacyUnpickler(f).load()
+
+    # Legacy TensorFlow pickle => convert.
+    if isinstance(data, tuple) and len(data) == 3 and all(isinstance(net, _TFNetworkStub) for net in data):
+        tf_G, tf_D, tf_Gs = data
+        G = convert_tf_generator(tf_G)
+        D = convert_tf_discriminator(tf_D)
+        G_ema = convert_tf_generator(tf_Gs)
+        data = dict(G=G, D=D, G_ema=G_ema)
+
+    # Add missing fields.
+    if 'training_set_kwargs' not in data:
+        data['training_set_kwargs'] = None
+    if 'augment_pipe' not in data:
+        data['augment_pipe'] = None
+
+    # Validate contents.
+    assert isinstance(data['G'], torch.nn.Module)
+    assert isinstance(data['D'], torch.nn.Module)
+    assert isinstance(data['G_ema'], torch.nn.Module)
+    assert isinstance(data['training_set_kwargs'], (dict, type(None)))
+    assert isinstance(data['augment_pipe'], (torch.nn.Module, type(None)))
+
+    # Force FP16.
+    if force_fp16:
+        for key in ['G', 'D', 'G_ema']:
+            old = data[key]
+            kwargs = copy.deepcopy(old.init_kwargs)
+            fp16_kwargs = kwargs.get('synthesis_kwargs', kwargs)
+            fp16_kwargs.num_fp16_res = 4
+            fp16_kwargs.conv_clamp = 256
+            if kwargs != old.init_kwargs:
+                new = type(old)(**kwargs).eval().requires_grad_(False)
+                misc.copy_params_and_buffers(old, new, require_all=True)
+                data[key] = new
+    return data
+
+#----------------------------------------------------------------------------
+
+class _TFNetworkStub(dnnlib.EasyDict):
+    pass
+
+class _LegacyUnpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        if module == 'dnnlib.tflib.network' and name == 'Network':
+            return _TFNetworkStub
+        return super().find_class(module, name)
+
+#----------------------------------------------------------------------------
+
+def _collect_tf_params(tf_net):
+    # pylint: disable=protected-access
+    tf_params = dict()
+    def recurse(prefix, tf_net):
+        for name, value in tf_net.variables:
+            tf_params[prefix + name] = value
+        for name, comp in tf_net.components.items():
+            recurse(prefix + name + '/', comp)
+    recurse('', tf_net)
+    return tf_params
+
+#----------------------------------------------------------------------------
+
+def _populate_module_params(module, *patterns):
+    for name, tensor in misc.named_params_and_buffers(module):
+        found = False
+        value = None
+        for pattern, value_fn in zip(patterns[0::2], patterns[1::2]):
+            match = re.fullmatch(pattern, name)
+            if match:
+                found = True
+                if value_fn is not None:
+                    value = value_fn(*match.groups())
+                break
+        try:
+            assert found
+            if value is not None:
+                tensor.copy_(torch.from_numpy(np.array(value)))
+        except:
+            print(name, list(tensor.shape))
+            raise
+
+#----------------------------------------------------------------------------
+
+def convert_tf_generator(tf_G):
+    if tf_G.version < 4:
+        raise ValueError('TensorFlow pickle version too low')
+
+    # Collect kwargs.
+    tf_kwargs = tf_G.static_kwargs
+    known_kwargs = set()
+    def kwarg(tf_name, default=None, none=None):
+        known_kwargs.add(tf_name)
+        val = tf_kwargs.get(tf_name, default)
+        return val if val is not None else none
+
+    # Convert kwargs.
+    from training import networks_stylegan2
+    network_class = networks_stylegan2.Generator
+    kwargs = dnnlib.EasyDict(
+        z_dim               = kwarg('latent_size',          512),
+        c_dim               = kwarg('label_size',           0),
+        w_dim               = kwarg('dlatent_size',         512),
+        img_resolution      = kwarg('resolution',           1024),
+        img_channels        = kwarg('num_channels',         3),
+        channel_base        = kwarg('fmap_base',            16384) * 2,
+        channel_max         = kwarg('fmap_max',             512),
+        num_fp16_res        = kwarg('num_fp16_res',         0),
+        conv_clamp          = kwarg('conv_clamp',           None),
+        architecture        = kwarg('architecture',         'skip'),
+        resample_filter     = kwarg('resample_kernel',      [1,3,3,1]),
+        use_noise           = kwarg('use_noise',            True),
+        activation          = kwarg('nonlinearity',         'lrelu'),
+        mapping_kwargs      = dnnlib.EasyDict(
+            num_layers      = kwarg('mapping_layers',       8),
+            embed_features  = kwarg('label_fmaps',          None),
+            layer_features  = kwarg('mapping_fmaps',        None),
+            activation      = kwarg('mapping_nonlinearity', 'lrelu'),
+            lr_multiplier   = kwarg('mapping_lrmul',        0.01),
+            w_avg_beta      = kwarg('w_avg_beta',           0.995,  none=1),
+        ),
+    )
+
+    # Check for unknown kwargs.
+    kwarg('truncation_psi')
+    kwarg('truncation_cutoff')
+    kwarg('style_mixing_prob')
+    kwarg('structure')
+    kwarg('conditioning')
+    kwarg('fused_modconv')
+    unknown_kwargs = list(set(tf_kwargs.keys()) - known_kwargs)
+    if len(unknown_kwargs) > 0:
+        raise ValueError('Unknown TensorFlow kwarg', unknown_kwargs[0])
+
+    # Collect params.
+    tf_params = _collect_tf_params(tf_G)
+    for name, value in list(tf_params.items()):
+        match = re.fullmatch(r'ToRGB_lod(\d+)/(.*)', name)
+        if match:
+            r = kwargs.img_resolution // (2 ** int(match.group(1)))
+            tf_params[f'{r}x{r}/ToRGB/{match.group(2)}'] = value
+            kwargs.synthesis.kwargs.architecture = 'orig'
+    #for name, value in tf_params.items(): print(f'{name:<50s}{list(value.shape)}')
+
+    # Convert params.
+    G = network_class(**kwargs).eval().requires_grad_(False)
+    # pylint: disable=unnecessary-lambda
+    # pylint: disable=f-string-without-interpolation
+    _populate_module_params(G,
+        r'mapping\.w_avg',                                  lambda:     tf_params[f'dlatent_avg'],
+        r'mapping\.embed\.weight',                          lambda:     tf_params[f'mapping/LabelEmbed/weight'].transpose(),
+        r'mapping\.embed\.bias',                            lambda:     tf_params[f'mapping/LabelEmbed/bias'],
+        r'mapping\.fc(\d+)\.weight',                        lambda i:   tf_params[f'mapping/Dense{i}/weight'].transpose(),
+        r'mapping\.fc(\d+)\.bias',                          lambda i:   tf_params[f'mapping/Dense{i}/bias'],
+        r'synthesis\.b4\.const',                            lambda:     tf_params[f'synthesis/4x4/Const/const'][0],
+        r'synthesis\.b4\.conv1\.weight',                    lambda:     tf_params[f'synthesis/4x4/Conv/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b4\.conv1\.bias',                      lambda:     tf_params[f'synthesis/4x4/Conv/bias'],
+        r'synthesis\.b4\.conv1\.noise_const',               lambda:     tf_params[f'synthesis/noise0'][0, 0],
+        r'synthesis\.b4\.conv1\.noise_strength',            lambda:     tf_params[f'synthesis/4x4/Conv/noise_strength'],
+        r'synthesis\.b4\.conv1\.affine\.weight',            lambda:     tf_params[f'synthesis/4x4/Conv/mod_weight'].transpose(),
+        r'synthesis\.b4\.conv1\.affine\.bias',              lambda:     tf_params[f'synthesis/4x4/Conv/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.conv0\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/weight'][::-1, ::-1].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.conv0\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/bias'],
+        r'synthesis\.b(\d+)\.conv0\.noise_const',           lambda r:   tf_params[f'synthesis/noise{int(np.log2(int(r)))*2-5}'][0, 0],
+        r'synthesis\.b(\d+)\.conv0\.noise_strength',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/noise_strength'],
+        r'synthesis\.b(\d+)\.conv0\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.conv0\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.conv1\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.conv1\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/bias'],
+        r'synthesis\.b(\d+)\.conv1\.noise_const',           lambda r:   tf_params[f'synthesis/noise{int(np.log2(int(r)))*2-4}'][0, 0],
+        r'synthesis\.b(\d+)\.conv1\.noise_strength',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/noise_strength'],
+        r'synthesis\.b(\d+)\.conv1\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.conv1\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.torgb\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.torgb\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/bias'],
+        r'synthesis\.b(\d+)\.torgb\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.torgb\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.skip\.weight',                 lambda r:   tf_params[f'synthesis/{r}x{r}/Skip/weight'][::-1, ::-1].transpose(3, 2, 0, 1),
+        r'.*\.resample_filter',                             None,
+        r'.*\.act_filter',                                  None,
+    )
+    return G
+
+#----------------------------------------------------------------------------
+
+def convert_tf_discriminator(tf_D):
+    if tf_D.version < 4:
+        raise ValueError('TensorFlow pickle version too low')
+
+    # Collect kwargs.
+    tf_kwargs = tf_D.static_kwargs
+    known_kwargs = set()
+    def kwarg(tf_name, default=None):
+        known_kwargs.add(tf_name)
+        return tf_kwargs.get(tf_name, default)
+
+    # Convert kwargs.
+    kwargs = dnnlib.EasyDict(
+        c_dim                   = kwarg('label_size',           0),
+        img_resolution          = kwarg('resolution',           1024),
+        img_channels            = kwarg('num_channels',         3),
+        architecture            = kwarg('architecture',         'resnet'),
+        channel_base            = kwarg('fmap_base',            16384) * 2,
+        channel_max             = kwarg('fmap_max',             512),
+        num_fp16_res            = kwarg('num_fp16_res',         0),
+        conv_clamp              = kwarg('conv_clamp',           None),
+        cmap_dim                = kwarg('mapping_fmaps',        None),
+        block_kwargs = dnnlib.EasyDict(
+            activation          = kwarg('nonlinearity',         'lrelu'),
+            resample_filter     = kwarg('resample_kernel',      [1,3,3,1]),
+            freeze_layers       = kwarg('freeze_layers',        0),
+        ),
+        mapping_kwargs = dnnlib.EasyDict(
+            num_layers          = kwarg('mapping_layers',       0),
+            embed_features      = kwarg('mapping_fmaps',        None),
+            layer_features      = kwarg('mapping_fmaps',        None),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+            lr_multiplier       = kwarg('mapping_lrmul',        0.1),
+        ),
+        epilogue_kwargs = dnnlib.EasyDict(
+            mbstd_group_size    = kwarg('mbstd_group_size',     None),
+            mbstd_num_channels  = kwarg('mbstd_num_features',   1),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+        ),
+    )
+
+    # Check for unknown kwargs.
+    kwarg('structure')
+    kwarg('conditioning')
+    unknown_kwargs = list(set(tf_kwargs.keys()) - known_kwargs)
+    if len(unknown_kwargs) > 0:
+        raise ValueError('Unknown TensorFlow kwarg', unknown_kwargs[0])
+
+    # Collect params.
+    tf_params = _collect_tf_params(tf_D)
+    for name, value in list(tf_params.items()):
+        match = re.fullmatch(r'FromRGB_lod(\d+)/(.*)', name)
+        if match:
+            r = kwargs.img_resolution // (2 ** int(match.group(1)))
+            tf_params[f'{r}x{r}/FromRGB/{match.group(2)}'] = value
+            kwargs.architecture = 'orig'
+    #for name, value in tf_params.items(): print(f'{name:<50s}{list(value.shape)}')
+
+    # Convert params.
+    from training import networks_stylegan2
+    D = networks_stylegan2.Discriminator(**kwargs).eval().requires_grad_(False)
+    # pylint: disable=unnecessary-lambda
+    # pylint: disable=f-string-without-interpolation
+    _populate_module_params(D,
+        r'b(\d+)\.fromrgb\.weight',     lambda r:       tf_params[f'{r}x{r}/FromRGB/weight'].transpose(3, 2, 0, 1),
+        r'b(\d+)\.fromrgb\.bias',       lambda r:       tf_params[f'{r}x{r}/FromRGB/bias'],
+        r'b(\d+)\.conv(\d+)\.weight',   lambda r, i:    tf_params[f'{r}x{r}/Conv{i}{["","_down"][int(i)]}/weight'].transpose(3, 2, 0, 1),
+        r'b(\d+)\.conv(\d+)\.bias',     lambda r, i:    tf_params[f'{r}x{r}/Conv{i}{["","_down"][int(i)]}/bias'],
+        r'b(\d+)\.skip\.weight',        lambda r:       tf_params[f'{r}x{r}/Skip/weight'].transpose(3, 2, 0, 1),
+        r'mapping\.embed\.weight',      lambda:         tf_params[f'LabelEmbed/weight'].transpose(),
+        r'mapping\.embed\.bias',        lambda:         tf_params[f'LabelEmbed/bias'],
+        r'mapping\.fc(\d+)\.weight',    lambda i:       tf_params[f'Mapping{i}/weight'].transpose(),
+        r'mapping\.fc(\d+)\.bias',      lambda i:       tf_params[f'Mapping{i}/bias'],
+        r'b4\.conv\.weight',            lambda:         tf_params[f'4x4/Conv/weight'].transpose(3, 2, 0, 1),
+        r'b4\.conv\.bias',              lambda:         tf_params[f'4x4/Conv/bias'],
+        r'b4\.fc\.weight',              lambda:         tf_params[f'4x4/Dense0/weight'].transpose(),
+        r'b4\.fc\.bias',                lambda:         tf_params[f'4x4/Dense0/bias'],
+        r'b4\.out\.weight',             lambda:         tf_params[f'Output/weight'].transpose(),
+        r'b4\.out\.bias',               lambda:         tf_params[f'Output/bias'],
+        r'.*\.resample_filter',         None,
+    )
+    return D
+
+#----------------------------------------------------------------------------
+
+@click.command()
+@click.option('--source', help='Input pickle', required=True, metavar='PATH')
+@click.option('--dest', help='Output pickle', required=True, metavar='PATH')
+@click.option('--force-fp16', help='Force the networks to use FP16', type=bool, default=False, metavar='BOOL', show_default=True)
+def convert_network_pickle(source, dest, force_fp16):
+    """Convert legacy network pickle into the native PyTorch format.
+
+    The tool is able to load the main network configurations exported using the TensorFlow version of StyleGAN2 or StyleGAN2-ADA.
+    It does not support e.g. StyleGAN2-ADA comparison methods, StyleGAN2 configs A-D, or StyleGAN1 networks.
+
+    Example:
+
+    \b
+    python legacy.py \\
+        --source=https://nvlabs-fi-cdn.nvidia.com/stylegan2/networks/stylegan2-cat-config-f.pkl \\
+        --dest=stylegan2-cat-config-f.pkl
+    """
+    print(f'Loading "{source}"...')
+    with dnnlib.util.open_url(source) as f:
+        data = load_network_pkl(f, force_fp16=force_fp16)
+    print(f'Saving "{dest}"...')
+    with open(dest, 'wb') as f:
+        pickle.dump(data, f)
+    print('Done.')
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    convert_network_pickle() # pylint: disable=no-value-for-parameter
+
+#----------------------------------------------------------------------------

models/__init__.py

@@ -0,0 +1,68 @@
+"""This package contains modules related to objective functions, optimizations, and network architectures.
+
+To add a custom model class called 'dummy', you need to add a file called 'dummy_model.py' and define a subclass DummyModel inherited from BaseModel.
+You need to implement the following five functions:
+    -- <__init__>:                      initialize the class; first call BaseModel.__init__(self, opt).
+    -- <set_input>:                     unpack data from dataset and apply preprocessing.
+    -- <forward>:                       produce intermediate results.
+    -- <optimize_parameters>:           calculate loss, gradients, and update network weights.
+    -- <modify_commandline_options>:    (optionally) add model-specific options and set default options.
+
+In the function <__init__>, you need to define four lists:
+    -- self.loss_names (str list):          specify the training losses that you want to plot and save.
+    -- self.model_names (str list):         define networks used in our training.
+    -- self.visual_names (str list):        specify the images that you want to display and save.
+    -- self.optimizers (optimizer list):    define and initialize optimizers. You can define one optimizer for each network. If two networks are updated at the same time, you can use itertools.chain to group them. See cycle_gan_model.py for an usage.
+
+Now you can use the model class by specifying flag '--model dummy'.
+See our template model class 'template_model.py' for more details.
+"""
+
+import importlib
+from models.base_model import BaseModel
+
+
+def find_model_using_name(model_name):
+    """Import the module "models/[model_name]_model.py".
+
+    In the file, the class called DatasetNameModel() will
+    be instantiated. It has to be a subclass of BaseModel,
+    and it is case-insensitive.
+    """
+    model_filename = "models." + model_name + "_model"
+    print(model_filename)
+    modellib = importlib.import_module(model_filename)
+    model = None
+    target_model_name = model_name.replace('_', '') + 'model'
+    for name, cls in modellib.__dict__.items():
+        if name.lower() == target_model_name.lower() \
+           and issubclass(cls, BaseModel):
+            model = cls
+
+    if model is None:
+        print("In %s.py, there should be a subclass of BaseModel with class name that matches %s in lowercase." % (model_filename, target_model_name))
+        exit(0)
+
+    return model
+
+
+def get_option_setter(model_name):
+    """Return the static method <modify_commandline_options> of the model class."""
+    model_class = find_model_using_name(model_name)
+    return model_class.modify_commandline_options
+
+
+def create_model(opt):
+    """Create a model given the option.
+
+    This function warps the class CustomDatasetDataLoader.
+    This is the main interface between this package and 'train.py'/'test.py'
+
+    Example:
+        >>> from models import create_model
+        >>> model = create_model(opt)
+    """
+    model = find_model_using_name(opt.model)
+    instance = model(opt)
+    print("model [%s] was created" % type(instance).__name__)
+    return instance

models/base_model.py

@@ -0,0 +1,234 @@
+import os
+import torch
+from collections import OrderedDict
+from abc import ABC, abstractmethod
+from . import networks
+
+
+class BaseModel(ABC):
+    """This class is an abstract base class (ABC) for models.
+    To create a subclass, you need to implement the following five functions:
+        -- <__init__>:                      initialize the class; first call BaseModel.__init__(self, opt).
+        -- <set_input>:                     unpack data from dataset and apply preprocessing.
+        -- <forward>:                       produce intermediate results.
+        -- <optimize_parameters>:           calculate losses, gradients, and update network weights.
+        -- <modify_commandline_options>:    (optionally) add model-specific options and set default options.
+    """
+
+    def __init__(self, opt):
+        """Initialize the BaseModel class.
+
+        Parameters:
+            opt (Option class)-- stores all the experiment flags; needs to be a subclass of BaseOptions
+
+        When creating your custom class, you need to implement your own initialization.
+        In this function, you should first call <BaseModel.__init__(self, opt)>
+        Then, you need to define four lists:
+            -- self.loss_names (str list):          specify the training losses that you want to plot and save.
+            -- self.model_names (str list):         define networks used in our training.
+            -- self.visual_names (str list):        specify the images that you want to display and save.
+            -- self.optimizers (optimizer list):    define and initialize optimizers. You can define one optimizer for each network. If two networks are updated at the same time, you can use itertools.chain to group them. See cycle_gan_model.py for an example.
+        """
+        self.opt = opt
+        self.gpu_ids = opt.gpu_ids
+        self.isTrain = opt.isTrain
+        self.device = torch.device('cuda:{}'.format(self.gpu_ids[0])) if self.gpu_ids else torch.device('cpu')  # get device name: CPU or GPU
+        self.save_dir = os.path.join(opt.checkpoints_dir, opt.name)  # save all the checkpoints to save_dir
+        if opt.preprocess != 'scale_width':  # with [scale_width], input images might have different sizes, which hurts the performance of cudnn.benchmark.
+            torch.backends.cudnn.benchmark = True
+        self.loss_names = []
+        self.model_names = []
+        self.visual_names = []
+        self.optimizers = []
+        self.image_paths = []
+        self.metric = 0  # used for learning rate policy 'plateau'
+
+    @staticmethod
+    def modify_commandline_options(parser, is_train):
+        """Add new model-specific options, and rewrite default values for existing options.
+
+        Parameters:
+            parser          -- original option parser
+            is_train (bool) -- whether training phase or test phase. You can use this flag to add training-specific or test-specific options.
+
+        Returns:
+            the modified parser.
+        """
+        return parser
+
+    @abstractmethod
+    def set_input(self, input):
+        """Unpack input data from the dataloader and perform necessary pre-processing steps.
+
+        Parameters:
+            input (dict): includes the data itself and its metadata information.
+        """
+        pass
+
+    @abstractmethod
+    def forward(self):
+        """Run forward pass; called by both functions <optimize_parameters> and <test>."""
+        pass
+
+    @abstractmethod
+    def optimize_parameters(self):
+        """Calculate losses, gradients, and update network weights; called in every training iteration"""
+        pass
+
+    def setup(self, opt):
+        """Load and print networks; create schedulers
+
+        Parameters:
+            opt (Option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        if self.isTrain:
+            self.schedulers = [networks.get_scheduler(optimizer, opt) for optimizer in self.optimizers]
+        if not self.isTrain or opt.continue_train:
+            load_suffix = 'iter_%d' % opt.load_iter if opt.load_iter > 0 else opt.epoch
+            self.load_networks(load_suffix)
+        self.print_networks(opt.verbose)
+
+    def eval(self):
+        """Make models eval mode during test time"""
+        for name in self.model_names:
+            if isinstance(name, str):
+                net = getattr(self, 'net' + name)
+                net.eval()
+
+    def test(self):
+        """Forward function used in test time.
+
+        This function wraps <forward> function in no_grad() so we don't save intermediate steps for backprop
+        It also calls <compute_visuals> to produce additional visualization results
+        """
+        with torch.no_grad():
+            self.forward()
+            self.compute_visuals()
+
+    def compute_visuals(self):
+        """Calculate additional output images for visdom and HTML visualization"""
+        pass
+
+    def get_image_paths(self):
+        """ Return image paths that are used to load current data"""
+        return self.image_paths
+
+    def update_learning_rate(self):
+        """Update learning rates for all the networks; called at the end of every epoch"""
+        old_lr = self.optimizers[0].param_groups[0]['lr']
+        for scheduler in self.schedulers:
+            if self.opt.lr_policy == 'plateau':
+                scheduler.step(self.metric)
+            else:
+                scheduler.step()
+
+        lr = self.optimizers[0].param_groups[0]['lr']
+        print('learning rate %.7f -> %.7f' % (old_lr, lr))
+
+    def get_current_visuals(self):
+        """Return visualization images. train.py will display these images with visdom, and save the images to a HTML"""
+        visual_ret = OrderedDict()
+        for name in self.visual_names:
+            if isinstance(name, str):
+                visual_ret[name] = getattr(self, name)
+        return visual_ret
+
+    def get_current_losses(self):
+        """Return traning losses / errors. train.py will print out these errors on console, and save them to a file"""
+        errors_ret = OrderedDict()
+        for name in self.loss_names:
+            if isinstance(name, str):
+                errors_ret[name] = float(getattr(self, 'loss_' + name))  # float(...) works for both scalar tensor and float number
+        return errors_ret
+
+    def save_networks(self, epoch):
+        """Save all the networks to the disk.
+
+        Parameters:
+            epoch (int) -- current epoch; used in the file name '%s_net_%s.pth' % (epoch, name)
+        """
+        for name in self.model_names:
+            if isinstance(name, str):
+                save_filename = '%s_net_%s.pth' % (epoch, name)
+                save_path = os.path.join(self.save_dir, save_filename)
+                net = getattr(self, 'net' + name)
+
+                if len(self.gpu_ids) > 0 and torch.cuda.is_available():
+                    if hasattr(net, 'module'):
+                        torch.save(net.module.cpu().state_dict(), save_path)
+                        net.cuda(self.gpu_ids[0])
+                    else:
+                        torch.save(net.cpu().state_dict(), save_path)
+                        net.cuda(self.gpu_ids[0])
+                else:
+                    torch.save(net.cpu().state_dict(), save_path)
+
+    def __patch_instance_norm_state_dict(self, state_dict, module, keys, i=0):
+        """Fix InstanceNorm checkpoints incompatibility (prior to 0.4)"""
+        key = keys[i]
+        if i + 1 == len(keys):  # at the end, pointing to a parameter/buffer
+            if module.__class__.__name__.startswith('InstanceNorm') and \
+                    (key == 'running_mean' or key == 'running_var'):
+                if getattr(module, key) is None:
+                    state_dict.pop('.'.join(keys))
+            if module.__class__.__name__.startswith('InstanceNorm') and \
+               (key == 'num_batches_tracked'):
+                state_dict.pop('.'.join(keys))
+        else:
+            self.__patch_instance_norm_state_dict(state_dict, getattr(module, key), keys, i + 1)
+
+    def load_networks(self, epoch):
+        """Load all the networks from the disk.
+
+        Parameters:
+            epoch (int) -- current epoch; used in the file name '%s_net_%s.pth' % (epoch, name)
+        """
+        for name in self.model_names:
+            if isinstance(name, str):
+                load_filename = '%s_net_%s.pth' % (epoch, name)
+                load_path = os.path.join(self.save_dir, load_filename)
+                net = getattr(self, 'net' + name)
+                if isinstance(net, torch.nn.DataParallel):
+                    net = net.module
+                print('loading the model from %s' % load_path)
+                # if you are using PyTorch newer than 0.4 (e.g., built from
+                # GitHub source), you can remove str() on self.device
+                state_dict = torch.load(load_path, map_location=str(self.device))
+                if hasattr(state_dict, '_metadata'):
+                    del state_dict._metadata
+
+                # patch InstanceNorm checkpoints prior to 0.4
+                for key in list(state_dict.keys()):  # need to copy keys here because we mutate in loop
+                    self.__patch_instance_norm_state_dict(state_dict, net, key.split('.'))
+                net.load_state_dict(state_dict)
+
+    def print_networks(self, verbose):
+        """Print the total number of parameters in the network and (if verbose) network architecture
+
+        Parameters:
+            verbose (bool) -- if verbose: print the network architecture
+        """
+        print('---------- Networks initialized -------------')
+        for name in self.model_names:
+            if isinstance(name, str):
+                net = getattr(self, 'net' + name)
+                num_params = 0
+                for param in net.parameters():
+                    num_params += param.numel()
+                if verbose:
+                    print(net)
+                print('[Network %s] Total number of parameters : %.3f M' % (name, num_params / 1e6))
+        print('-----------------------------------------------')
+
+    def set_requires_grad(self, nets, requires_grad=False):
+        """Set requies_grad=Fasle for all the networks to avoid unnecessary computations
+        Parameters:
+            nets (network list)   -- a list of networks
+            requires_grad (bool)  -- whether the networks require gradients or not
+        """
+        if not isinstance(nets, list):
+            nets = [nets]
+        for net in nets:
+            if net is not None:
+                for param in net.parameters():
+                    param.requires_grad = requires_grad

