init code

README.md

@@ -1,6 +1,6 @@
 ---
 title: EmojiGAN
-emoji: 📚
+emoji: 😎
 colorFrom: blue
 colorTo: gray
 sdk: gradio

app.py

@@ -0,0 +1,74 @@
+import gradio as gr
+import os
+import numpy as np
+import torch
+import pickle
+import types
+
+from huggingface_hub import hf_hub_url, cached_download
+
+TOKEN = os.environ['TOKEN']
+
+with open(cached_download(hf_hub_url('mfrashad/stylegan2_emoji_512', 'stylegan2_emoji_512.pkl'), use_auth_token=TOKEN), 'rb') as f:
+    G = pickle.load(f)['G_ema']# torch.nn.Module
+
+device = torch.device("cpu")
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+    G = G.to(device)
+else:
+    _old_forward = G.forward
+
+    def _new_forward(self, *args, **kwargs):
+        kwargs["force_fp32"] = True
+        return _old_forward(*args, **kwargs)
+
+    G.forward = types.MethodType(_new_forward, G)
+
+    _old_synthesis_forward = G.synthesis.forward
+
+    def _new_synthesis_forward(self, *args, **kwargs):
+        kwargs["force_fp32"] = True
+        return _old_synthesis_forward(*args, **kwargs)
+
+    G.synthesis.forward = types.MethodType(_new_synthesis_forward, G.synthesis)
+
+
+def generate(num_images, interpolate):
+    if interpolate:
+        z1 = torch.randn([1, G.z_dim])# latent codes
+        z2 = torch.randn([1, G.z_dim])# latent codes
+        zs = torch.cat([z1 + (z2 - z1) * i / (num_images-1) for i in range(num_images)], 0)
+    else:
+        zs = torch.randn([num_images, G.z_dim])# latent codes
+    with torch.no_grad():
+        zs = zs.to(device)
+        img = G(zs, None, force_fp32=True, noise_mode='const') 
+        img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
+    return img.cpu().numpy()
+
+demo = gr.Blocks()
+
+def infer(num_images, interpolate):
+    img = generate(round(num_images), interpolate)
+    imgs = list(img)
+    return imgs
+
+with demo:
+    gr.Markdown(
+    """
+    # EmojiGAN
+    Generate Emojis with AI (StyleGAN2-ADA). Made by [mfrashad](https://github.com/mfrashad)
+    """)
+    images_num = gr.inputs.Slider(default=1, label="Num Images", minimum=1, maximum=16, step=1)
+    interpolate = gr.inputs.Checkbox(default=False, label="Interpolate")
+    submit = gr.Button("Generate")
+    
+    
+    out = gr.Gallery()
+
+    submit.click(fn=infer, 
+               inputs=[images_num, interpolate], 
+               outputs=out)
+
+demo.launch()

