Add files from GH

app.py

@@ -0,0 +1,173 @@
+import time
+import torch
+import streamlit as st
+from PIL import Image, ImageDraw
+from streamlit_image_coordinates import streamlit_image_coordinates
+
+import draggan
+import utils
+
+
+## Default to CPU if no GPU is available
+if torch.cuda.is_available():
+  device = torch.device("cuda")
+else:
+    device = torch.device("cpu")
+
+
+### Streamlit setup ###
+
+st.set_page_config(
+    page_title="DragGAN Demo",
+    page_icon="🐉",
+    layout="wide",
+)
+
+
+message_container = st.empty()
+
+col1, col2 = st.columns([1, 2])
+
+def reset():
+    st.session_state.clear()
+
+def reset_rerun():
+    reset()
+    st.experimental_rerun()
+
+
+### Run/Reset buttons in right col ###
+with col2:
+    st.markdown("")
+    but_col1, but_col2 = st.columns([1,7])
+    run_button = but_col1.button("▶️ Run")
+    reset_button = but_col2.button("🔁 Reset")
+
+
+### Settings panel in left col ###
+with col1:
+    st.header("🐉 DragGAN")
+
+    settings_col1, settings_col2 = st.columns([1,1])
+    # Models from Self-Distilled SG https://github.com/self-distilled-stylegan/self-distilled-internet-photos
+    model_options = {
+        "Lions": "https://storage.googleapis.com/self-distilled-stylegan/lions_512_pytorch.pkl",
+        "Faces (FFHQ)": "https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/ffhq.pkl",
+        "Elephants": "https://storage.googleapis.com/self-distilled-stylegan/elephants_512_pytorch.pkl",
+        "Parrots": "https://storage.googleapis.com/self-distilled-stylegan/parrots_512_pytorch.pkl",
+        "Horses": "https://storage.googleapis.com/self-distilled-stylegan/horses_256_pytorch.pkl",
+        "Bicycles": "https://storage.googleapis.com/self-distilled-stylegan/bicycles_256_pytorch.pkl",
+        "Giraffes": "https://storage.googleapis.com/self-distilled-stylegan/giraffes_512_pytorch.pkl",
+        "Dogs (1)": "https://storage.googleapis.com/self-distilled-stylegan/dogs_1024_pytorch.pkl",
+        "Dogs (2)": "https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/afhqdog.pkl",
+        "Cats": "https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/afhqcat.pkl",
+        "Wildlife": "https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/afhqwild.pkl",
+        "MetFaces": "https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metfaces.pkl",
+    }
+    model_name = str(settings_col1.selectbox("Model", list(model_options.keys()), on_change=reset, help="StyleGAN2 model to use, downloaded and cached on first run"))
+    model_url = model_options[model_name]
+    seed = settings_col2.number_input("Seed", value=22, step=1, min_value=0, on_change=reset, help="Random seed for generating W+ latent")
+    target_resolution = int(settings_col1.selectbox("Resolution", [256, 512, 1024], index=1, on_change=reset, help="Resize generated image to this resolution (may be different than native model resolution)"))
+    n_iter = int(settings_col1.number_input("Iterations", value=200, step=5, help="Number of iterations to run optimization", on_change=reset))
+    step_size = settings_col2.number_input("Step Size", value=1e-3, step=1e-4, min_value=1e-4, format="%.4f", help="Step size (Learning Rate) for gradient descent")
+    multiplier = settings_col1.number_input("Speed", value=1.0, step=0.05, min_value=0.05, help="Multiplier for target patch movement")
+    tolerance = settings_col2.number_input("Tolerance", value=2, step=1, min_value=1, help="Number of pixels away from target to stop")
+    
+    display_every = settings_col2.number_input("Display Every", value=25, step=1, min_value=1, help="Display image during optimization every n iterations")
+    truncation_psi = settings_col1.number_input("Truncation", value=0.8, step=0.1, min_value=0.0, on_change=reset, help="Truncation trick value to control diversity (higher = more diverse)")
+    truncation_cutoff = settings_col2.number_input(
+        "Truncation Cutoff", value=8, step=1, min_value=-1, max_value=18, on_change=reset, help="Number of layers to apply truncation to (-1 = all layers)"
+    )    
+    
+
+    if reset_button:
+        reset_rerun()
+
+if "points" not in st.session_state:
+    st.session_state["points"] = []
+    st.session_state["points_types"] = []
+    # State variable to track whether the next click should be a 'handle' or 'target'
+    st.session_state["next_click"] = "handle"
+
+
+s = time.perf_counter()
+G = draggan.load_model(model_url, device=device)
+
+if "W" not in st.session_state:
+    W = draggan.generate_W(
+        G,
+        seed=int(seed),
+        truncation_psi=truncation_psi,
+        truncation_cutoff=int(truncation_cutoff),
+        network_pkl=model_url,
+        device=device,
+    )
+else:
+    W = st.session_state["W"]
+
+img, F0 = draggan.generate_image(W, G, network_pkl=model_url, device=device)
+if img.size[0] != target_resolution:
+    img = img.resize((target_resolution, target_resolution))
+print(f"Generated image in {(time.perf_counter() - s)*1000:.0f}ms")
+
+# Draw an ellipse at each coordinate in points
+if "points" in st.session_state and "points_types" in st.session_state:
+    handles, targets = [], []
+    for point, point_type in zip(
+        st.session_state["points"], st.session_state["points_types"]
+    ):
+        if point_type == "handle":
+            handles.append(point)
+        else:
+            targets.append(point)
+    if len(handles) > 0:
+        utils.draw_handle_target_points(img, handles, targets)
+
+
+### Right column image container ###
+with col2:
+    empty = st.empty()
+    with empty.container():
+        value = streamlit_image_coordinates(img, key="pil")
+        # New point is clicked
+        if value is not None:
+            point = value["x"], value["y"]
+            if point not in st.session_state["points"]:
+                # st.session_state["points"].append(point)
+                st.session_state["points"].append(point)
+                st.session_state["points_types"].append(st.session_state["next_click"])
+                st.session_state["next_click"] = (
+                    "target" if st.session_state["next_click"] == "handle" else "handle"
+                )
+                
+                st.experimental_rerun()
+
+## Optimization loop
+if run_button:
+    if len(handles) > 0 and len(targets) > 0 and len(handles) == len(targets) and all(targets):
+        W = draggan.optimize(
+            W,
+            G,
+            handle_points=handles,
+            target_points=targets,
+            r1=3,
+            r2=12,
+            tolerance=tolerance,
+            max_iter=n_iter,
+            lr=step_size,
+            multiplier=multiplier,
+            empty=empty,
+            display_every=display_every,
+            target_resolution=target_resolution,
+            device=device,
+        )
+        # st.write(handles)
+        # st.write(targets)
+
+        st.session_state.clear()
+        st.session_state["W"] = W
+        st.experimental_rerun()
+    else:
+        message_container.warning("Please add at least one handle and one target point.")