models/diy_networks.py

@@ -0,0 +1,918 @@
+import torch
+import torch.nn as nn
+from torch import Tensor
+from types import FunctionType
+from typing import Type, Any, Callable, Union, List, Optional
+
+def _log_api_usage_once(obj: Any) -> None:
+    if not obj.__module__.startswith("torchvision"):
+        return
+    name = obj.__class__.__name__
+    if isinstance(obj, FunctionType):
+        name = obj.__name__
+    torch._C._log_api_usage_once(f"{obj.__module__}.{name}")
+
+def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d:
+    """3x3 convolution with padding"""
+    return nn.Conv2d(
+        in_planes,
+        out_planes,
+        kernel_size=3,
+        stride=stride,
+        padding=dilation,
+        groups=groups,
+        bias=False,
+        dilation=dilation,
+    )
+
+
+def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
+
+class Bottleneck(nn.Module):
+    # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
+    # while original implementation places the stride at the first 1x1 convolution(self.conv1)
+    # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
+    # This variant is also known as ResNet V1.5 and improves accuracy according to
+    # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.
+
+    expansion: int = 4
+
+    def __init__(
+        self,
+        inplanes: int,
+        planes: int,
+        stride: int = 1,
+        downsample: Optional[nn.Module] = None,
+        groups: int = 1,
+        base_width: int = 64,
+        dilation: int = 1,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        width = int(planes * (base_width / 64.0)) * groups
+        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
+        self.conv1 = conv1x1(inplanes, width)
+        self.bn1 = norm_layer(width)
+        self.conv2 = conv3x3(width, width, stride, groups, dilation)
+        self.bn2 = norm_layer(width)
+        self.conv3 = conv1x1(width, planes * self.expansion)
+        self.bn3 = norm_layer(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x: Tensor) -> Tensor:
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+
+        return out
+
+class BasicBlock(nn.Module):
+    expansion: int = 1
+
+    def __init__(
+        self,
+        inplanes: int,
+        planes: int,
+        stride: int = 1,
+        downsample: Optional[nn.Module] = None,
+        groups: int = 1,
+        base_width: int = 64,
+        dilation: int = 1,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        if groups != 1 or base_width != 64:
+            raise ValueError("BasicBlock only supports groups=1 and base_width=64")
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = norm_layer(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = norm_layer(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x: Tensor) -> Tensor:
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+
+        return out
+
+
+class ResPoseNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        num_point: int = 12,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        block.expansion = 1
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 98
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.layer1 = self._make_layer(block, 98, 3, stride=2)
+        self.layer2 = self._make_layer(block, 49, 3, stride=2)
+        self.layer3 = self._make_layer(block, 1, 3, stride=2)
+
+        self.layer4 = self._make_layer(block, 32, 3, stride=2)
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(32 * block.expansion, num_point * 2)
+
+        # self.layer4_1 = self._make_layer(block, 16, 3, stride=2)
+        # self.layer4_2 = self._make_layer(block, 16, 3, stride=2)
+        # self.layer4_3 = self._make_layer(block, 16, 3, stride=2)
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc_1 = nn.Linear(16 * block.expansion, 2) # move
+        # self.fc_2 = nn.Linear(16 * block.expansion, 3) # pose
+        # self.fc_3 = nn.Linear(16 * block.expansion, 11) # attributes
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+
+        x = self.layer4(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        return x
+
+    def _forward_trans(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer4(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    # def _forward_impl(self, x: Tensor) -> Tensor:
+    #     # See note [TorchScript super()]
+    #     x = self.layer1(x)
+    #     x = self.layer2(x)
+    #     x = self.layer3(x)
+    #
+    #     x_1 = self.layer4_1(x)
+    #     x_1 = self.avgpool(x_1)
+    #     x_1 = torch.flatten(x_1, 1)
+    #     x_1 = self.fc_1(x_1)
+    #
+    #     x_2 = self.layer4_2(x)
+    #     x_2 = self.avgpool(x_2)
+    #     x_2 = torch.flatten(x_2, 1)
+    #     x_2 = self.fc_2(x_2)
+    #
+    #     x_3 = self.layer4_3(x)
+    #     x_3 = self.avgpool(x_3)
+    #     x_3 = torch.flatten(x_3, 1)
+    #     x_3 = self.fc_3(x_3)
+    #
+    #     return x_1, x_2, x_3
+    #
+    # def _forward_feature(self, x: Tensor) -> Tensor:
+    #     # See note [TorchScript super()]
+    #     x = self.layer1(x)
+    #     x = self.layer2(x)
+    #     x = self.layer3(x)
+    #     return x
+    #
+    # def _forward_trans(self, x: Tensor) -> Tensor:
+    #     # See note [TorchScript super()]
+    #     x_1 = self.layer4_1(x)
+    #     x_1 = self.avgpool(x_1)
+    #     x_1 = torch.flatten(x_1, 1)
+    #     x_1 = self.fc_1(x_1)
+    #
+    #     x_2 = self.layer4_2(x)
+    #     x_2 = self.avgpool(x_2)
+    #     x_2 = torch.flatten(x_2, 1)
+    #     x_2 = self.fc_2(x_2)
+    #
+    #     x_3 = self.layer4_3(x)
+    #     x_3 = self.avgpool(x_3)
+    #     x_3 = torch.flatten(x_3, 1)
+    #     x_3 = self.fc_3(x_3)
+    #
+    #     return x_1, x_2, x_3
+    #
+    def forward(self, x: Tensor, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x)
+
+class NormResPoseNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        num_point: int = 12,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        block.expansion = 1
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 98
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.layer1 = self._make_layer(block, 98, 3, stride=2)
+        self.layer2 = self._make_layer(block, 49, 3, stride=2)
+        self.layer3 = self._make_layer(block, 1, 3, stride=2)
+
+        self.layer4 = self._make_layer(block, 32, 3, stride=2)
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(32 * block.expansion, num_point * 2)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = torch.sigmoid(self.layer3(x))
+
+        x = self.layer4(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = torch.sigmoid(self.layer3(x))
+        return x
+
+    def _forward_trans(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer4(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    def forward(self, x: Tensor, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x)
+
+
+class ResPoseWNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        num_point: int = 12,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        # from .attention_networks import Self_Attn
+        # self.attention_layer = Self_Attn(64, 'relu')
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 3
+        self.dilation = 1
+        block.expansion = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 64, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 1, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        self.layer5 = self._make_layer(block, 32, 3, stride=2)
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(32 * block.expansion, num_point * 2)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = torch.tanh(self.layer4(x))
+
+        x = self.layer5(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = torch.tanh(self.layer4(x))
+
+        return x
+
+    def _forward_trans(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer5(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    # def _forward_impl(self, x: Tensor) -> Tensor:
+    #     # See note [TorchScript super()]
+    #     x = self.layer1(x)
+    #     x = self.layer2(x)
+    #     x = self.layer3(x)
+    #
+    #     x_1 = self.layer4_1(x)
+    #     x_1 = self.avgpool(x_1)
+    #     x_1 = torch.flatten(x_1, 1)
+    #     x_1 = self.fc_1(x_1)
+    #
+    #     x_2 = self.layer4_2(x)
+    #     x_2 = self.avgpool(x_2)
+    #     x_2 = torch.flatten(x_2, 1)
+    #     x_2 = self.fc_2(x_2)
+    #
+    #     x_3 = self.layer4_3(x)
+    #     x_3 = self.avgpool(x_3)
+    #     x_3 = torch.flatten(x_3, 1)
+    #     x_3 = self.fc_3(x_3)
+    #
+    #     return x_1, x_2, x_3
+    #
+    # def _forward_feature(self, x: Tensor) -> Tensor:
+    #     # See note [TorchScript super()]
+    #     x = self.layer1(x)
+    #     x = self.layer2(x)
+    #     x = self.layer3(x)
+    #     return x
+    #
+    # def _forward_trans(self, x: Tensor) -> Tensor:
+    #     # See note [TorchScript super()]
+    #     x_1 = self.layer4_1(x)
+    #     x_1 = self.avgpool(x_1)
+    #     x_1 = torch.flatten(x_1, 1)
+    #     x_1 = self.fc_1(x_1)
+    #
+    #     x_2 = self.layer4_2(x)
+    #     x_2 = self.avgpool(x_2)
+    #     x_2 = torch.flatten(x_2, 1)
+    #     x_2 = self.fc_2(x_2)
+    #
+    #     x_3 = self.layer4_3(x)
+    #     x_3 = self.avgpool(x_3)
+    #     x_3 = torch.flatten(x_3, 1)
+    #     x_3 = self.fc_3(x_3)
+    #
+    #     return x_1, x_2, x_3
+    #
+    def forward(self, x: Tensor, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x)
+
+class ResPose4Net(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        num_point: int = 12,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        block.expansion = 1
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 98
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.layer1 = self._make_layer(block, 98, 3, stride=2)
+        self.layer2 = self._make_layer(block, 49, 3, stride=2)
+        self.layer3 = self._make_layer(block, 7, 3, stride=2)
+        self.layer4 = self._make_layer(block, 1, 3, stride=2)
+
+        self.layer5 = self._make_layer(block, 32, 3, stride=2)
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(32 * block.expansion, num_point * 2)
+
+        # self.layer4_1 = self._make_layer(block, 16, 3, stride=2)
+        # self.layer4_2 = self._make_layer(block, 16, 3, stride=2)
+        # self.layer4_3 = self._make_layer(block, 16, 3, stride=2)
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc_1 = nn.Linear(16 * block.expansion, 2) # move
+        # self.fc_2 = nn.Linear(16 * block.expansion, 3) # pose
+        # self.fc_3 = nn.Linear(16 * block.expansion, 11) # attributes
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.layer5(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        return x
+
+    def _forward_trans(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.layer5(x)
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = torch.sigmoid(self.fc(x))
+
+        return x
+
+    def forward(self, x: Tensor, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x)
+
+def _normresposenet(**kwargs: Any) -> ResPoseNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return NormResPoseNet(Bottleneck, **kwargs)
+
+def _resposenet(**kwargs: Any) -> ResPoseNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return ResPoseNet(Bottleneck, **kwargs)
+
+def _respose4net(**kwargs: Any) -> ResPose4Net:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return ResPose4Net(Bottleneck, **kwargs)
+
+def _resposewnet(**kwargs: Any) -> ResPoseNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return ResPoseWNet(Bottleneck, layers = [3, 4, 6, 3], **kwargs)
+
+class PoseMapBN(nn.Module):
+    def __init__(self, input_num, output_num):
+        super().__init__()
+
+        self.ln1 = nn.Linear(input_num, 256)
+        self.bn1 = nn.BatchNorm1d(256)
+        self.ac1 = nn.LeakyReLU()
+
+        self.ln2 = nn.Linear(256, 256)
+        self.bn2 = nn.BatchNorm1d(256)
+        self.ac2 = nn.LeakyReLU()
+
+        self.ln3 = nn.Linear(256, output_num)
+
+    def forward(self, x):
+        x = self.ac1(self.bn1(self.ln1(x)))
+        x = self.ac2(self.bn2(self.ln2(x)))
+        out = self.ln3(x)
+        return out
+
+class PoseMap(nn.Module):
+    def __init__(self, input_num, output_num):
+        super().__init__()
+
+        self.ln1 = nn.Linear(input_num, 256)
+        self.ac1 = nn.LeakyReLU()
+
+        self.ln2 = nn.Linear(256, 256)
+        self.ac2 = nn.LeakyReLU()
+
+        self.ln3 = nn.Linear(256, output_num)
+
+    def forward(self, x):
+        x = self.ac1(self.ln1(x))
+        x = self.ac2(self.ln2(x))
+        out = self.ln3(x)
+        return out

models/lmcode_networks.py

@@ -0,0 +1,394 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+from collections import namedtuple
+from munch import Munch
+from copy import deepcopy
+from functools import partial
+
+IDXPAIR = namedtuple('IDXPAIR', 'start end')
+index_map = Munch(chin=IDXPAIR(0 + 8, 33 - 8),
+                  eyebrows=IDXPAIR(33, 51),
+                  eyebrowsedges=IDXPAIR(33, 46),
+                  nose=IDXPAIR(51, 55),
+                  nostrils=IDXPAIR(55, 60),
+                  eyes=IDXPAIR(60, 76),
+                  lipedges=IDXPAIR(76, 82),
+                  lipupper=IDXPAIR(77, 82),
+                  liplower=IDXPAIR(83, 88),
+                  lipinner=IDXPAIR(88, 96))
+OPPAIR = namedtuple('OPPAIR', 'shift resize')
+
+def conv3x3(in_planes, out_planes, strd=1, padding=1,
+            bias=False,dilation=1):
+    "3x3 convolution with padding"
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3,
+                     stride=strd, padding=padding, bias=bias,
+                     dilation=dilation)
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        # self.bn1 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        # self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        # out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        # out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+class HourGlass(nn.Module):
+    def __init__(self, num_modules, depth, num_features, first_one=False):
+        super(HourGlass, self).__init__()
+        self.num_modules = num_modules
+        self.depth = depth
+        self.features = num_features
+        self.coordconv = CoordConvTh(64, 64, True, True, 256, first_one,
+                                     out_channels=256,
+                                     kernel_size=1, stride=1, padding=0)
+        self._generate_network(self.depth)
+
+    def _generate_network(self, level):
+        self.add_module('b1_' + str(level), ConvBlock(256, 256))
+        self.add_module('b2_' + str(level), ConvBlock(256, 256))
+        if level > 1:
+            self._generate_network(level - 1)
+        else:
+            self.add_module('b2_plus_' + str(level), ConvBlock(256, 256))
+        self.add_module('b3_' + str(level), ConvBlock(256, 256))
+
+    def _forward(self, level, inp):
+        up1 = inp
+        up1 = self._modules['b1_' + str(level)](up1)
+        low1 = F.avg_pool2d(inp, 2, stride=2)
+        low1 = self._modules['b2_' + str(level)](low1)
+
+        if level > 1:
+            low2 = self._forward(level - 1, low1)
+        else:
+            low2 = low1
+            low2 = self._modules['b2_plus_' + str(level)](low2)
+        low3 = low2
+        low3 = self._modules['b3_' + str(level)](low3)
+        up2 = F.interpolate(low3, scale_factor=2, mode='nearest')
+
+        return up1 + up2
+
+    def forward(self, x, heatmap):
+        x, last_channel = self.coordconv(x, heatmap)
+        return self._forward(self.depth, x), last_channel
+
+class AddCoordsTh(nn.Module):
+    def __init__(self, height=64, width=64, with_r=False, with_boundary=False):
+        super(AddCoordsTh, self).__init__()
+        self.with_r = with_r
+        self.with_boundary = with_boundary
+        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+        with torch.no_grad():
+            x_coords = torch.arange(height).unsqueeze(1).expand(height, width).float()
+            y_coords = torch.arange(width).unsqueeze(0).expand(height, width).float()
+            x_coords = (x_coords / (height - 1)) * 2 - 1
+            y_coords = (y_coords / (width - 1)) * 2 - 1
+            coords = torch.stack([x_coords, y_coords], dim=0)  # (2, height, width)
+
+            if self.with_r:
+                rr = torch.sqrt(torch.pow(x_coords, 2) + torch.pow(y_coords, 2))  # (height, width)
+                rr = (rr / torch.max(rr)).unsqueeze(0)
+                coords = torch.cat([coords, rr], dim=0)
+
+            self.coords = coords.unsqueeze(0).to(device)  # (1, 2 or 3, height, width)
+            self.x_coords = x_coords.to(device)
+            self.y_coords = y_coords.to(device)
+
+    def forward(self, x, heatmap=None):
+        """
+        x: (batch, c, x_dim, y_dim)
+        """
+        coords = self.coords.repeat(x.size(0), 1, 1, 1).to(x.device)
+
+        if self.with_boundary and heatmap is not None:
+            boundary_channel = torch.clamp(heatmap[:, -1:, :, :], 0.0, 1.0)
+            zero_tensor = torch.zeros_like(self.x_coords)
+            xx_boundary_channel = torch.where(boundary_channel > 0.05, self.x_coords, zero_tensor).to(x.device)
+            yy_boundary_channel = torch.where(boundary_channel > 0.05, self.y_coords, zero_tensor).to(x.device)
+            coords = torch.cat([coords, xx_boundary_channel, yy_boundary_channel], dim=1)
+
+        x_and_coords = torch.cat([x, coords], dim=1)
+        return x_and_coords
+
+
+class CoordConvTh(nn.Module):
+    """CoordConv layer as in the paper."""
+    def __init__(self, height, width, with_r, with_boundary,
+                 in_channels, first_one=False, *args, **kwargs):
+        super(CoordConvTh, self).__init__()
+        self.addcoords = AddCoordsTh(height, width, with_r, with_boundary)
+        in_channels += 2
+        if with_r:
+            in_channels += 1
+        if with_boundary and not first_one:
+            in_channels += 2
+        self.conv = nn.Conv2d(in_channels=in_channels, *args, **kwargs)
+
+    def forward(self, input_tensor, heatmap=None):
+        ret = self.addcoords(input_tensor, heatmap)
+        last_channel = ret[:, -2:, :, :]
+        ret = self.conv(ret)
+        return ret, last_channel
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_planes, out_planes):
+        super(ConvBlock, self).__init__()
+        self.bn1 = nn.BatchNorm2d(in_planes)
+        conv3x3 = partial(nn.Conv2d, kernel_size=3, stride=1, padding=1, bias=False, dilation=1)
+        self.conv1 = conv3x3(in_planes, int(out_planes / 2))
+        self.bn2 = nn.BatchNorm2d(int(out_planes / 2))
+        self.conv2 = conv3x3(int(out_planes / 2), int(out_planes / 4))
+        self.bn3 = nn.BatchNorm2d(int(out_planes / 4))
+        self.conv3 = conv3x3(int(out_planes / 4), int(out_planes / 4))
+
+        self.downsample = None
+        if in_planes != out_planes:
+            self.downsample = nn.Sequential(nn.BatchNorm2d(in_planes),
+                                            nn.ReLU(True),
+                                            nn.Conv2d(in_planes, out_planes, 1, 1, bias=False))
+
+    def forward(self, x):
+        residual = x
+
+        out1 = self.bn1(x)
+        out1 = F.relu(out1, True)
+        out1 = self.conv1(out1)
+
+        out2 = self.bn2(out1)
+        out2 = F.relu(out2, True)
+        out2 = self.conv2(out2)
+
+        out3 = self.bn3(out2)
+        out3 = F.relu(out3, True)
+        out3 = self.conv3(out3)
+
+        out3 = torch.cat((out1, out2, out3), 1)
+        if self.downsample is not None:
+            residual = self.downsample(residual)
+        out3 += residual
+        return out3
+
+
+class FAN(nn.Module):
+    def __init__(self, num_modules=1, end_relu=False, num_landmarks=98, fname_pretrained=None):
+        super(FAN, self).__init__()
+        self.num_modules = num_modules
+        self.end_relu = end_relu
+
+        # Base part
+        self.conv1 = CoordConvTh(256, 256, True, False,
+                                 in_channels=3, out_channels=64,
+                                 kernel_size=7, stride=2, padding=3)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.conv2 = ConvBlock(64, 128)
+        self.conv3 = ConvBlock(128, 128)
+        self.conv4 = ConvBlock(128, 256)
+
+        # Stacking part
+        self.add_module('m0', HourGlass(1, 4, 256, first_one=True))
+        self.add_module('top_m_0', ConvBlock(256, 256))
+        self.add_module('conv_last0', nn.Conv2d(256, 256, 1, 1, 0))
+        self.add_module('bn_end0', nn.BatchNorm2d(256))
+        self.add_module('l0', nn.Conv2d(256, num_landmarks+1, 1, 1, 0))
+
+        if fname_pretrained is not None:
+            self.load_pretrained_weights(fname_pretrained)
+
+    def load_pretrained_weights(self, fname):
+        if torch.cuda.is_available():
+            checkpoint = torch.load(fname)
+        else:
+            checkpoint = torch.load(fname, map_location=torch.device('cpu'))
+        model_weights = self.state_dict()
+        model_weights.update({k: v for k, v in checkpoint['state_dict'].items()
+                              if k in model_weights})
+        self.load_state_dict(model_weights)
+
+    def forward(self, x):
+        x, _ = self.conv1(x)
+        x = F.relu(self.bn1(x), True)
+        x = F.avg_pool2d(self.conv2(x), 2, stride=2)
+        x = self.conv3(x)
+        x = self.conv4(x)
+
+        outputs = []
+        boundary_channels = []
+        tmp_out = None
+        ll, boundary_channel = self._modules['m0'](x, tmp_out)
+        ll = self._modules['top_m_0'](ll)
+        ll = F.relu(self._modules['bn_end0']
+                    (self._modules['conv_last0'](ll)), True)
+
+        # Predict heatmaps
+        tmp_out = self._modules['l0'](ll)
+        if self.end_relu:
+            tmp_out = F.relu(tmp_out)  # HACK: Added relu
+        outputs.append(tmp_out)
+        boundary_channels.append(boundary_channel)
+        return outputs, boundary_channels
+
+    @torch.no_grad()
+    def get_heatmap(self, x, b_preprocess=True):
+        ''' outputs 0-1 normalized heatmap '''
+        x = F.interpolate(x, size=256, mode='bilinear')
+        x_01 = x*0.5 + 0.5
+        outputs, _ = self(x_01)
+        heatmaps = outputs[-1][:, :-1, :, :]
+        scale_factor = x.size(2) // heatmaps.size(2)
+        if b_preprocess:
+            heatmaps = F.interpolate(heatmaps, scale_factor=scale_factor,
+                                     mode='bilinear', align_corners=True)
+            heatmaps = preprocess(heatmaps)
+        return heatmaps
+
+    @torch.no_grad()
+    def get_landmark(self, x):
+        ''' outputs landmarks of x.shape '''
+        heatmaps = self.get_heatmap(x, b_preprocess=False)
+        landmarks = []
+        for i in range(x.size(0)):
+            pred_landmarks = get_preds_fromhm(heatmaps[i].cpu().unsqueeze(0))
+            landmarks.append(pred_landmarks)
+        scale_factor = x.size(2) // heatmaps.size(2)
+        landmarks = torch.cat(landmarks) * scale_factor
+        return landmarks
+
+
+def get_preds_fromhm(hm):
+    max, idx = torch.max(
+        hm.view(hm.size(0), hm.size(1), hm.size(2) * hm.size(3)), 2)
+    idx += 1
+    preds = idx.view(idx.size(0), idx.size(1), 1).repeat(1, 1, 2).float()
+    preds[..., 0].apply_(lambda x: (x - 1) % hm.size(3) + 1)
+    preds[..., 1].add_(-1).div_(hm.size(2)).floor_().add_(1)
+
+    for i in range(preds.size(0)):
+        for j in range(preds.size(1)):
+            hm_ = hm[i, j, :]
+            pX, pY = int(preds[i, j, 0]) - 1, int(preds[i, j, 1]) - 1
+            if pX > 0 and pX < 63 and pY > 0 and pY < 63:
+                diff = torch.FloatTensor(
+                    [hm_[pY, pX + 1] - hm_[pY, pX - 1],
+                     hm_[pY + 1, pX] - hm_[pY - 1, pX]])
+                preds[i, j].add_(diff.sign_().mul_(.25))
+
+    preds.add_(-0.5)
+    return preds
+
+def truncate(x, thres=0.1):
+    """Remove small values in heatmaps."""
+    return torch.where(x < thres, torch.zeros_like(x), x)
+
+def normalize(x, eps=1e-6):
+    """Apply min-max normalization."""
+    x = x.contiguous()
+    N, C, H, W = x.size()
+    x_ = x.view(N*C, -1)
+    max_val = torch.max(x_, dim=1, keepdim=True)[0]
+    min_val = torch.min(x_, dim=1, keepdim=True)[0]
+    x_ = (x_ - min_val) / (max_val - min_val + eps)
+    out = x_.view(N, C, H, W)
+    return out
+
+def resize(x, p=2):
+    """Resize heatmaps."""
+    return x**p
+
+
+def shift(x, N):
+    """Shift N pixels up or down."""
+    up = N >= 0
+    N = abs(N)
+    _, _, H, W = x.size()
+    head = torch.arange(N)
+    tail = torch.arange(H-N)
+
+    if up:
+        head = torch.arange(H-N)+N
+        tail = torch.arange(N)
+    else:
+        head = torch.arange(N) + (H-N)
+        tail = torch.arange(H-N)
+
+    # permutation indices
+    perm = torch.cat([head, tail]).to(x.device)
+    out = x[:, :, perm, :]
+    return out
+
+def preprocess(x):
+    """Preprocess 98-dimensional heatmaps."""
+    N, C, H, W = x.size()
+    x = truncate(x)
+    x = normalize(x)
+
+    sw = H // 256
+    operations = Munch(chin=OPPAIR(0, 3),
+                       eyebrows=OPPAIR(-7*sw, 2),
+                       nostrils=OPPAIR(8*sw, 4),
+                       lipupper=OPPAIR(-8*sw, 4),
+                       liplower=OPPAIR(8*sw, 4),
+                       lipinner=OPPAIR(-2*sw, 3))
+
+    for part, ops in operations.items():
+        start, end = index_map[part]
+        x[:, start:end] = resize(shift(x[:, start:end], ops.shift), ops.resize)
+
+    zero_out = torch.cat([torch.arange(0, index_map.chin.start),
+                          torch.arange(index_map.chin.end, 33),
+                          torch.LongTensor([index_map.eyebrowsedges.start,
+                                            index_map.eyebrowsedges.end,
+                                            index_map.lipedges.start,
+                                            index_map.lipedges.end])])
+    x[:, zero_out] = 0
+
+    start, end = index_map.nose
+    x[:, start+1:end] = shift(x[:, start+1:end], 4*sw)
+    x[:, start:end] = resize(x[:, start:end], 1)
+
+    start, end = index_map.eyes
+    x[:, start:end] = resize(x[:, start:end], 1)
+    x[:, start:end] = resize(shift(x[:, start:end], -8), 3) + \
+        shift(x[:, start:end], -24)
+
+    # Second-level mask
+    x2 = deepcopy(x)
+    x2[:, index_map.chin.start:index_map.chin.end] = 0  # start:end was 0:33
+    x2[:, index_map.lipedges.start:index_map.lipinner.end] = 0  # start:end was 76:96
+    x2[:, index_map.eyebrows.start:index_map.eyebrows.end] = 0  # start:end was 33:51
+
+    x = torch.sum(x, dim=1, keepdim=True)  # (N, 1, H, W)
+    x2 = torch.sum(x2, dim=1, keepdim=True)  # mask without faceline and mouth
+
+    x[x != x] = 0  # set nan to zero
+    x2[x != x] = 0  # set nan to zero
+    return x.clamp_(0, 1), x2.clamp_(0, 1)

models/resnet.py

@@ -0,0 +1,1452 @@
+from typing import Type, Any, Callable, Union, List, Optional
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+try:
+    from torch.hub import load_state_dict_from_url  # noqa: 401
+except ImportError:
+    from torch.utils.model_zoo import load_url as load_state_dict_from_url  # noqa: 401
+from types import FunctionType
+
+def _log_api_usage_once(obj: Any) -> None:
+    if not obj.__module__.startswith("torchvision"):
+        return
+    name = obj.__class__.__name__
+    if isinstance(obj, FunctionType):
+        name = obj.__name__
+    torch._C._log_api_usage_once(f"{obj.__module__}.{name}")
+
+__all__ = [
+    "ResNet",
+    "resnet18",
+    "resnet34",
+    "resnet50",
+    "resnet101",
+    "resnet152",
+    "resnext50_32x4d",
+    "resnext101_32x8d",
+    "wide_resnet50_2",
+    "wide_resnet101_2",
+]
+
+
+model_urls = {
+    "resnet18": "https://download.pytorch.org/models/resnet18-f37072fd.pth",
+    "resnet34": "https://download.pytorch.org/models/resnet34-b627a593.pth",
+    "resnet50": "https://download.pytorch.org/models/resnet50-0676ba61.pth",
+    "resnet101": "https://download.pytorch.org/models/resnet101-63fe2227.pth",
+    "resnet152": "https://download.pytorch.org/models/resnet152-394f9c45.pth",
+    "resnext50_32x4d": "https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth",
+    "resnext101_32x8d": "https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth",
+    "wide_resnet50_2": "https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth",
+    "wide_resnet101_2": "https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth",
+}
+
+
+def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d:
+    """3x3 convolution with padding"""
+    return nn.Conv2d(
+        in_planes,
+        out_planes,
+        kernel_size=3,
+        stride=stride,
+        padding=dilation,
+        groups=groups,
+        bias=False,
+        dilation=dilation,
+    )
+
+
+def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d:
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
+
+
+class BasicBlock(nn.Module):
+    expansion: int = 1
+
+    def __init__(
+        self,
+        inplanes: int,
+        planes: int,
+        stride: int = 1,
+        downsample: Optional[nn.Module] = None,
+        groups: int = 1,
+        base_width: int = 64,
+        dilation: int = 1,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        if groups != 1 or base_width != 64:
+            raise ValueError("BasicBlock only supports groups=1 and base_width=64")
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = norm_layer(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = norm_layer(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x: Tensor) -> Tensor:
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
+    # while original implementation places the stride at the first 1x1 convolution(self.conv1)
+    # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
+    # This variant is also known as ResNet V1.5 and improves accuracy according to
+    # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.
+
+    expansion: int = 4
+
+    def __init__(
+        self,
+        inplanes: int,
+        planes: int,
+        stride: int = 1,
+        downsample: Optional[nn.Module] = None,
+        groups: int = 1,
+        base_width: int = 64,
+        dilation: int = 1,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        width = int(planes * (base_width / 64.0)) * groups
+        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
+        self.conv1 = conv1x1(inplanes, width)
+        self.bn1 = norm_layer(width)
+        self.conv2 = conv3x3(width, width, stride, groups, dilation)
+        self.bn2 = norm_layer(width)
+        self.conv3 = conv1x1(width, planes * self.expansion)
+        self.bn3 = norm_layer(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x: Tensor) -> Tensor:
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+        x = self.fc(x)
+
+        return x
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self._forward_impl(x)
+
+
+class ResAppNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=1, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        self.layer5 = self._make_layer(block, 1, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer1_a = self._make_layer(block, 64, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        self.layer2_a = self._make_layer(block, 128, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        self.avgpool_a = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc_a = nn.Linear(128 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+        # print(x.shape, flush = True)
+
+        x = self.layer1_a(x)
+        x = self.layer2_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+
+        return x
+
+    def _forward_trans(self, x: Tensor) -> Tensor:
+
+        x = self.layer1_a(x)
+        x = self.layer2_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def forward(self, x: Tensor, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x)
+
+
+class ResDisNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        # from .attention_networks import Self_Attn
+        # self.attention_layer = Self_Attn(64, 'relu')
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        self.layer5 = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.layer1_a = self._make_layer(block, 64, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer2_a = self._make_layer(block, 128, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer3_a = self._make_layer(block, 256, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer4_a = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+
+        self.inplanes = 1
+        self.layer5_b = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.avgpool_a = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc_a = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor, y: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+        # print(x.shape, flush = True)
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(x*y)
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+
+        return x
+
+    def _forward_trans(self, x: Tensor, y: Tensor) -> Tensor:
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(x*y)
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def forward(self, x: Tensor, y: Tensor=None, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x, y)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x, y)
+
+class ResMDisNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        # from .attention_networks import Self_Attn
+        # self.attention_layer = Self_Attn(64, 'relu')
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        self.layer5 = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.layer1_a = self._make_layer(block, 64, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer2_a = self._make_layer(block, 128, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer3_a = self._make_layer(block, 256, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer4_a = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+
+        self.inplanes = 2
+        self.layer5_b = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.avgpool_a = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc_a = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor, y: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+        # print(x.shape, flush = True)
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(x*y)
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+
+        return x
+
+    def _forward_trans(self, x: Tensor, y: Tensor) -> Tensor:
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(x*y)
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def forward(self, x: Tensor, y: Tensor=None, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x, y)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x, y)
+
+class ResDis2Net(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        # from .attention_networks import Self_Attn
+        # self.attention_layer = Self_Attn(64, 'relu')
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        self.layer5 = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.inplanes *= 2
+        self.layer1_a = self._make_layer(block, 64, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer2_a = self._make_layer(block, 128, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer3_a = self._make_layer(block, 256, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer4_a = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+
+        self.inplanes = 1
+        self.layer5_b = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.avgpool_a = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc_a = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor, y: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+        # print(x.shape, flush = True)
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(torch.cat([x,y], 1))
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+
+        return x
+
+    def _forward_trans(self, x: Tensor, y: Tensor) -> Tensor:
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(torch.cat([x,y], 1))
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def forward(self, x: Tensor, y: Tensor=None, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x, y)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x, y)
+
+class ResDisAttNet(nn.Module):
+    def __init__(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        layers: List[int],
+        num_classes: int = 1000,
+        zero_init_residual: bool = False,
+        groups: int = 1,
+        width_per_group: int = 64,
+        replace_stride_with_dilation: Optional[List[bool]] = None,
+        norm_layer: Optional[Callable[..., nn.Module]] = None,
+    ) -> None:
+        super().__init__()
+        _log_api_usage_once(self)
+
+        from .attention_networks import Pose_Attn
+        self.attention_layer = Pose_Attn(64, 'relu')
+
+        if norm_layer is None:
+            norm_layer = nn.BatchNorm2d
+        self._norm_layer = norm_layer
+
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            # each element in the tuple indicates if we should replace
+            # the 2x2 stride with a dilated convolution instead
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError(
+                "replace_stride_with_dilation should be None "
+                f"or a 3-element tuple, got {replace_stride_with_dilation}"
+            )
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = norm_layer(self.inplanes)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+        # self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        # self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        self.layer5 = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.layer1_a = self._make_layer(block, 64, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer2_a = self._make_layer(block, 128, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer3_a = self._make_layer(block, 256, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+        self.layer4_a = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2])
+
+        self.inplanes = 1
+        self.layer5_b = self._make_layer(block, 16, layers[3], stride=1, dilate=replace_stride_with_dilation[2])
+
+        self.avgpool_a = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc_a = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        # Zero-initialize the last BN in each residual branch,
+        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
+        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, Bottleneck):
+                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
+                elif isinstance(m, BasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]
+
+    def _make_layer(
+        self,
+        block: Type[Union[BasicBlock, Bottleneck]],
+        planes: int,
+        blocks: int,
+        stride: int = 1,
+        dilate: bool = False,
+    ) -> nn.Sequential:
+        norm_layer = self._norm_layer
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                norm_layer(planes * block.expansion),
+            )
+
+        layers = []
+        layers.append(
+            block(
+                self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    groups=self.groups,
+                    base_width=self.base_width,
+                    dilation=self.dilation,
+                    norm_layer=norm_layer,
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def _forward_impl(self, x: Tensor, y: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+        # print(x.shape, flush = True)
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(self.attention_layer(x, y)[0])
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def _forward_feature(self, x: Tensor) -> Tensor:
+        # See note [TorchScript super()]
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.layer5(x)
+
+        return x
+
+    def _forward_trans(self, x: Tensor, y: Tensor) -> Tensor:
+
+        y = self.layer5_b(y)
+
+        x = self.layer1_a(self.attention_layer(x, y)[0])
+        x = self.layer2_a(x)
+        # x = self.layer2_a(self.attention_layer(x))
+        x = self.layer3_a(x)
+        x = self.layer4_a(x)
+        x = self.avgpool_a(x)
+        x = torch.flatten(x, 1)
+        x = self.fc_a(x)
+
+        return x
+
+    def forward(self, x: Tensor, y: Tensor=None, mode: int = 0) -> Tensor:
+        if mode == 0:
+            return self._forward_impl(x, y)
+        elif mode == 1:
+            return self._forward_feature(x)
+        elif mode == 2:
+            return self._forward_trans(x, y)
+
+def _resnet(
+    arch: str,
+    block: Type[Union[BasicBlock, Bottleneck]],
+    layers: List[int],
+    pretrained: bool,
+    progress: bool,
+    **kwargs: Any,
+) -> ResNet:
+    model = ResNet(block, layers, **kwargs)
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+def _resappnet(
+    arch: str,
+    block: Type[Union[BasicBlock, Bottleneck]],
+    layers: List[int],
+    pretrained: bool,
+    progress: bool,
+    **kwargs: Any,
+) -> ResAppNet:
+    model = ResAppNet(block, layers, **kwargs)
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+def _resdisnet(
+    arch: str,
+    block: Type[Union[BasicBlock, Bottleneck]],
+    layers: List[int],
+    pretrained: bool,
+    progress: bool,
+    **kwargs: Any,
+) -> ResDisNet:
+    model = ResDisNet(block, layers, **kwargs)
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+def _resmdisnet(
+    arch: str,
+    block: Type[Union[BasicBlock, Bottleneck]],
+    layers: List[int],
+    pretrained: bool,
+    progress: bool,
+    **kwargs: Any,
+) -> ResMDisNet:
+    model = ResMDisNet(block, layers, **kwargs)
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+def _resdis2net(
+    arch: str,
+    block: Type[Union[BasicBlock, Bottleneck]],
+    layers: List[int],
+    pretrained: bool,
+    progress: bool,
+    **kwargs: Any,
+) -> ResDis2Net:
+    model = ResDis2Net(block, layers, **kwargs)
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+def _resdisattnet(
+    arch: str,
+    block: Type[Union[BasicBlock, Bottleneck]],
+    layers: List[int],
+    pretrained: bool,
+    progress: bool,
+    **kwargs: Any,
+) -> ResDisAttNet:
+    model = ResDisAttNet(block, layers, **kwargs)
+    if pretrained:
+        state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
+        model.load_state_dict(state_dict)
+    return model
+
+def resnet18(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNet-18 model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    return _resnet("resnet18", BasicBlock, [2, 2, 2, 2], pretrained, progress, **kwargs)
+
+
+def resnet34(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNet-34 model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    return _resnet("resnet34", BasicBlock, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+
+def resnet50(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNet-50 model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    return _resnet("resnet50", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+
+def resnet101(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNet-101 model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    return _resnet("resnet101", Bottleneck, [3, 4, 23, 3], pretrained, progress, **kwargs)
+
+
+def resnet152(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNet-152 model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    return _resnet("resnet152", Bottleneck, [3, 8, 36, 3], pretrained, progress, **kwargs)
+
+
+def resnext50_32x4d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNeXt-50 32x4d model from
+    `"Aggregated Residual Transformation for Deep Neural Networks" <https://arxiv.org/pdf/1611.05431.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["groups"] = 32
+    kwargs["width_per_group"] = 4
+    return _resnet("resnext50_32x4d", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+
+def resnext101_32x8d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNeXt-101 32x8d model from
+    `"Aggregated Residual Transformation for Deep Neural Networks" <https://arxiv.org/pdf/1611.05431.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["groups"] = 32
+    kwargs["width_per_group"] = 8
+    return _resnet("resnext101_32x8d", Bottleneck, [3, 4, 23, 3], pretrained, progress, **kwargs)
+
+
+def wide_resnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resnet("wide_resnet50_2", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+
+def wide_resnet101_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-101-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resnet("wide_resnet101_2", Bottleneck, [3, 4, 23, 3], pretrained, progress, **kwargs)
+
+
+def wide_resappnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resappnet("wide_resappnet50_2", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+def wide_resdisnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resdisnet("wide_resdisnet50_2", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+def wide_resmdisnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resmdisnet("wide_resmdisnet50_2", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+def wide_resdis2net50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resdis2net("wide_resdis2net50_2", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+def wide_resdisattnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""Wide ResNet-50-2 model from
+    `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_.
+    The model is the same as ResNet except for the bottleneck number of channels
+    which is twice larger in every block. The number of channels in outer 1x1
+    convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048
+    channels, and in Wide ResNet-50-2 has 2048-1024-2048.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    kwargs["width_per_group"] = 64 * 2
+    return _resdisattnet("wide_resdisattnet50_2", Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs)
+
+def resappnet34(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
+    r"""ResNet-34 model from
+    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+        progress (bool): If True, displays a progress bar of the download to stderr
+    """
+    return _resappnet("resappnet34", BasicBlock, [3, 4, 6, 3], pretrained, progress, **kwargs)