requirements.txt

@@ -0,0 +1,5 @@
+torch
+numpy
+Pillow
+imageio
+scipy

torch_utils/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty

torch_utils/custom_ops.py

@@ -0,0 +1,126 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import glob
+import torch
+import torch.utils.cpp_extension
+import importlib
+import hashlib
+import shutil
+from pathlib import Path
+
+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
+
+#----------------------------------------------------------------------------
+# Main entry point for compiling and loading C++/CUDA plugins.
+
+_cached_plugins = dict()
+
+def get_plugin(module_name, sources, **build_kwargs):
+    assert verbosity in ['none', 'brief', 'full']
+
+    # 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)
+
+    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
+
+        # Compile and load.
+        verbose_build = (verbosity == 'full')
+
+        # 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.)
+        source_dirs_set = set(os.path.dirname(source) for source in sources)
+        if len(source_dirs_set) == 1 and ('TORCH_EXTENSIONS_DIR' in os.environ):
+            all_source_files = sorted(list(x for x in Path(list(source_dirs_set)[0]).iterdir() if x.is_file()))
+
+            # Compute a combined hash digest for all source files in the same
+            # custom op directory (usually .cu, .cpp, .py and .h files).
+            hash_md5 = hashlib.md5()
+            for src in all_source_files:
+                with open(src, 'rb') as f:
+                    hash_md5.update(f.read())
+            build_dir = torch.utils.cpp_extension._get_build_directory(module_name, verbose=verbose_build) # pylint: disable=protected-access
+            digest_build_dir = os.path.join(build_dir, hash_md5.hexdigest())
+
+            if not os.path.isdir(digest_build_dir):
+                os.makedirs(digest_build_dir, exist_ok=True)
+                baton = FileBaton(os.path.join(digest_build_dir, 'lock'))
+                if baton.try_acquire():
+                    try:
+                        for src in all_source_files:
+                            shutil.copyfile(src, os.path.join(digest_build_dir, os.path.basename(src)))
+                    finally:
+                        baton.release()
+                else:
+                    # Someone else is copying source files under the digest dir,
+                    # wait until done and continue.
+                    baton.wait()
+            digest_sources = [os.path.join(digest_build_dir, os.path.basename(x)) for x in sources]
+            torch.utils.cpp_extension.load(name=module_name, build_directory=build_dir,
+                verbose=verbose_build, sources=digest_sources, **build_kwargs)
+        else:
+            torch.utils.cpp_extension.load(name=module_name, verbose=verbose_build, sources=sources, **build_kwargs)
+        module = importlib.import_module(module_name)
+
+    except:
+        if verbosity == 'brief':
+            print('Failed!')
+        raise
+
+    # Print status and add to cache.
+    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 +1,262 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import 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 suppress known warnings in torch.jit.trace().
+
+class suppress_tracer_warnings(warnings.catch_warnings):
+    def __enter__(self):
+        super().__enter__()
+        warnings.simplefilter('ignore', category=torch.jit.TracerWarning)
+        return self
+
+#----------------------------------------------------------------------------
+# 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 = {name: tensor for name, tensor in 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()
+        other = tensor.clone()
+        torch.distributed.broadcast(tensor=other, src=0)
+        assert (nan_to_num(tensor) == nan_to_num(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(e.outputs[0].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.  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.  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.  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.  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 +1,212 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom PyTorch ops for efficient bias and activation."""
+
+import os
+import warnings
+import numpy as np
+import torch
+import dnnlib
+import traceback
+
+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),
+}
+
+#----------------------------------------------------------------------------
+
+_inited = False
+_plugin = None
+_null_tensor = torch.empty([0])
+
+def _init():
+    global _inited, _plugin
+    if not _inited:
+        _inited = True
+        sources = ['bias_act.cpp', 'bias_act.cu']
+        sources = [os.path.join(os.path.dirname(__file__), s) for s in sources]
+        try:
+            _plugin = custom_ops.get_plugin('bias_act_plugin', sources=sources, extra_cuda_cflags=['--use_fast_math'])
+        except:
+            warnings.warn('Failed to build CUDA kernels for bias_act. Falling back to slow reference implementation. Details:\n\n' + traceback.format_exc())
+    return _plugin is not None
+
+#----------------------------------------------------------------------------
+
+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 +1,170 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom replacement for `torch.nn.functional.conv2d` that supports
+arbitrarily high order gradients with zero performance penalty."""
+
+import warnings
+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():
+    global weight_gradients_disabled
+    old = weight_gradients_disabled
+    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
+    if any(torch.__version__.startswith(x) for x in ['1.7.', '1.8.', '1.9']):
+        return True
+    warnings.warn(f'conv2d_gradfix not supported on PyTorch {torch.__version__}. Falling back to torch.nn.functional.conv2d().')
+    return False
+
+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()
+
+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
+            if not transpose:
+                output = torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, **common_kwargs)
+            else: # transpose
+                output = torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, output_padding=output_padding, **common_kwargs)
+            ctx.save_for_backward(input, weight)
+            return output
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            input, weight = ctx.saved_tensors
+            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)
+                grad_input = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs).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):
+            op = torch._C._jit_get_operation('aten::cudnn_convolution_backward_weight' if not transpose else 'aten::cudnn_convolution_transpose_backward_weight')
+            flags = [torch.backends.cudnn.benchmark, torch.backends.cudnn.deterministic, torch.backends.cudnn.allow_tf32]
+            grad_weight = op(weight_shape, grad_output, input, padding, stride, dilation, groups, *flags)
+            assert grad_weight.shape == weight_shape
+            ctx.save_for_backward(grad_output, input)
+            return grad_weight
+
+        @staticmethod
+        def backward(ctx, grad2_grad_weight):
+            grad_output, input = ctx.saved_tensors
+            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)
+                grad2_input = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs).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 +1,156 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""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.
+    if not flip_weight: # conv2d() actually performs correlation (flip_weight=True) not convolution (flip_weight=False).
+        w = w.flip([2, 3])
+
+    # Workaround performance pitfall in cuDNN 8.0.5, triggered when using
+    # 1x1 kernel + memory_format=channels_last + less than 64 channels.
+    if kw == 1 and kh == 1 and stride == 1 and padding in [0, [0, 0], (0, 0)] and not transpose:
+        if x.stride()[1] == 1 and min(out_channels, in_channels_per_group) < 64:
+            if out_channels <= 4 and groups == 1:
+                in_shape = x.shape
+                x = w.squeeze(3).squeeze(2) @ x.reshape([in_shape[0], in_channels_per_group, -1])
+                x = x.reshape([in_shape[0], out_channels, in_shape[2], in_shape[3]])
+            else:
+                x = x.to(memory_format=torch.contiguous_format)
+                w = w.to(memory_format=torch.contiguous_format)
+                x = conv2d_gradfix.conv2d(x, w, groups=groups)
+            return x.to(memory_format=torch.channels_last)
+
+    # Otherwise => 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/fma.py