+
+

draggan.py

@@ -0,0 +1,463 @@
+import os
+import sys
+import time
+from typing import List, Optional, Tuple
+import math
+
+import numpy as np
+import PIL
+import torch
+import streamlit as st
+
+stylegan2_dir = os.path.abspath("stylegan2")
+sys.path.insert(0, stylegan2_dir)
+import dnnlib
+import legacy
+
+import utils
+
+
+@st.cache_resource
+def load_model(
+    network_pkl: str = "https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/afhqdog.pkl",
+    device: torch.device = torch.device("cuda"),
+    fp16: bool = True,
+) -> torch.nn.Module:
+    """
+    Loads a pretrained StyleGAN2-ADA generator network from a pickle file.
+
+    Args:
+        network_pkl (str): The URL or local path to the network pickle file.
+        device (torch.device): The device to use for the computation.
+        fp16 (bool): Whether to use half-precision floating point format for the network weights.
+
+    Returns:
+        The pretrained generator network.
+    """
+    print('Loading networks from "%s"...' % network_pkl)
+    with dnnlib.util.open_url(network_pkl) as f:
+        chkpt = legacy.load_network_pkl(f, force_fp16=fp16)
+    G = chkpt["G_ema"].to(device).eval()
+    for param in G.parameters():
+        param.requires_grad_(False)
+
+    # Create a new attribute called "activations" for the Generator class
+    # This will be a list of activations from each layer
+    G.__setattr__("activations", None)
+
+    # Forward hook to collect features
+    def hook(module, input, output):
+        G.activations = output
+
+    # Apply the hook to the 7th layer (256x256)
+    for i, (name, module) in enumerate(G.synthesis.named_children()):
+        if i == 6:
+            print("Registering hook for:", name)
+            module.register_forward_hook(hook)
+
+    return G
+
+
+@st.cache_data()
+def generate_W(
+    _G: torch.nn.Module,
+    seed: int = 0,
+    network_pkl: Optional[str] = None,
+    truncation_psi: float = 1.0,
+    truncation_cutoff: Optional[int] = None,
+    device: torch.device = torch.device("cuda"),
+) -> np.ndarray:
+    """
+    Generates a latent code tensor in W+ space from a pretrained StyleGAN2-ADA generator network.
+
+    Args:
+        _G (torch.nn.Module): The generator network, with underscore to avoid streamlit cache error
+        seed (int): The random seed to use for generating the latent code.
+        network_pkl (Optional[str]): The path to the network pickle file. If None, the default network will be used.
+        truncation_psi (float): The truncation psi value to use for the mapping network.
+        truncation_cutoff (Optional[int]): The number of layers to use for the truncation trick. If None, all layers will be used.
+        device (torch.device): The device to use for the computation.
+
+    Returns:
+        The W+ latent as a numpy array of shape [1, num_layers, 512].
+    """
+    G = _G
+    torch.manual_seed(seed)
+    z = torch.randn(1, G.z_dim).to(device)
+    num_layers = G.synthesis.num_ws
+    if truncation_cutoff == -1:
+        truncation_cutoff = None
+    elif truncation_cutoff is not None:
+        truncation_cutoff = min(num_layers, truncation_cutoff)
+    W = G.mapping(
+        z,
+        None,
+        truncation_psi=truncation_psi,
+        truncation_cutoff=truncation_cutoff,
+    )
+    return W.cpu().numpy()
+
+
+def forward_G(
+    G: torch.nn.Module,
+    W: torch.Tensor,
+    device: torch.device,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Forward pass through the generator network.
+
+    Args:
+        G (torch.nn.Module): The generator network.
+        W (torch.Tensor): The latent code tensor of shape [batch_size, latent_dim, 512].
+        device (torch.device): The device to use for the computation.
+
+    Returns:
+        A tuple containing the generated image tensor of shape [batch_size, 3, height, width]
+        and the feature maps tensor of shape [batch_size, num_channels, height, width].
+    """
+    if not isinstance(W, torch.Tensor):
+        W = torch.from_numpy(W).to(device)
+
+    img = G.synthesis(W, noise_mode="const", force_fp32=True)
+
+    return img, G.activations[0]
+
+
+@st.cache_data()
+def generate_image(
+    W,
+    _G: Optional[torch.nn.Module] = None,
+    network_pkl: Optional[str] = None,
+    class_idx=None,
+    device=torch.device("cuda"),
+) -> Tuple[PIL.Image.Image, torch.Tensor]:
+    """
+    Generates an image using a pretrained generator network.
+
+    Args:
+        W (torch.Tensor): A tensor of latent codes of shape [batch_size, latent_dim, 512].
+        _G (Optional[torch.nn.Module]): The generator network. If None, the network will be loaded from `network_pkl`.
+        network_pkl (Optional[str]): The path to the network pickle file. If None, the default network will be used.
+        class_idx (Optional[int]): The class index to use for conditional generation. If None, unconditional generation will be used.
+        device (str): The device to use for the computation.
+
+    Returns:
+        A tuple containing the generated image as a PIL Image object and the feature maps tensor of shape [batch_size, num_channels, height, width].
+    """
+    if _G is None:
+        assert network_pkl is not None
+        _G = load_model(network_pkl, device)
+    G = _G
+
+    # Labels.