models/rnn_net.py

@@ -0,0 +1,99 @@
+"""
+Copyright Snap Inc. 2021. This sample code is made available by Snap Inc. for informational purposes only.
+No license, whether implied or otherwise, is granted in or to such code (including any rights to copy, modify,
+publish, distribute and/or commercialize such code), unless you have entered into a separate agreement for such rights.
+Such code is provided as-is, without warranty of any kind, express or implied, including any warranties of merchantability,
+title, fitness for a particular purpose, non-infringement, or that such code is free of defects, errors or viruses.
+In no event will Snap Inc. be liable for any damages or losses of any kind arising from the sample code or your use thereof.
+"""
+
+import torch
+import torch.nn as nn
+from torch.nn import init
+import torch.optim as optim
+
+
+class RNNModule(nn.Module):
+    def __init__(self,
+                 z_dim=64,
+                 h_dim=64,
+                 w_residual=0.2):
+        super(RNNModule, self).__init__()
+
+        self.z_dim = z_dim
+        self.h_dim = h_dim
+        self.w_residual = w_residual
+
+        self.enc_cell = nn.LSTMCell(z_dim, h_dim)
+        self.cell = nn.LSTMCell(z_dim, h_dim)
+        self.w = nn.Parameter(torch.FloatTensor(h_dim, h_dim))
+        self.b = nn.Parameter(torch.FloatTensor(h_dim))
+        self.fc1 = nn.Linear(h_dim * 2, z_dim)
+        self.relu = nn.ReLU()
+        self.fc2 = nn.Linear(z_dim, z_dim)
+
+        self.init_weights()
+
+    def init_optim(self, lr, beta1, beta2):
+        self.optim = optim.Adam(params=self.parameters(),
+                                lr=lr,
+                                betas=(beta1, beta2),
+                                weight_decay=0,
+                                eps=1e-8)
+
+    def init_weights(self):
+        for module in self.modules():
+            if (isinstance(module, nn.LSTMCell)):
+                for name, param in module.named_parameters():
+                    if ('weight_ih' in name) or ('weight_hh' in name):
+                        mul = param.shape[0] // 4
+                        for idx in range(4):
+                            init.orthogonal_(param[idx * mul:(idx + 1) * mul])
+                    elif 'bias' in name:
+                        param.data.fill_(0)
+            if (isinstance(module, nn.Linear)):
+                init.orthogonal_(module.weight)
+
+        nn.init.normal_(self.w, std=0.02)
+        self.b.data.fill_(0.0)
+
+    def forward(self, z, n_frame):
+
+        out = [z]
+        h_, c_ = self.enc_cell(z)
+        h = [h_]
+        c = [c_]
+        e = []
+        for i in range(n_frame - 1):
+            e_ = self.get_initial_state_z(z.shape[0])
+            h_, c_ = self.cell(e_, (h[-1], c[-1]))
+            mul = torch.matmul(h_, self.w) + self.b
+            mul = torch.tanh(mul)
+            e.append(e_)
+            h.append(h_)
+            c.append(c_)
+            out_ = out[-1] + self.w_residual * mul
+            out.append(out_)
+
+        out = [item.unsqueeze(1) for item in out]
+
+        out = torch.cat(out, dim=1).view(-1, self.z_dim)
+
+        e = [item.unsqueeze(1) for item in e]
+        e = torch.cat(e, dim=1).view(-1, self.z_dim)
+
+        hh = h[1:]
+        hh = [item.unsqueeze(1) for item in hh]
+        hh = torch.cat(hh, dim=1).view(-1, self.h_dim)
+
+        cc = c[1:]
+        cc = [item.unsqueeze(1) for item in cc]
+        cc = torch.cat(cc, dim=1).view(-1, self.h_dim)
+
+        hc = torch.cat((hh, cc), dim=1)
+        e_rec = self.fc2(self.relu(self.fc1(hc)))
+
+        return out, e, e_rec
+
+    def get_initial_state_z(self, batchSize):
+        return torch.cuda.FloatTensor(batchSize, self.z_dim).normal_()

models/sample_model.py

@@ -0,0 +1,243 @@
+import torch
+from .base_model import BaseModel
+from . import networks
+from . import lmcode_networks
+from . import diy_networks
+from . import resnet
+
+import dnnlib
+import legacy
+import torch.nn.functional as F
+import numpy as np
+import random
+import os
+
+from . import rnn_net
+
+def make_transform(translate, angle):
+    m = np.eye(3)
+    s = np.sin(angle/360.0*np.pi*2)
+    c = np.cos(angle/360.0*np.pi*2)
+    m[0][0] = c
+    m[0][1] = s
+    m[0][2] = translate[0]
+    m[1][0] = -s
+    m[1][1] = c
+    m[1][2] = translate[1]
+    return m
+
+class SampleModel(BaseModel):
+    """ This class implements the pix2pix model, for learning a mapping from input images to output images given paired data.
+
+    The model training requires '--dataset_mode aligned' dataset.
+    By default, it uses a '--netG unet256' U-Net generator,
+    a '--netD basic' discriminator (PatchGAN),
+    and a '--gan_mode' vanilla GAN loss (the cross-entropy objective used in the orignal GAN paper).
+
+    pix2pix paper: https://arxiv.org/pdf/1611.07004.pdf
+    """
+    @staticmethod
+    def modify_commandline_options(parser, is_train=True):
+        """Add new dataset-specific options, and rewrite default values for existing options.
+
+        Parameters:
+            parser          -- original option parser
+            is_train (bool) -- whether training phase or test phase. You can use this flag to add training-specific or test-specific options.
+
+        Returns:
+            the modified parser.
+
+        For pix2pix, we do not use image buffer
+        The training objective is: GAN Loss + lambda_L1 * ||G(A)-B||_1
+        By default, we use vanilla GAN loss, UNet with batchnorm, and aligned datasets.
+        """
+        # changing the default values to match the pix2pix paper (https://phillipi.github.io/pix2pix/)
+        parser.set_defaults(norm='batch', netG='unet_256', dataset_mode='noiseshufflevideo', num_test = 32)
+        parser.add_argument('--pose_path', type=str, default='', help='path for pose net')
+        parser.add_argument('--rnn_path', type=str, default='', help='path for rnn net')
+        parser.add_argument('--n_frames_G', type=int, default=60)
+        parser.add_argument('--w_residual', type=float, default=0.2)
+        parser.add_argument('--num_point', type=int, default=14)
+        parser.add_argument('--model_names', type=str, default='')
+        if is_train:
+            parser.set_defaults(pool_size=0, gan_mode='vanilla')
+            parser.add_argument('--lambda_L1', type=float, default=1.0, help='weight for L1 loss')
+
+        return parser
+
+    def __init__(self, opt):
+        """Initialize the pix2pix class.
+
+        Parameters:
+            opt (Option class)-- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        BaseModel.__init__(self, opt)
+        # specify the training losses you want to print out. The training/test scripts will call <BaseModel.get_current_losses>
+        self.loss_names = ['G_L1', 'G_VGG', 'G_W']
+        # specify the images you want to save/display. The training/test scripts will call <BaseModel.get_current_visuals>
+        self.visual_names = ['real_vid_B', 'fake_vid_AR', 'fake_vid_BR', 'fake_vid_AR1', 'fake_vid_BR1', 'fake_vid_AR2', 'fake_vid_BR2', 'fake_vid_AB', 'fake_vid_B', 'fake_vid']
+        # specify the models you want to save to the disk. The training/test scripts will call <BaseModel.save_networks> and <BaseModel.load_networks>
+        if self.isTrain:
+            self.model_names = ['FE']
+        else:  # during test time, only load G
+            self.model_names = ['FE']
+        if opt.model_names != '':
+            str_models = opt.model_names.split(',')
+            self.model_names = []
+            for str_model in str_models:
+                self.model_names.append(str_model)
+        # define networks (both generator and discriminator)
+        with dnnlib.util.open_url(opt.network_pkl) as f:
+            self.netG = legacy.load_network_pkl(f)['G_ema'].eval().to(self.gpu_ids[0])  # type: ignore
+
+        lm_path = 'pretrained_models/wing.ckpt'
+        self.netFE_lm = lmcode_networks.FAN(fname_pretrained=lm_path).eval().to(self.gpu_ids[0])
+        self.netFE_pose = diy_networks._resposenet(num_point=opt.num_point).eval().to(self.gpu_ids[0])
+        if opt.pose_path != '':
+            self.netFE_pose.load_state_dict(torch.load(opt.pose_path))
+
+        self.netFE = resnet.wide_resdisnet50_2(num_classes=512 * 16).to(self.gpu_ids[0])
+        self.netFE = networks.init_net(self.netFE, opt.init_type, opt.init_gain, self.gpu_ids)
+
+        self.netR = rnn_net.RNNModule(w_residual = opt.w_residual).to(self.gpu_ids[0])
+        if opt.rnn_path != '':
+            self.netR.load_state_dict(torch.load(opt.rnn_path))
+        self.n_frames_G = opt.n_frames_G
+        self.style_gan_size = 8
+
+        self.m_zero = make_transform((0.0,0.0),(0.0))
+        self.count = 0
+
+
+    def set_input(self, input):
+        """Unpack input data from the dataloader and perform necessary pre-processing steps.
+
+        Parameters:
+            input (dict): include the data itself and its metadata information.
+
+        The option 'direction' can be used to swap images in domain A and domain B.
+        """
+        self.real_Bs = input['A'].to(self.device)
+        self.image_paths = input['A_paths']
+        self.count += 1
+        self.image_paths[0] = os.path.split(self.image_paths[0])[0] + '/' + str(self.count) + '.png'
+
+        real_v_list = []
+        with torch.no_grad():
+            for i in range(self.real_Bs.shape[1]):
+                real_v_list.append(self.netFE_pose(self.netFE_lm.get_heatmap(self.real_Bs[:,i,...], b_preprocess=False), mode = 1).unsqueeze(1))
+
+        self.real_v = torch.cat(real_v_list, 1).detach()
+
+        self.real_z = input['B'].to(self.device)
+
+    def forward(self):
+        """Run forward pass; called by both functions <optimize_parameters> and <test>."""
+
+        self.real_A_w = self.netG.mapping(self.real_z, None)
+        self.real_A = self.netG.synthesis(self.real_A_w, noise_mode='const').detach().clamp(-1, 1)
+        if self.real_A.shape[2] != 256:
+            self.real_A = F.interpolate(self.real_A, size=(256, 256), mode='area')
+        self.real_A_heat = self.netFE_lm.get_heatmap(self.real_A, b_preprocess=False)
+        self.real_A_pose = self.netFE_pose(self.real_A_heat, mode=1).detach()
+        self.real_A_app = self.netFE(self.real_A, mode=1).detach()
+        self.fake_A_w = self.netFE(self.real_A_app, self.real_A_pose, mode=2).view(-1, 16, 512)
+        self.fake_A = self.netG.synthesis(self.fake_A_w, noise_mode='const')  # G(A)
+
+        self.real_B_app = self.netFE(self.real_Bs[:, 0, ...], mode=1)
+
+        x_fake, self.rand_in, self.rand_rec = self.netR(self.real_v[:, 0].view(self.opt.batch_size, self.style_gan_size * self.style_gan_size), self.n_frames_G)
+        x_fake = x_fake.view(self.opt.batch_size, self.n_frames_G, 1, self.style_gan_size,
+                             self.style_gan_size)
+
+        self.real_R_pose = x_fake.clone()
+
+        x_fake, self.rand_in, self.rand_rec = self.netR(self.real_v[:, 29].view(self.opt.batch_size, self.style_gan_size * self.style_gan_size), self.n_frames_G)
+        x_fake = x_fake.view(self.opt.batch_size, self.n_frames_G, 1, self.style_gan_size,
+                             self.style_gan_size)
+
+        self.real_R1_pose = x_fake.clone()
+
+        x_fake, self.rand_in, self.rand_rec = self.netR(self.real_v[:, 59].view(self.opt.batch_size, self.style_gan_size * self.style_gan_size), self.n_frames_G)
+        x_fake = x_fake.view(self.opt.batch_size, self.n_frames_G, 1, self.style_gan_size,
+                             self.style_gan_size)
+
+        self.real_R2_pose = x_fake.clone()
+
+        x_fake_A, self.rand_in, self.rand_rec = self.netR(self.real_A_pose.view(self.opt.batch_size, self.style_gan_size * self.style_gan_size), self.n_frames_G)
+        x_fake_A = x_fake_A.view(self.opt.batch_size, self.n_frames_G, 1, self.style_gan_size,
+                             self.style_gan_size)
+
+        self.real_R_pose_A = x_fake_A
+
+        if hasattr(self.netG.synthesis, 'input'):
+            self.netG.synthesis.input.transform.copy_(torch.from_numpy(self.m_zero))
+
+        self.real_A_list = []
+        self.real_B_list = []
+        self.fake_AR_list = []
+        self.fake_BR_list = []
+        self.fake_AR1_list = []
+        self.fake_BR1_list = []
+        self.fake_AR2_list = []
+        self.fake_BR2_list = []
+        self.fake_AB_list = []
+        self.fake_B_list = []
+        # for i in range(self.real_Bs.shape[1]):
+        self.real_B_app = self.netFE(self.real_Bs[:,0,...], mode=1)
+        for i in range(self.n_frames_G):
+            self.real_B = self.real_Bs[:,i,...]
+            if self.real_B.shape[2] != 256:
+                self.real_B = F.interpolate(self.real_B, size=(256, 256), mode='area')
+
+            self.fake_AR_w = self.netFE(self.real_A_app, self.real_R_pose[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_BR_w = self.netFE(self.real_B_app, self.real_R_pose[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_AR1_w = self.netFE(self.real_A_app, self.real_R1_pose[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_BR1_w = self.netFE(self.real_B_app, self.real_R1_pose[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_AR2_w = self.netFE(self.real_A_app, self.real_R2_pose[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_BR2_w = self.netFE(self.real_B_app, self.real_R2_pose[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_AB_w = self.netFE(self.real_A_app, self.real_R_pose_A[:,i,...], mode=2).view(-1, 16, 512)
+            self.fake_B_w = self.netFE(self.real_B_app, self.real_R_pose_A[:,i,...], mode=2).view(-1, 16, 512)
+
+            self.fake_AR = self.netG.synthesis(self.fake_AR_w, noise_mode='const')  # G(A)
+            self.fake_BR = self.netG.synthesis(self.fake_BR_w, noise_mode='const')  # G(A)
+            self.fake_AR1 = self.netG.synthesis(self.fake_AR1_w, noise_mode='const')  # G(A)
+            self.fake_BR1 = self.netG.synthesis(self.fake_BR1_w, noise_mode='const')  # G(A)
+            self.fake_AR2 = self.netG.synthesis(self.fake_AR2_w, noise_mode='const')  # G(A)
+            self.fake_BR2 = self.netG.synthesis(self.fake_BR2_w, noise_mode='const')  # G(A)
+            self.fake_AB = self.netG.synthesis(self.fake_AB_w, noise_mode='const')  # G(A)
+            self.fake_B = self.netG.synthesis(self.fake_B_w, noise_mode='const')  # G(A)
+
+            self.real_A_list.append(self.real_A.clamp(-1, 1))
+            self.real_B_list.append(self.real_B.clamp(-1, 1))
+            self.fake_AR_list.append(self.fake_AR.clamp(-1, 1))
+            self.fake_BR_list.append(self.fake_BR.clamp(-1, 1))
+            self.fake_AR1_list.append(self.fake_AR1.clamp(-1, 1))
+            self.fake_BR1_list.append(self.fake_BR1.clamp(-1, 1))
+            self.fake_AR2_list.append(self.fake_AR2.clamp(-1, 1))
+            self.fake_BR2_list.append(self.fake_BR2.clamp(-1, 1))
+            self.fake_AB_list.append(self.fake_AB.clamp(-1, 1))
+            self.fake_B_list.append(self.fake_B.clamp(-1, 1))
+            
+    def optimize_parameters(self):
+        self.forward()                   # compute fake images: G(A)
+        # update G
+        self.optimizer_FE.zero_grad()        # set G's gradients to zero
+        self.backward_G()                   # calculate graidents for G
+        self.optimizer_FE.step()  # udpate G's weights
+
+    def compute_visuals(self):
+
+        self.real_vid_A = torch.cat(self.real_A_list, 0)
+        self.real_vid_B = torch.cat(self.real_B_list, 0)
+        self.fake_vid_AR = torch.cat(self.fake_AR_list, 0)
+        self.fake_vid_BR = torch.cat(self.fake_BR_list, 0)
+        self.fake_vid_AR1 = torch.cat(self.fake_AR1_list, 0)
+        self.fake_vid_BR1 = torch.cat(self.fake_BR1_list, 0)
+        self.fake_vid_AR2 = torch.cat(self.fake_AR2_list, 0)
+        self.fake_vid_BR2 = torch.cat(self.fake_BR2_list, 0)
+        self.fake_vid_AB = torch.cat(self.fake_AB_list, 0)
+        self.fake_vid_B = torch.cat(self.fake_B_list, 0)
+
+        self.fake_vid = torch.cat([torch.cat([self.fake_vid_BR, self.fake_vid_BR1, self.fake_vid_BR2, self.fake_vid_B], dim = 3), torch.cat([self.fake_vid_AR, self.fake_vid_AR1, self.fake_vid_AR2, self.fake_vid_AB], dim = 3)], dim = 2)
+

options/__init__.py

@@ -0,0 +1 @@
+"""This package options includes option modules: training options, test options, and basic options (used in both training and test)."""

options/base_options.py

@@ -0,0 +1,138 @@
+import argparse
+import os
+from util import util
+import torch
+import models
+import data
+
+
+class BaseOptions():
+    """This class defines options used during both training and test time.
+
+    It also implements several helper functions such as parsing, printing, and saving the options.
+    It also gathers additional options defined in <modify_commandline_options> functions in both dataset class and model class.
+    """
+
+    def __init__(self):
+        """Reset the class; indicates the class hasn't been initailized"""
+        self.initialized = False
+
+    def initialize(self, parser):
+        """Define the common options that are used in both training and test."""
+        # basic parameters
+        parser.add_argument('--dataroot', required=True, help='path to images (should have subfolders trainA, trainB, valA, valB, etc)')
+        parser.add_argument('--name', type=str, default='experiment_name', help='name of the experiment. It decides where to store samples and models')
+        parser.add_argument('--network_pkl', type=str, help='Network pickle filename')
+        parser.add_argument('--use_wandb', action='store_true', help='use wandb')
+        parser.add_argument('--gpu_ids', type=str, default='0', help='gpu ids: e.g. 0  0,1,2, 0,2. use -1 for CPU')
+        parser.add_argument('--checkpoints_dir', type=str, default='./checkpoints', help='models are saved here')
+        # model parameters
+        parser.add_argument('--model', type=str, default='cycle_gan', help='chooses which model to use. [cycle_gan | pix2pix | test | colorization]')
+        parser.add_argument('--input_nc', type=int, default=3, help='# of input image channels: 3 for RGB and 1 for grayscale')
+        parser.add_argument('--output_nc', type=int, default=3, help='# of output image channels: 3 for RGB and 1 for grayscale')
+        parser.add_argument('--ngf', type=int, default=64, help='# of gen filters in the last conv layer')
+        parser.add_argument('--ndf', type=int, default=64, help='# of discrim filters in the first conv layer')
+        parser.add_argument('--netD', type=str, default='basic', help='specify discriminator architecture [basic | n_layers | pixel]. The basic model is a 70x70 PatchGAN. n_layers allows you to specify the layers in the discriminator')
+        parser.add_argument('--netG', type=str, default='resnet_9blocks', help='specify generator architecture [resnet_9blocks | resnet_6blocks | unet_256 | unet_128]')
+        parser.add_argument('--n_layers_D', type=int, default=3, help='only used if netD==n_layers')
+        parser.add_argument('--norm', type=str, default='instance', help='instance normalization or batch normalization [instance | batch | none]')
+        parser.add_argument('--init_type', type=str, default='normal', help='network initialization [normal | xavier | kaiming | orthogonal]')
+        parser.add_argument('--init_gain', type=float, default=0.02, help='scaling factor for normal, xavier and orthogonal.')
+        parser.add_argument('--no_dropout', action='store_true', help='no dropout for the generator')
+        # dataset parameters
+        parser.add_argument('--dataset_mode', type=str, default='unaligned', help='chooses how datasets are loaded. [unaligned | aligned | single | colorization]')
+        parser.add_argument('--direction', type=str, default='AtoB', help='AtoB or BtoA')
+        parser.add_argument('--serial_batches', action='store_true', help='if true, takes images in order to make batches, otherwise takes them randomly')
+        parser.add_argument('--num_threads', default=4, type=int, help='# threads for loading data')
+        parser.add_argument('--batch_size', type=int, default=1, help='input batch size')
+        parser.add_argument('--load_size', type=int, default=256, help='scale images to this size')
+        parser.add_argument('--crop_size', type=int, default=256, help='then crop to this size')
+        parser.add_argument('--max_dataset_size', type=int, default=20000, help='Maximum number of samples allowed per dataset. If the dataset directory contains more than max_dataset_size, only a subset is loaded.')
+        parser.add_argument('--preprocess', type=str, default='resize_and_crop', help='scaling and cropping of images at load time [resize_and_crop | crop | scale_width | scale_width_and_crop | none]')
+        parser.add_argument('--no_flip', action='store_true', help='if specified, do not flip the images for data augmentation')
+        parser.add_argument('--display_winsize', type=int, default=256, help='display window size for both visdom and HTML')
+        # additional parameters
+        parser.add_argument('--epoch', type=str, default='latest', help='which epoch to load? set to latest to use latest cached model')
+        parser.add_argument('--load_iter', type=int, default='0', help='which iteration to load? if load_iter > 0, the code will load models by iter_[load_iter]; otherwise, the code will load models by [epoch]')
+        parser.add_argument('--verbose', action='store_true', help='if specified, print more debugging information')
+        parser.add_argument('--suffix', default='', type=str, help='customized suffix: opt.name = opt.name + suffix: e.g., {model}_{netG}_size{load_size}')
+        self.initialized = True
+        return parser
+
+    def gather_options(self):
+        """Initialize our parser with basic options(only once).
+        Add additional model-specific and dataset-specific options.
+        These options are defined in the <modify_commandline_options> function
+        in model and dataset classes.
+        """
+        if not self.initialized:  # check if it has been initialized
+            parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+            parser = self.initialize(parser)
+
+        # get the basic options
+        opt, _ = parser.parse_known_args()
+
+        # modify model-related parser options
+        model_name = opt.model
+        model_option_setter = models.get_option_setter(model_name)
+        parser = model_option_setter(parser, self.isTrain)
+        opt, _ = parser.parse_known_args()  # parse again with new defaults
+
+        # modify dataset-related parser options
+        dataset_name = opt.dataset_mode
+        dataset_option_setter = data.get_option_setter(dataset_name)
+        parser = dataset_option_setter(parser, self.isTrain)
+
+        # save and return the parser
+        self.parser = parser
+        return parser.parse_args()
+
+    def print_options(self, opt):
+        """Print and save options
+
+        It will print both current options and default values(if different).
+        It will save options into a text file / [checkpoints_dir] / opt.txt
+        """
+        message = ''
+        message += '----------------- Options ---------------\n'
+        for k, v in sorted(vars(opt).items()):
+            comment = ''
+            default = self.parser.get_default(k)
+            if v != default:
+                comment = '\t[default: %s]' % str(default)
+            message += '{:>25}: {:<30}{}\n'.format(str(k), str(v), comment)
+        message += '----------------- End -------------------'
+        print(message)
+
+        # save to the disk
+        expr_dir = os.path.join(opt.checkpoints_dir, opt.name)
+        util.mkdirs(expr_dir)
+        file_name = os.path.join(expr_dir, '{}_opt.txt'.format(opt.phase))
+        with open(file_name, 'wt') as opt_file:
+            opt_file.write(message)
+            opt_file.write('\n')
+
+    def parse(self):
+        """Parse our options, create checkpoints directory suffix, and set up gpu device."""
+        opt = self.gather_options()
+        opt.isTrain = self.isTrain   # train or test
+
+        # process opt.suffix
+        if opt.suffix:
+            suffix = ('_' + opt.suffix.format(**vars(opt))) if opt.suffix != '' else ''
+            opt.name = opt.name + suffix
+
+        self.print_options(opt)
+
+        # set gpu ids
+        str_ids = opt.gpu_ids.split(',')
+        opt.gpu_ids = []
+        for str_id in str_ids:
+            id = int(str_id)
+            if id >= 0:
+                opt.gpu_ids.append(id)
+        if len(opt.gpu_ids) > 0:
+            torch.cuda.set_device(opt.gpu_ids[0])
+
+        self.opt = opt
+        return self.opt

options/test_options.py

@@ -0,0 +1,23 @@
+from .base_options import BaseOptions
+
+
+class TestOptions(BaseOptions):
+    """This class includes test options.
+
+    It also includes shared options defined in BaseOptions.
+    """
+
+    def initialize(self, parser):
+        parser = BaseOptions.initialize(self, parser)  # define shared options
+        parser.add_argument('--results_dir', type=str, default='./results/', help='saves results here.')
+        parser.add_argument('--aspect_ratio', type=float, default=1.0, help='aspect ratio of result images')
+        parser.add_argument('--phase', type=str, default='test', help='train, val, test, etc')
+        # Dropout and Batchnorm has different behavioir during training and test.
+        parser.add_argument('--eval', action='store_true', help='use eval mode during test time.')
+        parser.add_argument('--num_test', type=int, default=50, help='how many test images to run')
+        # rewrite devalue values
+        parser.set_defaults(model='test')
+        # To avoid cropping, the load_size should be the same as crop_size
+        parser.set_defaults(load_size=parser.get_default('crop_size'))
+        self.isTrain = False
+        return parser

options/train_options.py

@@ -0,0 +1,43 @@
+from .base_options import BaseOptions
+
+
+class TrainOptions(BaseOptions):
+    """This class includes training options.
+
+    It also includes shared options defined in BaseOptions.
+    """
+
+    def initialize(self, parser):
+        parser = BaseOptions.initialize(self, parser)
+        # visdom and HTML visualization parameters
+        parser.add_argument('--display_freq', type=int, default=400, help='frequency of showing training results on screen')
+        parser.add_argument('--display_ncols', type=int, default=4, help='if positive, display all images in a single visdom web panel with certain number of images per row.')
+        parser.add_argument('--display_id', type=int, default=0, help='window id of the web display')
+        parser.add_argument('--display_server', type=str, default="http://localhost", help='visdom server of the web display')
+        parser.add_argument('--display_env', type=str, default='main', help='visdom display environment name (default is "main")')
+        parser.add_argument('--display_port', type=int, default=8097, help='visdom port of the web display')
+        parser.add_argument('--update_html_freq', type=int, default=1000, help='frequency of saving training results to html')
+        parser.add_argument('--print_freq', type=int, default=100, help='frequency of showing training results on console')
+        parser.add_argument('--no_html', action='store_true', help='do not save intermediate training results to [opt.checkpoints_dir]/[opt.name]/web/')
+        # network saving and loading parameters
+        parser.add_argument('--save_latest_freq', type=int, default=5000, help='frequency of saving the latest results')
+        parser.add_argument('--save_epoch_freq', type=int, default=5, help='frequency of saving checkpoints at the end of epochs')
+        parser.add_argument('--save_by_iter', action='store_true', help='whether saves model by iteration')
+        parser.add_argument('--continue_train', action='store_true', help='continue training: load the latest model')
+        parser.add_argument('--epoch_count', type=int, default=1, help='the starting epoch count, we save the model by <epoch_count>, <epoch_count>+<save_latest_freq>, ...')
+        parser.add_argument('--phase', type=str, default='train', help='train, val, test, etc')
+        # training parameters
+        parser.add_argument('--n_epochs', type=int, default=50, help='number of epochs with the initial learning rate')
+        parser.add_argument('--n_epochs_decay', type=int, default=50, help='number of epochs to linearly decay learning rate to zero')
+        parser.add_argument('--beta1', type=float, default=0.5, help='momentum term of adam')
+        parser.add_argument('--lr', type=float, default=0.0002, help='initial learning rate for adam')
+        parser.add_argument('--gan_mode', type=str, default='lsgan', help='the type of GAN objective. [vanilla| lsgan | wgangp]. vanilla GAN loss is the cross-entropy objective used in the original GAN paper.')
+        parser.add_argument('--pool_size', type=int, default=50, help='the size of image buffer that stores previously generated images')
+        parser.add_argument('--lr_policy', type=str, default='linear', help='learning rate policy. [linear | step | plateau | cosine]')
+        parser.add_argument('--lr_decay_iters', type=int, default=50, help='multiply by a gamma every lr_decay_iters iterations')
+
+        parser.add_argument('--epoch_gan', type=int, default=0,
+                            help='finetune the whole model with GAN loss finally')
+
+        self.isTrain = True
+        return parser

pretrained_models/motion_net.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59d0dc583c1979aba667b74761f704f073427400448b2347f2766bd0e317095b
+size 336251

pretrained_models/network-snapshot-005000.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f4b61e718d80495ad0864e4aa1107e4b731a423baedc5ebb0d5eb813fa990f0
+size 222508337

pretrained_models/wing.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbfd137307a4c7debd5c283b9b0ce539466cee417ac0a155e184d857f9f2899c
+size 193670248

torch_utils/__init__.py

@@ -0,0 +1,9 @@
+# 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

@@ -0,0 +1,157 @@
+# 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 (x86)/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/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