@@ -0,0 +1,60 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""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 +1,83 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""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 warnings
+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():
+    if not enabled:
+        return False
+    if any(torch.__version__.startswith(x) for x in ['1.7.', '1.8.', '1.9']):
+        return True
+    warnings.warn(f'grid_sample_gradfix not supported on PyTorch {torch.__version__}. Falling back to torch.nn.functional.grid_sample().')
+    return False
+
+#----------------------------------------------------------------------------
+
+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 +1,103 @@
+// Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#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.dim() == 4, "x must be rank 4");
+    TORCH_CHECK(f.dim() == 2, "f must be rank 2");
+    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");
+
+    // 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 +1,350 @@
+// Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+//
+// NVIDIA CORPORATION and its licensors retain all intellectual property
+// and proprietary rights in and to this software, related documentation
+// and any modifications thereto.  Any use, reproduction, disclosure or
+// distribution of this software and related documentation without an express
+// license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+#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
+
+    if (s != 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1) // contiguous
+    {
+        if (fx <= 7  && fy <= 7 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 7,7,  64,16,1>, 64,16,1, 1};
+        if (fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 6,6,  64,16,1>, 64,16,1, 1};
+        if (fx <= 5  && fy <= 5 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 5,5,  64,16,1>, 64,16,1, 1};
+        if (fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,  64,16,1>, 64,16,1, 1};
+        if (fx <= 3  && fy <= 3 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 3,3,  64,16,1>, 64,16,1, 1};
+        if (fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 20 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 20,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 12 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 12,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 8,1,  128,8,1>, 128,8,1, 1};
+        if (fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,24, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 20) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,20, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,16, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 12) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,12, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,8,  32,32,1>, 32,32,1, 1};
+    }
+    if (s == 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1) // channels_last
+    {
+        if (fx <= 7  && fy <= 7 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 7,7,  16,16,8>,  16,16,8,  1};
+        if (fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,  16,16,8>,  16,16,8,  1};
+        if (fx <= 5  && fy <= 5 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,  16,16,8>,  16,16,8,  1};
+        if (fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,  16,16,8>,  16,16,8,  1};
+        if (fx <= 3  && fy <= 3 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 4,4,  16,16,8>,  16,16,8,  1};
+        if (fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 24,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 20 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 20,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 16,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 12 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 12,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 8,1,  128,1,16>, 128,1,16, 1};
+        if (fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,24, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 20) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,20, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,16, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 12) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,12, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,1, 1,8,  1,128,16>, 1,128,16, 1};
+    }
+    if (s != 1 && p.up.x == 2 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1) // contiguous
+    {
+        if (fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 8,8, 64,16,1>, 64,16,1, 1};
+        if (fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 6,6, 64,16,1>, 64,16,1, 1};
+        if (fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 4,4, 64,16,1>, 64,16,1, 1};
+        if (fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 2,2, 64,16,1>, 64,16,1, 1};
+    }
+    if (s == 1 && p.up.x == 2 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1) // channels_last
+    {
+        if (fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 8,8, 16,16,8>, 16,16,8, 1};
+        if (fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 6,6, 16,16,8>, 16,16,8, 1};
+        if (fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 4,4, 16,16,8>, 16,16,8, 1};
+        if (fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,2, 1,1, 2,2, 16,16,8>, 16,16,8, 1};
+    }
+    if (s != 1 && p.up.x == 2 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1) // contiguous
+    {
+        if (fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 24,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 20 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 20,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 16,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 12 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 12,1, 128,8,1>, 128,8,1, 1};
+        if (fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 8,1,  128,8,1>, 128,8,1, 1};
+    }
+    if (s == 1 && p.up.x == 2 && p.up.y == 1 && p.down.x == 1 && p.down.y == 1) // channels_last
+    {
+        if (fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 24,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 20 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 20,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 16,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 12 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 12,1, 128,1,16>, 128,1,16, 1};
+        if (fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 2,1, 1,1, 8,1,  128,1,16>, 128,1,16, 1};
+    }
+    if (s != 1 && p.up.x == 1 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1) // contiguous
+    {
+        if (fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,24, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 20) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,20, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,16, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 12) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,12, 32,32,1>, 32,32,1, 1};
+        if (fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,8,  32,32,1>, 32,32,1, 1};
+    }
+    if (s == 1 && p.up.x == 1 && p.up.y == 2 && p.down.x == 1 && p.down.y == 1) // channels_last
+    {
+        if (fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,24, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 20) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,20, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,16, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 12) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,12, 1,128,16>, 1,128,16, 1};
+        if (fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,2, 1,1, 1,8,  1,128,16>, 1,128,16, 1};
+    }
+    if (s != 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 2) // contiguous
+    {
+        if (fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 8,8,  32,8,1>, 32,8,1, 1};
+        if (fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 6,6,  32,8,1>, 32,8,1, 1};
+        if (fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 4,4,  32,8,1>, 32,8,1, 1};
+        if (fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 2,2,  32,8,1>, 32,8,1, 1};
+    }
+    if (s == 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 2) // channels_last
+    {
+        if (fx <= 8  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 8,8,  8,8,8>, 8,8,8, 1};
+        if (fx <= 6  && fy <= 6 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 6,6,  8,8,8>, 8,8,8, 1};
+        if (fx <= 4  && fy <= 4 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 4,4,  8,8,8>, 8,8,8, 1};
+        if (fx <= 2  && fy <= 2 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,2, 2,2,  8,8,8>, 8,8,8, 1};
+    }
+    if (s != 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 1) // contiguous
+    {
+        if (fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 24,1, 64,8,1>, 64,8,1, 1};
+        if (fx <= 20 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 20,1, 64,8,1>, 64,8,1, 1};
+        if (fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 16,1, 64,8,1>, 64,8,1, 1};
+        if (fx <= 12 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 12,1, 64,8,1>, 64,8,1, 1};
+        if (fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 8,1,  64,8,1>, 64,8,1, 1};
+    }
+    if (s == 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 2 && p.down.y == 1) // channels_last
+    {
+        if (fx <= 24 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 24,1, 64,1,8>, 64,1,8, 1};
+        if (fx <= 20 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 20,1, 64,1,8>, 64,1,8, 1};
+        if (fx <= 16 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 16,1, 64,1,8>, 64,1,8, 1};
+        if (fx <= 12 && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 12,1, 64,1,8>, 64,1,8, 1};
+        if (fx <= 8  && fy <= 1 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 2,1, 8,1,  64,1,8>, 64,1,8, 1};
+    }
+    if (s != 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 2) // contiguous
+    {
+        if (fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,24, 32,16,1>, 32,16,1, 1};
+        if (fx <= 1  && fy <= 20) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,20, 32,16,1>, 32,16,1, 1};
+        if (fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,16, 32,16,1>, 32,16,1, 1};
+        if (fx <= 1  && fy <= 12) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,12, 32,16,1>, 32,16,1, 1};
+        if (fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,8,  32,16,1>, 32,16,1, 1};
+    }
+    if (s == 1 && p.up.x == 1 && p.up.y == 1 && p.down.x == 1 && p.down.y == 2) // channels_last
+    {
+        if (fx <= 1  && fy <= 24) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,24, 1,64,8>, 1,64,8, 1};
+        if (fx <= 1  && fy <= 20) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,20, 1,64,8>, 1,64,8, 1};
+        if (fx <= 1  && fy <= 16) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,16, 1,64,8>, 1,64,8, 1};
+        if (fx <= 1  && fy <= 12) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,12, 1,64,8>, 1,64,8, 1};
+        if (fx <= 1  && fy <= 8 ) spec = {(void*)upfirdn2d_kernel_small<T, 1,1, 1,2, 1,8,  1,64,8>, 1,64,8, 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.  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);
+
+//------------------------------------------------------------------------