+    label = torch.zeros([1, G.c_dim], device=device)
+    if G.c_dim != 0:
+        if class_idx is None:
+            raise Exception(
+                "Must specify class label with --class when using a conditional network"
+            )
+        label[:, class_idx] = 1
+    else:
+        if class_idx is not None:
+            print("warn: --class=lbl ignored when running on an unconditional network")
+
+    ## Generate image
+    img, features = forward_G(G, W, device)
+
+    img = utils.tensor_to_PIL(img)
+
+    return img, features
+
+
+def optimize(
+    W: np.ndarray,
+    G: torch.nn.Module,
+    handle_points: List[Tuple[int, int]],
+    target_points: List[Tuple[int, int]],
+    r1: int = 3,
+    r2: int = 12,
+    tolerance: int = 2,
+    max_iter: int = 200,
+    lr: float = 0.1,
+    multiplier: float = 1.0,
+    lambda_: float = 0.1,
+    device: torch.device = torch.device("cuda"),
+    empty=None,
+    display_every: int = 10,
+    target_resolution: int = 512,
+) -> np.ndarray:
+    """
+    Optimizes the latent code tensor W to generate an image that matches the target points.
+
+    Args:
+        W (np.ndarray): The initial latent code tensor of shape [1, num_layers, 512].
+        G (torch.nn.Module): The generator network.
+        handle_points (List[Tuple[int, int]]): The initial handle points as a list of (x, y) tuples.
+        target_points (List[Tuple[int, int]]): The target points as a list of (x, y) tuples.
+        r1 (int): The radius of the motion supervision loss.
+        r2 (int): The radius of the point tracking.
+        d (int): The tolerance for the handle points to reach the target points.
+        max_iter (int): The maximum number of optimization iterations.
+        lr (float): The learning rate for the optimizer.
+        multiplier (float): The speed multiplier for the motion supervision loss.
+        lambda_ (float): The weight of the motion supervision loss.
+        device (torch.device): The device to use for the computation.
+        empty: The st.empty object to display the intermediate images.
+        display_every (int): The number of iterations between displaying intermediate images.
+        target_resolution (int): The target resolution for the generated image.
+
+    Returns:
+        The optimized latent code tensor W as a numpy array of shape [1, num_layers, 512].
+    """
+    img, F0 = forward_G(G, W, device)
+    empty.image(
+        utils.tensor_to_PIL(img), caption="Initial image", width=target_resolution
+    )
+
+    F0_resized = torch.nn.functional.interpolate(
+        F0,
+        size=(target_resolution, target_resolution),
+        mode="bilinear",
+        align_corners=True,
+    ).detach()
+
+    # Convert handle/target points to tensors and reorder to [y, x]
+    handle_points: torch.tensor = (
+        torch.tensor(handle_points, device=device).flip(-1).float()
+    )
+    handle_points_0 = handle_points.clone()
+    target_points: torch.tensor = (
+        torch.tensor(target_points, device=device).flip(-1).float()
+    )
+
+    W = torch.from_numpy(W).to(device).float()
+    W.requires_grad_(False)
+
+    # Only optimize the first 6 layers of W
+    W_layers_to_optimize = W[:, :6].clone()
+    W_layers_to_optimize.requires_grad_(True)
+
+    optimizer = torch.optim.Adam([W_layers_to_optimize], lr=lr)
+
+    for i in range(max_iter):
+        start = time.perf_counter()
+
+        # # Check if the handle points have reached the target points
+        if torch.allclose(handle_points, target_points, atol=tolerance):
+            break
+
+        optimizer.zero_grad()
+
+        # Detach only the unoptimized layers
+        W_combined = torch.cat([W_layers_to_optimize, W[:, 6:].detach()], dim=1)
+
+        # Run the generator to get the image and feature maps
+        img, F = forward_G(G, W_combined, device)
+
+        ## Bilinear interpolate F to be same size as img
+        F_resized = torch.nn.functional.interpolate(
+            F,
+            size=(target_resolution, target_resolution),
+            mode="bilinear",
+            align_corners=True,
+        )
+
+        # Compute the motion supervision loss
+        loss, all_shifted_coordinates = motion_supervision(
+            F_resized,
+            F0_resized,
+            handle_points,
+            target_points,
+            r1,
+            lambda_,
+            device,
+            multiplier=multiplier,
+        )
+
+        # Backpropagate the loss and update the latent code
+        loss.backward()
+
+        # # Clip gradients if their norm exceeds max_grad_norm
+        # torch.nn.utils.clip_grad_norm_(W_layers_to_optimize, 1.0)
+
+        # # Compute the L2 regularization term
+        # l2_regularization = 100 * torch.norm(W_layers_to_optimize - W[:, :6]) ** 2
+        # print(l2_regularization.item())
+        # # Add the regularization term to the loss
+        # loss += l2_regularization
+
+        optimizer.step()
+
+        print(
+            f"{i}\tLoss: {loss.item():0.2f}\tTime: {(time.perf_counter() - start) * 1000:.0f}ms"
+        )
+
+        if i % display_every == 0 or i == max_iter - 1:
+            # Draw d_i intermediate target as orange ellipse
+            img = utils.tensor_to_PIL(img)
+            if img.size[0] != target_resolution:
+                img = img.resize((target_resolution, target_resolution))
+
+            utils.draw_handle_target_points(img, handle_points.flip(-1).cpu().long().numpy().tolist(), target_points.flip(-1).cpu().long().numpy().tolist())
+            # draw = PIL.ImageDraw.Draw(img)
+
+
+            empty.image(
+                img, caption=f"iter: {i}, loss: {loss:.2f}", width=target_resolution
+            )
+
+        # Update the handle points with point tracking
+        handle_points = point_tracking(
+            F_resized,
+            F0_resized,
+            handle_points,
+            handle_points_0,
+            r2,
+            device,
+        )
+
+    return torch.cat([W_layers_to_optimize, W[:, 6:]], dim=1).detach().cpu().numpy()
+
+
+def motion_supervision(
+    F: torch.Tensor,
+    F0: torch.Tensor,
+    handle_points: torch.Tensor,
+    target_points: torch.Tensor,
+    r1: int = 3,
+    lambda_: float = 20.0,
+    device: torch.device = torch.device("cuda"),
+    multiplier: float = 1.0,
+) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+    """
+    Computes the motion supervision loss and the shifted coordinates for each handle point.
+
+    Args:
+        F (torch.Tensor): The feature map tensor of shape [batch_size, num_channels, height, width].