@@ -0,0 +1,266 @@
+# 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

@@ -0,0 +1,9 @@
+# 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/ops/bias_act.cpp

@@ -0,0 +1,99 @@
+// 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.
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "bias_act.h"
+
+//------------------------------------------------------------------------
+
+static bool has_same_layout(torch::Tensor x, torch::Tensor y)
+{
+    if (x.dim() != y.dim())
+        return false;
+    for (int64_t i = 0; i < x.dim(); i++)
+    {
+        if (x.size(i) != y.size(i))
+            return false;
+        if (x.size(i) >= 2 && x.stride(i) != y.stride(i))
+            return false;
+    }
+    return true;
+}
+
+//------------------------------------------------------------------------
+
+static torch::Tensor bias_act(torch::Tensor x, torch::Tensor b, torch::Tensor xref, torch::Tensor yref, torch::Tensor dy, int grad, int dim, int act, float alpha, float gain, float clamp)
+{
+    // Validate arguments.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    TORCH_CHECK(b.numel() == 0 || (b.dtype() == x.dtype() && b.device() == x.device()), "b must have the same dtype and device as x");
+    TORCH_CHECK(xref.numel() == 0 || (xref.sizes() == x.sizes() && xref.dtype() == x.dtype() && xref.device() == x.device()), "xref must have the same shape, dtype, and device as x");
+    TORCH_CHECK(yref.numel() == 0 || (yref.sizes() == x.sizes() && yref.dtype() == x.dtype() && yref.device() == x.device()), "yref must have the same shape, dtype, and device as x");
+    TORCH_CHECK(dy.numel() == 0 || (dy.sizes() == x.sizes() && dy.dtype() == x.dtype() && dy.device() == x.device()), "dy must have the same dtype and device as x");
+    TORCH_CHECK(x.numel() <= INT_MAX, "x is too large");
+    TORCH_CHECK(b.dim() == 1, "b must have rank 1");
+    TORCH_CHECK(b.numel() == 0 || (dim >= 0 && dim < x.dim()), "dim is out of bounds");
+    TORCH_CHECK(b.numel() == 0 || b.numel() == x.size(dim), "b has wrong number of elements");
+    TORCH_CHECK(grad >= 0, "grad must be non-negative");
+
+    // Validate layout.
+    TORCH_CHECK(x.is_non_overlapping_and_dense(), "x must be non-overlapping and dense");
+    TORCH_CHECK(b.is_contiguous(), "b must be contiguous");
+    TORCH_CHECK(xref.numel() == 0 || has_same_layout(xref, x), "xref must have the same layout as x");
+    TORCH_CHECK(yref.numel() == 0 || has_same_layout(yref, x), "yref must have the same layout as x");
+    TORCH_CHECK(dy.numel() == 0 || has_same_layout(dy, x), "dy must have the same layout as x");
+
+    // Create output tensor.
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    torch::Tensor y = torch::empty_like(x);
+    TORCH_CHECK(has_same_layout(y, x), "y must have the same layout as x");
+
+    // Initialize CUDA kernel parameters.
+    bias_act_kernel_params p;
+    p.x     = x.data_ptr();
+    p.b     = (b.numel()) ? b.data_ptr() : NULL;
+    p.xref  = (xref.numel()) ? xref.data_ptr() : NULL;
+    p.yref  = (yref.numel()) ? yref.data_ptr() : NULL;
+    p.dy    = (dy.numel()) ? dy.data_ptr() : NULL;
+    p.y     = y.data_ptr();
+    p.grad  = grad;
+    p.act   = act;
+    p.alpha = alpha;
+    p.gain  = gain;
+    p.clamp = clamp;
+    p.sizeX = (int)x.numel();
+    p.sizeB = (int)b.numel();
+    p.stepB = (b.numel()) ? (int)x.stride(dim) : 1;
+
+    // Choose CUDA kernel.
+    void* kernel;
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&]
+    {
+        kernel = choose_bias_act_kernel<scalar_t>(p);
+    });
+    TORCH_CHECK(kernel, "no CUDA kernel found for the specified activation func");
+
+    // Launch CUDA kernel.
+    p.loopX = 4;
+    int blockSize = 4 * 32;
+    int gridSize = (p.sizeX - 1) / (p.loopX * blockSize) + 1;
+    void* args[] = {&p};
+    AT_CUDA_CHECK(cudaLaunchKernel(kernel, gridSize, blockSize, args, 0, at::cuda::getCurrentCUDAStream()));
+    return y;
+}
+
+//------------------------------------------------------------------------
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("bias_act", &bias_act);
+}
+
+//------------------------------------------------------------------------

torch_utils/ops/bias_act.cu

@@ -0,0 +1,173 @@
+// 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.
+
+#include <c10/util/Half.h>
+#include "bias_act.h"
+
+//------------------------------------------------------------------------
+// Helpers.
+
+template <class T> struct InternalType;
+template <> struct InternalType<double>     { typedef double scalar_t; };
+template <> struct InternalType<float>      { typedef float  scalar_t; };
+template <> struct InternalType<c10::Half>  { typedef float  scalar_t; };
+
+//------------------------------------------------------------------------
+// CUDA kernel.
+
+template <class T, int A>
+__global__ void bias_act_kernel(bias_act_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+    int G                 = p.grad;
+    scalar_t alpha        = (scalar_t)p.alpha;
+    scalar_t gain         = (scalar_t)p.gain;
+    scalar_t clamp        = (scalar_t)p.clamp;
+    scalar_t one          = (scalar_t)1;
+    scalar_t two          = (scalar_t)2;
+    scalar_t expRange     = (scalar_t)80;
+    scalar_t halfExpRange = (scalar_t)40;
+    scalar_t seluScale    = (scalar_t)1.0507009873554804934193349852946;
+    scalar_t seluAlpha    = (scalar_t)1.6732632423543772848170429916717;
+
+    // Loop over elements.
+    int xi = blockIdx.x * p.loopX * blockDim.x + threadIdx.x;
+    for (int loopIdx = 0; loopIdx < p.loopX && xi < p.sizeX; loopIdx++, xi += blockDim.x)
+    {
+        // Load.
+        scalar_t x = (scalar_t)((const T*)p.x)[xi];
+        scalar_t b = (p.b) ? (scalar_t)((const T*)p.b)[(xi / p.stepB) % p.sizeB] : 0;
+        scalar_t xref = (p.xref) ? (scalar_t)((const T*)p.xref)[xi] : 0;
+        scalar_t yref = (p.yref) ? (scalar_t)((const T*)p.yref)[xi] : 0;
+        scalar_t dy = (p.dy) ? (scalar_t)((const T*)p.dy)[xi] : one;
+        scalar_t yy = (gain != 0) ? yref / gain : 0;
+        scalar_t y = 0;
+
+        // Apply bias.
+        ((G == 0) ? x : xref) += b;
+
+        // linear
+        if (A == 1)
+        {
+            if (G == 0) y = x;
+            if (G == 1) y = x;
+        }
+
+        // relu
+        if (A == 2)
+        {
+            if (G == 0) y = (x > 0) ? x : 0;
+            if (G == 1) y = (yy > 0) ? x : 0;
+        }
+
+        // lrelu
+        if (A == 3)
+        {
+            if (G == 0) y = (x > 0) ? x : x * alpha;
+            if (G == 1) y = (yy > 0) ? x : x * alpha;
+        }
+
+        // tanh
+        if (A == 4)
+        {
+            if (G == 0) { scalar_t c = exp(x); scalar_t d = one / c; y = (x < -expRange) ? -one : (x > expRange) ? one : (c - d) / (c + d); }
+            if (G == 1) y = x * (one - yy * yy);
+            if (G == 2) y = x * (one - yy * yy) * (-two * yy);
+        }
+
+        // sigmoid
+        if (A == 5)
+        {
+            if (G == 0) y = (x < -expRange) ? 0 : one / (exp(-x) + one);
+            if (G == 1) y = x * yy * (one - yy);
+            if (G == 2) y = x * yy * (one - yy) * (one - two * yy);
+        }
+
+        // elu
+        if (A == 6)
+        {
+            if (G == 0) y = (x >= 0) ? x : exp(x) - one;
+            if (G == 1) y = (yy >= 0) ? x : x * (yy + one);
+            if (G == 2) y = (yy >= 0) ? 0 : x * (yy + one);
+        }
+
+        // selu
+        if (A == 7)
+        {
+            if (G == 0) y = (x >= 0) ? seluScale * x : (seluScale * seluAlpha) * (exp(x) - one);
+            if (G == 1) y = (yy >= 0) ? x * seluScale : x * (yy + seluScale * seluAlpha);
+            if (G == 2) y = (yy >= 0) ? 0 : x * (yy + seluScale * seluAlpha);
+        }
+
+        // softplus
+        if (A == 8)
+        {
+            if (G == 0) y = (x > expRange) ? x : log(exp(x) + one);
+            if (G == 1) y = x * (one - exp(-yy));
+            if (G == 2) { scalar_t c = exp(-yy); y = x * c * (one - c); }
+        }
+
+        // swish
+        if (A == 9)
+        {
+            if (G == 0)
+                y = (x < -expRange) ? 0 : x / (exp(-x) + one);
+            else
+            {
+                scalar_t c = exp(xref);
+                scalar_t d = c + one;
+                if (G == 1)
+                    y = (xref > halfExpRange) ? x : x * c * (xref + d) / (d * d);
+                else
+                    y = (xref > halfExpRange) ? 0 : x * c * (xref * (two - d) + two * d) / (d * d * d);
+                yref = (xref < -expRange) ? 0 : xref / (exp(-xref) + one) * gain;
+            }
+        }
+
+        // Apply gain.
+        y *= gain * dy;
+
+        // Clamp.
+        if (clamp >= 0)
+        {
+            if (G == 0)
+                y = (y > -clamp & y < clamp) ? y : (y >= 0) ? clamp : -clamp;
+            else
+                y = (yref > -clamp & yref < clamp) ? y : 0;
+        }
+
+        // Store.
+        ((T*)p.y)[xi] = (T)y;
+    }
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> void* choose_bias_act_kernel(const bias_act_kernel_params& p)
+{
+    if (p.act == 1) return (void*)bias_act_kernel<T, 1>;
+    if (p.act == 2) return (void*)bias_act_kernel<T, 2>;
+    if (p.act == 3) return (void*)bias_act_kernel<T, 3>;
+    if (p.act == 4) return (void*)bias_act_kernel<T, 4>;
+    if (p.act == 5) return (void*)bias_act_kernel<T, 5>;
+    if (p.act == 6) return (void*)bias_act_kernel<T, 6>;
+    if (p.act == 7) return (void*)bias_act_kernel<T, 7>;
+    if (p.act == 8) return (void*)bias_act_kernel<T, 8>;
+    if (p.act == 9) return (void*)bias_act_kernel<T, 9>;
+    return NULL;
+}
+
+//------------------------------------------------------------------------
+// Template specializations.
+
+template void* choose_bias_act_kernel<double>       (const bias_act_kernel_params& p);
+template void* choose_bias_act_kernel<float>        (const bias_act_kernel_params& p);
+template void* choose_bias_act_kernel<c10::Half>    (const bias_act_kernel_params& p);
+
+//------------------------------------------------------------------------

torch_utils/ops/bias_act.h

@@ -0,0 +1,38 @@
+// 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.
+
+//------------------------------------------------------------------------
+// CUDA kernel parameters.
+
+struct bias_act_kernel_params
+{
+    const void* x;      // [sizeX]
+    const void* b;      // [sizeB] or NULL
+    const void* xref;   // [sizeX] or NULL
+    const void* yref;   // [sizeX] or NULL
+    const void* dy;     // [sizeX] or NULL
+    void*       y;      // [sizeX]
+
+    int         grad;
+    int         act;
+    float       alpha;
+    float       gain;
+    float       clamp;
+
+    int         sizeX;
+    int         sizeB;
+    int         stepB;
+    int         loopX;
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> void* choose_bias_act_kernel(const bias_act_kernel_params& p);
+
+//------------------------------------------------------------------------

torch_utils/ops/bias_act.py

@@ -0,0 +1,209 @@
+# 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.
+
+"""Custom PyTorch ops for efficient bias and activation."""
+
+import os
+import numpy as np
+import torch
+import dnnlib
+
+from .. import custom_ops
+from .. import misc
+
+#----------------------------------------------------------------------------
+
+activation_funcs = {
+    'linear':   dnnlib.EasyDict(func=lambda x, **_:         x,                                          def_alpha=0,    def_gain=1,             cuda_idx=1, ref='',  has_2nd_grad=False),
+    'relu':     dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.relu(x),                def_alpha=0,    def_gain=np.sqrt(2),    cuda_idx=2, ref='y', has_2nd_grad=False),
+    'lrelu':    dnnlib.EasyDict(func=lambda x, alpha, **_:  torch.nn.functional.leaky_relu(x, alpha),   def_alpha=0.2,  def_gain=np.sqrt(2),    cuda_idx=3, ref='y', has_2nd_grad=False),
+    'tanh':     dnnlib.EasyDict(func=lambda x, **_:         torch.tanh(x),                              def_alpha=0,    def_gain=1,             cuda_idx=4, ref='y', has_2nd_grad=True),
+    'sigmoid':  dnnlib.EasyDict(func=lambda x, **_:         torch.sigmoid(x),                           def_alpha=0,    def_gain=1,             cuda_idx=5, ref='y', has_2nd_grad=True),
+    'elu':      dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.elu(x),                 def_alpha=0,    def_gain=1,             cuda_idx=6, ref='y', has_2nd_grad=True),
+    'selu':     dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.selu(x),                def_alpha=0,    def_gain=1,             cuda_idx=7, ref='y', has_2nd_grad=True),
+    'softplus': dnnlib.EasyDict(func=lambda x, **_:         torch.nn.functional.softplus(x),            def_alpha=0,    def_gain=1,             cuda_idx=8, ref='y', has_2nd_grad=True),
+    'swish':    dnnlib.EasyDict(func=lambda x, **_:         torch.sigmoid(x) * x,                       def_alpha=0,    def_gain=np.sqrt(2),    cuda_idx=9, ref='x', has_2nd_grad=True),
+}
+
+#----------------------------------------------------------------------------
+
+_plugin = None
+_null_tensor = torch.empty([0])
+
+def _init():
+    global _plugin
+    if _plugin is None:
+        _plugin = custom_ops.get_plugin(
+            module_name='bias_act_plugin',
+            sources=['bias_act.cpp', 'bias_act.cu'],
+            headers=['bias_act.h'],
+            source_dir=os.path.dirname(__file__),
+            extra_cuda_cflags=['--use_fast_math'],
+        )
+    return True
+
+#----------------------------------------------------------------------------
+
+def bias_act(x, b=None, dim=1, act='linear', alpha=None, gain=None, clamp=None, impl='cuda'):
+    r"""Fused bias and activation function.
+
+    Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
+    and scales the result by `gain`. Each of the steps is optional. In most cases,
+    the fused op is considerably more efficient than performing the same calculation
+    using standard PyTorch ops. It supports first and second order gradients,
+    but not third order gradients.
+
+    Args:
+        x:      Input activation tensor. Can be of any shape.
+        b:      Bias vector, or `None` to disable. Must be a 1D tensor of the same type
+                as `x`. The shape must be known, and it must match the dimension of `x`
+                corresponding to `dim`.
+        dim:    The dimension in `x` corresponding to the elements of `b`.
+                The value of `dim` is ignored if `b` is not specified.
+        act:    Name of the activation function to evaluate, or `"linear"` to disable.
+                Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
+                See `activation_funcs` for a full list. `None` is not allowed.
+        alpha:  Shape parameter for the activation function, or `None` to use the default.
+        gain:   Scaling factor for the output tensor, or `None` to use default.
+                See `activation_funcs` for the default scaling of each activation function.
+                If unsure, consider specifying 1.
+        clamp:  Clamp the output values to `[-clamp, +clamp]`, or `None` to disable
+                the clamping (default).
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert impl in ['ref', 'cuda']
+    if impl == 'cuda' and x.device.type == 'cuda' and _init():
+        return _bias_act_cuda(dim=dim, act=act, alpha=alpha, gain=gain, clamp=clamp).apply(x, b)
+    return _bias_act_ref(x=x, b=b, dim=dim, act=act, alpha=alpha, gain=gain, clamp=clamp)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def _bias_act_ref(x, b=None, dim=1, act='linear', alpha=None, gain=None, clamp=None):
+    """Slow reference implementation of `bias_act()` using standard TensorFlow ops.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert clamp is None or clamp >= 0
+    spec = activation_funcs[act]
+    alpha = float(alpha if alpha is not None else spec.def_alpha)
+    gain = float(gain if gain is not None else spec.def_gain)
+    clamp = float(clamp if clamp is not None else -1)
+
+    # Add bias.
+    if b is not None:
+        assert isinstance(b, torch.Tensor) and b.ndim == 1
+        assert 0 <= dim < x.ndim
+        assert b.shape[0] == x.shape[dim]
+        x = x + b.reshape([-1 if i == dim else 1 for i in range(x.ndim)])
+
+    # Evaluate activation function.
+    alpha = float(alpha)
+    x = spec.func(x, alpha=alpha)
+
+    # Scale by gain.
+    gain = float(gain)
+    if gain != 1:
+        x = x * gain
+
+    # Clamp.
+    if clamp >= 0:
+        x = x.clamp(-clamp, clamp) # pylint: disable=invalid-unary-operand-type
+    return x
+
+#----------------------------------------------------------------------------
+
+_bias_act_cuda_cache = dict()
+
+def _bias_act_cuda(dim=1, act='linear', alpha=None, gain=None, clamp=None):
+    """Fast CUDA implementation of `bias_act()` using custom ops.
+    """
+    # Parse arguments.
+    assert clamp is None or clamp >= 0
+    spec = activation_funcs[act]
+    alpha = float(alpha if alpha is not None else spec.def_alpha)
+    gain = float(gain if gain is not None else spec.def_gain)
+    clamp = float(clamp if clamp is not None else -1)
+
+    # Lookup from cache.
+    key = (dim, act, alpha, gain, clamp)
+    if key in _bias_act_cuda_cache:
+        return _bias_act_cuda_cache[key]
+
+    # Forward op.
+    class BiasActCuda(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, b): # pylint: disable=arguments-differ
+            ctx.memory_format = torch.channels_last if x.ndim > 2 and x.stride(1) == 1 else torch.contiguous_format
+            x = x.contiguous(memory_format=ctx.memory_format)
+            b = b.contiguous() if b is not None else _null_tensor
+            y = x
+            if act != 'linear' or gain != 1 or clamp >= 0 or b is not _null_tensor:
+                y = _plugin.bias_act(x, b, _null_tensor, _null_tensor, _null_tensor, 0, dim, spec.cuda_idx, alpha, gain, clamp)
+            ctx.save_for_backward(
+                x if 'x' in spec.ref or spec.has_2nd_grad else _null_tensor,
+                b if 'x' in spec.ref or spec.has_2nd_grad else _null_tensor,
+                y if 'y' in spec.ref else _null_tensor)
+            return y
+
+        @staticmethod
+        def backward(ctx, dy): # pylint: disable=arguments-differ
+            dy = dy.contiguous(memory_format=ctx.memory_format)
+            x, b, y = ctx.saved_tensors
+            dx = None
+            db = None
+
+            if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]:
+                dx = dy
+                if act != 'linear' or gain != 1 or clamp >= 0:
+                    dx = BiasActCudaGrad.apply(dy, x, b, y)
+
+            if ctx.needs_input_grad[1]:
+                db = dx.sum([i for i in range(dx.ndim) if i != dim])
+
+            return dx, db
+
+    # Backward op.
+    class BiasActCudaGrad(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, dy, x, b, y): # pylint: disable=arguments-differ
+            ctx.memory_format = torch.channels_last if dy.ndim > 2 and dy.stride(1) == 1 else torch.contiguous_format
+            dx = _plugin.bias_act(dy, b, x, y, _null_tensor, 1, dim, spec.cuda_idx, alpha, gain, clamp)
+            ctx.save_for_backward(
+                dy if spec.has_2nd_grad else _null_tensor,
+                x, b, y)
+            return dx
+
+        @staticmethod
+        def backward(ctx, d_dx): # pylint: disable=arguments-differ
+            d_dx = d_dx.contiguous(memory_format=ctx.memory_format)
+            dy, x, b, y = ctx.saved_tensors
+            d_dy = None
+            d_x = None
+            d_b = None
+            d_y = None
+
+            if ctx.needs_input_grad[0]:
+                d_dy = BiasActCudaGrad.apply(d_dx, x, b, y)
+
+            if spec.has_2nd_grad and (ctx.needs_input_grad[1] or ctx.needs_input_grad[2]):
+                d_x = _plugin.bias_act(d_dx, b, x, y, dy, 2, dim, spec.cuda_idx, alpha, gain, clamp)
+
+            if spec.has_2nd_grad and ctx.needs_input_grad[2]:
+                d_b = d_x.sum([i for i in range(d_x.ndim) if i != dim])
+
+            return d_dy, d_x, d_b, d_y
+
+    # Add to cache.
+    _bias_act_cuda_cache[key] = BiasActCuda
+    return BiasActCuda
+
+#----------------------------------------------------------------------------

torch_utils/ops/conv2d_gradfix.py

@@ -0,0 +1,198 @@
+# 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.
+
+"""Custom replacement for `torch.nn.functional.conv2d` that supports
+arbitrarily high order gradients with zero performance penalty."""
+
+import contextlib
+import torch
+
+# pylint: disable=redefined-builtin
+# pylint: disable=arguments-differ
+# pylint: disable=protected-access
+
+#----------------------------------------------------------------------------
+
+enabled = False                     # Enable the custom op by setting this to true.
+weight_gradients_disabled = False   # Forcefully disable computation of gradients with respect to the weights.
+
+@contextlib.contextmanager
+def no_weight_gradients(disable=True):
+    global weight_gradients_disabled
+    old = weight_gradients_disabled
+    if disable:
+        weight_gradients_disabled = True
+    yield
+    weight_gradients_disabled = old
+
+#----------------------------------------------------------------------------
+
+def conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1):
+    if _should_use_custom_op(input):
+        return _conv2d_gradfix(transpose=False, weight_shape=weight.shape, stride=stride, padding=padding, output_padding=0, dilation=dilation, groups=groups).apply(input, weight, bias)
+    return torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)
+
+def conv_transpose2d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1):
+    if _should_use_custom_op(input):
+        return _conv2d_gradfix(transpose=True, weight_shape=weight.shape, stride=stride, padding=padding, output_padding=output_padding, groups=groups, dilation=dilation).apply(input, weight, bias)
+    return torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, output_padding=output_padding, groups=groups, dilation=dilation)
+
+#----------------------------------------------------------------------------
+
+def _should_use_custom_op(input):
+    assert isinstance(input, torch.Tensor)
+    if (not enabled) or (not torch.backends.cudnn.enabled):
+        return False
+    if input.device.type != 'cuda':
+        return False
+    return True
+
+def _tuple_of_ints(xs, ndim):
+    xs = tuple(xs) if isinstance(xs, (tuple, list)) else (xs,) * ndim
+    assert len(xs) == ndim
+    assert all(isinstance(x, int) for x in xs)
+    return xs
+
+#----------------------------------------------------------------------------
+
+_conv2d_gradfix_cache = dict()
+_null_tensor = torch.empty([0])
+
+def _conv2d_gradfix(transpose, weight_shape, stride, padding, output_padding, dilation, groups):
+    # Parse arguments.
+    ndim = 2
+    weight_shape = tuple(weight_shape)
+    stride = _tuple_of_ints(stride, ndim)
+    padding = _tuple_of_ints(padding, ndim)
+    output_padding = _tuple_of_ints(output_padding, ndim)
+    dilation = _tuple_of_ints(dilation, ndim)
+
+    # Lookup from cache.
+    key = (transpose, weight_shape, stride, padding, output_padding, dilation, groups)
+    if key in _conv2d_gradfix_cache:
+        return _conv2d_gradfix_cache[key]
+
+    # Validate arguments.
+    assert groups >= 1
+    assert len(weight_shape) == ndim + 2
+    assert all(stride[i] >= 1 for i in range(ndim))
+    assert all(padding[i] >= 0 for i in range(ndim))
+    assert all(dilation[i] >= 0 for i in range(ndim))
+    if not transpose:
+        assert all(output_padding[i] == 0 for i in range(ndim))
+    else: # transpose
+        assert all(0 <= output_padding[i] < max(stride[i], dilation[i]) for i in range(ndim))
+
+    # Helpers.
+    common_kwargs = dict(stride=stride, padding=padding, dilation=dilation, groups=groups)
+    def calc_output_padding(input_shape, output_shape):
+        if transpose:
+            return [0, 0]
+        return [
+            input_shape[i + 2]
+            - (output_shape[i + 2] - 1) * stride[i]
+            - (1 - 2 * padding[i])
+            - dilation[i] * (weight_shape[i + 2] - 1)
+            for i in range(ndim)
+        ]
+
+    # Forward & backward.
+    class Conv2d(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, input, weight, bias):
+            assert weight.shape == weight_shape
+            ctx.save_for_backward(
+                input if weight.requires_grad else _null_tensor,
+                weight if input.requires_grad else _null_tensor,
+            )
+            ctx.input_shape = input.shape
+
+            # Simple 1x1 convolution => cuBLAS (only on Volta, not on Ampere).
+            if weight_shape[2:] == stride == dilation == (1, 1) and padding == (0, 0) and torch.cuda.get_device_capability(input.device) < (8, 0):
+                a = weight.reshape(groups, weight_shape[0] // groups, weight_shape[1])
+                b = input.reshape(input.shape[0], groups, input.shape[1] // groups, -1)
+                c = (a.transpose(1, 2) if transpose else a) @ b.permute(1, 2, 0, 3).flatten(2)
+                c = c.reshape(-1, input.shape[0], *input.shape[2:]).transpose(0, 1)
+                c = c if bias is None else c + bias.unsqueeze(0).unsqueeze(2).unsqueeze(3)
+                return c.contiguous(memory_format=(torch.channels_last if input.stride(1) == 1 else torch.contiguous_format))
+
+            # General case => cuDNN.
+            if transpose:
+                return torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, output_padding=output_padding, **common_kwargs)
+            return torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, **common_kwargs)
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            input, weight = ctx.saved_tensors
+            input_shape = ctx.input_shape
+            grad_input = None
+            grad_weight = None
+            grad_bias = None
+
+            if ctx.needs_input_grad[0]:
+                p = calc_output_padding(input_shape=input_shape, output_shape=grad_output.shape)
+                op = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs)
+                grad_input = op.apply(grad_output, weight, None)
+                assert grad_input.shape == input_shape
+
+            if ctx.needs_input_grad[1] and not weight_gradients_disabled:
+                grad_weight = Conv2dGradWeight.apply(grad_output, input)
+                assert grad_weight.shape == weight_shape
+
+            if ctx.needs_input_grad[2]:
+                grad_bias = grad_output.sum([0, 2, 3])
+
+            return grad_input, grad_weight, grad_bias
+
+    # Gradient with respect to the weights.
+    class Conv2dGradWeight(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, grad_output, input):
+            ctx.save_for_backward(
+                grad_output if input.requires_grad else _null_tensor,
+                input if grad_output.requires_grad else _null_tensor,
+            )
+            ctx.grad_output_shape = grad_output.shape
+            ctx.input_shape = input.shape
+
+            # Simple 1x1 convolution => cuBLAS (on both Volta and Ampere).
+            if weight_shape[2:] == stride == dilation == (1, 1) and padding == (0, 0):
+                a = grad_output.reshape(grad_output.shape[0], groups, grad_output.shape[1] // groups, -1).permute(1, 2, 0, 3).flatten(2)
+                b = input.reshape(input.shape[0], groups, input.shape[1] // groups, -1).permute(1, 2, 0, 3).flatten(2)
+                c = (b @ a.transpose(1, 2) if transpose else a @ b.transpose(1, 2)).reshape(weight_shape)
+                return c.contiguous(memory_format=(torch.channels_last if input.stride(1) == 1 else torch.contiguous_format))
+
+            # General case => cuDNN.
+            name = 'aten::cudnn_convolution_transpose_backward_weight' if transpose else 'aten::cudnn_convolution_backward_weight'
+            flags = [torch.backends.cudnn.benchmark, torch.backends.cudnn.deterministic, torch.backends.cudnn.allow_tf32]
+            return torch._C._jit_get_operation(name)(weight_shape, grad_output, input, padding, stride, dilation, groups, *flags)
+
+        @staticmethod
+        def backward(ctx, grad2_grad_weight):
+            grad_output, input = ctx.saved_tensors
+            grad_output_shape = ctx.grad_output_shape
+            input_shape = ctx.input_shape
+            grad2_grad_output = None
+            grad2_input = None
+
+            if ctx.needs_input_grad[0]:
+                grad2_grad_output = Conv2d.apply(input, grad2_grad_weight, None)
+                assert grad2_grad_output.shape == grad_output_shape
+
+            if ctx.needs_input_grad[1]:
+                p = calc_output_padding(input_shape=input_shape, output_shape=grad_output_shape)
+                op = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs)
+                grad2_input = op.apply(grad_output, grad2_grad_weight, None)
+                assert grad2_input.shape == input_shape
+
+            return grad2_grad_output, grad2_input
+
+    _conv2d_gradfix_cache[key] = Conv2d
+    return Conv2d
+
+#----------------------------------------------------------------------------