torch_utils/ops/upfirdn2d.py

@@ -0,0 +1,384 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Custom PyTorch ops for efficient resampling of 2D images."""
+
+import os
+import warnings
+import numpy as np
+import torch
+import traceback
+
+from .. import custom_ops
+from .. import misc
+from . import conv2d_gradfix
+
+#----------------------------------------------------------------------------
+
+_inited = False
+_plugin = None
+
+def _init():
+    global _inited, _plugin
+    if not _inited:
+        sources = ['upfirdn2d.cpp', 'upfirdn2d.cu']
+        sources = [os.path.join(os.path.dirname(__file__), s) for s in sources]
+        try:
+            _plugin = custom_ops.get_plugin('upfirdn2d_plugin', sources=sources, extra_cuda_cflags=['--use_fast_math'])
+        except:
+            warnings.warn('Failed to build CUDA kernels for upfirdn2d. Falling back to slow reference implementation. Details:\n\n' + traceback.format_exc())
+    return _plugin is not None
+
+def _parse_scaling(scaling):
+    if isinstance(scaling, int):
+        scaling = [scaling, scaling]
+    assert isinstance(scaling, (list, tuple))
+    assert all(isinstance(x, int) for x in scaling)
+    sx, sy = scaling
+    assert sx >= 1 and sy >= 1
+    return sx, sy
+
+def _parse_padding(padding):
+    if isinstance(padding, int):
+        padding = [padding, padding]
+    assert isinstance(padding, (list, tuple))
+    assert all(isinstance(x, int) for x in padding)
+    if len(padding) == 2:
+        padx, pady = padding
+        padding = [padx, padx, pady, pady]
+    padx0, padx1, pady0, pady1 = padding
+    return padx0, padx1, pady0, pady1
+
+def _get_filter_size(f):
+    if f is None:
+        return 1, 1
+    assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
+    fw = f.shape[-1]
+    fh = f.shape[0]
+    with misc.suppress_tracer_warnings():
+        fw = int(fw)
+        fh = int(fh)
+    misc.assert_shape(f, [fh, fw][:f.ndim])
+    assert fw >= 1 and fh >= 1
+    return fw, fh
+
+#----------------------------------------------------------------------------
+
+def setup_filter(f, device=torch.device('cpu'), normalize=True, flip_filter=False, gain=1, separable=None):
+    r"""Convenience function to setup 2D FIR filter for `upfirdn2d()`.
+
+    Args:
+        f:           Torch tensor, numpy array, or python list of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable),
+                     `[]` (impulse), or
+                     `None` (identity).
+        device:      Result device (default: cpu).
+        normalize:   Normalize the filter so that it retains the magnitude
+                     for constant input signal (DC)? (default: True).
+        flip_filter: Flip the filter? (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        separable:   Return a separable filter? (default: select automatically).
+
+    Returns:
+        Float32 tensor of the shape
+        `[filter_height, filter_width]` (non-separable) or
+        `[filter_taps]` (separable).
+    """
+    # Validate.
+    if f is None:
+        f = 1
+    f = torch.as_tensor(f, dtype=torch.float32)
+    assert f.ndim in [0, 1, 2]
+    assert f.numel() > 0
+    if f.ndim == 0:
+        f = f[np.newaxis]
+
+    # Separable?
+    if separable is None:
+        separable = (f.ndim == 1 and f.numel() >= 8)
+    if f.ndim == 1 and not separable:
+        f = f.ger(f)
+    assert f.ndim == (1 if separable else 2)
+
+    # Apply normalize, flip, gain, and device.
+    if normalize:
+        f /= f.sum()
+    if flip_filter:
+        f = f.flip(list(range(f.ndim)))
+    f = f * (gain ** (f.ndim / 2))
+    f = f.to(device=device)
+    return f
+
+#----------------------------------------------------------------------------
+
+def upfirdn2d(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Pad, upsample, filter, and downsample a batch of 2D images.
+
+    Performs the following sequence of operations for each channel:
+
+    1. Upsample the image by inserting N-1 zeros after each pixel (`up`).
+
+    2. Pad the image with the specified number of zeros on each side (`padding`).
+       Negative padding corresponds to cropping the image.
+
+    3. Convolve the image with the specified 2D FIR filter (`f`), shrinking it
+       so that the footprint of all output pixels lies within the input image.
+
+    4. Downsample the image by keeping every Nth pixel (`down`).
+
+    This sequence of operations bears close resemblance to scipy.signal.upfirdn().
+    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/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        up:          Integer upsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        down:        Integer downsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (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).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        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 _upfirdn2d_cuda(up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain).apply(x, f)
+    return _upfirdn2d_ref(x, f, up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain)
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def _upfirdn2d_ref(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1):
+    """Slow reference implementation of `upfirdn2d()` using standard PyTorch ops.
+    """
+    # Validate arguments.
+    assert isinstance(x, torch.Tensor) and x.ndim == 4
+    if f is None:
+        f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+    assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
+    assert f.dtype == torch.float32 and not f.requires_grad
+    batch_size, num_channels, in_height, in_width = x.shape
+    upx, upy = _parse_scaling(up)
+    downx, downy = _parse_scaling(down)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+
+    # Upsample by inserting zeros.
+    x = x.reshape([batch_size, num_channels, in_height, 1, in_width, 1])
+    x = torch.nn.functional.pad(x, [0, upx - 1, 0, 0, 0, upy - 1])
+    x = x.reshape([batch_size, num_channels, in_height * upy, in_width * upx])
+
+    # Pad or crop.
+    x = torch.nn.functional.pad(x, [max(padx0, 0), max(padx1, 0), max(pady0, 0), max(pady1, 0)])
+    x = x[:, :, max(-pady0, 0) : x.shape[2] - max(-pady1, 0), max(-padx0, 0) : x.shape[3] - max(-padx1, 0)]
+
+    # Setup filter.
+    f = f * (gain ** (f.ndim / 2))
+    f = f.to(x.dtype)
+    if not flip_filter:
+        f = f.flip(list(range(f.ndim)))
+
+    # Convolve with the filter.
+    f = f[np.newaxis, np.newaxis].repeat([num_channels, 1] + [1] * f.ndim)