+        F0 (torch.Tensor): The original feature map tensor of shape [batch_size, num_channels, height, width].
+        handle_points (torch.Tensor): The handle points tensor of shape [num_handle_points, 2].
+        target_points (torch.Tensor): The target points tensor of shape [num_handle_points, 2].
+        r1 (int): The radius of the circular mask around each handle point.
+        lambda_ (float): The weight of the reconstruction loss for the unmasked region.
+        device (torch.device): The device to use for the computation.
+        multiplier (float): The multiplier to use for the direction vector.
+
+    Returns:
+        A tuple containing the motion supervision loss tensor and a list of shifted coordinates
+        for each handle point, where each element in the list is a tensor of shape [num_points, 2].
+    """
+    n = handle_points.shape[0]  # Number of handle points
+    loss = 0.0
+    all_shifted_coordinates = []  # List of shifted patches
+
+    for i in range(n):
+        # Compute direction vector
+        target2handle = target_points[i] - handle_points[i]
+        d_i = target2handle / (torch.norm(target2handle) + 1e-7) * multiplier
+        if torch.norm(d_i) > torch.norm(target2handle):
+            d_i = target2handle
+
+        # Compute the mask for the pixels within radius r1 of the handle point
+        mask = utils.create_circular_mask(
+            F.shape[2], F.shape[3], center=handle_points[i].tolist(), radius=r1
+        ).to(device)
+        # mask = utils.create_square_mask(F.shape[2], F.shape[3], center=handle_points[i].tolist(), radius=r1).to(device)
+
+        # Find indices where mask is True
+        coordinates = torch.nonzero(mask).float()  # shape [num_points, 2]
+
+        # Shift the coordinates in the direction d_i
+        shifted_coordinates = coordinates + d_i[None]
+        all_shifted_coordinates.append(shifted_coordinates)
+
+        h, w = F.shape[2], F.shape[3]
+
+        # Extract features in the mask region and compute the loss
+        F_qi = F[:, :, mask]  # shape: [C, H*W]
+
+        # Sample shifted patch from F
+        normalized_shifted_coordinates = shifted_coordinates.clone()
+        normalized_shifted_coordinates[:, 0] = (
+            2.0 * shifted_coordinates[:, 0] / (h - 1)
+        ) - 1  # for height
+        normalized_shifted_coordinates[:, 1] = (
+            2.0 * shifted_coordinates[:, 1] / (w - 1)
+        ) - 1  # for width
+        # Add extra dimensions for batch and channels (required by grid_sample)
+        normalized_shifted_coordinates = normalized_shifted_coordinates.unsqueeze(
+            0
+        ).unsqueeze(
+            0
+        )  # shape [1, 1, num_points, 2]
+        normalized_shifted_coordinates = normalized_shifted_coordinates.flip(
+            -1
+        )  # grid_sample expects [x, y] instead of [y, x]
+        normalized_shifted_coordinates = normalized_shifted_coordinates.clamp(-1, 1)
+
+        # Use grid_sample to interpolate the feature map F at the shifted patch coordinates
+        F_qi_plus_di = torch.nn.functional.grid_sample(
+            F, normalized_shifted_coordinates, mode="bilinear", align_corners=True
+        )
+        # Output has shape [1, C, 1, num_points] so squeeze it
+        F_qi_plus_di = F_qi_plus_di.squeeze(2)  # shape [1, C, num_points]
+
+        loss += torch.nn.functional.l1_loss(F_qi.detach(), F_qi_plus_di)
+
+    # TODO: add reconstruction loss for the unmasked region
+    # # Add reconstruction loss for the unmasked region
+    # loss += lambda_ * torch.norm((F - F0) * (1 - mask_total), p=1)
+
+    return loss, all_shifted_coordinates
+
+
+def point_tracking(
+    F: torch.Tensor,
+    F0: torch.Tensor,
+    handle_points: torch.Tensor,  # [N, y, x]
+    handle_points_0: torch.Tensor,  # [N, y, x]
+    r2: int = 3,
+    device: torch.device = torch.device("cuda"),
+) -> torch.Tensor:
+    """
+    Tracks the movement of handle points in an image using feature matching.
+
+    Args:
+        F (torch.Tensor): The feature maps tensor of shape [batch_size, num_channels, height, width].
+        F0 (torch.Tensor): The feature maps tensor of shape [batch_size, num_channels, height, width] for the initial image.
+        handle_points (torch.Tensor): The handle points tensor of shape [N, y, x].
+        handle_points_0 (torch.Tensor): The handle points tensor of shape [N, y, x] for the initial image.
+        r2 (int): The radius of the patch around each handle point to use for feature matching.
+        device (torch.device): The device to use for the computation.
+
+    Returns:
+        The new handle points tensor of shape [N, y, x].
+    """
+    n = handle_points.shape[0]  # Number of handle points
+    new_handle_points = torch.zeros_like(handle_points)
+
+    for i in range(n):
+        # Compute the patch around the handle point
+        patch = utils.create_square_mask(
+            F.shape[2], F.shape[3], center=handle_points[i].tolist(), radius=r2
+        ).to(device)
+
+        # Find indices where the patch is True
+        patch_coordinates = torch.nonzero(patch)  # shape [num_points, 2]
+
+        # Extract features in the patch 
+        F_qi = F[
+            :, :, patch_coordinates[:, 0], patch_coordinates[:, 1]
+        ]
+        # Extract feature of the initial handle point
+        f_i = F0[
+            :, :, handle_points_0[i][0].long(), handle_points_0[i][1].long()
+        ]
+
+        # Compute the L1 distance between the patch features and the initial handle point feature
+        distances = torch.norm(F_qi - f_i[:, :, None], p=1, dim=1)
+
+        # Find the new handle point as the one with minimum distance
+        min_index = torch.argmin(distances)
+        new_handle_points[i] = patch_coordinates[min_index]
+
+    return new_handle_points

requirements.txt

@@ -0,0 +1,5 @@
+torch
+torchvision
+streamlit==1.22.0
+# streamlit-drawable-canvas==0.9.2
+streamlit-image-coordinates==0.1.3

stylegan2/LICENSE.txt

@@ -0,0 +1,97 @@
+Copyright (c) 2021, NVIDIA Corporation. All rights reserved.