torch_utils/ops/conv2d_resample.py

@@ -0,0 +1,143 @@
+# 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.
+
+"""2D convolution with optional up/downsampling."""
+
+import torch
+
+from .. import misc
+from . import conv2d_gradfix
+from . import upfirdn2d
+from .upfirdn2d import _parse_padding
+from .upfirdn2d import _get_filter_size
+
+#----------------------------------------------------------------------------
+
+def _get_weight_shape(w):
+    with misc.suppress_tracer_warnings(): # this value will be treated as a constant
+        shape = [int(sz) for sz in w.shape]
+    misc.assert_shape(w, shape)
+    return shape
+
+#----------------------------------------------------------------------------
+
+def _conv2d_wrapper(x, w, stride=1, padding=0, groups=1, transpose=False, flip_weight=True):
+    """Wrapper for the underlying `conv2d()` and `conv_transpose2d()` implementations.
+    """
+    _out_channels, _in_channels_per_group, kh, kw = _get_weight_shape(w)
+
+    # Flip weight if requested.
+    # Note: conv2d() actually performs correlation (flip_weight=True) not convolution (flip_weight=False).
+    if not flip_weight and (kw > 1 or kh > 1):
+        w = w.flip([2, 3])
+
+    # Execute using conv2d_gradfix.
+    op = conv2d_gradfix.conv_transpose2d if transpose else conv2d_gradfix.conv2d
+    return op(x, w, stride=stride, padding=padding, groups=groups)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def conv2d_resample(x, w, f=None, up=1, down=1, padding=0, groups=1, flip_weight=True, flip_filter=False):
+    r"""2D convolution with optional up/downsampling.
+
+    Padding is performed only once at the beginning, not between the operations.
+
+    Args:
+        x:              Input tensor of shape
+                        `[batch_size, in_channels, in_height, in_width]`.
+        w:              Weight tensor of shape
+                        `[out_channels, in_channels//groups, kernel_height, kernel_width]`.
+        f:              Low-pass filter for up/downsampling. Must be prepared beforehand by
+                        calling upfirdn2d.setup_filter(). None = identity (default).
+        up:             Integer upsampling factor (default: 1).
+        down:           Integer downsampling factor (default: 1).
+        padding:        Padding with respect to the upsampled image. Can be a single number
+                        or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                        (default: 0).
+        groups:         Split input channels into N groups (default: 1).
+        flip_weight:    False = convolution, True = correlation (default: True).
+        flip_filter:    False = convolution, True = correlation (default: False).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    # Validate arguments.
+    assert isinstance(x, torch.Tensor) and (x.ndim == 4)
+    assert isinstance(w, torch.Tensor) and (w.ndim == 4) and (w.dtype == x.dtype)
+    assert f is None or (isinstance(f, torch.Tensor) and f.ndim in [1, 2] and f.dtype == torch.float32)
+    assert isinstance(up, int) and (up >= 1)
+    assert isinstance(down, int) and (down >= 1)
+    assert isinstance(groups, int) and (groups >= 1)
+    out_channels, in_channels_per_group, kh, kw = _get_weight_shape(w)
+    fw, fh = _get_filter_size(f)
+    px0, px1, py0, py1 = _parse_padding(padding)
+
+    # Adjust padding to account for up/downsampling.
+    if up > 1:
+        px0 += (fw + up - 1) // 2
+        px1 += (fw - up) // 2
+        py0 += (fh + up - 1) // 2
+        py1 += (fh - up) // 2
+    if down > 1:
+        px0 += (fw - down + 1) // 2
+        px1 += (fw - down) // 2
+        py0 += (fh - down + 1) // 2
+        py1 += (fh - down) // 2
+
+    # Fast path: 1x1 convolution with downsampling only => downsample first, then convolve.
+    if kw == 1 and kh == 1 and (down > 1 and up == 1):
+        x = upfirdn2d.upfirdn2d(x=x, f=f, down=down, padding=[px0,px1,py0,py1], flip_filter=flip_filter)
+        x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
+        return x
+
+    # Fast path: 1x1 convolution with upsampling only => convolve first, then upsample.
+    if kw == 1 and kh == 1 and (up > 1 and down == 1):
+        x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
+        x = upfirdn2d.upfirdn2d(x=x, f=f, up=up, padding=[px0,px1,py0,py1], gain=up**2, flip_filter=flip_filter)
+        return x
+
+    # Fast path: downsampling only => use strided convolution.
+    if down > 1 and up == 1:
+        x = upfirdn2d.upfirdn2d(x=x, f=f, padding=[px0,px1,py0,py1], flip_filter=flip_filter)
+        x = _conv2d_wrapper(x=x, w=w, stride=down, groups=groups, flip_weight=flip_weight)
+        return x
+
+    # Fast path: upsampling with optional downsampling => use transpose strided convolution.
+    if up > 1:
+        if groups == 1:
+            w = w.transpose(0, 1)
+        else:
+            w = w.reshape(groups, out_channels // groups, in_channels_per_group, kh, kw)
+            w = w.transpose(1, 2)
+            w = w.reshape(groups * in_channels_per_group, out_channels // groups, kh, kw)
+        px0 -= kw - 1
+        px1 -= kw - up
+        py0 -= kh - 1
+        py1 -= kh - up
+        pxt = max(min(-px0, -px1), 0)
+        pyt = max(min(-py0, -py1), 0)
+        x = _conv2d_wrapper(x=x, w=w, stride=up, padding=[pyt,pxt], groups=groups, transpose=True, flip_weight=(not flip_weight))
+        x = upfirdn2d.upfirdn2d(x=x, f=f, padding=[px0+pxt,px1+pxt,py0+pyt,py1+pyt], gain=up**2, flip_filter=flip_filter)
+        if down > 1:
+            x = upfirdn2d.upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
+        return x
+
+    # Fast path: no up/downsampling, padding supported by the underlying implementation => use plain conv2d.
+    if up == 1 and down == 1:
+        if px0 == px1 and py0 == py1 and px0 >= 0 and py0 >= 0:
+            return _conv2d_wrapper(x=x, w=w, padding=[py0,px0], groups=groups, flip_weight=flip_weight)
+
+    # Fallback: Generic reference implementation.
+    x = upfirdn2d.upfirdn2d(x=x, f=(f if up > 1 else None), up=up, padding=[px0,px1,py0,py1], gain=up**2, flip_filter=flip_filter)
+    x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
+    if down > 1:
+        x = upfirdn2d.upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
+    return x
+
+#----------------------------------------------------------------------------

torch_utils/ops/filtered_lrelu.cpp

@@ -0,0 +1,300 @@
+// 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.
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "filtered_lrelu.h"
+
+//------------------------------------------------------------------------
+
+static std::tuple<torch::Tensor, torch::Tensor, int> filtered_lrelu(
+    torch::Tensor x, torch::Tensor fu, torch::Tensor fd, torch::Tensor b, torch::Tensor si,
+    int up, int down, int px0, int px1, int py0, int py1, int sx, int sy, float gain, float slope, float clamp, bool flip_filters, bool writeSigns)
+{
+    // Set CUDA device.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+
+    // Validate arguments.
+    TORCH_CHECK(fu.device() == x.device() && fd.device() == x.device() && b.device() == x.device(), "all input tensors must reside on the same device");
+    TORCH_CHECK(fu.dtype() == torch::kFloat && fd.dtype() == torch::kFloat, "fu and fd must be float32");
+    TORCH_CHECK(b.dtype() == x.dtype(), "x and b must have the same dtype");
+    TORCH_CHECK(x.dtype() == torch::kHalf || x.dtype() == torch::kFloat, "x and b must be float16 or float32");
+    TORCH_CHECK(x.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(x.size(0) * x.size(1) <= INT_MAX && x.size(2) <= INT_MAX && x.size(3) <= INT_MAX, "x is too large");
+    TORCH_CHECK(x.numel() > 0, "x is empty");
+    TORCH_CHECK((fu.dim() == 1 || fu.dim() == 2) && (fd.dim() == 1 || fd.dim() == 2), "fu and fd must be rank 1 or 2");
+    TORCH_CHECK(fu.size(0) <= INT_MAX && fu.size(-1) <= INT_MAX, "fu is too large");
+    TORCH_CHECK(fd.size(0) <= INT_MAX && fd.size(-1) <= INT_MAX, "fd is too large");
+    TORCH_CHECK(fu.numel() > 0, "fu is empty");
+    TORCH_CHECK(fd.numel() > 0, "fd is empty");
+    TORCH_CHECK(b.dim() == 1 && b.size(0) == x.size(1), "b must be a vector with the same number of channels as x");
+    TORCH_CHECK(up >= 1 && down >= 1, "up and down must be at least 1");
+
+    // Figure out how much shared memory is available on the device.
+    int maxSharedBytes = 0;
+    AT_CUDA_CHECK(cudaDeviceGetAttribute(&maxSharedBytes, cudaDevAttrMaxSharedMemoryPerBlockOptin, x.device().index()));
+    int sharedKB = maxSharedBytes >> 10;
+
+    // Populate enough launch parameters to check if a CUDA kernel exists.
+    filtered_lrelu_kernel_params p;
+    p.up      = up;
+    p.down    = down;
+    p.fuShape = make_int2((int)fu.size(-1), fu.dim() == 2 ? (int)fu.size(0) : 0); // shape [n, 0] indicates separable filter.
+    p.fdShape = make_int2((int)fd.size(-1), fd.dim() == 2 ? (int)fd.size(0) : 0);
+    filtered_lrelu_kernel_spec test_spec = choose_filtered_lrelu_kernel<float, int32_t, false, false>(p, sharedKB);
+    if (!test_spec.exec)
+    {
+        // No kernel found - return empty tensors and indicate missing kernel with return code of -1.
+        return std::make_tuple(torch::Tensor(), torch::Tensor(), -1);
+    }
+
+    // Input/output element size.
+    int64_t sz = (x.dtype() == torch::kHalf) ? 2 : 4;
+
+    // Input sizes.
+    int64_t xw = (int)x.size(3);
+    int64_t xh = (int)x.size(2);
+    int64_t fut_w = (int)fu.size(-1) - 1;
+    int64_t fut_h = (int)fu.size(0)  - 1;
+    int64_t fdt_w = (int)fd.size(-1) - 1;
+    int64_t fdt_h = (int)fd.size(0)  - 1;
+
+    // Logical size of upsampled buffer.
+    int64_t cw = xw * up + (px0 + px1) - fut_w;
+    int64_t ch = xh * up + (py0 + py1) - fut_h;
+    TORCH_CHECK(cw > fdt_w && ch > fdt_h, "upsampled buffer must be at least the size of downsampling filter");
+    TORCH_CHECK(cw <= INT_MAX && ch <= INT_MAX, "upsampled buffer is too large");
+
+    // Compute output size and allocate.
+    int64_t yw = (cw - fdt_w + (down - 1)) / down;
+    int64_t yh = (ch - fdt_h + (down - 1)) / down;
+    TORCH_CHECK(yw > 0 && yh > 0, "output must be at least 1x1");
+    TORCH_CHECK(yw <= INT_MAX && yh <= INT_MAX, "output is too large");
+    torch::Tensor y = torch::empty({x.size(0), x.size(1), yh, yw}, x.options(), x.suggest_memory_format());
+
+    // Allocate sign tensor.
+    torch::Tensor so;
+    torch::Tensor s = si;
+    bool readSigns = !!s.numel();
+    int64_t sw_active = 0; // Active width of sign tensor.
+    if (writeSigns)
+    {
+        sw_active = yw * down - (down - 1) + fdt_w;     // Active width in elements.
+        int64_t sh = yh * down - (down - 1) + fdt_h;    // Height = active height.
+        int64_t sw = (sw_active + 15) & ~15;            // Width  = active width in elements, rounded up to multiple of 16.
+        TORCH_CHECK(sh <= INT_MAX && (sw >> 2) <= INT_MAX, "signs is too large");
+        s = so = torch::empty({x.size(0), x.size(1), sh, sw >> 2}, x.options().dtype(torch::kUInt8), at::MemoryFormat::Contiguous);
+    }
+    else if (readSigns)
+        sw_active = s.size(3) << 2;
+
+    // Validate sign tensor if in use.
+    if (readSigns || writeSigns)
+    {
+        TORCH_CHECK(s.is_contiguous(), "signs must be contiguous");
+        TORCH_CHECK(s.dtype() == torch::kUInt8, "signs must be uint8");
+        TORCH_CHECK(s.device() == x.device(), "signs must reside on the same device as x");
+        TORCH_CHECK(s.dim() == 4, "signs must be rank 4");
+        TORCH_CHECK(s.size(0) == x.size(0) && s.size(1) == x.size(1), "signs must have same batch & channels as x");
+        TORCH_CHECK(s.size(2) <= INT_MAX && s.size(3) <= INT_MAX, "signs is too large");
+    }
+
+    // Populate rest of CUDA kernel parameters.
+    p.x         = x.data_ptr();
+    p.y         = y.data_ptr();
+    p.b         = b.data_ptr();
+    p.s         = (readSigns || writeSigns) ? s.data_ptr<unsigned char>() : 0;
+    p.fu        = fu.data_ptr<float>();
+    p.fd        = fd.data_ptr<float>();
+    p.pad0      = make_int2(px0, py0);
+    p.gain      = gain;
+    p.slope     = slope;
+    p.clamp     = clamp;
+    p.flip      = (flip_filters) ? 1 : 0;
+    p.xShape    = make_int4((int)x.size(3), (int)x.size(2), (int)x.size(1), (int)x.size(0));
+    p.yShape    = make_int4((int)y.size(3), (int)y.size(2), (int)y.size(1), (int)y.size(0));
+    p.sShape    = (readSigns || writeSigns) ? make_int2((int)s.size(3), (int)s.size(2)) : make_int2(0, 0); // Width is in bytes. Contiguous.
+    p.sOfs      = make_int2(sx, sy);
+    p.swLimit   = (sw_active + 3) >> 2; // Rounded up to bytes.
+
+    // x, y, b strides are in bytes.
+    p.xStride   = make_longlong4(sz * x.stride(3), sz * x.stride(2), sz * x.stride(1), sz * x.stride(0));
+    p.yStride   = make_longlong4(sz * y.stride(3), sz * y.stride(2), sz * y.stride(1), sz * y.stride(0));
+    p.bStride   = sz * b.stride(0);
+
+    // fu, fd strides are in elements.
+    p.fuStride  = make_longlong3(fu.stride(-1), fu.dim() == 2 ? fu.stride(0) : 0, 0);
+    p.fdStride  = make_longlong3(fd.stride(-1), fd.dim() == 2 ? fd.stride(0) : 0, 0);
+
+    // Determine if indices don't fit in int32. Support negative strides although Torch currently never produces those.
+    bool index64b = false;
+    if (std::abs(p.bStride * x.size(1)) > INT_MAX) index64b = true;
+    if (std::min(x.size(0) * p.xStride.w, 0ll) + std::min(x.size(1) * p.xStride.z, 0ll) + std::min(x.size(2) * p.xStride.y, 0ll) + std::min(x.size(3) * p.xStride.x, 0ll) < -INT_MAX) index64b = true;
+    if (std::max(x.size(0) * p.xStride.w, 0ll) + std::max(x.size(1) * p.xStride.z, 0ll) + std::max(x.size(2) * p.xStride.y, 0ll) + std::max(x.size(3) * p.xStride.x, 0ll) >  INT_MAX) index64b = true;
+    if (std::min(y.size(0) * p.yStride.w, 0ll) + std::min(y.size(1) * p.yStride.z, 0ll) + std::min(y.size(2) * p.yStride.y, 0ll) + std::min(y.size(3) * p.yStride.x, 0ll) < -INT_MAX) index64b = true;
+    if (std::max(y.size(0) * p.yStride.w, 0ll) + std::max(y.size(1) * p.yStride.z, 0ll) + std::max(y.size(2) * p.yStride.y, 0ll) + std::max(y.size(3) * p.yStride.x, 0ll) >  INT_MAX) index64b = true;
+    if (s.numel() > INT_MAX) index64b = true;
+
+    // Choose CUDA kernel.
+    filtered_lrelu_kernel_spec spec = { 0 };
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "filtered_lrelu_cuda", [&]
+    {
+        if constexpr (sizeof(scalar_t) <= 4) // Exclude doubles. constexpr prevents template instantiation.
+        {
+            // Choose kernel based on index type, datatype and sign read/write modes.
+            if      (!index64b &&  writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int32_t, true,  false>(p, sharedKB);
+            else if (!index64b && !writeSigns &&  readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int32_t, false, true >(p, sharedKB);
+            else if (!index64b && !writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int32_t, false, false>(p, sharedKB);
+            else if ( index64b &&  writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int64_t, true,  false>(p, sharedKB);
+            else if ( index64b && !writeSigns &&  readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int64_t, false, true >(p, sharedKB);
+            else if ( index64b && !writeSigns && !readSigns) spec = choose_filtered_lrelu_kernel<scalar_t, int64_t, false, false>(p, sharedKB);
+        }
+    });
+    TORCH_CHECK(spec.exec, "internal error - CUDA kernel not found") // This should not happen because we tested earlier that kernel exists.
+
+    // Launch CUDA kernel.
+    void* args[] = {&p};
+    int bx = spec.numWarps * 32;
+    int gx = (p.yShape.x - 1) / spec.tileOut.x + 1;
+    int gy = (p.yShape.y - 1) / spec.tileOut.y + 1;
+    int gz = p.yShape.z * p.yShape.w;
+
+    // Repeat multiple horizontal tiles in a CTA?
+    if (spec.xrep)
+    {
+        p.tilesXrep = spec.xrep;
+        p.tilesXdim = gx;
+
+        gx = (gx + p.tilesXrep - 1) / p.tilesXrep;
+        std::swap(gx, gy);
+    }
+    else
+    {
+        p.tilesXrep = 0;
+        p.tilesXdim = 0;
+    }
+
+    // Launch filter setup kernel.
+    AT_CUDA_CHECK(cudaLaunchKernel(spec.setup, 1, 1024, args, 0, at::cuda::getCurrentCUDAStream()));
+
+    // Copy kernels to constant memory.
+    if      ( writeSigns && !readSigns) AT_CUDA_CHECK((copy_filters<true,  false>(at::cuda::getCurrentCUDAStream())));
+    else if (!writeSigns &&  readSigns) AT_CUDA_CHECK((copy_filters<false, true >(at::cuda::getCurrentCUDAStream())));
+    else if (!writeSigns && !readSigns) AT_CUDA_CHECK((copy_filters<false, false>(at::cuda::getCurrentCUDAStream())));
+
+    // Set cache and shared memory configurations for main kernel.
+    AT_CUDA_CHECK(cudaFuncSetCacheConfig(spec.exec, cudaFuncCachePreferShared));
+    if (spec.dynamicSharedKB) // Need dynamically allocated shared memory?
+        AT_CUDA_CHECK(cudaFuncSetAttribute(spec.exec, cudaFuncAttributeMaxDynamicSharedMemorySize, spec.dynamicSharedKB << 10));
+    AT_CUDA_CHECK(cudaFuncSetSharedMemConfig(spec.exec, cudaSharedMemBankSizeFourByte));
+
+    // Launch main kernel.
+    const int maxSubGz = 65535; // CUDA maximum for block z dimension.
+    for (int zofs=0; zofs < gz; zofs += maxSubGz) // Do multiple launches if gz is too big.
+    {
+        p.blockZofs = zofs;
+        int subGz = std::min(maxSubGz, gz - zofs);
+        AT_CUDA_CHECK(cudaLaunchKernel(spec.exec, dim3(gx, gy, subGz), bx, args, spec.dynamicSharedKB << 10, at::cuda::getCurrentCUDAStream()));
+    }
+
+    // Done.
+    return std::make_tuple(y, so, 0);
+}
+
+//------------------------------------------------------------------------
+
+static torch::Tensor filtered_lrelu_act(torch::Tensor x, torch::Tensor si, int sx, int sy, float gain, float slope, float clamp, bool writeSigns)
+{
+    // Set CUDA device.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+
+    // Validate arguments.
+    TORCH_CHECK(x.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(x.size(0) * x.size(1) <= INT_MAX && x.size(2) <= INT_MAX && x.size(3) <= INT_MAX, "x is too large");
+    TORCH_CHECK(x.numel() > 0, "x is empty");
+    TORCH_CHECK(x.dtype() == torch::kHalf || x.dtype() == torch::kFloat || x.dtype() == torch::kDouble, "x must be float16, float32 or float64");
+
+    // Output signs if we don't have sign input.
+    torch::Tensor so;
+    torch::Tensor s = si;
+    bool readSigns = !!s.numel();
+    if (writeSigns)
+    {
+        int64_t sw = x.size(3);
+        sw = (sw + 15) & ~15; // Round to a multiple of 16 for coalescing.
+        s = so = torch::empty({x.size(0), x.size(1), x.size(2), sw >> 2}, x.options().dtype(torch::kUInt8), at::MemoryFormat::Contiguous);
+    }
+
+    // Validate sign tensor if in use.
+    if (readSigns || writeSigns)
+    {
+        TORCH_CHECK(s.is_contiguous(), "signs must be contiguous");
+        TORCH_CHECK(s.dtype() == torch::kUInt8, "signs must be uint8");
+        TORCH_CHECK(s.device() == x.device(), "signs must reside on the same device as x");
+        TORCH_CHECK(s.dim() == 4, "signs must be rank 4");
+        TORCH_CHECK(s.size(0) == x.size(0) && s.size(1) == x.size(1), "signs must have same batch & channels as x");
+        TORCH_CHECK(s.size(2) <= INT_MAX && (s.size(3) << 2) <= INT_MAX, "signs tensor is too large");
+    }
+
+    // Initialize CUDA kernel parameters.
+    filtered_lrelu_act_kernel_params p;
+    p.x         = x.data_ptr();
+    p.s         = (readSigns || writeSigns) ? s.data_ptr<unsigned char>() : 0;
+    p.gain      = gain;
+    p.slope     = slope;
+    p.clamp     = clamp;
+    p.xShape    = make_int4((int)x.size(3), (int)x.size(2), (int)x.size(1), (int)x.size(0));
+    p.xStride   = make_longlong4(x.stride(3), x.stride(2), x.stride(1), x.stride(0));
+    p.sShape    = (readSigns || writeSigns) ? make_int2((int)s.size(3) << 2, (int)s.size(2)) : make_int2(0, 0); // Width is in elements. Contiguous.
+    p.sOfs      = make_int2(sx, sy);
+
+    // Choose CUDA kernel.
+    void* func = 0;
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "filtered_lrelu_act_cuda", [&]
+    {
+        if (writeSigns)
+            func = choose_filtered_lrelu_act_kernel<scalar_t, true, false>();
+        else if (readSigns)
+            func = choose_filtered_lrelu_act_kernel<scalar_t, false, true>();
+        else
+            func = choose_filtered_lrelu_act_kernel<scalar_t, false, false>();
+    });
+    TORCH_CHECK(func, "internal error - CUDA kernel not found");
+
+    // Launch CUDA kernel.
+    void* args[] = {&p};
+    int bx = 128; // 4 warps per block.
+
+    // Logical size of launch = writeSigns ? p.s : p.x
+    uint32_t gx = writeSigns ? p.sShape.x : p.xShape.x;
+    uint32_t gy = writeSigns ? p.sShape.y : p.xShape.y;
+    uint32_t gz = p.xShape.z * p.xShape.w; // Same as in p.sShape if signs are in use.
+    gx = (gx - 1) / bx + 1;
+
+    // Make sure grid y and z dimensions are within CUDA launch limits. Kernel loops internally to do the rest.
+    const uint32_t gmax = 65535;
+    gy = std::min(gy, gmax);
+    gz = std::min(gz, gmax);
+
+    // Launch.
+    AT_CUDA_CHECK(cudaLaunchKernel(func, dim3(gx, gy, gz), bx, args, 0, at::cuda::getCurrentCUDAStream()));
+    return so;
+}
+
+//------------------------------------------------------------------------
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("filtered_lrelu",      &filtered_lrelu);      // The whole thing.
+    m.def("filtered_lrelu_act_", &filtered_lrelu_act);  // Activation and sign tensor handling only. Modifies data tensor in-place.
+}
+
+//------------------------------------------------------------------------