+    if f.ndim == 4:
+        x = conv2d_gradfix.conv2d(input=x, weight=f, groups=num_channels)
+    else:
+        x = conv2d_gradfix.conv2d(input=x, weight=f.unsqueeze(2), groups=num_channels)
+        x = conv2d_gradfix.conv2d(input=x, weight=f.unsqueeze(3), groups=num_channels)
+
+    # Downsample by throwing away pixels.
+    x = x[:, :, ::downy, ::downx]
+    return x
+
+#----------------------------------------------------------------------------
+
+_upfirdn2d_cuda_cache = dict()
+
+def _upfirdn2d_cuda(up=1, down=1, padding=0, flip_filter=False, gain=1):
+    """Fast CUDA implementation of `upfirdn2d()` using custom ops.
+    """
+    # Parse arguments.
+    upx, upy = _parse_scaling(up)
+    downx, downy = _parse_scaling(down)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+
+    # Lookup from cache.
+    key = (upx, upy, downx, downy, padx0, padx1, pady0, pady1, flip_filter, gain)
+    if key in _upfirdn2d_cuda_cache:
+        return _upfirdn2d_cuda_cache[key]
+
+    # Forward op.
+    class Upfirdn2dCuda(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, f): # pylint: disable=arguments-differ
+            assert isinstance(x, torch.Tensor) and x.ndim == 4
+            if f is None:
+                f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
+            assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
+            y = x
+            if f.ndim == 2:
+                y = _plugin.upfirdn2d(y, f, upx, upy, downx, downy, padx0, padx1, pady0, pady1, flip_filter, gain)
+            else:
+                y = _plugin.upfirdn2d(y, f.unsqueeze(0), upx, 1, downx, 1, padx0, padx1, 0, 0, flip_filter, np.sqrt(gain))
+                y = _plugin.upfirdn2d(y, f.unsqueeze(1), 1, upy, 1, downy, 0, 0, pady0, pady1, flip_filter, np.sqrt(gain))
+            ctx.save_for_backward(f)
+            ctx.x_shape = x.shape
+            return y
+
+        @staticmethod
+        def backward(ctx, dy): # pylint: disable=arguments-differ
+            f, = ctx.saved_tensors
+            _, _, ih, iw = ctx.x_shape
+            _, _, oh, ow = dy.shape
+            fw, fh = _get_filter_size(f)
+            p = [
+                fw - padx0 - 1,
+                iw * upx - ow * downx + padx0 - upx + 1,
+                fh - pady0 - 1,
+                ih * upy - oh * downy + pady0 - upy + 1,
+            ]
+            dx = None
+            df = None
+
+            if ctx.needs_input_grad[0]:
+                dx = _upfirdn2d_cuda(up=down, down=up, padding=p, flip_filter=(not flip_filter), gain=gain).apply(dy, f)
+
+            assert not ctx.needs_input_grad[1]
+            return dx, df
+
+    # Add to cache.
+    _upfirdn2d_cuda_cache[key] = Upfirdn2dCuda
+    return Upfirdn2dCuda
+
+#----------------------------------------------------------------------------
+
+def filter2d(x, f, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Filter a batch of 2D images using the given 2D FIR filter.
+
+    By default, the result is padded so that its shape matches the input.
+    User-specified padding is applied on top of that, with negative values
+    indicating cropping. Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        padding:     Padding with respect to the output. Can be a single number or a
+                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+    fw, fh = _get_filter_size(f)
+    p = [
+        padx0 + fw // 2,
+        padx1 + (fw - 1) // 2,
+        pady0 + fh // 2,
+        pady1 + (fh - 1) // 2,
+    ]
+    return upfirdn2d(x, f, padding=p, flip_filter=flip_filter, gain=gain, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample2d(x, f, up=2, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Upsample a batch of 2D images using the given 2D FIR filter.
+
+    By default, the result is padded so that its shape is a multiple of the input.
+    User-specified padding is applied on top of that, with negative values
+    indicating cropping. Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        up:          Integer upsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        padding:     Padding with respect to the output. Can be a single number or a
+                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    upx, upy = _parse_scaling(up)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+    fw, fh = _get_filter_size(f)
+    p = [
+        padx0 + (fw + upx - 1) // 2,
+        padx1 + (fw - upx) // 2,
+        pady0 + (fh + upy - 1) // 2,
+        pady1 + (fh - upy) // 2,
+    ]
+    return upfirdn2d(x, f, up=up, padding=p, flip_filter=flip_filter, gain=gain*upx*upy, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def downsample2d(x, f, down=2, padding=0, flip_filter=False, gain=1, impl='cuda'):
+    r"""Downsample a batch of 2D images using the given 2D FIR filter.
+
+    By default, the result is padded so that its shape is a fraction of the input.
+    User-specified padding is applied on top of that, with negative values
+    indicating cropping. Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:           Float32/float64/float16 input tensor of the shape
+                     `[batch_size, num_channels, in_height, in_width]`.
+        f:           Float32 FIR filter of the shape
+                     `[filter_height, filter_width]` (non-separable),
+                     `[filter_taps]` (separable), or
+                     `None` (identity).
+        down:        Integer downsampling factor. Can be a single int or a list/tuple
+                     `[x, y]` (default: 1).
+        padding:     Padding with respect to the input. Can be a single number or a
+                     list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
+                     (default: 0).
+        flip_filter: False = convolution, True = correlation (default: False).
+        gain:        Overall scaling factor for signal magnitude (default: 1).
+        impl:        Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
+
+    Returns:
+        Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
+    """
+    downx, downy = _parse_scaling(down)
+    padx0, padx1, pady0, pady1 = _parse_padding(padding)
+    fw, fh = _get_filter_size(f)
+    p = [
+        padx0 + (fw - downx + 1) // 2,
+        padx1 + (fw - downx) // 2,
+        pady0 + (fh - downy + 1) // 2,
+        pady1 + (fh - downy) // 2,
+    ]
+    return upfirdn2d(x, f, down=down, padding=p, flip_filter=flip_filter, gain=gain, impl=impl)
+
+#----------------------------------------------------------------------------