+
+
+NVIDIA Source Code License for StyleGAN2 with Adaptive Discriminator Augmentation (ADA)
+
+
+=======================================================================
+
+1. Definitions
+
+"Licensor" means any person or entity that distributes its Work.
+
+"Software" means the original work of authorship made available under
+this License.
+
+"Work" means the Software and any additions to or derivative works of
+the Software that are made available under this License.
+
+The terms "reproduce," "reproduction," "derivative works," and
+"distribution" have the meaning as provided under U.S. copyright law;
+provided, however, that for the purposes of this License, derivative
+works shall not include works that remain separable from, or merely
+link (or bind by name) to the interfaces of, the Work.
+
+Works, including the Software, are "made available" under this License
+by including in or with the Work either (a) a copyright notice
+referencing the applicability of this License to the Work, or (b) a
+copy of this License.
+
+2. License Grants
+
+    2.1 Copyright Grant. Subject to the terms and conditions of this
+    License, each Licensor grants to you a perpetual, worldwide,
+    non-exclusive, royalty-free, copyright license to reproduce,
+    prepare derivative works of, publicly display, publicly perform,
+    sublicense and distribute its Work and any resulting derivative
+    works in any form.
+
+3. Limitations
+
+    3.1 Redistribution. You may reproduce or distribute the Work only
+    if (a) you do so under this License, (b) you include a complete
+    copy of this License with your distribution, and (c) you retain
+    without modification any copyright, patent, trademark, or
+    attribution notices that are present in the Work.
+
+    3.2 Derivative Works. You may specify that additional or different
+    terms apply to the use, reproduction, and distribution of your
+    derivative works of the Work ("Your Terms") only if (a) Your Terms
+    provide that the use limitation in Section 3.3 applies to your
+    derivative works, and (b) you identify the specific derivative
+    works that are subject to Your Terms. Notwithstanding Your Terms,
+    this License (including the redistribution requirements in Section
+    3.1) will continue to apply to the Work itself.
+
+    3.3 Use Limitation. The Work and any derivative works thereof only
+    may be used or intended for use non-commercially. Notwithstanding
+    the foregoing, NVIDIA and its affiliates may use the Work and any
+    derivative works commercially. As used herein, "non-commercially"
+    means for research or evaluation purposes only.
+
+    3.4 Patent Claims. If you bring or threaten to bring a patent claim
+    against any Licensor (including any claim, cross-claim or
+    counterclaim in a lawsuit) to enforce any patents that you allege
+    are infringed by any Work, then your rights under this License from
+    such Licensor (including the grant in Section 2.1) will terminate
+    immediately.
+
+    3.5 Trademarks. This License does not grant any rights to use any
+    Licensor’s or its affiliates’ names, logos, or trademarks, except
+    as necessary to reproduce the notices described in this License.
+
+    3.6 Termination. If you violate any term of this License, then your
+    rights under this License (including the grant in Section 2.1) will
+    terminate immediately.
+
+4. Disclaimer of Warranty.
+
+THE WORK IS PROVIDED "AS IS" WITHOUT WARRANTIES OR CONDITIONS OF ANY
+KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR
+NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER
+THIS LICENSE.
+
+5. Limitation of Liability.
+
+EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL
+THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE
+SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT,
+INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF
+OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK
+(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION,
+LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER
+COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF
+THE POSSIBILITY OF SUCH DAMAGES.
+
+=======================================================================

stylegan2/dnnlib/__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.
+
+from .util import EasyDict, make_cache_dir_path

stylegan2/dnnlib/util.py

@@ -0,0 +1,477 @@
+# 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.
+
+"""Miscellaneous utility classes and functions."""
+
+import ctypes
+import fnmatch
+import importlib
+import inspect
+import numpy as np
+import os
+import shutil
+import sys
+import types
+import io
+import pickle
+import re
+import requests
+import html
+import hashlib
+import glob
+import tempfile
+import urllib
+import urllib.request
+import uuid
+
+from distutils.util import strtobool
+from typing import Any, List, Tuple, Union
+
+
+# Util classes
+# ------------------------------------------------------------------------------------------
+
+
+class EasyDict(dict):
+    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self[name]
+        except KeyError:
+            raise AttributeError(name)
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        self[name] = value
+
+    def __delattr__(self, name: str) -> None:
+        del self[name]
+
+
+class Logger(object):
+    """Redirect stderr to stdout, optionally print stdout to a file, and optionally force flushing on both stdout and the file."""
+
+    def __init__(self, file_name: str = None, file_mode: str = "w", should_flush: bool = True):
+        self.file = None
+
+        if file_name is not None:
+            self.file = open(file_name, file_mode)
+
+        self.should_flush = should_flush
+        self.stdout = sys.stdout
+        self.stderr = sys.stderr
+
+        sys.stdout = self
+        sys.stderr = self
+
+    def __enter__(self) -> "Logger":
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.close()
+
+    def write(self, text: Union[str, bytes]) -> None:
+        """Write text to stdout (and a file) and optionally flush."""
+        if isinstance(text, bytes):
+            text = text.decode()
+        if len(text) == 0: # workaround for a bug in VSCode debugger: sys.stdout.write(''); sys.stdout.flush() => crash
+            return
+
+        if self.file is not None:
+            self.file.write(text)
+
+        self.stdout.write(text)
+
+        if self.should_flush:
+            self.flush()
+
+    def flush(self) -> None:
+        """Flush written text to both stdout and a file, if open."""
+        if self.file is not None:
+            self.file.flush()
+
+        self.stdout.flush()
+
+    def close(self) -> None:
+        """Flush, close possible files, and remove stdout/stderr mirroring."""