torch_utils/ops/filtered_lrelu.cu

@@ -0,0 +1,1284 @@
+// 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.
+
+#include <c10/util/Half.h>
+#include "filtered_lrelu.h"
+#include <cstdint>
+
+//------------------------------------------------------------------------
+// Helpers.
+
+enum // Filter modes.
+{
+    MODE_SUSD = 0,  // Separable upsampling, separable downsampling.
+    MODE_FUSD = 1,  // Full upsampling, separable downsampling.
+    MODE_SUFD = 2,  // Separable upsampling, full downsampling.
+    MODE_FUFD = 3,  // Full upsampling, full downsampling.
+};
+
+template <class T> struct InternalType;
+template <> struct InternalType<double>
+{
+    typedef double scalar_t; typedef double2 vec2_t; typedef double4 vec4_t;
+    __device__ __forceinline__ static vec2_t zero_vec2(void) { return make_double2(0, 0); }
+    __device__ __forceinline__ static vec4_t zero_vec4(void) { return make_double4(0, 0, 0, 0); }
+    __device__ __forceinline__ static double clamp(double x, double c) { return fmin(fmax(x, -c), c); }
+};
+template <> struct InternalType<float>
+{
+    typedef float scalar_t; typedef float2 vec2_t; typedef float4 vec4_t;
+    __device__ __forceinline__ static vec2_t zero_vec2(void) { return make_float2(0, 0); }
+    __device__ __forceinline__ static vec4_t zero_vec4(void) { return make_float4(0, 0, 0, 0); }
+    __device__ __forceinline__ static float clamp(float x, float c) { return fminf(fmaxf(x, -c), c); }
+};
+template <> struct InternalType<c10::Half>
+{
+    typedef float scalar_t; typedef float2 vec2_t; typedef float4 vec4_t;
+    __device__ __forceinline__ static vec2_t zero_vec2(void) { return make_float2(0, 0); }
+    __device__ __forceinline__ static vec4_t zero_vec4(void) { return make_float4(0, 0, 0, 0); }
+    __device__ __forceinline__ static float clamp(float x, float c) { return fminf(fmaxf(x, -c), c); }
+};
+
+#define MIN(A, B)       ((A) < (B) ? (A) : (B))
+#define MAX(A, B)       ((A) > (B) ? (A) : (B))
+#define CEIL_DIV(A, B) (((B)==1) ? (A) : \
+                        ((B)==2) ? ((int)((A)+1) >> 1) : \
+                        ((B)==4) ? ((int)((A)+3) >> 2) : \
+                        (((A) + ((A) > 0 ? (B) - 1 : 0)) / (B)))
+
+// This works only up to blocks of size 256 x 256 and for all N that are powers of two.
+template <int N> __device__ __forceinline__ void fast_div_mod(int& x, int& y, unsigned int i)
+{
+    if ((N & (N-1)) && N <= 256)
+        y = (i * ((1<<24)/N + 1)) >> 24; // Assumes N <= 256, i < N*256.
+    else
+        y = i/N;
+
+    x = i - y*N;
+}
+
+// Type cast stride before reading it.
+template <class T> __device__ __forceinline__ T get_stride(const int64_t& x)
+{
+    return *reinterpret_cast<const T*>(&x);
+}
+
+//------------------------------------------------------------------------
+// Filters, setup kernel, copying function.
+
+#define MAX_FILTER_SIZE 32
+
+// Combined up/down filter buffers so that transfer can be done with one copy.
+__device__              float g_fbuf[2 * MAX_FILTER_SIZE * MAX_FILTER_SIZE]; // Filters in global memory, written by setup kernel.
+__device__ __constant__ float c_fbuf[2 * MAX_FILTER_SIZE * MAX_FILTER_SIZE]; // Filters in constant memory, read by main kernel.
+
+// Accessors to combined buffers to index up/down filters individually.
+#define c_fu (c_fbuf)
+#define c_fd (c_fbuf + MAX_FILTER_SIZE * MAX_FILTER_SIZE)
+#define g_fu (g_fbuf)
+#define g_fd (g_fbuf + MAX_FILTER_SIZE * MAX_FILTER_SIZE)
+
+// Set up filters into global memory buffer.
+static __global__ void setup_filters_kernel(filtered_lrelu_kernel_params p)
+{
+    for (int idx = threadIdx.x; idx < MAX_FILTER_SIZE * MAX_FILTER_SIZE; idx += blockDim.x)
+    {
+        int x, y;
+        fast_div_mod<MAX_FILTER_SIZE>(x, y, idx);
+
+        int fu_x = p.flip ? x : (p.fuShape.x - 1 - x);
+        int fu_y = p.flip ? y : (p.fuShape.y - 1 - y);
+        if (p.fuShape.y > 0)
+            g_fu[idx] = (x >= p.fuShape.x || y >= p.fuShape.y) ? 0.0f : p.fu[fu_x * p.fuStride.x + fu_y * p.fuStride.y];
+        else
+            g_fu[idx] = (x >= p.fuShape.x || y > 0) ? 0.0f : p.fu[fu_x * p.fuStride.x];
+
+        int fd_x = p.flip ? x : (p.fdShape.x - 1 - x);
+        int fd_y = p.flip ? y : (p.fdShape.y - 1 - y);
+        if (p.fdShape.y > 0)
+            g_fd[idx] = (x >= p.fdShape.x || y >= p.fdShape.y) ? 0.0f : p.fd[fd_x * p.fdStride.x + fd_y * p.fdStride.y];
+        else
+            g_fd[idx] = (x >= p.fdShape.x || y > 0) ? 0.0f : p.fd[fd_x * p.fdStride.x];
+    }
+}
+
+// Host function to copy filters written by setup kernel into constant buffer for main kernel.
+template <bool, bool> static cudaError_t copy_filters(cudaStream_t stream)
+{
+    void* src = 0;
+    cudaError_t err = cudaGetSymbolAddress(&src, g_fbuf);
+    if (err) return err;
+    return cudaMemcpyToSymbolAsync(c_fbuf, src, 2 * MAX_FILTER_SIZE * MAX_FILTER_SIZE * sizeof(float), 0, cudaMemcpyDeviceToDevice, stream);
+}
+
+//------------------------------------------------------------------------
+// Coordinate spaces:
+// - Relative to input tensor:      inX, inY, tileInX, tileInY
+// - Relative to input tile:        relInX, relInY, tileInW, tileInH
+// - Relative to upsampled tile:    relUpX, relUpY, tileUpW, tileUpH
+// - Relative to output tile:       relOutX, relOutY, tileOutW, tileOutH
+// - Relative to output tensor:     outX, outY, tileOutX, tileOutY
+//
+// Relationships between coordinate spaces:
+// - inX = tileInX + relInX
+// - inY = tileInY + relInY
+// - relUpX = relInX * up + phaseInX
+// - relUpY = relInY * up + phaseInY
+// - relUpX = relOutX * down
+// - relUpY = relOutY * down
+// - outX = tileOutX + relOutX
+// - outY = tileOutY + relOutY
+
+extern __shared__ char s_buf_raw[]; // When sharedKB <= 48, allocate shared memory statically inside the kernel, otherwise use the externally allocated shared memory buffer.
+
+template <class T, class index_t, int sharedKB, bool signWrite, bool signRead, int filterMode, int up, int fuSize, int down, int fdSize, int tileOutW, int tileOutH, int threadsPerBlock, bool enableXrep, bool enableWriteSkip>
+static __global__ void filtered_lrelu_kernel(filtered_lrelu_kernel_params p)
+{
+    // Check that we don't try to support non-existing filter modes.
+    static_assert(up   == 1 || up   == 2 || up   == 4, "only up=1, up=2, up=4 scales supported");
+    static_assert(down == 1 || down == 2 || down == 4, "only down=1, down=2, down=4 scales supported");
+    static_assert(fuSize >= up,   "upsampling filter size must be at least upsampling factor");
+    static_assert(fdSize >= down, "downsampling filter size must be at least downsampling factor");
+    static_assert(fuSize % up   == 0, "upsampling filter size must be divisible with upsampling factor");
+    static_assert(fdSize % down == 0, "downsampling filter size must be divisible with downsampling factor");
+    static_assert(fuSize <= MAX_FILTER_SIZE && fdSize <= MAX_FILTER_SIZE, "filter size greater than MAX_FILTER_SIZE");
+    static_assert(up   != 1 || (fuSize == 1 && (filterMode == MODE_FUFD || filterMode == MODE_FUSD)), "up=1 supported only for 1x1 full filters");
+    static_assert(down != 1 || (fdSize == 1 && (filterMode == MODE_FUFD || filterMode == MODE_SUFD)), "down=1 supported only for 1x1 full filters");
+    static_assert(!(up   == 4 && (filterMode == MODE_FUFD || filterMode == MODE_FUSD)), "full filters not supported for up=4");
+    static_assert(!(down == 4 && (filterMode == MODE_FUFD || filterMode == MODE_SUFD)), "full filters not supported for down=4");
+
+    // Static definitions.
+    typedef typename InternalType<T>::scalar_t scalar_t;
+    typedef typename InternalType<T>::vec2_t vec2_t;
+    typedef typename InternalType<T>::vec4_t vec4_t;
+    const int tileUpW    = (tileOutW * down + (fdSize - 1) - (down - 1) + 3) & ~3;  // Upsampled tile width, rounded up to multiple of 4.
+    const int tileUpH    = tileOutH * down + (fdSize - 1) - (down - 1);             // Upsampled tile height.
+    const int tileInW    = CEIL_DIV(tileUpW  + (fuSize - 1), up);                   // Input tile width.
+    const int tileInH    = CEIL_DIV(tileUpH  + (fuSize - 1), up);                   // Input tile height.
+    const int tileUpH_up = CEIL_DIV(tileUpH, up) * up;                              // Upsampled tile height rounded up to a multiple of up.
+    const int tileInH_up = CEIL_DIV(tileUpH_up + (fuSize - 1), up);                 // For allocations only, to avoid shared memory read overruns with up=2 and up=4.
+
+    // Merge 1x1 downsampling into last upsampling step for upf1 and ups2.
+    const bool downInline = (down == 1) && ((up == 1 && filterMode == MODE_FUFD) || (up == 2 && filterMode == MODE_SUFD));
+
+    // Sizes of logical buffers.
+    const int szIn    = tileInH_up * tileInW;
+    const int szUpX   = tileInH_up * tileUpW;
+    const int szUpXY  = downInline ? 0 : (tileUpH * tileUpW);
+    const int szDownX = tileUpH * tileOutW;
+
+    // Sizes for shared memory arrays.
+    const int s_buf0_size_base =
+        (filterMode == MODE_SUSD) ? MAX(szIn, szUpXY) :
+        (filterMode == MODE_FUSD) ? MAX(szIn, szDownX) :
+        (filterMode == MODE_SUFD) ? MAX(szIn, szUpXY) :
+        (filterMode == MODE_FUFD) ? szIn :
+        -1;
+    const int s_buf1_size_base =
+        (filterMode == MODE_SUSD) ? MAX(szUpX, szDownX) :
+        (filterMode == MODE_FUSD) ? szUpXY :
+        (filterMode == MODE_SUFD) ? szUpX  :
+        (filterMode == MODE_FUFD) ? szUpXY :
+        -1;
+
+    // Ensure U128 alignment.
+    const int s_buf0_size = (s_buf0_size_base + 3) & ~3;
+    const int s_buf1_size = (s_buf1_size_base + 3) & ~3;
+
+    // Check at compile time that we don't use too much shared memory.
+    static_assert((s_buf0_size + s_buf1_size) * sizeof(scalar_t) <= (sharedKB << 10), "shared memory overflow");
+
+    // Declare shared memory arrays.
+    scalar_t* s_buf0;
+    scalar_t* s_buf1;
+    if (sharedKB <= 48)
+    {
+        // Allocate shared memory arrays here.
+        __shared__ scalar_t s_buf0_st[(sharedKB > 48) ? (1<<24) : (s_buf0_size + s_buf1_size)]; // Prevent launching if this isn't optimized away when unused.
+        s_buf0 = s_buf0_st;
+        s_buf1 = s_buf0 + s_buf0_size;
+    }
+    else
+    {
+        // Use the dynamically allocated shared memory array.
+        s_buf0 = (scalar_t*)s_buf_raw;
+        s_buf1 = s_buf0 + s_buf0_size;
+    }
+
+    // Pointers to the buffers.
+    scalar_t* s_tileIn;       // Input tile:                      [relInX * tileInH + relInY]
+    scalar_t* s_tileUpX;      // After horizontal upsampling:     [relInY * tileUpW + relUpX]
+    scalar_t* s_tileUpXY;     // After upsampling:                [relUpY * tileUpW + relUpX]
+    scalar_t* s_tileDownX;    // After horizontal downsampling:   [relUpY * tileOutW + relOutX]
+    if (filterMode == MODE_SUSD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpX   = s_buf1;
+        s_tileUpXY  = s_buf0;
+        s_tileDownX = s_buf1;
+    }
+    else if (filterMode == MODE_FUSD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpXY  = s_buf1;
+        s_tileDownX = s_buf0;
+    }
+    else if (filterMode == MODE_SUFD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpX   = s_buf1;
+        s_tileUpXY  = s_buf0;
+    }
+    else if (filterMode == MODE_FUFD)
+    {
+        s_tileIn    = s_buf0;
+        s_tileUpXY  = s_buf1;
+    }
+
+    // Allow large grids in z direction via per-launch offset.
+    int channelIdx = blockIdx.z + p.blockZofs;
+    int batchIdx = channelIdx / p.yShape.z;
+    channelIdx -= batchIdx * p.yShape.z;
+
+    // Offset to output feature map. In bytes.
+    index_t mapOfsOut = channelIdx * get_stride<index_t>(p.yStride.z) + batchIdx * get_stride<index_t>(p.yStride.w);
+
+    // Sign shift amount.
+    uint32_t signXo = ((threadIdx.x + p.sOfs.x) << 1) & 6;
+
+    // Inner tile loop.
+    #pragma unroll 1
+    for (int tileIdx = 0; !enableXrep || (tileIdx < MIN(p.tilesXrep, p.tilesXdim - p.tilesXrep * blockIdx.y)); tileIdx++)
+    {
+        // Locate output tile.
+        int tileX = enableXrep ? blockIdx.y * p.tilesXrep + tileIdx : blockIdx.x;
+        int tileOutX = tileX * tileOutW;
+        int tileOutY = (enableXrep ? blockIdx.x : blockIdx.y) * tileOutH;
+
+        // Locate input tile.
+        int tmpX = tileOutX * down - p.pad0.x;
+        int tmpY = tileOutY * down - p.pad0.y;
+        int tileInX = CEIL_DIV(tmpX, up);
+        int tileInY = CEIL_DIV(tmpY, up);
+        const int phaseInX = tileInX * up - tmpX;
+        const int phaseInY = tileInY * up - tmpY;
+
+        // Extra sync if input and output buffers are the same and we are not on first tile.
+        if (enableXrep && tileIdx > 0 && (filterMode == MODE_FUSD || (filterMode == MODE_SUFD && !downInline) || (filterMode == MODE_FUFD && downInline)))
+            __syncthreads();
+
+        // Load input tile & apply bias. Unrolled.
+        scalar_t b = (scalar_t)*(const T*)((const char*)p.b + (channelIdx * get_stride<index_t>(p.bStride)));
+        index_t mapOfsIn = channelIdx * get_stride<index_t>(p.xStride.z) + batchIdx * get_stride<index_t>(p.xStride.w);
+        int idx = threadIdx.x;
+        const int loopCountIN = CEIL_DIV(tileInW * tileInH, threadsPerBlock);
+        #pragma unroll
+        for (int loop = 0; loop < loopCountIN; loop++)
+        {
+            int relInX, relInY;
+            fast_div_mod<tileInW>(relInX, relInY, idx);
+            int inX = tileInX + relInX;
+            int inY = tileInY + relInY;
+            scalar_t v = 0;
+
+            if ((uint32_t)inX < p.xShape.x && (uint32_t)inY < p.xShape.y)
+                v = (scalar_t)*((const T*)((const char*)p.x + (inX * get_stride<index_t>(p.xStride.x) + inY * get_stride<index_t>(p.xStride.y) + mapOfsIn))) + b;
+
+            bool skip = (loop == loopCountIN-1) && (idx >= tileInW * tileInH);
+            if (!skip)
+                s_tileIn[idx] = v;
+
+            idx += threadsPerBlock;
+        }
+
+        if (filterMode == MODE_SUSD || filterMode == MODE_SUFD) // Separable upsampling filter.
+        {
+            // Horizontal upsampling.
+            __syncthreads();
+            if (up == 4)
+            {
+                for (int idx = threadIdx.x*up; idx < tileUpW * tileInH; idx += blockDim.x*up)
+                {
+                    int relUpX0, relInY;
+                    fast_div_mod<tileUpW>(relUpX0, relInY, idx);
+                    int relInX0 = relUpX0 / up;
+                    int src0 = relInX0 + tileInW * relInY;
+                    int dst = relInY * tileUpW + relUpX0;
+                    vec4_t v = InternalType<T>::zero_vec4();
+                    scalar_t a = s_tileIn[src0];
+                    if (phaseInX == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 3];
+                            v.z += a * (scalar_t)c_fu[step * up + 2];
+                            v.w += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else if (phaseInX == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.z += a * (scalar_t)c_fu[step * up + 3];
+                            v.w += a * (scalar_t)c_fu[step * up + 2];
+                        }
+                    }
+                    else if (phaseInX == 2)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 2];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                            v.z += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.w += a * (scalar_t)c_fu[step * up + 3];
+                        }
+                    }
+                    else // (phaseInX == 3)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 3];
+                            v.y += a * (scalar_t)c_fu[step * up + 2];
+                            v.z += a * (scalar_t)c_fu[step * up + 1];
+                            v.w += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                        }
+                    }
+                    s_tileUpX[dst+0] = v.x;
+                    s_tileUpX[dst+1] = v.y;
+                    s_tileUpX[dst+2] = v.z;
+                    s_tileUpX[dst+3] = v.w;
+                }
+            }
+            else if (up == 2)
+            {
+                bool p0 = (phaseInX == 0);
+                for (int idx = threadIdx.x*up; idx < tileUpW * tileInH; idx += blockDim.x*up)
+                {
+                    int relUpX0, relInY;
+                    fast_div_mod<tileUpW>(relUpX0, relInY, idx);
+                    int relInX0 = relUpX0 / up;
+                    int src0 = relInX0 + tileInW * relInY;
+                    int dst = relInY * tileUpW + relUpX0;
+                    vec2_t v = InternalType<T>::zero_vec2();
+                    scalar_t a = s_tileIn[src0];
+                    if (p0) // (phaseInX == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else // (phaseInX == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileIn[src0 + step + 1];
+                        }
+                    }
+                    s_tileUpX[dst+0] = v.x;
+                    s_tileUpX[dst+1] = v.y;
+                }
+            }
+
+            // Vertical upsampling & nonlinearity.
+
+            __syncthreads();
+            int groupMask = 15 << ((threadIdx.x & 31) & ~3);
+            int minY = tileOutY ? (tileOutY - tileOutH) * down + tileUpH : 0; // Skip already written signs.
+            int sShapeMaxY = MIN(p.sShape.y, tileOutY * down + tileUpH); // Avoid out-of-tile sign writes.
+            if (up == 4)
+            {
+                minY -= 3; // Adjust according to block height.
+                for (int idx = threadIdx.x; idx < tileUpW * tileUpH_up / up; idx += blockDim.x)
+                {
+                    int relUpX, relInY0;
+                    fast_div_mod<tileUpW>(relUpX, relInY0, idx);
+                    int relUpY0 = relInY0 * up;
+                    int src0 = relInY0 * tileUpW + relUpX;
+                    int dst = relUpY0 * tileUpW + relUpX;
+                    vec4_t v = InternalType<T>::zero_vec4();
+
+                    scalar_t a = s_tileUpX[src0];
+                    if (phaseInY == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.y += a * (scalar_t)c_fu[step * up + 3];
+                            v.z += a * (scalar_t)c_fu[step * up + 2];
+                            v.w += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else if (phaseInY == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.z += a * (scalar_t)c_fu[step * up + 3];
+                            v.w += a * (scalar_t)c_fu[step * up + 2];
+                        }
+                    }
+                    else if (phaseInY == 2)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 2];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                            v.z += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.w += a * (scalar_t)c_fu[step * up + 3];
+                        }
+                    }
+                    else // (phaseInY == 3)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 3];
+                            v.y += a * (scalar_t)c_fu[step * up + 2];
+                            v.z += a * (scalar_t)c_fu[step * up + 1];
+                            v.w += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                        }
+                    }
+
+                    int x = tileOutX * down + relUpX;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si0 = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+                    index_t si1 = si0 + p.sShape.x;
+                    index_t si2 = si0 + p.sShape.x * 2;
+                    index_t si3 = si0 + p.sShape.x * 3;
+
+                    v.x *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.y *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.z *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.w *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write signs.
+                            int sx = __float_as_uint(v.x) >> 31 <<  0;
+                            int sy = __float_as_uint(v.y) >> 31 <<  8;
+                            int sz = __float_as_uint(v.z) >> 31 << 16;
+                            int sw = __float_as_uint(v.w) >> 31 << 24;
+                            if (sx) v.x *= p.slope;
+                            if (sy) v.y *= p.slope;
+                            if (sz) v.z *= p.slope;
+                            if (sw) v.w *= p.slope;
+                            if (fabsf(v.x) > p.clamp) { sx = 2 <<  0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                            if (fabsf(v.y) > p.clamp) { sy = 2 <<  8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                            if (fabsf(v.z) > p.clamp) { sz = 2 << 16; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                            if (fabsf(v.w) > p.clamp) { sw = 2 << 24; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                // Combine signs.
+                                uint32_t s = sx + sy + sw + sz;
+                                s <<= (signX & 3) << 1;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                                if ((uint32_t)(signY + 2) < sShapeMaxY) { p.s[si2] = (unsigned char)(s >> 16); }
+                                if ((uint32_t)(signY + 3) < sShapeMaxY) { p.s[si3] = (unsigned char)(s >> 24); }
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write signs.
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                int sx = __float_as_uint(v.x) >> 31 <<  0;
+                                int sy = __float_as_uint(v.y) >> 31 <<  8;
+                                int sz = __float_as_uint(v.z) >> 31 << 16;
+                                int sw = __float_as_uint(v.w) >> 31 << 24;
+                                if (sx) v.x *= p.slope;
+                                if (sy) v.y *= p.slope;
+                                if (sz) v.z *= p.slope;
+                                if (sw) v.w *= p.slope;
+                                if (fabsf(v.x) > p.clamp) { sx = 2 <<  0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                                if (fabsf(v.y) > p.clamp) { sy = 2 <<  8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                                if (fabsf(v.z) > p.clamp) { sz = 2 << 16; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                                if (fabsf(v.w) > p.clamp) { sw = 2 << 24; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                                // Combine signs.
+                                uint32_t s = sx + sy + sw + sz;
+                                s <<= (signX & 3) << 1;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                                if ((uint32_t)(signY + 2) < sShapeMaxY) { p.s[si2] = (unsigned char)(s >> 16); }
+                                if ((uint32_t)(signY + 3) < sShapeMaxY) { p.s[si3] = (unsigned char)(s >> 24); }
+                            }
+                            else
+                            {
+                                // Just compute the values.
+                                if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                                if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                                if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                                if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead) // Read signs and apply.
+                    {
+                        if ((uint32_t)signXb < p.swLimit)
+                        {
+                            int ss = (signX & 3) << 1;
+                            if ((uint32_t)(signY + 0) < p.sShape.y) { int s = p.s[si0] >> ss; if (s & 1) v.x *= p.slope; if (s & 2) v.x = 0.f; }
+                            if ((uint32_t)(signY + 1) < p.sShape.y) { int s = p.s[si1] >> ss; if (s & 1) v.y *= p.slope; if (s & 2) v.y = 0.f; }
+                            if ((uint32_t)(signY + 2) < p.sShape.y) { int s = p.s[si2] >> ss; if (s & 1) v.z *= p.slope; if (s & 2) v.z = 0.f; }
+                            if ((uint32_t)(signY + 3) < p.sShape.y) { int s = p.s[si3] >> ss; if (s & 1) v.w *= p.slope; if (s & 2) v.w = 0.f; }
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                        if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                        if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                        if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                    }
+
+                    s_tileUpXY[dst + 0 * tileUpW] = v.x;
+                    if (relUpY0 + 1 < tileUpH) s_tileUpXY[dst + 1 * tileUpW] = v.y;
+                    if (relUpY0 + 2 < tileUpH) s_tileUpXY[dst + 2 * tileUpW] = v.z;
+                    if (relUpY0 + 3 < tileUpH) s_tileUpXY[dst + 3 * tileUpW] = v.w;
+                }
+            }
+            else if (up == 2)
+            {
+                minY -= 1; // Adjust according to block height.
+                for (int idx = threadIdx.x; idx < tileUpW * tileUpH_up / up; idx += blockDim.x)
+                {
+                    int relUpX, relInY0;
+                    fast_div_mod<tileUpW>(relUpX, relInY0, idx);
+                    int relUpY0 = relInY0 * up;
+                    int src0 = relInY0 * tileUpW + relUpX;
+                    int dst = relUpY0 * tileUpW + relUpX;
+                    vec2_t v = InternalType<T>::zero_vec2();
+
+                    scalar_t a = s_tileUpX[src0];
+                    if (phaseInY == 0)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                            v.y += a * (scalar_t)c_fu[step * up + 1];
+                        }
+                    }
+                    else // (phaseInY == 1)
+                    {
+                        #pragma unroll
+                        for (int step = 0; step < fuSize / up; step++)
+                        {
+                            v.x += a * (scalar_t)c_fu[step * up + 1];
+                            v.y += a * (scalar_t)c_fu[step * up + 0];
+                            a = s_tileUpX[src0 + (step + 1) * tileUpW];
+                        }
+                    }
+
+                    int x = tileOutX * down + relUpX;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si0 = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+                    index_t si1 = si0 + p.sShape.x;
+
+                    v.x *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.y *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write signs.
+                            int sx = __float_as_uint(v.x) >> 31 << 0;
+                            int sy = __float_as_uint(v.y) >> 31 << 8;
+                            if (sx) v.x *= p.slope;
+                            if (sy) v.y *= p.slope;
+                            if (fabsf(v.x) > p.clamp) { sx = 2 << 0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                            if (fabsf(v.y) > p.clamp) { sy = 2 << 8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                // Combine signs.
+                                int s = sx + sy;
+                                s <<= signXo;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write signs.
+                            if ((uint32_t)signXb < p.swLimit && signY >= minY)
+                            {
+                                int sx = __float_as_uint(v.x) >> 31 << 0;
+                                int sy = __float_as_uint(v.y) >> 31 << 8;
+                                if (sx) v.x *= p.slope;
+                                if (sy) v.y *= p.slope;
+                                if (fabsf(v.x) > p.clamp) { sx = 2 << 0; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                                if (fabsf(v.y) > p.clamp) { sy = 2 << 8; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+
+                                // Combine signs.
+                                int s = sx + sy;
+                                s <<= signXo;
+                                s |= __shfl_xor_sync(groupMask, s, 1);
+                                s |= __shfl_xor_sync(groupMask, s, 2);
+
+                                // Write signs.
+                                if ((uint32_t)(signY + 0) < sShapeMaxY) { p.s[si0] = (unsigned char)(s >>  0); }
+                                if ((uint32_t)(signY + 1) < sShapeMaxY) { p.s[si1] = (unsigned char)(s >>  8); }
+                            }
+                            else
+                            {
+                                // Just compute the values.
+                                if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                                if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead) // Read signs and apply.
+                    {
+                        if ((uint32_t)signXb < p.swLimit)
+                        {
+                            if ((uint32_t)(signY + 0) < p.sShape.y) { int s = p.s[si0] >> signXo; if (s & 1) v.x *= p.slope; if (s & 2) v.x = 0.f; }
+                            if ((uint32_t)(signY + 1) < p.sShape.y) { int s = p.s[si1] >> signXo; if (s & 1) v.y *= p.slope; if (s & 2) v.y = 0.f; }
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                        if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                    }
+
+                    if (!downInline)
+                    {
+                        // Write into temporary buffer.
+                        s_tileUpXY[dst] = v.x;
+                        if (relUpY0 < tileUpH - 1)
+                            s_tileUpXY[dst + tileUpW] = v.y;
+                    }
+                    else
+                    {
+                        // Write directly into output buffer.
+                        if ((uint32_t)x < p.yShape.x)
+                        {
+                            int ymax = MIN(p.yShape.y, tileUpH + tileOutY * down);
+                            index_t ofs = x * get_stride<index_t>(p.yStride.x) + y * get_stride<index_t>(p.yStride.y) + mapOfsOut;
+                            if ((uint32_t)y + 0 < p.yShape.y) *((T*)((char*)p.y + ofs)) = (T)(v.x * (scalar_t)c_fd[0]);
+                            if ((uint32_t)y + 1 < ymax) *((T*)((char*)p.y + ofs + get_stride<index_t>(p.yStride.y))) = (T)(v.y * (scalar_t)c_fd[0]);
+                        }
+                    }
+                }
+            }
+        }
+        else if (filterMode == MODE_FUSD || filterMode == MODE_FUFD)
+        {
+            // Full upsampling filter.
+
+            if (up == 2)
+            {
+                // 2 x 2-wide.
+                __syncthreads();
+                int minY = tileOutY ? (tileOutY - tileOutH) * down + tileUpH + p.sOfs.y : 0; // Skip already written signs.
+                for (int idx = threadIdx.x * 4; idx < tileUpW * tileUpH; idx += blockDim.x * 4)
+                {
+                    int relUpX0, relUpY0;
+                    fast_div_mod<tileUpW>(relUpX0, relUpY0, idx);
+                    int relInX0 = CEIL_DIV(relUpX0 - phaseInX, up);
+                    int relInY0 = CEIL_DIV(relUpY0 - phaseInY, up);
+                    int src0 = relInX0 + tileInW * relInY0;
+                    int tap0y = (relInY0 * up + phaseInY - relUpY0);
+
+                    #define X_LOOP(TAPY, PX) \
+                        for (int sx = 0; sx < fuSize / up; sx++) \
+                        { \
+                            v.x += a * (scalar_t)c_fu[(sx * up + (((PX) - 0) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; \
+                            v.z += b * (scalar_t)c_fu[(sx * up + (((PX) - 0) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; if ((PX) == 0) { a = b; b = s_tileIn[src0 + 2 + sx + sy * tileInW]; } \
+                            v.y += a * (scalar_t)c_fu[(sx * up + (((PX) - 1) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; \
+                            v.w += b * (scalar_t)c_fu[(sx * up + (((PX) - 1) & (up - 1))) + (sy * up + (TAPY)) * MAX_FILTER_SIZE]; if ((PX) == 1) { a = b; b = s_tileIn[src0 + 2 + sx + sy * tileInW]; } \
+                        }
+
+                    vec4_t v = InternalType<T>::zero_vec4();
+                    if (tap0y == 0 && phaseInX == 0)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(0, 0) }
+                    if (tap0y == 0 && phaseInX == 1)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(0, 1) }
+                    if (tap0y == 1 && phaseInX == 0)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(1, 0) }
+                    if (tap0y == 1 && phaseInX == 1)
+                        #pragma unroll
+                        for (int sy = 0; sy < fuSize / up; sy++) { scalar_t a = s_tileIn[src0 + sy * tileInW]; scalar_t b = s_tileIn[src0 + sy * tileInW + 1];
+                            #pragma unroll
+                            X_LOOP(1, 1) }
+
+                    #undef X_LOOP
+
+                    int x = tileOutX * down + relUpX0;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+
+                    v.x *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.y *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.z *= (scalar_t)((float)up * (float)up * p.gain);
+                    v.w *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write signs.
+                            int sx = __float_as_uint(v.x) >> 31;
+                            int sy = __float_as_uint(v.y) >> 31;
+                            int sz = __float_as_uint(v.z) >> 31;
+                            int sw = __float_as_uint(v.w) >> 31;
+                            if (sx) v.x *= p.slope; if (fabsf(v.x) > p.clamp) { sx = 2; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                            if (sy) v.y *= p.slope; if (fabsf(v.y) > p.clamp) { sy = 2; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                            if (sz) v.z *= p.slope; if (fabsf(v.z) > p.clamp) { sz = 2; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                            if (sw) v.w *= p.slope; if (fabsf(v.w) > p.clamp) { sw = 2; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                p.s[si] = sx + (sy << 2) + (sz << 4) + (sw << 6);
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write signs.
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                int sx = __float_as_uint(v.x) >> 31;
+                                int sy = __float_as_uint(v.y) >> 31;
+                                int sz = __float_as_uint(v.z) >> 31;
+                                int sw = __float_as_uint(v.w) >> 31;
+                                if (sx) v.x *= p.slope; if (fabsf(v.x) > p.clamp) { sx = 2; v.x = InternalType<T>::clamp(v.x, p.clamp); }
+                                if (sy) v.y *= p.slope; if (fabsf(v.y) > p.clamp) { sy = 2; v.y = InternalType<T>::clamp(v.y, p.clamp); }
+                                if (sz) v.z *= p.slope; if (fabsf(v.z) > p.clamp) { sz = 2; v.z = InternalType<T>::clamp(v.z, p.clamp); }
+                                if (sw) v.w *= p.slope; if (fabsf(v.w) > p.clamp) { sw = 2; v.w = InternalType<T>::clamp(v.w, p.clamp); }
+
+                                p.s[si] = sx + (sy << 2) + (sz << 4) + (sw << 6);
+                            }
+                            else
+                            {
+                                // Just compute the values.
+                                if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                                if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                                if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                                if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead) // Read sign and apply.
+                    {
+                        if ((uint32_t)signY < p.sShape.y)
+                        {
+                            int s = 0;
+                            if ((uint32_t)signXb     < p.swLimit) s  = p.s[si];
+                            if ((uint32_t)signXb + 1 < p.swLimit) s |= p.s[si + 1] << 8;
+                            s >>= (signX & 3) << 1;
+                            if (s & 0x01) v.x *= p.slope; if (s & 0x02) v.x = 0.f;
+                            if (s & 0x04) v.y *= p.slope; if (s & 0x08) v.y = 0.f;
+                            if (s & 0x10) v.z *= p.slope; if (s & 0x20) v.z = 0.f;
+                            if (s & 0x40) v.w *= p.slope; if (s & 0x80) v.w = 0.f;
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v.x < 0.f) v.x *= p.slope; v.x = InternalType<T>::clamp(v.x, p.clamp);
+                        if (v.y < 0.f) v.y *= p.slope; v.y = InternalType<T>::clamp(v.y, p.clamp);
+                        if (v.z < 0.f) v.z *= p.slope; v.z = InternalType<T>::clamp(v.z, p.clamp);
+                        if (v.w < 0.f) v.w *= p.slope; v.w = InternalType<T>::clamp(v.w, p.clamp);
+                    }
+
+                    s_tileUpXY[idx + 0] = v.x;
+                    s_tileUpXY[idx + 1] = v.y;
+                    s_tileUpXY[idx + 2] = v.z;
+                    s_tileUpXY[idx + 3] = v.w;
+                }
+            }
+            else if (up == 1)
+            {
+                __syncthreads();
+                uint32_t groupMask = 15 << ((threadIdx.x & 31) & ~3);
+                int minY = tileOutY ? (tileOutY - tileOutH) * down + tileUpH : 0; // Skip already written signs.
+                for (int idx = threadIdx.x; idx < tileUpW * tileUpH; idx += blockDim.x)
+                {
+                    int relUpX0, relUpY0;
+                    fast_div_mod<tileUpW>(relUpX0, relUpY0, idx);
+                    scalar_t v = s_tileIn[idx] * (scalar_t)c_fu[0]; // 1x1 filter.
+
+                    int x = tileOutX * down + relUpX0;
+                    int y = tileOutY * down + relUpY0;
+                    int signX = x + p.sOfs.x;
+                    int signY = y + p.sOfs.y;
+                    int signZ = blockIdx.z + p.blockZofs;
+                    int signXb = signX >> 2;
+                    index_t si = signXb + p.sShape.x * (signY + (index_t)p.sShape.y * signZ);
+                    v *= (scalar_t)((float)up * (float)up * p.gain);
+
+                    if (signWrite)
+                    {
+                        if (!enableWriteSkip)
+                        {
+                            // Determine and write sign.
+                            uint32_t s = 0;
+                            uint32_t signXbit = (1u << signXo);
+                            if (v < 0.f)
+                            {
+                                s = signXbit;
+                                v *= p.slope;
+                            }
+                            if (fabsf(v) > p.clamp)
+                            {
+                                s = signXbit * 2;
+                                v = InternalType<T>::clamp(v, p.clamp);
+                            }
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                s += __shfl_xor_sync(groupMask, s, 1);  // Coalesce.
+                                s += __shfl_xor_sync(groupMask, s, 2);  // Coalesce.
+                                p.s[si] = s;                            // Write.
+                            }
+                        }
+                        else
+                        {
+                            // Determine and write sign.
+                            if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y && signY >= minY)
+                            {
+                                uint32_t s = 0;
+                                uint32_t signXbit = (1u << signXo);
+                                if (v < 0.f)
+                                {
+                                    s = signXbit;
+                                    v *= p.slope;
+                                }
+                                if (fabsf(v) > p.clamp)
+                                {
+                                    s = signXbit * 2;
+                                    v = InternalType<T>::clamp(v, p.clamp);
+                                }
+                                s += __shfl_xor_sync(groupMask, s, 1);  // Coalesce.
+                                s += __shfl_xor_sync(groupMask, s, 2);  // Coalesce.
+                                p.s[si] = s;                            // Write.
+                            }
+                            else
+                            {
+                                // Just compute the value.
+                                if (v < 0.f) v *= p.slope;
+                                v = InternalType<T>::clamp(v, p.clamp);
+                            }
+                        }
+                    }
+                    else if (signRead)
+                    {
+                        // Read sign and apply if within sign tensor bounds.
+                        if ((uint32_t)signXb < p.swLimit && (uint32_t)signY < p.sShape.y)
+                        {
+                            int s = p.s[si];
+                            s >>= signXo;
+                            if (s & 1) v *= p.slope;
+                            if (s & 2) v = 0.f;
+                        }
+                    }
+                    else // Forward pass with no sign write.
+                    {
+                        if (v < 0.f) v *= p.slope;
+                        v = InternalType<T>::clamp(v, p.clamp);
+                    }
+
+                    if (!downInline) // Write into temporary buffer.
+                        s_tileUpXY[idx] = v;
+                    else if ((uint32_t)x < p.yShape.x && (uint32_t)y < p.yShape.y) // Write directly into output buffer
+                        *((T*)((char*)p.y + (x * get_stride<index_t>(p.yStride.x) + y * get_stride<index_t>(p.yStride.y) + mapOfsOut))) = (T)(v * (scalar_t)c_fd[0]);
+                }
+            }
+        }
+
+        // Downsampling.
+        if (filterMode == MODE_SUSD || filterMode == MODE_FUSD)
+        {
+            // Horizontal downsampling.
+            __syncthreads();
+            if (down == 4 && tileOutW % 4 == 0)
+            {
+                // Calculate 4 pixels at a time.
+                for (int idx = threadIdx.x * 4; idx < tileOutW * tileUpH; idx += blockDim.x * 4)
+                {
+                    int relOutX0, relUpY;
+                    fast_div_mod<tileOutW>(relOutX0, relUpY, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int src0 = relUpY * tileUpW + relUpX0;
+                    vec4_t v = InternalType<T>::zero_vec4();
+                    #pragma unroll
+                    for (int step = 0; step < fdSize; step++)
+                    {
+                        v.x += s_tileUpXY[src0 +  0 + step] * (scalar_t)c_fd[step];
+                        v.y += s_tileUpXY[src0 +  4 + step] * (scalar_t)c_fd[step];
+                        v.z += s_tileUpXY[src0 +  8 + step] * (scalar_t)c_fd[step];
+                        v.w += s_tileUpXY[src0 + 12 + step] * (scalar_t)c_fd[step];
+                    }
+                    s_tileDownX[idx+0] = v.x;
+                    s_tileDownX[idx+1] = v.y;
+                    s_tileDownX[idx+2] = v.z;
+                    s_tileDownX[idx+3] = v.w;
+                }
+            }
+            else if ((down == 2 || down == 4) && (tileOutW % 2 == 0))
+            {
+                // Calculate 2 pixels at a time.
+                for (int idx = threadIdx.x * 2; idx < tileOutW * tileUpH; idx += blockDim.x * 2)
+                {
+                    int relOutX0, relUpY;
+                    fast_div_mod<tileOutW>(relOutX0, relUpY, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int src0 = relUpY * tileUpW + relUpX0;
+                    vec2_t v = InternalType<T>::zero_vec2();
+                    #pragma unroll
+                    for (int step = 0; step < fdSize; step++)
+                    {
+                        v.x += s_tileUpXY[src0 +    0 + step] * (scalar_t)c_fd[step];
+                        v.y += s_tileUpXY[src0 + down + step] * (scalar_t)c_fd[step];
+                    }
+                    s_tileDownX[idx+0] = v.x;
+                    s_tileDownX[idx+1] = v.y;
+                }
+            }
+            else
+            {
+                // Calculate 1 pixel at a time.
+                for (int idx = threadIdx.x; idx < tileOutW * tileUpH; idx += blockDim.x)
+                {
+                    int relOutX0, relUpY;
+                    fast_div_mod<tileOutW>(relOutX0, relUpY, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int src = relUpY * tileUpW + relUpX0;
+                    scalar_t v = 0.f;
+                    #pragma unroll
+                    for (int step = 0; step < fdSize; step++)
+                        v += s_tileUpXY[src + step] * (scalar_t)c_fd[step];
+                    s_tileDownX[idx] = v;
+                }
+            }
+
+            // Vertical downsampling & store output tile.
+            __syncthreads();
+            for (int idx = threadIdx.x; idx < tileOutW * tileOutH; idx += blockDim.x)
+            {
+                int relOutX, relOutY0;
+                fast_div_mod<tileOutW>(relOutX, relOutY0, idx);
+                int relUpY0 = relOutY0 * down;
+                int src0 = relUpY0 * tileOutW + relOutX;
+                scalar_t v = 0;
+                #pragma unroll
+                for (int step = 0; step < fdSize; step++)
+                    v += s_tileDownX[src0 + step * tileOutW] * (scalar_t)c_fd[step];
+
+                int outX = tileOutX + relOutX;
+                int outY = tileOutY + relOutY0;
+
+                if (outX < p.yShape.x & outY < p.yShape.y)
+                    *((T*)((char*)p.y + (outX * get_stride<index_t>(p.yStride.x) + outY * get_stride<index_t>(p.yStride.y) + mapOfsOut))) = (T)v;
+            }
+        }
+        else if (filterMode == MODE_SUFD || filterMode == MODE_FUFD)
+        {
+            // Full downsampling filter.
+            if (down == 2)
+            {
+                // 2-wide.
+                __syncthreads();
+                for (int idx = threadIdx.x * 2; idx < tileOutW * tileOutH; idx += blockDim.x * 2)
+                {
+                    int relOutX0, relOutY0;
+                    fast_div_mod<tileOutW>(relOutX0, relOutY0, idx);
+                    int relUpX0 = relOutX0 * down;
+                    int relUpY0 = relOutY0 * down;
+                    int src0 = relUpY0 * tileUpW + relUpX0;
+                    vec2_t v = InternalType<T>::zero_vec2();
+                    #pragma unroll
+                    for (int sy = 0; sy < fdSize; sy++)
+                    #pragma unroll
+                    for (int sx = 0; sx < fdSize; sx++)
+                    {
+                        v.x += s_tileUpXY[src0 + 0 + sx + sy * tileUpW] * (scalar_t)c_fd[sx + sy * MAX_FILTER_SIZE];
+                        v.y += s_tileUpXY[src0 + 2 + sx + sy * tileUpW] * (scalar_t)c_fd[sx + sy * MAX_FILTER_SIZE];
+                    }
+
+                    int outX = tileOutX + relOutX0;
+                    int outY = tileOutY + relOutY0;
+                    if ((uint32_t)outY < p.yShape.y)
+                    {
+                        index_t ofs = outX * get_stride<index_t>(p.yStride.x) + outY * get_stride<index_t>(p.yStride.y) + mapOfsOut;
+                        if (outX + 0 < p.yShape.x) *((T*)((char*)p.y + ofs)) = (T)v.x;
+                        if (outX + 1 < p.yShape.x) *((T*)((char*)p.y + ofs + get_stride<index_t>(p.yStride.x))) = (T)v.y;
+                    }
+                }
+            }
+            else if (down == 1 && !downInline)
+            {
+                // Thread per pixel.
+                __syncthreads();
+                for (int idx = threadIdx.x; idx < tileOutW * tileOutH; idx += blockDim.x)
+                {
+                    int relOutX0, relOutY0;
+                    fast_div_mod<tileOutW>(relOutX0, relOutY0, idx);
+                    scalar_t v = s_tileUpXY[idx] * (scalar_t)c_fd[0]; // 1x1 filter.
+
+                    int outX = tileOutX + relOutX0;
+                    int outY = tileOutY + relOutY0;
+                    if ((uint32_t)outX < p.yShape.x && (uint32_t)outY < p.yShape.y)
+                        *((T*)((char*)p.y + (outX * get_stride<index_t>(p.yStride.x) + outY * get_stride<index_t>(p.yStride.y) + mapOfsOut))) = (T)v;
+                }
+            }
+        }
+
+        if (!enableXrep)
+            break;
+    }
+}
+
+//------------------------------------------------------------------------
+// Compute activation function and signs for upsampled data tensor, modifying data tensor in-place. Used for accelerating the generic variant.
+// Sign tensor is known to be contiguous, and p.x and p.s have the same z, w dimensions. 64-bit indexing is always used.
+
+template <class T, bool signWrite, bool signRead>
+static __global__ void filtered_lrelu_act_kernel(filtered_lrelu_act_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+
+    // Indexing.
+    int32_t x = threadIdx.x + blockIdx.x * blockDim.x;
+    int32_t ymax = signWrite ? p.sShape.y : p.xShape.y;
+    int32_t qmax = p.xShape.z * p.xShape.w; // Combined minibatch*channel maximum index.
+
+    // Loop to accommodate oversized tensors.
+    for (int32_t q = blockIdx.z; q < qmax; q += gridDim.z)
+    for (int32_t y = blockIdx.y; y < ymax; y += gridDim.y)
+    {
+        // Extract z and w (channel, minibatch index).
+        int32_t w = q / p.xShape.z;
+        int32_t z = q - w * p.xShape.z;
+
+        // Choose behavior based on sign read/write mode.
+        if (signWrite)
+        {
+            // Process value if in p.x.
+            uint32_t s = 0;
+            if (x < p.xShape.x && y < p.xShape.y)
+            {
+                int64_t ix = x * p.xStride.x + y * p.xStride.y + z * p.xStride.z + w * p.xStride.w;
+                T* pv = ((T*)p.x) + ix;
+                scalar_t v = (scalar_t)(*pv);
+
+                // Gain, LReLU, clamp.
+                v *= p.gain;
+                if (v < 0.f)
+                {
+                    v *= p.slope;
+                    s = 1; // Sign.
+                }
+                if (fabsf(v) > p.clamp)
+                {
+                    v = InternalType<T>::clamp(v, p.clamp);
+                    s = 2; // Clamp.
+                }
+
+                *pv = (T)v; // Write value.
+            }
+
+            // Coalesce into threads 0 and 16 of warp.
+            uint32_t m = (threadIdx.x & 16) ? 0xffff0000u : 0x0000ffffu;
+            s <<= ((threadIdx.x & 15) << 1); // Shift into place.
+            s |= __shfl_xor_sync(m, s, 1); // Distribute.
+            s |= __shfl_xor_sync(m, s, 2);
+            s |= __shfl_xor_sync(m, s, 4);
+            s |= __shfl_xor_sync(m, s, 8);
+
+            // Write signs if leader and in p.s.
+            if (!(threadIdx.x & 15) && x < p.sShape.x) // y is always in.
+            {
+                uint64_t is = x + p.sShape.x * (y + (int64_t)p.sShape.y * q); // Contiguous.
+                ((uint32_t*)p.s)[is >> 4] = s;
+            }
+        }
+        else if (signRead)
+        {
+            // Process value if in p.x.
+            if (x < p.xShape.x) // y is always in.
+            {
+                int64_t ix = x * p.xStride.x + y * p.xStride.y + z * p.xStride.z + w * p.xStride.w;
+                T* pv = ((T*)p.x) + ix;
+                scalar_t v = (scalar_t)(*pv);
+                v *= p.gain;
+
+                // Apply sign buffer offset.
+                uint32_t sx = x + p.sOfs.x;
+                uint32_t sy = y + p.sOfs.y;
+
+                // Read and apply signs if we land inside valid region of sign buffer.
+                if (sx < p.sShape.x && sy < p.sShape.y)
+                {
+                    uint64_t is = (sx >> 2) + (p.sShape.x >> 2) * (sy + (uint64_t)p.sShape.y * q); // Contiguous.
+                    unsigned char s = p.s[is];
+                    s >>= (sx & 3) << 1; // Shift into place.
+                    if (s & 1) // Sign?
+                        v *= p.slope;
+                    if (s & 2) // Clamp?
+                        v = 0.f;
+                }
+
+                *pv = (T)v; // Write value.
+            }
+        }
+        else
+        {
+            // Forward pass with no sign write. Process value if in p.x.
+            if (x < p.xShape.x) // y is always in.
+            {
+                int64_t ix = x * p.xStride.x + y * p.xStride.y + z * p.xStride.z + w * p.xStride.w;
+                T* pv = ((T*)p.x) + ix;
+                scalar_t v = (scalar_t)(*pv);
+                v *= p.gain;
+                if (v < 0.f)
+                    v *= p.slope;
+                if (fabsf(v) > p.clamp)
+                    v = InternalType<T>::clamp(v, p.clamp);
+                *pv = (T)v; // Write value.
+            }
+        }
+    }
+}
+
+template <class T, bool signWrite, bool signRead> void* choose_filtered_lrelu_act_kernel(void)
+{
+    return (void*)filtered_lrelu_act_kernel<T, signWrite, signRead>;
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T, class index_t, bool signWrite, bool signRead> filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel(const filtered_lrelu_kernel_params& p, int sharedKB)
+{
+    filtered_lrelu_kernel_spec s = { 0 };
+
+    // Return the first matching kernel.
+#define CASE(SH, U, FU, D, FD, MODE, TW, TH, W, XR, WS) \
+    if (sharedKB >= SH) \
+    if ((p.fuShape.y == 0 && (MODE == MODE_SUSD || MODE == MODE_SUFD)) || (p.fuShape.y > 0 && (MODE == MODE_FUSD || MODE == MODE_FUFD))) \
+    if ((p.fdShape.y == 0 && (MODE == MODE_SUSD || MODE == MODE_FUSD)) || (p.fdShape.y > 0 && (MODE == MODE_SUFD || MODE == MODE_FUFD))) \
+    if (p.up == U && p.fuShape.x <= FU && p.fuShape.y <= FU && p.down == D && p.fdShape.x <= FD && p.fdShape.y <= FD) \
+    { \
+        static_assert((D*TW % 4) == 0, "down * tileWidth must be divisible by 4"); \
+        static_assert(FU % U == 0, "upscaling filter size must be multiple of upscaling factor"); \
+        static_assert(FD % D == 0, "downscaling filter size must be multiple of downscaling factor"); \
+        s.setup = (void*)setup_filters_kernel; \
+        s.exec = (void*)filtered_lrelu_kernel<T, index_t, SH, signWrite, signRead, MODE, U, FU, D, FD, TW, TH, W*32, !!XR, !!WS>; \
+        s.tileOut = make_int2(TW, TH); \
+        s.numWarps = W; \
+        s.xrep = XR; \
+        s.dynamicSharedKB = (SH == 48) ? 0 : SH; \
+        return s; \
+    }
+
+    // Launch parameters for various kernel specializations.
+    // Small filters must be listed before large filters, otherwise the kernel for larger filter will always match first.
+    // Kernels that use more shared memory must be listed before those that use less, for the same reason.
+
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/1,1,  /*mode*/MODE_FUFD, /*tw,th,warps,xrep,wskip*/64,  178, 32,  0,  0) // 1t-upf1-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/1,1,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/152, 95,  16,  0,  0) // 4t-ups2-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/2,8,  /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/56,  22,  16,  0,  0) // 4t-upf1-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/2,8,  /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/56,  29,  16,  11, 0) // 4t-ups2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/2,8,  /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/60,  28,  16,  0,  0) // 4t-upf2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/2,8,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/56,  28,  16,  0,  0) // 4t-ups2-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/4,16, /*down,fd*/2,8,  /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/56,  31,  16,  11, 0) // 4t-ups4-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/4,16, /*down,fd*/2,8,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/56,  36,  16,  0,  0) // 4t-ups4-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/4,16, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/16,  22,  16,  12, 0) // 4t-ups2-downs4
+    CASE(/*sharedKB*/48, /*up,fu*/2,8,  /*down,fd*/4,16, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/29,  15,  16,  0,  0) // 4t-upf2-downs4
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/1,1,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/96,  150, 28,  0,  0) // 6t-ups2-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/2,12, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/32,  35,  24,  0,  0) // 6t-upf1-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/2,12, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  46,  16,  10, 0) // 6t-ups2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/2,12, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/58,  28,  24,  8,  0) // 6t-upf2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/2,12, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/52,  28,  16,  0,  0) // 6t-ups2-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/4,24, /*down,fd*/2,12, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  51,  16,  5,  0) // 6t-ups4-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/4,24, /*down,fd*/2,12, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/32,  56,  16,  6,  0) // 6t-ups4-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/4,24, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/16,  18,  16,  12, 0) // 6t-ups2-downs4
+    CASE(/*sharedKB*/96, /*up,fu*/2,12, /*down,fd*/4,24, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/27,  31,  32,  6,  0) // 6t-upf2-downs4 96kB
+    CASE(/*sharedKB*/48, /*up,fu*/2,12, /*down,fd*/4,24, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/27,  13,  24,  0,  0) // 6t-upf2-downs4
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/1,1,  /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/148, 89,  24,  0,  0) // 8t-ups2-downf1
+    CASE(/*sharedKB*/48, /*up,fu*/1,1,  /*down,fd*/2,16, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/32,  31,  16,  5,  0) // 8t-upf1-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/2,16, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  41,  16,  9,  0) // 8t-ups2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/2,16, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/56,  26,  24,  0,  0) // 8t-upf2-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/2,16, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/32,  40,  16,  0,  0) // 8t-ups2-downf2
+    CASE(/*sharedKB*/48, /*up,fu*/4,32, /*down,fd*/2,16, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/32,  46,  24,  5,  0) // 8t-ups4-downs2
+    CASE(/*sharedKB*/48, /*up,fu*/4,32, /*down,fd*/2,16, /*mode*/MODE_SUFD, /*tw,th,warps,xrep,wskip*/32,  50,  16,  0,  0) // 8t-ups4-downf2
+    CASE(/*sharedKB*/96, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/24,  24,  32,  12, 1) // 8t-ups2-downs4 96kB
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_SUSD, /*tw,th,warps,xrep,wskip*/16,  13,  16,  10, 1) // 8t-ups2-downs4
+    CASE(/*sharedKB*/96, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/25,  28,  28,  4,  0) // 8t-upf2-downs4 96kB
+    CASE(/*sharedKB*/48, /*up,fu*/2,16, /*down,fd*/4,32, /*mode*/MODE_FUSD, /*tw,th,warps,xrep,wskip*/25,  10,  24,  0,  0) // 8t-upf2-downs4
+
+    #undef CASE
+    return s; // No kernel found.
+}
+
+//------------------------------------------------------------------------