torch_utils/persistence.py

@@ -0,0 +1,251 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Facilities for pickling Python code alongside other data.
+
+The pickled code is automatically imported into a separate Python module
+during unpickling. This way, any previously exported pickles will remain
+usable even if the original code is no longer available, or if the current
+version of the code is not consistent with what was originally pickled."""
+
+import sys
+import pickle
+import io
+import inspect
+import copy
+import uuid
+import types
+import dnnlib
+
+#----------------------------------------------------------------------------
+
+_version            = 6         # internal version number
+_decorators         = set()     # {decorator_class, ...}
+_import_hooks       = []        # [hook_function, ...]
+_module_to_src_dict = dict()    # {module: src, ...}
+_src_to_module_dict = dict()    # {src: module, ...}
+
+#----------------------------------------------------------------------------
+
+def persistent_class(orig_class):
+    r"""Class decorator that extends a given class to save its source code
+    when pickled.
+
+    Example:
+
+        from torch_utils import persistence
+
+        @persistence.persistent_class
+        class MyNetwork(torch.nn.Module):
+            def __init__(self, num_inputs, num_outputs):
+                super().__init__()
+                self.fc = MyLayer(num_inputs, num_outputs)
+                ...
+
+        @persistence.persistent_class
+        class MyLayer(torch.nn.Module):
+            ...
+
+    When pickled, any instance of `MyNetwork` and `MyLayer` will save its
+    source code alongside other internal state (e.g., parameters, buffers,
+    and submodules). This way, any previously exported pickle will remain
+    usable even if the class definitions have been modified or are no
+    longer available.
+
+    The decorator saves the source code of the entire Python module
+    containing the decorated class. It does *not* save the source code of
+    any imported modules. Thus, the imported modules must be available
+    during unpickling, also including `torch_utils.persistence` itself.
+
+    It is ok to call functions defined in the same module from the
+    decorated class. However, if the decorated class depends on other
+    classes defined in the same module, they must be decorated as well.
+    This is illustrated in the above example in the case of `MyLayer`.
+
+    It is also possible to employ the decorator just-in-time before
+    calling the constructor. For example:
+
+        cls = MyLayer
+        if want_to_make_it_persistent:
+            cls = persistence.persistent_class(cls)
+        layer = cls(num_inputs, num_outputs)
+
+    As an additional feature, the decorator also keeps track of the
+    arguments that were used to construct each instance of the decorated
+    class. The arguments can be queried via `obj.init_args` and
+    `obj.init_kwargs`, and they are automatically pickled alongside other
+    object state. A typical use case is to first unpickle a previous
+    instance of a persistent class, and then upgrade it to use the latest
+    version of the source code:
+
+        with open('old_pickle.pkl', 'rb') as f:
+            old_net = pickle.load(f)
+        new_net = MyNetwork(*old_obj.init_args, **old_obj.init_kwargs)
+        misc.copy_params_and_buffers(old_net, new_net, require_all=True)
+    """
+    assert isinstance(orig_class, type)
+    if is_persistent(orig_class):
+        return orig_class
+
+    assert orig_class.__module__ in sys.modules
+    orig_module = sys.modules[orig_class.__module__]
+    orig_module_src = _module_to_src(orig_module)
+
+    class Decorator(orig_class):
+        _orig_module_src = orig_module_src
+        _orig_class_name = orig_class.__name__
+
+        def __init__(self, *args, **kwargs):
+            super().__init__(*args, **kwargs)
+            self._init_args = copy.deepcopy(args)
+            self._init_kwargs = copy.deepcopy(kwargs)
+            assert orig_class.__name__ in orig_module.__dict__
+            _check_pickleable(self.__reduce__())
+
+        @property
+        def init_args(self):
+            return copy.deepcopy(self._init_args)
+
+        @property
+        def init_kwargs(self):
+            return dnnlib.EasyDict(copy.deepcopy(self._init_kwargs))
+
+        def __reduce__(self):
+            fields = list(super().__reduce__())
+            fields += [None] * max(3 - len(fields), 0)
+            if fields[0] is not _reconstruct_persistent_obj:
+                meta = dict(type='class', version=_version, module_src=self._orig_module_src, class_name=self._orig_class_name, state=fields[2])
+                fields[0] = _reconstruct_persistent_obj # reconstruct func
+                fields[1] = (meta,) # reconstruct args
+                fields[2] = None # state dict
+            return tuple(fields)
+
+    Decorator.__name__ = orig_class.__name__
+    _decorators.add(Decorator)
+    return Decorator
+
+#----------------------------------------------------------------------------
+
+def is_persistent(obj):
+    r"""Test whether the given object or class is persistent, i.e.,
+    whether it will save its source code when pickled.
+    """
+    try:
+        if obj in _decorators:
+            return True
+    except TypeError:
+        pass
+    return type(obj) in _decorators # pylint: disable=unidiomatic-typecheck
+
+#----------------------------------------------------------------------------
+
+def import_hook(hook):
+    r"""Register an import hook that is called whenever a persistent object
+    is being unpickled. A typical use case is to patch the pickled source
+    code to avoid errors and inconsistencies when the API of some imported
+    module has changed.
+
+    The hook should have the following signature:
+
+        hook(meta) -> modified meta
+
+    `meta` is an instance of `dnnlib.EasyDict` with the following fields:
+
+        type:       Type of the persistent object, e.g. `'class'`.
+        version:    Internal version number of `torch_utils.persistence`.
+        module_src  Original source code of the Python module.
+        class_name: Class name in the original Python module.
+        state:      Internal state of the object.
+
+    Example:
+
+        @persistence.import_hook
+        def wreck_my_network(meta):
+            if meta.class_name == 'MyNetwork':
+                print('MyNetwork is being imported. I will wreck it!')
+                meta.module_src = meta.module_src.replace("True", "False")
+            return meta
+    """
+    assert callable(hook)
+    _import_hooks.append(hook)
+
+#----------------------------------------------------------------------------
+
+def _reconstruct_persistent_obj(meta):
+    r"""Hook that is called internally by the `pickle` module to unpickle
+    a persistent object.
+    """
+    meta = dnnlib.EasyDict(meta)
+    meta.state = dnnlib.EasyDict(meta.state)
+    for hook in _import_hooks:
+        meta = hook(meta)
+        assert meta is not None
+
+    assert meta.version == _version
+    module = _src_to_module(meta.module_src)
+
+    assert meta.type == 'class'
+    orig_class = module.__dict__[meta.class_name]
+    decorator_class = persistent_class(orig_class)
+    obj = decorator_class.__new__(decorator_class)
+
+    setstate = getattr(obj, '__setstate__', None)
+    if callable(setstate):
+        setstate(meta.state) # pylint: disable=not-callable
+    else:
+        obj.__dict__.update(meta.state)
+    return obj
+
+#----------------------------------------------------------------------------
+
+def _module_to_src(module):
+    r"""Query the source code of a given Python module.
+    """
+    src = _module_to_src_dict.get(module, None)
+    if src is None:
+        src = inspect.getsource(module)
+        _module_to_src_dict[module] = src
+        _src_to_module_dict[src] = module
+    return src
+
+def _src_to_module(src):
+    r"""Get or create a Python module for the given source code.
+    """
+    module = _src_to_module_dict.get(src, None)
+    if module is None:
+        module_name = "_imported_module_" + uuid.uuid4().hex
+        module = types.ModuleType(module_name)
+        sys.modules[module_name] = module
+        _module_to_src_dict[module] = src
+        _src_to_module_dict[src] = module
+        exec(src, module.__dict__) # pylint: disable=exec-used
+    return module
+
+#----------------------------------------------------------------------------
+
+def _check_pickleable(obj):
+    r"""Check that the given object is pickleable, raising an exception if
+    it is not. This function is expected to be considerably more efficient
+    than actually pickling the object.
+    """
+    def recurse(obj):
+        if isinstance(obj, (list, tuple, set)):
+            return [recurse(x) for x in obj]
+        if isinstance(obj, dict):
+            return [[recurse(x), recurse(y)] for x, y in obj.items()]
+        if isinstance(obj, (str, int, float, bool, bytes, bytearray)):
+            return None # Python primitive types are pickleable.
+        if f'{type(obj).__module__}.{type(obj).__name__}' in ['numpy.ndarray', 'torch.Tensor']:
+            return None # NumPy arrays and PyTorch tensors are pickleable.
+        if is_persistent(obj):
+            return None # Persistent objects are pickleable, by virtue of the constructor check.
+        return obj
+    with io.BytesIO() as f:
+        pickle.dump(recurse(obj), f)
+
+#----------------------------------------------------------------------------