+        self.flush()
+
+        # if using multiple loggers, prevent closing in wrong order
+        if sys.stdout is self:
+            sys.stdout = self.stdout
+        if sys.stderr is self:
+            sys.stderr = self.stderr
+
+        if self.file is not None:
+            self.file.close()
+            self.file = None
+
+
+# Cache directories
+# ------------------------------------------------------------------------------------------
+
+_dnnlib_cache_dir = None
+
+def set_cache_dir(path: str) -> None:
+    global _dnnlib_cache_dir
+    _dnnlib_cache_dir = path
+
+def make_cache_dir_path(*paths: str) -> str:
+    if _dnnlib_cache_dir is not None:
+        return os.path.join(_dnnlib_cache_dir, *paths)
+    if 'DNNLIB_CACHE_DIR' in os.environ:
+        return os.path.join(os.environ['DNNLIB_CACHE_DIR'], *paths)
+    if 'HOME' in os.environ:
+        return os.path.join(os.environ['HOME'], '.cache', 'dnnlib', *paths)
+    if 'USERPROFILE' in os.environ:
+        return os.path.join(os.environ['USERPROFILE'], '.cache', 'dnnlib', *paths)
+    return os.path.join(tempfile.gettempdir(), '.cache', 'dnnlib', *paths)
+
+# Small util functions
+# ------------------------------------------------------------------------------------------
+
+
+def format_time(seconds: Union[int, float]) -> str:
+    """Convert the seconds to human readable string with days, hours, minutes and seconds."""
+    s = int(np.rint(seconds))
+
+    if s < 60:
+        return "{0}s".format(s)
+    elif s < 60 * 60:
+        return "{0}m {1:02}s".format(s // 60, s % 60)
+    elif s < 24 * 60 * 60:
+        return "{0}h {1:02}m {2:02}s".format(s // (60 * 60), (s // 60) % 60, s % 60)
+    else:
+        return "{0}d {1:02}h {2:02}m".format(s // (24 * 60 * 60), (s // (60 * 60)) % 24, (s // 60) % 60)
+
+
+def ask_yes_no(question: str) -> bool:
+    """Ask the user the question until the user inputs a valid answer."""
+    while True:
+        try:
+            print("{0} [y/n]".format(question))
+            return strtobool(input().lower())
+        except ValueError:
+            pass
+
+
+def tuple_product(t: Tuple) -> Any:
+    """Calculate the product of the tuple elements."""
+    result = 1
+
+    for v in t:
+        result *= v
+
+    return result
+
+
+_str_to_ctype = {
+    "uint8": ctypes.c_ubyte,
+    "uint16": ctypes.c_uint16,
+    "uint32": ctypes.c_uint32,
+    "uint64": ctypes.c_uint64,
+    "int8": ctypes.c_byte,
+    "int16": ctypes.c_int16,
+    "int32": ctypes.c_int32,
+    "int64": ctypes.c_int64,
+    "float32": ctypes.c_float,
+    "float64": ctypes.c_double
+}
+
+
+def get_dtype_and_ctype(type_obj: Any) -> Tuple[np.dtype, Any]:
+    """Given a type name string (or an object having a __name__ attribute), return matching Numpy and ctypes types that have the same size in bytes."""
+    type_str = None
+
+    if isinstance(type_obj, str):
+        type_str = type_obj
+    elif hasattr(type_obj, "__name__"):
+        type_str = type_obj.__name__
+    elif hasattr(type_obj, "name"):
+        type_str = type_obj.name
+    else:
+        raise RuntimeError("Cannot infer type name from input")
+
+    assert type_str in _str_to_ctype.keys()
+
+    my_dtype = np.dtype(type_str)
+    my_ctype = _str_to_ctype[type_str]
+
+    assert my_dtype.itemsize == ctypes.sizeof(my_ctype)
+
+    return my_dtype, my_ctype
+
+
+def is_pickleable(obj: Any) -> bool:
+    try:
+        with io.BytesIO() as stream:
+            pickle.dump(obj, stream)
+        return True
+    except:
+        return False
+
+
+# Functionality to import modules/objects by name, and call functions by name
+# ------------------------------------------------------------------------------------------
+
+def get_module_from_obj_name(obj_name: str) -> Tuple[types.ModuleType, str]:
+    """Searches for the underlying module behind the name to some python object.
+    Returns the module and the object name (original name with module part removed)."""
+
+    # allow convenience shorthands, substitute them by full names
+    obj_name = re.sub("^np.", "numpy.", obj_name)
+    obj_name = re.sub("^tf.", "tensorflow.", obj_name)
+
+    # list alternatives for (module_name, local_obj_name)
+    parts = obj_name.split(".")
+    name_pairs = [(".".join(parts[:i]), ".".join(parts[i:])) for i in range(len(parts), 0, -1)]
+
+    # try each alternative in turn
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+            return module, local_obj_name
+        except:
+            pass
+
+    # maybe some of the modules themselves contain errors?
+    for module_name, _local_obj_name in name_pairs:
+        try:
+            importlib.import_module(module_name) # may raise ImportError
+        except ImportError:
+            if not str(sys.exc_info()[1]).startswith("No module named '" + module_name + "'"):
+                raise
+
+    # maybe the requested attribute is missing?
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+        except ImportError:
+            pass
+
+    # we are out of luck, but we have no idea why
+    raise ImportError(obj_name)
+
+
+def get_obj_from_module(module: types.ModuleType, obj_name: str) -> Any:
+    """Traverses the object name and returns the last (rightmost) python object."""
+    if obj_name == '':
+        return module
+    obj = module
+    for part in obj_name.split("."):
+        obj = getattr(obj, part)
+    return obj
+
+
+def get_obj_by_name(name: str) -> Any:
+    """Finds the python object with the given name."""
+    module, obj_name = get_module_from_obj_name(name)
+    return get_obj_from_module(module, obj_name)
+
+
+def call_func_by_name(*args, func_name: str = None, **kwargs) -> Any:
+    """Finds the python object with the given name and calls it as a function."""
+    assert func_name is not None
+    func_obj = get_obj_by_name(func_name)
+    assert callable(func_obj)
+    return func_obj(*args, **kwargs)
+
+
+def construct_class_by_name(*args, class_name: str = None, **kwargs) -> Any:
+    """Finds the python class with the given name and constructs it with the given arguments."""
+    return call_func_by_name(*args, func_name=class_name, **kwargs)
+
+
+def get_module_dir_by_obj_name(obj_name: str) -> str:
+    """Get the directory path of the module containing the given object name."""
+    module, _ = get_module_from_obj_name(obj_name)
+    return os.path.dirname(inspect.getfile(module))
+
+
+def is_top_level_function(obj: Any) -> bool:
+    """Determine whether the given object is a top-level function, i.e., defined at module scope using 'def'."""