torch_utils/ops/filtered_lrelu.h

@@ -0,0 +1,90 @@
+// 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.
+
+#include <cuda_runtime.h>
+
+//------------------------------------------------------------------------
+// CUDA kernel parameters.
+
+struct filtered_lrelu_kernel_params
+{
+    // These parameters decide which kernel to use.
+    int             up;         // upsampling ratio (1, 2, 4)
+    int             down;       // downsampling ratio (1, 2, 4)
+    int2            fuShape;    // [size, 1] | [size, size]
+    int2            fdShape;    // [size, 1] | [size, size]
+
+    int             _dummy;     // Alignment.
+
+    // Rest of the parameters.
+    const void*     x;          // Input tensor.
+    void*           y;          // Output tensor.
+    const void*     b;          // Bias tensor.
+    unsigned char*  s;          // Sign tensor in/out. NULL if unused.
+    const float*    fu;         // Upsampling filter.
+    const float*    fd;         // Downsampling filter.
+
+    int2            pad0;       // Left/top padding.
+    float           gain;       // Additional gain factor.
+    float           slope;      // Leaky ReLU slope on negative side.
+    float           clamp;      // Clamp after nonlinearity.
+    int             flip;       // Filter kernel flip for gradient computation.
+
+    int             tilesXdim;  // Original number of horizontal output tiles.
+    int             tilesXrep;  // Number of horizontal tiles per CTA.
+    int             blockZofs;  // Block z offset to support large minibatch, channel dimensions.
+
+    int4            xShape;     // [width, height, channel, batch]
+    int4            yShape;     // [width, height, channel, batch]
+    int2            sShape;     // [width, height] - width is in bytes. Contiguous. Zeros if unused.
+    int2            sOfs;       // [ofs_x, ofs_y] - offset between upsampled data and sign tensor.
+    int             swLimit;    // Active width of sign tensor in bytes.
+
+    longlong4       xStride;    // Strides of all tensors except signs, same component order as shapes.
+    longlong4       yStride;    //
+    int64_t         bStride;    //
+    longlong3       fuStride;   //
+    longlong3       fdStride;   //
+};
+
+struct filtered_lrelu_act_kernel_params
+{
+    void*           x;          // Input/output, modified in-place.
+    unsigned char*  s;          // Sign tensor in/out. NULL if unused.
+
+    float           gain;       // Additional gain factor.
+    float           slope;      // Leaky ReLU slope on negative side.
+    float           clamp;      // Clamp after nonlinearity.
+
+    int4            xShape;     // [width, height, channel, batch]
+    longlong4       xStride;    // Input/output tensor strides, same order as in shape.
+    int2            sShape;     // [width, height] - width is in elements. Contiguous. Zeros if unused.
+    int2            sOfs;       // [ofs_x, ofs_y] - offset between upsampled data and sign tensor.
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel specialization.
+
+struct filtered_lrelu_kernel_spec
+{
+    void*   setup;              // Function for filter kernel setup.
+    void*   exec;               // Function for main operation.
+    int2    tileOut;            // Width/height of launch tile.
+    int     numWarps;           // Number of warps per thread block, determines launch block size.
+    int     xrep;               // For processing multiple horizontal tiles per thread block.
+    int     dynamicSharedKB;    // How much dynamic shared memory the exec kernel wants.
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T, class index_t, bool signWrite, bool signRead> filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel(const filtered_lrelu_kernel_params& p, int sharedKB);
+template <class T, bool signWrite, bool signRead> void* choose_filtered_lrelu_act_kernel(void);
+template <bool signWrite, bool signRead> cudaError_t copy_filters(cudaStream_t stream);
+
+//------------------------------------------------------------------------

torch_utils/ops/filtered_lrelu.py

@@ -0,0 +1,274 @@
+# 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 os
+import numpy as np
+import torch
+import warnings
+
+from .. import custom_ops
+from .. import misc
+from . import upfirdn2d
+from . import bias_act
+
+#----------------------------------------------------------------------------
+
+_plugin = None
+
+def _init():
+    global _plugin
+    if _plugin is None:
+        _plugin = custom_ops.get_plugin(
+            module_name='filtered_lrelu_plugin',
+            sources=['filtered_lrelu.cpp', 'filtered_lrelu_wr.cu', 'filtered_lrelu_rd.cu', 'filtered_lrelu_ns.cu'],
+            headers=['filtered_lrelu.h', 'filtered_lrelu.cu'],
+            source_dir=os.path.dirname(__file__),
+            extra_cuda_cflags=['--use_fast_math'],
+        )
+    return True
+
+def _get_filter_size(f):
+    if f is None:
+        return 1, 1
+    assert isinstance(f, torch.Tensor)
+    assert 1 <= f.ndim <= 2
+    return f.shape[-1], f.shape[0] # width, height
+
+def _parse_padding(padding):
+    if isinstance(padding, int):
+        padding = [padding, padding]
+    assert isinstance(padding, (list, tuple))
+    assert all(isinstance(x, (int, np.integer)) for x in padding)
+    padding = [int(x) for x in padding]
+    if len(padding) == 2:
+        px, py = padding
+        padding = [px, px, py, py]
+    px0, px1, py0, py1 = padding
+    return px0, px1, py0, py1
+
+#----------------------------------------------------------------------------
+
+def filtered_lrelu(x, fu=None, fd=None, b=None, up=1, down=1, padding=0, gain=np.sqrt(2), slope=0.2, clamp=None, flip_filter=False, impl='cuda'):
+    r"""Filtered leaky ReLU for a batch of 2D images.
+
+    Performs the following sequence of operations for each channel:
+
+    1. Add channel-specific bias if provided (`b`).
+
+    2. Upsample the image by inserting N-1 zeros after each pixel (`up`).
+
+    3. Pad the image with the specified number of zeros on each side (`padding`).
+       Negative padding corresponds to cropping the image.
+
+    4. Convolve the image with the specified upsampling FIR filter (`fu`), shrinking it
+       so that the footprint of all output pixels lies within the input image.
+
+    5. Multiply each value by the provided gain factor (`gain`).
+
+    6. Apply leaky ReLU activation function to each value.
+
+    7. Clamp each value between -clamp and +clamp, if `clamp` parameter is provided.
+
+    8. Convolve the image with the specified downsampling FIR filter (`fd`), shrinking
+       it so that the footprint of all output pixels lies within the input image.
+
+    9. Downsample the image by keeping every Nth pixel (`down`).
+
+    The fused op is considerably more efficient than performing the same calculation
+    using standard PyTorch ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:           Float32/float16/float64 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        fu:          Float32 upsampling FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        fd:          Float32 downsampling FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        b:           Bias vector, or `None` to disable. Must be a 1D tensor of the same type
+                     as `x`. The length of vector must must match the channel dimension of `x`.
+        up:          Integer upsampling factor (default: 1).
+        down:        Integer downsampling factor. (default: 1).
+        padding:     Padding with respect to the upsampled image. Can be a single number
+                     or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        gain:        Overall scaling factor for signal magnitude (default: sqrt(2)).
+        slope:       Slope on the negative side of leaky ReLU (default: 0.2).
+        clamp:       Maximum magnitude for leaky ReLU output (default: None).
+        flip_filter: False = convolution, True = correlation (default: False).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    assert isinstance(x, torch.Tensor)
+    assert impl in ['ref', 'cuda']
+    if impl == 'cuda' and x.device.type == 'cuda' and _init():
+        return _filtered_lrelu_cuda(up=up, down=down, padding=padding, gain=gain, slope=slope, clamp=clamp, flip_filter=flip_filter).apply(x, fu, fd, b, None, 0, 0)
+    return _filtered_lrelu_ref(x, fu=fu, fd=fd, b=b, up=up, down=down, padding=padding, gain=gain, slope=slope, clamp=clamp, flip_filter=flip_filter)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def _filtered_lrelu_ref(x, fu=None, fd=None, b=None, up=1, down=1, padding=0, gain=np.sqrt(2), slope=0.2, clamp=None, flip_filter=False):
+    """Slow and memory-inefficient reference implementation of `filtered_lrelu()` using
+    existing `upfirdn2n()` and `bias_act()` ops.
+    """
+    assert isinstance(x, torch.Tensor) and x.ndim == 4
+    fu_w, fu_h = _get_filter_size(fu)
+    fd_w, fd_h = _get_filter_size(fd)
+    if b is not None:
+        assert isinstance(b, torch.Tensor) and b.dtype == x.dtype
+        misc.assert_shape(b, [x.shape[1]])
+    assert isinstance(up, int) and up >= 1
+    assert isinstance(down, int) and down >= 1
+    px0, px1, py0, py1 = _parse_padding(padding)
+    assert gain == float(gain) and gain > 0
+    assert slope == float(slope) and slope >= 0
+    assert clamp is None or (clamp == float(clamp) and clamp >= 0)
+
+    # Calculate output size.
+    batch_size, channels, in_h, in_w = x.shape
+    in_dtype = x.dtype
+    out_w = (in_w * up + (px0 + px1) - (fu_w - 1) - (fd_w - 1) + (down - 1)) // down
+    out_h = (in_h * up + (py0 + py1) - (fu_h - 1) - (fd_h - 1) + (down - 1)) // down
+
+    # Compute using existing ops.
+    x = bias_act.bias_act(x=x, b=b) # Apply bias.
+    x = upfirdn2d.upfirdn2d(x=x, f=fu, up=up, padding=[px0, px1, py0, py1], gain=up**2, flip_filter=flip_filter) # Upsample.
+    x = bias_act.bias_act(x=x, act='lrelu', alpha=slope, gain=gain, clamp=clamp) # Bias, leaky ReLU, clamp.
+    x = upfirdn2d.upfirdn2d(x=x, f=fd, down=down, flip_filter=flip_filter) # Downsample.
+
+    # Check output shape & dtype.
+    misc.assert_shape(x, [batch_size, channels, out_h, out_w])
+    assert x.dtype == in_dtype
+    return x
+
+#----------------------------------------------------------------------------
+
+_filtered_lrelu_cuda_cache = dict()
+
+def _filtered_lrelu_cuda(up=1, down=1, padding=0, gain=np.sqrt(2), slope=0.2, clamp=None, flip_filter=False):
+    """Fast CUDA implementation of `filtered_lrelu()` using custom ops.
+    """
+    assert isinstance(up, int) and up >= 1
+    assert isinstance(down, int) and down >= 1
+    px0, px1, py0, py1 = _parse_padding(padding)
+    assert gain == float(gain) and gain > 0
+    gain = float(gain)
+    assert slope == float(slope) and slope >= 0
+    slope = float(slope)
+    assert clamp is None or (clamp == float(clamp) and clamp >= 0)
+    clamp = float(clamp if clamp is not None else 'inf')
+
+    # Lookup from cache.
+    key = (up, down, px0, px1, py0, py1, gain, slope, clamp, flip_filter)
+    if key in _filtered_lrelu_cuda_cache:
+        return _filtered_lrelu_cuda_cache[key]
+
+    # Forward op.
+    class FilteredLReluCuda(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, fu, fd, b, si, sx, sy): # pylint: disable=arguments-differ
+            assert isinstance(x, torch.Tensor) and x.ndim == 4
+
+            # Replace empty up/downsample kernels with full 1x1 kernels (faster than separable).
+            if fu is None:
+                fu = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+            if fd is None:
+                fd = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+            assert 1 <= fu.ndim <= 2
+            assert 1 <= fd.ndim <= 2
+
+            # Replace separable 1x1 kernels with full 1x1 kernels when scale factor is 1.
+            if up == 1 and fu.ndim == 1 and fu.shape[0] == 1:
+                fu = fu.square()[None]
+            if down == 1 and fd.ndim == 1 and fd.shape[0] == 1:
+                fd = fd.square()[None]
+
+            # Missing sign input tensor.
+            if si is None:
+                si = torch.empty([0])
+
+            # Missing bias tensor.
+            if b is None:
+                b = torch.zeros([x.shape[1]], dtype=x.dtype, device=x.device)
+
+            # Construct internal sign tensor only if gradients are needed.
+            write_signs = (si.numel() == 0) and (x.requires_grad or b.requires_grad)
+
+            # Warn if input storage strides are not in decreasing order due to e.g. channels-last layout.
+            strides = [x.stride(i) for i in range(x.ndim) if x.size(i) > 1]
+            if any(a < b for a, b in zip(strides[:-1], strides[1:])):
+                warnings.warn("low-performance memory layout detected in filtered_lrelu input", RuntimeWarning)
+
+            # Call C++/Cuda plugin if datatype is supported.
+            if x.dtype in [torch.float16, torch.float32]:
+                if torch.cuda.current_stream(x.device) != torch.cuda.default_stream(x.device):
+                    warnings.warn("filtered_lrelu called with non-default cuda stream but concurrent execution is not supported", RuntimeWarning)
+                y, so, return_code = _plugin.filtered_lrelu(x, fu, fd, b, si, up, down, px0, px1, py0, py1, sx, sy, gain, slope, clamp, flip_filter, write_signs)
+            else:
+                return_code = -1
+
+            # No Cuda kernel found? Fall back to generic implementation. Still more memory efficient than the reference implementation because
+            # only the bit-packed sign tensor is retained for gradient computation.
+            if return_code < 0:
+                warnings.warn("filtered_lrelu called with parameters that have no optimized CUDA kernel, using generic fallback", RuntimeWarning)
+
+                y = x.add(b.unsqueeze(-1).unsqueeze(-1)) # Add bias.
+                y = upfirdn2d.upfirdn2d(x=y, f=fu, up=up, padding=[px0, px1, py0, py1], gain=up**2, flip_filter=flip_filter) # Upsample.
+                so = _plugin.filtered_lrelu_act_(y, si, sx, sy, gain, slope, clamp, write_signs) # Activation function and sign handling. Modifies y in-place.
+                y = upfirdn2d.upfirdn2d(x=y, f=fd, down=down, flip_filter=flip_filter) # Downsample.
+
+            # Prepare for gradient computation.
+            ctx.save_for_backward(fu, fd, (si if si.numel() else so))
+            ctx.x_shape = x.shape
+            ctx.y_shape = y.shape
+            ctx.s_ofs = sx, sy
+            return y
+
+        @staticmethod
+        def backward(ctx, dy): # pylint: disable=arguments-differ
+            fu, fd, si = ctx.saved_tensors
+            _, _, xh, xw = ctx.x_shape
+            _, _, yh, yw = ctx.y_shape
+            sx, sy = ctx.s_ofs
+            dx  = None # 0
+            dfu = None; assert not ctx.needs_input_grad[1]
+            dfd = None; assert not ctx.needs_input_grad[2]
+            db  = None # 3
+            dsi = None; assert not ctx.needs_input_grad[4]
+            dsx = None; assert not ctx.needs_input_grad[5]
+            dsy = None; assert not ctx.needs_input_grad[6]
+
+            if ctx.needs_input_grad[0] or ctx.needs_input_grad[3]:
+                pp = [
+                    (fu.shape[-1] - 1) + (fd.shape[-1] - 1) - px0,
+                    xw * up - yw * down + px0 - (up - 1),
+                    (fu.shape[0] - 1) + (fd.shape[0] - 1) - py0,
+                    xh * up - yh * down + py0 - (up - 1),
+                ]
+                gg = gain * (up ** 2) / (down ** 2)
+                ff = (not flip_filter)
+                sx = sx - (fu.shape[-1] - 1) + px0
+                sy = sy - (fu.shape[0]  - 1) + py0
+                dx = _filtered_lrelu_cuda(up=down, down=up, padding=pp, gain=gg, slope=slope, clamp=None, flip_filter=ff).apply(dy, fd, fu, None, si, sx, sy)
+
+            if ctx.needs_input_grad[3]:
+                db = dx.sum([0, 2, 3])
+
+            return dx, dfu, dfd, db, dsi, dsx, dsy
+
+    # Add to cache.
+    _filtered_lrelu_cuda_cache[key] = FilteredLReluCuda
+    return FilteredLReluCuda
+
+#----------------------------------------------------------------------------