torch_utils/training_stats.py

@@ -0,0 +1,268 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Facilities for reporting and collecting training statistics across
+multiple processes and devices. The interface is designed to minimize
+synchronization overhead as well as the amount of boilerplate in user
+code."""
+
+import re
+import numpy as np
+import torch
+import dnnlib
+
+from . import misc
+
+#----------------------------------------------------------------------------
+
+_num_moments    = 3             # [num_scalars, sum_of_scalars, sum_of_squares]
+_reduce_dtype   = torch.float32 # Data type to use for initial per-tensor reduction.
+_counter_dtype  = torch.float64 # Data type to use for the internal counters.
+_rank           = 0             # Rank of the current process.
+_sync_device    = None          # Device to use for multiprocess communication. None = single-process.
+_sync_called    = False         # Has _sync() been called yet?
+_counters       = dict()        # Running counters on each device, updated by report(): name => device => torch.Tensor
+_cumulative     = dict()        # Cumulative counters on the CPU, updated by _sync(): name => torch.Tensor
+
+#----------------------------------------------------------------------------
+
+def init_multiprocessing(rank, sync_device):
+    r"""Initializes `torch_utils.training_stats` for collecting statistics
+    across multiple processes.
+
+    This function must be called after
+    `torch.distributed.init_process_group()` and before `Collector.update()`.
+    The call is not necessary if multi-process collection is not needed.
+
+    Args:
+        rank:           Rank of the current process.
+        sync_device:    PyTorch device to use for inter-process
+                        communication, or None to disable multi-process
+                        collection. Typically `torch.device('cuda', rank)`.
+    """
+    global _rank, _sync_device
+    assert not _sync_called
+    _rank = rank
+    _sync_device = sync_device
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def report(name, value):
+    r"""Broadcasts the given set of scalars to all interested instances of
+    `Collector`, across device and process boundaries.
+
+    This function is expected to be extremely cheap and can be safely
+    called from anywhere in the training loop, loss function, or inside a
+    `torch.nn.Module`.
+
+    Warning: The current implementation expects the set of unique names to
+    be consistent across processes. Please make sure that `report()` is
+    called at least once for each unique name by each process, and in the
+    same order. If a given process has no scalars to broadcast, it can do
+    `report(name, [])` (empty list).
+
+    Args:
+        name:   Arbitrary string specifying the name of the statistic.
+                Averages are accumulated separately for each unique name.
+        value:  Arbitrary set of scalars. Can be a list, tuple,
+                NumPy array, PyTorch tensor, or Python scalar.
+
+    Returns:
+        The same `value` that was passed in.
+    """
+    if name not in _counters:
+        _counters[name] = dict()
+
+    elems = torch.as_tensor(value)
+    if elems.numel() == 0:
+        return value
+
+    elems = elems.detach().flatten().to(_reduce_dtype)
+    moments = torch.stack([
+        torch.ones_like(elems).sum(),
+        elems.sum(),
+        elems.square().sum(),
+    ])
+    assert moments.ndim == 1 and moments.shape[0] == _num_moments
+    moments = moments.to(_counter_dtype)
+
+    device = moments.device
+    if device not in _counters[name]:
+        _counters[name][device] = torch.zeros_like(moments)
+    _counters[name][device].add_(moments)
+    return value
+
+#----------------------------------------------------------------------------
+
+def report0(name, value):
+    r"""Broadcasts the given set of scalars by the first process (`rank = 0`),
+    but ignores any scalars provided by the other processes.
+    See `report()` for further details.
+    """
+    report(name, value if _rank == 0 else [])
+    return value
+
+#----------------------------------------------------------------------------
+
+class Collector:
+    r"""Collects the scalars broadcasted by `report()` and `report0()` and
+    computes their long-term averages (mean and standard deviation) over
+    user-defined periods of time.
+
+    The averages are first collected into internal counters that are not
+    directly visible to the user. They are then copied to the user-visible
+    state as a result of calling `update()` and can then be queried using
+    `mean()`, `std()`, `as_dict()`, etc. Calling `update()` also resets the
+    internal counters for the next round, so that the user-visible state
+    effectively reflects averages collected between the last two calls to
+    `update()`.
+
+    Args:
+        regex:          Regular expression defining which statistics to
+                        collect. The default is to collect everything.
+        keep_previous:  Whether to retain the previous averages if no
+                        scalars were collected on a given round
+                        (default: True).
+    """
+    def __init__(self, regex='.*', keep_previous=True):