+    return callable(obj) and obj.__name__ in sys.modules[obj.__module__].__dict__
+
+
+def get_top_level_function_name(obj: Any) -> str:
+    """Return the fully-qualified name of a top-level function."""
+    assert is_top_level_function(obj)
+    module = obj.__module__
+    if module == '__main__':
+        module = os.path.splitext(os.path.basename(sys.modules[module].__file__))[0]
+    return module + "." + obj.__name__
+
+
+# File system helpers
+# ------------------------------------------------------------------------------------------
+
+def list_dir_recursively_with_ignore(dir_path: str, ignores: List[str] = None, add_base_to_relative: bool = False) -> List[Tuple[str, str]]:
+    """List all files recursively in a given directory while ignoring given file and directory names.
+    Returns list of tuples containing both absolute and relative paths."""
+    assert os.path.isdir(dir_path)
+    base_name = os.path.basename(os.path.normpath(dir_path))
+
+    if ignores is None:
+        ignores = []
+
+    result = []
+
+    for root, dirs, files in os.walk(dir_path, topdown=True):
+        for ignore_ in ignores:
+            dirs_to_remove = [d for d in dirs if fnmatch.fnmatch(d, ignore_)]
+
+            # dirs need to be edited in-place
+            for d in dirs_to_remove:
+                dirs.remove(d)
+
+            files = [f for f in files if not fnmatch.fnmatch(f, ignore_)]
+
+        absolute_paths = [os.path.join(root, f) for f in files]
+        relative_paths = [os.path.relpath(p, dir_path) for p in absolute_paths]
+
+        if add_base_to_relative:
+            relative_paths = [os.path.join(base_name, p) for p in relative_paths]
+
+        assert len(absolute_paths) == len(relative_paths)
+        result += zip(absolute_paths, relative_paths)
+
+    return result
+
+
+def copy_files_and_create_dirs(files: List[Tuple[str, str]]) -> None:
+    """Takes in a list of tuples of (src, dst) paths and copies files.
+    Will create all necessary directories."""
+    for file in files:
+        target_dir_name = os.path.dirname(file[1])
+
+        # will create all intermediate-level directories
+        if not os.path.exists(target_dir_name):
+            os.makedirs(target_dir_name)
+
+        shutil.copyfile(file[0], file[1])
+
+
+# URL helpers
+# ------------------------------------------------------------------------------------------
+
+def is_url(obj: Any, allow_file_urls: bool = False) -> bool:
+    """Determine whether the given object is a valid URL string."""
+    if not isinstance(obj, str) or not "://" in obj:
+        return False
+    if allow_file_urls and obj.startswith('file://'):
+        return True
+    try:
+        res = requests.compat.urlparse(obj)
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+        res = requests.compat.urlparse(requests.compat.urljoin(obj, "/"))
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+    except:
+        return False
+    return True
+
+
+def open_url(url: str, cache_dir: str = None, num_attempts: int = 10, verbose: bool = True, return_filename: bool = False, cache: bool = True) -> Any:
+    """Download the given URL and return a binary-mode file object to access the data."""
+    assert num_attempts >= 1
+    assert not (return_filename and (not cache))
+
+    # Doesn't look like an URL scheme so interpret it as a local filename.
+    if not re.match('^[a-z]+://', url):
+        return url if return_filename else open(url, "rb")
+
+    # Handle file URLs.  This code handles unusual file:// patterns that
+    # arise on Windows:
+    #
+    # file:///c:/foo.txt
+    #
+    # which would translate to a local '/c:/foo.txt' filename that's
+    # invalid.  Drop the forward slash for such pathnames.
+    #
+    # If you touch this code path, you should test it on both Linux and
+    # Windows.
+    #
+    # Some internet resources suggest using urllib.request.url2pathname() but
+    # but that converts forward slashes to backslashes and this causes
+    # its own set of problems.
+    if url.startswith('file://'):
+        filename = urllib.parse.urlparse(url).path
+        if re.match(r'^/[a-zA-Z]:', filename):
+            filename = filename[1:]
+        return filename if return_filename else open(filename, "rb")
+
+    assert is_url(url)
+
+    # Lookup from cache.
+    if cache_dir is None:
+        cache_dir = make_cache_dir_path('downloads')
+
+    url_md5 = hashlib.md5(url.encode("utf-8")).hexdigest()
+    if cache:
+        cache_files = glob.glob(os.path.join(cache_dir, url_md5 + "_*"))
+        if len(cache_files) == 1:
+            filename = cache_files[0]
+            return filename if return_filename else open(filename, "rb")
+
+    # Download.
+    url_name = None
+    url_data = None
+    with requests.Session() as session:
+        if verbose:
+            print("Downloading %s ..." % url, end="", flush=True)
+        for attempts_left in reversed(range(num_attempts)):
+            try:
+                with session.get(url) as res:
+                    res.raise_for_status()
+                    if len(res.content) == 0:
+                        raise IOError("No data received")
+
+                    if len(res.content) < 8192:
+                        content_str = res.content.decode("utf-8")
+                        if "download_warning" in res.headers.get("Set-Cookie", ""):
+                            links = [html.unescape(link) for link in content_str.split('"') if "export=download" in link]
+                            if len(links) == 1:
+                                url = requests.compat.urljoin(url, links[0])
+                                raise IOError("Google Drive virus checker nag")
+                        if "Google Drive - Quota exceeded" in content_str:
+                            raise IOError("Google Drive download quota exceeded -- please try again later")
+
+                    match = re.search(r'filename="([^"]*)"', res.headers.get("Content-Disposition", ""))
+                    url_name = match[1] if match else url
+                    url_data = res.content
+                    if verbose:
+                        print(" done")
+                    break
+            except KeyboardInterrupt:
+                raise
+            except:
+                if not attempts_left:
+                    if verbose:
+                        print(" failed")
+                    raise
+                if verbose:
+                    print(".", end="", flush=True)
+
+    # Save to cache.