torch_utils/ops/filtered_lrelu_ns.cu

@@ -0,0 +1,27 @@
+// 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.
+
+#include "filtered_lrelu.cu"
+
+// Template/kernel specializations for no signs mode (no gradients required).
+
+// Full op, 32-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int32_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int32_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Full op, 64-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int64_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int64_t, false, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Activation/signs only for generic variant. 64-bit indexing.
+template void* choose_filtered_lrelu_act_kernel<c10::Half, false, false>(void);
+template void* choose_filtered_lrelu_act_kernel<float,     false, false>(void);
+template void* choose_filtered_lrelu_act_kernel<double,    false, false>(void);
+
+// Copy filters to constant memory.
+template cudaError_t copy_filters<false, false>(cudaStream_t stream);

torch_utils/ops/filtered_lrelu_rd.cu

@@ -0,0 +1,27 @@
+// 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.
+
+#include "filtered_lrelu.cu"
+
+// Template/kernel specializations for sign read mode.
+
+// Full op, 32-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int32_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int32_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Full op, 64-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int64_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int64_t, false, true>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Activation/signs only for generic variant. 64-bit indexing.
+template void* choose_filtered_lrelu_act_kernel<c10::Half, false, true>(void);
+template void* choose_filtered_lrelu_act_kernel<float,     false, true>(void);
+template void* choose_filtered_lrelu_act_kernel<double,    false, true>(void);
+
+// Copy filters to constant memory.
+template cudaError_t copy_filters<false, true>(cudaStream_t stream);

torch_utils/ops/filtered_lrelu_wr.cu

@@ -0,0 +1,27 @@
+// 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.
+
+#include "filtered_lrelu.cu"
+
+// Template/kernel specializations for sign write mode.
+
+// Full op, 32-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int32_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int32_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Full op, 64-bit indexing.
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<c10::Half, int64_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+template filtered_lrelu_kernel_spec choose_filtered_lrelu_kernel<float,     int64_t, true, false>(const filtered_lrelu_kernel_params& p, int sharedKB);
+
+// Activation/signs only for generic variant. 64-bit indexing.
+template void* choose_filtered_lrelu_act_kernel<c10::Half, true, false>(void);
+template void* choose_filtered_lrelu_act_kernel<float,     true, false>(void);
+template void* choose_filtered_lrelu_act_kernel<double,    true, false>(void);
+
+// Copy filters to constant memory.
+template cudaError_t copy_filters<true, false>(cudaStream_t stream);

torch_utils/ops/fma.py

@@ -0,0 +1,60 @@
+# 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.
+
+"""Fused multiply-add, with slightly faster gradients than `torch.addcmul()`."""
+
+import torch
+
+#----------------------------------------------------------------------------
+
+def fma(a, b, c): # => a * b + c
+    return _FusedMultiplyAdd.apply(a, b, c)
+
+#----------------------------------------------------------------------------
+
+class _FusedMultiplyAdd(torch.autograd.Function): # a * b + c
+    @staticmethod
+    def forward(ctx, a, b, c): # pylint: disable=arguments-differ
+        out = torch.addcmul(c, a, b)
+        ctx.save_for_backward(a, b)
+        ctx.c_shape = c.shape
+        return out
+
+    @staticmethod
+    def backward(ctx, dout): # pylint: disable=arguments-differ
+        a, b = ctx.saved_tensors
+        c_shape = ctx.c_shape
+        da = None
+        db = None
+        dc = None
+
+        if ctx.needs_input_grad[0]:
+            da = _unbroadcast(dout * b, a.shape)
+
+        if ctx.needs_input_grad[1]:
+            db = _unbroadcast(dout * a, b.shape)
+
+        if ctx.needs_input_grad[2]:
+            dc = _unbroadcast(dout, c_shape)
+
+        return da, db, dc
+
+#----------------------------------------------------------------------------
+
+def _unbroadcast(x, shape):
+    extra_dims = x.ndim - len(shape)
+    assert extra_dims >= 0
+    dim = [i for i in range(x.ndim) if x.shape[i] > 1 and (i < extra_dims or shape[i - extra_dims] == 1)]
+    if len(dim):
+        x = x.sum(dim=dim, keepdim=True)
+    if extra_dims:
+        x = x.reshape(-1, *x.shape[extra_dims+1:])
+    assert x.shape == shape
+    return x
+
+#----------------------------------------------------------------------------

torch_utils/ops/grid_sample_gradfix.py

@@ -0,0 +1,77 @@
+# 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.
+
+"""Custom replacement for `torch.nn.functional.grid_sample` that
+supports arbitrarily high order gradients between the input and output.
+Only works on 2D images and assumes
+`mode='bilinear'`, `padding_mode='zeros'`, `align_corners=False`."""
+
+import torch
+
+# pylint: disable=redefined-builtin
+# pylint: disable=arguments-differ
+# pylint: disable=protected-access
+
+#----------------------------------------------------------------------------
+
+enabled = False  # Enable the custom op by setting this to true.
+
+#----------------------------------------------------------------------------
+
+def grid_sample(input, grid):
+    if _should_use_custom_op():
+        return _GridSample2dForward.apply(input, grid)
+    return torch.nn.functional.grid_sample(input=input, grid=grid, mode='bilinear', padding_mode='zeros', align_corners=False)
+
+#----------------------------------------------------------------------------
+
+def _should_use_custom_op():
+    return enabled
+
+#----------------------------------------------------------------------------
+
+class _GridSample2dForward(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input, grid):
+        assert input.ndim == 4
+        assert grid.ndim == 4
+        output = torch.nn.functional.grid_sample(input=input, grid=grid, mode='bilinear', padding_mode='zeros', align_corners=False)
+        ctx.save_for_backward(input, grid)
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        input, grid = ctx.saved_tensors
+        grad_input, grad_grid = _GridSample2dBackward.apply(grad_output, input, grid)
+        return grad_input, grad_grid
+
+#----------------------------------------------------------------------------
+
+class _GridSample2dBackward(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, grad_output, input, grid):
+        op = torch._C._jit_get_operation('aten::grid_sampler_2d_backward')
+        grad_input, grad_grid = op(grad_output, input, grid, 0, 0, False)
+        ctx.save_for_backward(grid)
+        return grad_input, grad_grid
+
+    @staticmethod
+    def backward(ctx, grad2_grad_input, grad2_grad_grid):
+        _ = grad2_grad_grid # unused
+        grid, = ctx.saved_tensors
+        grad2_grad_output = None
+        grad2_input = None
+        grad2_grid = None
+
+        if ctx.needs_input_grad[0]:
+            grad2_grad_output = _GridSample2dForward.apply(grad2_grad_input, grid)
+
+        assert not ctx.needs_input_grad[2]
+        return grad2_grad_output, grad2_input, grad2_grid
+
+#----------------------------------------------------------------------------

torch_utils/ops/upfirdn2d.cpp

@@ -0,0 +1,107 @@
+// 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.
+
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "upfirdn2d.h"
+
+//------------------------------------------------------------------------
+
+static torch::Tensor upfirdn2d(torch::Tensor x, torch::Tensor f, int upx, int upy, int downx, int downy, int padx0, int padx1, int pady0, int pady1, bool flip, float gain)
+{
+    // Validate arguments.
+    TORCH_CHECK(x.is_cuda(), "x must reside on CUDA device");
+    TORCH_CHECK(f.device() == x.device(), "f must reside on the same device as x");
+    TORCH_CHECK(f.dtype() == torch::kFloat, "f must be float32");
+    TORCH_CHECK(x.numel() <= INT_MAX, "x is too large");
+    TORCH_CHECK(f.numel() <= INT_MAX, "f is too large");
+    TORCH_CHECK(x.numel() > 0, "x has zero size");
+    TORCH_CHECK(f.numel() > 0, "f has zero size");
+    TORCH_CHECK(x.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(f.dim() == 2, "f must be rank 2");
+    TORCH_CHECK((x.size(0)-1)*x.stride(0) + (x.size(1)-1)*x.stride(1) + (x.size(2)-1)*x.stride(2) + (x.size(3)-1)*x.stride(3) <= INT_MAX, "x memory footprint is too large");
+    TORCH_CHECK(f.size(0) >= 1 && f.size(1) >= 1, "f must be at least 1x1");
+    TORCH_CHECK(upx >= 1 && upy >= 1, "upsampling factor must be at least 1");
+    TORCH_CHECK(downx >= 1 && downy >= 1, "downsampling factor must be at least 1");
+
+    // Create output tensor.
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    int outW = ((int)x.size(3) * upx + padx0 + padx1 - (int)f.size(1) + downx) / downx;
+    int outH = ((int)x.size(2) * upy + pady0 + pady1 - (int)f.size(0) + downy) / downy;
+    TORCH_CHECK(outW >= 1 && outH >= 1, "output must be at least 1x1");
+    torch::Tensor y = torch::empty({x.size(0), x.size(1), outH, outW}, x.options(), x.suggest_memory_format());
+    TORCH_CHECK(y.numel() <= INT_MAX, "output is too large");
+    TORCH_CHECK((y.size(0)-1)*y.stride(0) + (y.size(1)-1)*y.stride(1) + (y.size(2)-1)*y.stride(2) + (y.size(3)-1)*y.stride(3) <= INT_MAX, "output memory footprint is too large");
+
+    // Initialize CUDA kernel parameters.
+    upfirdn2d_kernel_params p;
+    p.x             = x.data_ptr();
+    p.f             = f.data_ptr<float>();
+    p.y             = y.data_ptr();
+    p.up            = make_int2(upx, upy);
+    p.down          = make_int2(downx, downy);
+    p.pad0          = make_int2(padx0, pady0);
+    p.flip          = (flip) ? 1 : 0;
+    p.gain          = gain;
+    p.inSize        = make_int4((int)x.size(3), (int)x.size(2), (int)x.size(1), (int)x.size(0));
+    p.inStride      = make_int4((int)x.stride(3), (int)x.stride(2), (int)x.stride(1), (int)x.stride(0));
+    p.filterSize    = make_int2((int)f.size(1), (int)f.size(0));
+    p.filterStride  = make_int2((int)f.stride(1), (int)f.stride(0));
+    p.outSize       = make_int4((int)y.size(3), (int)y.size(2), (int)y.size(1), (int)y.size(0));
+    p.outStride     = make_int4((int)y.stride(3), (int)y.stride(2), (int)y.stride(1), (int)y.stride(0));
+    p.sizeMajor     = (p.inStride.z == 1) ? p.inSize.w : p.inSize.w * p.inSize.z;
+    p.sizeMinor     = (p.inStride.z == 1) ? p.inSize.z : 1;
+
+    // Choose CUDA kernel.
+    upfirdn2d_kernel_spec spec;
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "upfirdn2d_cuda", [&]
+    {
+        spec = choose_upfirdn2d_kernel<scalar_t>(p);
+    });
+
+    // Set looping options.
+    p.loopMajor     = (p.sizeMajor - 1) / 16384 + 1;
+    p.loopMinor     = spec.loopMinor;
+    p.loopX         = spec.loopX;
+    p.launchMinor   = (p.sizeMinor - 1) / p.loopMinor + 1;
+    p.launchMajor   = (p.sizeMajor - 1) / p.loopMajor + 1;
+
+    // Compute grid size.
+    dim3 blockSize, gridSize;
+    if (spec.tileOutW < 0) // large
+    {
+        blockSize = dim3(4, 32, 1);
+        gridSize = dim3(
+            ((p.outSize.y - 1) / blockSize.x + 1) * p.launchMinor,
+            (p.outSize.x - 1) / (blockSize.y * p.loopX) + 1,
+            p.launchMajor);
+    }
+    else // small
+    {
+        blockSize = dim3(256, 1, 1);
+        gridSize = dim3(
+            ((p.outSize.y - 1) / spec.tileOutH + 1) * p.launchMinor,
+            (p.outSize.x - 1) / (spec.tileOutW * p.loopX) + 1,
+            p.launchMajor);
+    }
+
+    // Launch CUDA kernel.
+    void* args[] = {&p};
+    AT_CUDA_CHECK(cudaLaunchKernel(spec.kernel, gridSize, blockSize, args, 0, at::cuda::getCurrentCUDAStream()));
+    return y;
+}
+
+//------------------------------------------------------------------------
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("upfirdn2d", &upfirdn2d);
+}
+
+//------------------------------------------------------------------------

torch_utils/ops/upfirdn2d.cu

@@ -0,0 +1,384 @@
+// 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.
+
+#include <c10/util/Half.h>
+#include "upfirdn2d.h"
+
+//------------------------------------------------------------------------
+// Helpers.
+
+template <class T> struct InternalType;
+template <> struct InternalType<double>     { typedef double scalar_t; };
+template <> struct InternalType<float>      { typedef float  scalar_t; };
+template <> struct InternalType<c10::Half>  { typedef float  scalar_t; };
+
+static __device__ __forceinline__ int floor_div(int a, int b)
+{
+    int t = 1 - a / b;
+    return (a + t * b) / b - t;
+}
+
+//------------------------------------------------------------------------
+// Generic CUDA implementation for large filters.
+
+template <class T> static __global__ void upfirdn2d_kernel_large(upfirdn2d_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+
+    // Calculate thread index.
+    int minorBase = blockIdx.x * blockDim.x + threadIdx.x;
+    int outY = minorBase / p.launchMinor;
+    minorBase -= outY * p.launchMinor;
+    int outXBase = blockIdx.y * p.loopX * blockDim.y + threadIdx.y;
+    int majorBase = blockIdx.z * p.loopMajor;
+    if (outXBase >= p.outSize.x | outY >= p.outSize.y | majorBase >= p.sizeMajor)
+        return;
+
+    // Setup Y receptive field.
+    int midY = outY * p.down.y + p.up.y - 1 - p.pad0.y;
+    int inY = min(max(floor_div(midY, p.up.y), 0), p.inSize.y);
+    int h = min(max(floor_div(midY + p.filterSize.y, p.up.y), 0), p.inSize.y) - inY;
+    int filterY = midY + p.filterSize.y - (inY + 1) * p.up.y;
+    if (p.flip)
+        filterY = p.filterSize.y - 1 - filterY;
+
+    // Loop over major, minor, and X.
+    for (int majorIdx = 0, major = majorBase; majorIdx < p.loopMajor & major < p.sizeMajor; majorIdx++, major++)
+    for (int minorIdx = 0, minor = minorBase; minorIdx < p.loopMinor & minor < p.sizeMinor; minorIdx++, minor += p.launchMinor)
+    {
+        int nc = major * p.sizeMinor + minor;
+        int n = nc / p.inSize.z;
+        int c = nc - n * p.inSize.z;
+        for (int loopX = 0, outX = outXBase; loopX < p.loopX & outX < p.outSize.x; loopX++, outX += blockDim.y)
+        {
+            // Setup X receptive field.
+            int midX = outX * p.down.x + p.up.x - 1 - p.pad0.x;
+            int inX = min(max(floor_div(midX, p.up.x), 0), p.inSize.x);
+            int w = min(max(floor_div(midX + p.filterSize.x, p.up.x), 0), p.inSize.x) - inX;
+            int filterX = midX + p.filterSize.x - (inX + 1) * p.up.x;
+            if (p.flip)
+                filterX = p.filterSize.x - 1 - filterX;
+
+            // Initialize pointers.
+            const T* xp = &((const T*)p.x)[inX * p.inStride.x + inY * p.inStride.y + c * p.inStride.z + n * p.inStride.w];
+            const float* fp = &p.f[filterX * p.filterStride.x + filterY * p.filterStride.y];
+            int filterStepX = ((p.flip) ? p.up.x : -p.up.x) * p.filterStride.x;
+            int filterStepY = ((p.flip) ? p.up.y : -p.up.y) * p.filterStride.y;
+
+            // Inner loop.
+            scalar_t v = 0;
+            for (int y = 0; y < h; y++)
+            {
+                for (int x = 0; x < w; x++)
+                {
+                    v += (scalar_t)(*xp) * (scalar_t)(*fp);
+                    xp += p.inStride.x;
+                    fp += filterStepX;
+                }
+                xp += p.inStride.y - w * p.inStride.x;
+                fp += filterStepY - w * filterStepX;
+            }
+
+            // Store result.
+            v *= p.gain;
+            ((T*)p.y)[outX * p.outStride.x + outY * p.outStride.y + c * p.outStride.z + n * p.outStride.w] = (T)v;
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// Specialized CUDA implementation for small filters.
+
+template <class T, int upx, int upy, int downx, int downy, int filterW, int filterH, int tileOutW, int tileOutH, int loopMinor>
+static __global__ void upfirdn2d_kernel_small(upfirdn2d_kernel_params p)
+{
+    typedef typename InternalType<T>::scalar_t scalar_t;
+    const int tileInW = ((tileOutW - 1) * downx + filterW - 1) / upx + 1;
+    const int tileInH = ((tileOutH - 1) * downy + filterH - 1) / upy + 1;
+    __shared__ volatile scalar_t sf[filterH][filterW];
+    __shared__ volatile scalar_t sx[tileInH][tileInW][loopMinor];
+
+    // Calculate tile index.
+    int minorBase = blockIdx.x;
+    int tileOutY = minorBase / p.launchMinor;
+    minorBase -= tileOutY * p.launchMinor;
+    minorBase *= loopMinor;
+    tileOutY *= tileOutH;
+    int tileOutXBase = blockIdx.y * p.loopX * tileOutW;
+    int majorBase = blockIdx.z * p.loopMajor;
+    if (tileOutXBase >= p.outSize.x | tileOutY >= p.outSize.y | majorBase >= p.sizeMajor)
+        return;
+
+    // Load filter (flipped).
+    for (int tapIdx = threadIdx.x; tapIdx < filterH * filterW; tapIdx += blockDim.x)
+    {
+        int fy = tapIdx / filterW;
+        int fx = tapIdx - fy * filterW;
+        scalar_t v = 0;
+        if (fx < p.filterSize.x & fy < p.filterSize.y)
+        {
+            int ffx = (p.flip) ? fx : p.filterSize.x - 1 - fx;
+            int ffy = (p.flip) ? fy : p.filterSize.y - 1 - fy;
+            v = (scalar_t)p.f[ffx * p.filterStride.x + ffy * p.filterStride.y];
+        }
+        sf[fy][fx] = v;
+    }
+
+    // Loop over major and X.
+    for (int majorIdx = 0, major = majorBase; majorIdx < p.loopMajor & major < p.sizeMajor; majorIdx++, major++)
+    {
+        int baseNC = major * p.sizeMinor + minorBase;
+        int n = baseNC / p.inSize.z;
+        int baseC = baseNC - n * p.inSize.z;
+        for (int loopX = 0, tileOutX = tileOutXBase; loopX < p.loopX & tileOutX < p.outSize.x; loopX++, tileOutX += tileOutW)
+        {
+            // Load input pixels.
+            int tileMidX = tileOutX * downx + upx - 1 - p.pad0.x;
+            int tileMidY = tileOutY * downy + upy - 1 - p.pad0.y;
+            int tileInX = floor_div(tileMidX, upx);
+            int tileInY = floor_div(tileMidY, upy);
+            __syncthreads();
+            for (int inIdx = threadIdx.x; inIdx < tileInH * tileInW * loopMinor; inIdx += blockDim.x)
+            {
+                int relC = inIdx;
+                int relInX = relC / loopMinor;
+                int relInY = relInX / tileInW;
+                relC -= relInX * loopMinor;
+                relInX -= relInY * tileInW;
+                int c = baseC + relC;
+                int inX = tileInX + relInX;
+                int inY = tileInY + relInY;
+                scalar_t v = 0;
+                if (inX >= 0 & inY >= 0 & inX < p.inSize.x & inY < p.inSize.y & c < p.inSize.z)
+                    v = (scalar_t)((const T*)p.x)[inX * p.inStride.x + inY * p.inStride.y + c * p.inStride.z + n * p.inStride.w];
+                sx[relInY][relInX][relC] = v;
+            }
+
+            // Loop over output pixels.
+            __syncthreads();
+            for (int outIdx = threadIdx.x; outIdx < tileOutH * tileOutW * loopMinor; outIdx += blockDim.x)
+            {
+                int relC = outIdx;
+                int relOutX = relC / loopMinor;
+                int relOutY = relOutX / tileOutW;
+                relC -= relOutX * loopMinor;
+                relOutX -= relOutY * tileOutW;
+                int c = baseC + relC;
+                int outX = tileOutX + relOutX;
+                int outY = tileOutY + relOutY;
+
+                // Setup receptive field.
+                int midX = tileMidX + relOutX * downx;
+                int midY = tileMidY + relOutY * downy;
+                int inX = floor_div(midX, upx);
+                int inY = floor_div(midY, upy);
+                int relInX = inX - tileInX;
+                int relInY = inY - tileInY;
+                int filterX = (inX + 1) * upx - midX - 1; // flipped
+                int filterY = (inY + 1) * upy - midY - 1; // flipped
+
+                // Inner loop.
+                if (outX < p.outSize.x & outY < p.outSize.y & c < p.outSize.z)
+                {
+                    scalar_t v = 0;
+                    #pragma unroll
+                    for (int y = 0; y < filterH / upy; y++)
+                        #pragma unroll
+                        for (int x = 0; x < filterW / upx; x++)
+                            v += sx[relInY + y][relInX + x][relC] * sf[filterY + y * upy][filterX + x * upx];
+                    v *= p.gain;
+                    ((T*)p.y)[outX * p.outStride.x + outY * p.outStride.y + c * p.outStride.z + n * p.outStride.w] = (T)v;
+                }
+            }
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> upfirdn2d_kernel_spec choose_upfirdn2d_kernel(const upfirdn2d_kernel_params& p)
+{
+    int s = p.inStride.z, fx = p.filterSize.x, fy = p.filterSize.y;
+    upfirdn2d_kernel_spec spec = {(void*)upfirdn2d_kernel_large<T>, -1,-1,1, 4}; // contiguous
+    if (s == 1)           spec = {(void*)upfirdn2d_kernel_large<T>, -1,-1,4, 1}; // channels_last
+
+    // No up/downsampling.
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,24, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,16, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 7  && fy <= 7 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 7,7,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 6,6,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 5  && fy <= 5 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 5,5,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 3  && fy <= 3 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 3,3,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,1,  128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,1,  128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 8,1,   128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,24,  32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,16,  32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,8,   32,32,1>, 32,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,24, 32,32,1>,  32,32,1,  1};
+        if (s == 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,16, 32,32,1>,  32,32,1,  1};
+        if (s == 1 && fx <= 7  && fy <= 7 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 7,7,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 6,6,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 5  && fy <= 5 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 5,5,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 3  && fy <= 3 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 3,3,   16,16,8>,  16,16,8,  1};
+        if (s == 1 && fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,1,  128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,1,  128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 8,1,   128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,24,  1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,16,  1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,8,   1,128,16>, 1,128,16, 1};
+    }
+
+    // 2x upsampling.
+    if (p.up.x == 2 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 24,24, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 16,16, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 8,8,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 6,6,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 4,4,   64,16,1>, 64,16,1, 1};
+        if (s != 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 2,2,   64,16,1>, 64,16,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 24,24, 32,32,1>, 32,32,1, 1};
+        if (s == 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 16,16, 32,32,1>, 32,32,1, 1};
+        if (s == 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 8,8,   16,16,8>, 16,16,8, 1};
+        if (s == 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 6,6,   16,16,8>, 16,16,8, 1};
+        if (s == 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 4,4,   16,16,8>, 16,16,8, 1};
+        if (s == 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 2,2,   16,16,8>, 16,16,8, 1};
+    }
+    if (p.up.x == 2 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 24,1, 128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 16,1, 128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 8,1,  128,8,1>, 128,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 24,1, 128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 16,1, 128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 8,1,  128,1,16>, 128,1,16, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,24, 32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,16, 32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,8,  32,32,1>, 32,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,24, 1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,16, 1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1 && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,8,  1,128,16>, 1,128,16, 1};
+    }
+
+    // 2x downsampling.
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 2)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 24,24, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 16,16, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 8,8,   32,8,1>,  32,8,1,  1};
+        if (s != 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 6,6,   32,8,1>,  32,8,1,  1};
+        if (s != 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 4,4,   32,8,1>,  32,8,1,  1};
+        if (s != 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 2,2,   32,8,1>,  32,8,1,  1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 24,24, 16,16,1>, 16,16,1, 1};
+        if (s == 1 && fx <= 16 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 16,16, 16,16,1>, 16,16,1, 1};
+        if (s == 1 && fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 8,8,   8,8,8>,   8,8,8,   1};
+        if (s == 1 && fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 6,6,   8,8,8>,   8,8,8,   1};
+        if (s == 1 && fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 4,4,   8,8,8>,   8,8,8,   1};
+        if (s == 1 && fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 2,2,   8,8,8>,   8,8,8,   1};
+    }
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 24,1, 64,8,1>, 64,8,1, 1};
+        if (s != 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 16,1, 64,8,1>, 64,8,1, 1};
+        if (s != 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 8,1,  64,8,1>, 64,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 24 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 24,1, 64,1,8>, 64,1,8, 1};
+        if (s == 1 && fx <= 16 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 16,1, 64,1,8>, 64,1,8, 1};
+        if (s == 1 && fx <= 8  && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 8,1,  64,1,8>, 64,1,8, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 2)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,24, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,16, 32,16,1>, 32,16,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,8,  32,16,1>, 32,16,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,24, 1,64,8>, 1,64,8, 1};
+        if (s == 1 && fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,16, 1,64,8>, 1,64,8, 1};
+        if (s == 1 && fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,8,  1,64,8>, 1,64,8, 1};
+    }
+
+    // 4x upsampling.
+    if (p.up.x == 4 && p.up.y == 4 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 48 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 48,48, 64,32,1>, 64,32,1, 1};
+        if (s != 1 && fx <= 32 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 32,32, 64,32,1>, 64,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 48 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 48,48, 32,32,1>, 32,32,1, 1};
+        if (s == 1 && fx <= 32 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 4,4, 1,1, 32,32, 32,32,1>, 32,32,1, 1};
+    }
+    if (p.up.x == 4 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 48,1, 128,8,1>, 128,8,1, 1};
+        if (s != 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 32,1, 128,8,1>, 128,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 48,1, 128,1,16>, 128,1,16, 1};
+        if (s == 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 4,1, 1,1, 32,1, 128,1,16>, 128,1,16, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 4 && p.down.x == 1 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,48, 32,32,1>, 32,32,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,32, 32,32,1>, 32,32,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,48, 1,128,16>, 1,128,16, 1};
+        if (s == 1 && fx <= 1 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,4, 1,1, 1,32, 1,128,16>, 1,128,16, 1};
+    }
+
+    // 4x downsampling (inefficient).
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 4 && p.down.y == 1)
+    {
+        // contiguous
+        if (s != 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 48,1, 32,8,1>, 32,8,1, 1};
+        if (s != 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 32,1, 32,8,1>, 32,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 48 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 48,1, 32,1,8>, 32,1,8, 1};
+        if (s == 1 && fx <= 32 && fy <= 1) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 4,1, 32,1, 32,1,8>, 32,1,8, 1};
+    }
+    if (p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 4)
+    {
+        // contiguous
+        if (s != 1 && fx <= 1 && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,48, 32,8,1>, 32,8,1, 1};
+        if (s != 1 && fx <= 1 && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,32, 32,8,1>, 32,8,1, 1};
+        // channels_last
+        if (s == 1 && fx <= 1  && fy <= 48) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,48, 1,32,8>, 1,32,8, 1};
+        if (s == 1 && fx <= 1  && fy <= 32) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,4, 1,32, 1,32,8>, 1,32,8, 1};
+    }
+    return spec;
+}
+
+//------------------------------------------------------------------------
+// Template specializations.
+
+template upfirdn2d_kernel_spec choose_upfirdn2d_kernel<double>   (const upfirdn2d_kernel_params& p);
+template upfirdn2d_kernel_spec choose_upfirdn2d_kernel<float>    (const upfirdn2d_kernel_params& p);
+template upfirdn2d_kernel_spec choose_upfirdn2d_kernel<c10::Half>(const upfirdn2d_kernel_params& p);
+
+//------------------------------------------------------------------------

torch_utils/ops/upfirdn2d.h

@@ -0,0 +1,59 @@
+// 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.
+
+#include <cuda_runtime.h>
+
+//------------------------------------------------------------------------
+// CUDA kernel parameters.
+
+struct upfirdn2d_kernel_params
+{
+    const void*     x;
+    const float*    f;
+    void*           y;
+
+    int2            up;
+    int2            down;
+    int2            pad0;
+    int             flip;
+    float           gain;
+
+    int4            inSize;         // [width, height, channel, batch]
+    int4            inStride;
+    int2            filterSize;     // [width, height]
+    int2            filterStride;
+    int4            outSize;        // [width, height, channel, batch]
+    int4            outStride;
+    int             sizeMinor;
+    int             sizeMajor;
+
+    int             loopMinor;
+    int             loopMajor;
+    int             loopX;
+    int             launchMinor;
+    int             launchMajor;
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel specialization.
+
+struct upfirdn2d_kernel_spec
+{
+    void*   kernel;
+    int     tileOutW;
+    int     tileOutH;
+    int     loopMinor;
+    int     loopX;
+};
+
+//------------------------------------------------------------------------
+// CUDA kernel selection.
+
+template <class T> upfirdn2d_kernel_spec choose_upfirdn2d_kernel(const upfirdn2d_kernel_params& p);
+
+//------------------------------------------------------------------------