+        self._regex = re.compile(regex)
+        self._keep_previous = keep_previous
+        self._cumulative = dict()
+        self._moments = dict()
+        self.update()
+        self._moments.clear()
+
+    def names(self):
+        r"""Returns the names of all statistics broadcasted so far that
+        match the regular expression specified at construction time.
+        """
+        return [name for name in _counters if self._regex.fullmatch(name)]
+
+    def update(self):
+        r"""Copies current values of the internal counters to the
+        user-visible state and resets them for the next round.
+
+        If `keep_previous=True` was specified at construction time, the
+        operation is skipped for statistics that have received no scalars
+        since the last update, retaining their previous averages.
+
+        This method performs a number of GPU-to-CPU transfers and one
+        `torch.distributed.all_reduce()`. It is intended to be called
+        periodically in the main training loop, typically once every
+        N training steps.
+        """
+        if not self._keep_previous:
+            self._moments.clear()
+        for name, cumulative in _sync(self.names()):
+            if name not in self._cumulative:
+                self._cumulative[name] = torch.zeros([_num_moments], dtype=_counter_dtype)
+            delta = cumulative - self._cumulative[name]
+            self._cumulative[name].copy_(cumulative)
+            if float(delta[0]) != 0:
+                self._moments[name] = delta
+
+    def _get_delta(self, name):
+        r"""Returns the raw moments that were accumulated for the given
+        statistic between the last two calls to `update()`, or zero if
+        no scalars were collected.
+        """
+        assert self._regex.fullmatch(name)
+        if name not in self._moments:
+            self._moments[name] = torch.zeros([_num_moments], dtype=_counter_dtype)
+        return self._moments[name]
+
+    def num(self, name):
+        r"""Returns the number of scalars that were accumulated for the given
+        statistic between the last two calls to `update()`, or zero if
+        no scalars were collected.
+        """
+        delta = self._get_delta(name)
+        return int(delta[0])
+
+    def mean(self, name):
+        r"""Returns the mean of the scalars that were accumulated for the
+        given statistic between the last two calls to `update()`, or NaN if
+        no scalars were collected.
+        """
+        delta = self._get_delta(name)
+        if int(delta[0]) == 0:
+            return float('nan')
+        return float(delta[1] / delta[0])
+
+    def std(self, name):
+        r"""Returns the standard deviation of the scalars that were
+        accumulated for the given statistic between the last two calls to
+        `update()`, or NaN if no scalars were collected.
+        """
+        delta = self._get_delta(name)
+        if int(delta[0]) == 0 or not np.isfinite(float(delta[1])):
+            return float('nan')
+        if int(delta[0]) == 1:
+            return float(0)
+        mean = float(delta[1] / delta[0])
+        raw_var = float(delta[2] / delta[0])
+        return np.sqrt(max(raw_var - np.square(mean), 0))
+
+    def as_dict(self):
+        r"""Returns the averages accumulated between the last two calls to
+        `update()` as an `dnnlib.EasyDict`. The contents are as follows:
+
+            dnnlib.EasyDict(
+                NAME = dnnlib.EasyDict(num=FLOAT, mean=FLOAT, std=FLOAT),
+                ...
+            )
+        """
+        stats = dnnlib.EasyDict()
+        for name in self.names():
+            stats[name] = dnnlib.EasyDict(num=self.num(name), mean=self.mean(name), std=self.std(name))
+        return stats
+
+    def __getitem__(self, name):
+        r"""Convenience getter.
+        `collector[name]` is a synonym for `collector.mean(name)`.
+        """
+        return self.mean(name)
+
+#----------------------------------------------------------------------------
+
+def _sync(names):
+    r"""Synchronize the global cumulative counters across devices and
+    processes. Called internally by `Collector.update()`.
+    """
+    if len(names) == 0:
+        return []
+    global _sync_called
+    _sync_called = True
+
+    # Collect deltas within current rank.
+    deltas = []
+    device = _sync_device if _sync_device is not None else torch.device('cpu')
+    for name in names:
+        delta = torch.zeros([_num_moments], dtype=_counter_dtype, device=device)
+        for counter in _counters[name].values():
+            delta.add_(counter.to(device))
+            counter.copy_(torch.zeros_like(counter))
+        deltas.append(delta)
+    deltas = torch.stack(deltas)
+
+    # Sum deltas across ranks.
+    if _sync_device is not None:
+        torch.distributed.all_reduce(deltas)
+
+    # Update cumulative values.
+    deltas = deltas.cpu()
+    for idx, name in enumerate(names):
+        if name not in _cumulative:
+            _cumulative[name] = torch.zeros([_num_moments], dtype=_counter_dtype)
+        _cumulative[name].add_(deltas[idx])
+
+    # Return name-value pairs.
+    return [(name, _cumulative[name]) for name in names]
+
+#----------------------------------------------------------------------------