+    if cache:
+        safe_name = re.sub(r"[^0-9a-zA-Z-._]", "_", url_name)
+        cache_file = os.path.join(cache_dir, url_md5 + "_" + safe_name)
+        temp_file = os.path.join(cache_dir, "tmp_" + uuid.uuid4().hex + "_" + url_md5 + "_" + safe_name)
+        os.makedirs(cache_dir, exist_ok=True)
+        with open(temp_file, "wb") as f:
+            f.write(url_data)
+        os.replace(temp_file, cache_file) # atomic
+        if return_filename:
+            return cache_file
+
+    # Return data as file object.
+    assert not return_filename
+    return io.BytesIO(url_data)

stylegan2/legacy.py

@@ -0,0 +1,320 @@
+# 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 click
+import pickle
+import re
+import copy
+import numpy as np
+import torch
+import dnnlib
+from torch_utils import misc
+
+#----------------------------------------------------------------------------
+
+def load_network_pkl(f, force_fp16=False):
+    data = _LegacyUnpickler(f).load()
+
+    # Legacy TensorFlow pickle => convert.
+    if isinstance(data, tuple) and len(data) == 3 and all(isinstance(net, _TFNetworkStub) for net in data):
+        tf_G, tf_D, tf_Gs = data
+        G = convert_tf_generator(tf_G)
+        D = convert_tf_discriminator(tf_D)
+        G_ema = convert_tf_generator(tf_Gs)
+        data = dict(G=G, D=D, G_ema=G_ema)
+
+    # Add missing fields.
+    if 'training_set_kwargs' not in data:
+        data['training_set_kwargs'] = None
+    if 'augment_pipe' not in data:
+        data['augment_pipe'] = None
+
+    # Validate contents.
+    assert isinstance(data['G'], torch.nn.Module)
+    assert isinstance(data['D'], torch.nn.Module)
+    assert isinstance(data['G_ema'], torch.nn.Module)
+    assert isinstance(data['training_set_kwargs'], (dict, type(None)))
+    assert isinstance(data['augment_pipe'], (torch.nn.Module, type(None)))
+
+    # Force FP16.
+    if force_fp16:
+        for key in ['G', 'D', 'G_ema']:
+            old = data[key]
+            kwargs = copy.deepcopy(old.init_kwargs)
+            if key.startswith('G'):
+                kwargs.synthesis_kwargs = dnnlib.EasyDict(kwargs.get('synthesis_kwargs', {}))
+                kwargs.synthesis_kwargs.num_fp16_res = 4
+                kwargs.synthesis_kwargs.conv_clamp = 256
+            if key.startswith('D'):
+                kwargs.num_fp16_res = 4
+                kwargs.conv_clamp = 256
+            if kwargs != old.init_kwargs:
+                new = type(old)(**kwargs).eval().requires_grad_(False)
+                misc.copy_params_and_buffers(old, new, require_all=True)
+                data[key] = new
+    return data
+
+#----------------------------------------------------------------------------
+
+class _TFNetworkStub(dnnlib.EasyDict):
+    pass
+
+class _LegacyUnpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        if module == 'dnnlib.tflib.network' and name == 'Network':
+            return _TFNetworkStub
+        return super().find_class(module, name)
+
+#----------------------------------------------------------------------------
+
+def _collect_tf_params(tf_net):
+    # pylint: disable=protected-access
+    tf_params = dict()
+    def recurse(prefix, tf_net):
+        for name, value in tf_net.variables:
+            tf_params[prefix + name] = value
+        for name, comp in tf_net.components.items():
+            recurse(prefix + name + '/', comp)
+    recurse('', tf_net)
+    return tf_params
+
+#----------------------------------------------------------------------------
+
+def _populate_module_params(module, *patterns):
+    for name, tensor in misc.named_params_and_buffers(module):
+        found = False
+        value = None
+        for pattern, value_fn in zip(patterns[0::2], patterns[1::2]):
+            match = re.fullmatch(pattern, name)
+            if match:
+                found = True
+                if value_fn is not None:
+                    value = value_fn(*match.groups())
+                break
+        try:
+            assert found
+            if value is not None:
+                tensor.copy_(torch.from_numpy(np.array(value)))
+        except:
+            print(name, list(tensor.shape))
+            raise
+
+#----------------------------------------------------------------------------
+
+def convert_tf_generator(tf_G):
+    if tf_G.version < 4:
+        raise ValueError('TensorFlow pickle version too low')
+
+    # Collect kwargs.
+    tf_kwargs = tf_G.static_kwargs
+    known_kwargs = set()
+    def kwarg(tf_name, default=None, none=None):
+        known_kwargs.add(tf_name)
+        val = tf_kwargs.get(tf_name, default)
+        return val if val is not None else none
+
+    # Convert kwargs.
+    kwargs = dnnlib.EasyDict(
+        z_dim                   = kwarg('latent_size',          512),
+        c_dim                   = kwarg('label_size',           0),
+        w_dim                   = kwarg('dlatent_size',         512),
+        img_resolution          = kwarg('resolution',           1024),
+        img_channels            = kwarg('num_channels',         3),
+        mapping_kwargs = dnnlib.EasyDict(
+            num_layers          = kwarg('mapping_layers',       8),
+            embed_features      = kwarg('label_fmaps',          None),
+            layer_features      = kwarg('mapping_fmaps',        None),
+            activation          = kwarg('mapping_nonlinearity', 'lrelu'),
+            lr_multiplier       = kwarg('mapping_lrmul',        0.01),
+            w_avg_beta          = kwarg('w_avg_beta',           0.995,  none=1),
+        ),
+        synthesis_kwargs = dnnlib.EasyDict(
+            channel_base        = kwarg('fmap_base',            16384) * 2,
+            channel_max         = kwarg('fmap_max',             512),
+            num_fp16_res        = kwarg('num_fp16_res',         0),
+            conv_clamp          = kwarg('conv_clamp',           None),
+            architecture        = kwarg('architecture',         'skip'),
+            resample_filter     = kwarg('resample_kernel',      [1,3,3,1]),
+            use_noise           = kwarg('use_noise',            True),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+        ),
+    )
+
+    # Check for unknown kwargs.
+    kwarg('truncation_psi')
+    kwarg('truncation_cutoff')
+    kwarg('style_mixing_prob')
+    kwarg('structure')
+    unknown_kwargs = list(set(tf_kwargs.keys()) - known_kwargs)
+    if len(unknown_kwargs) > 0:
+        raise ValueError('Unknown TensorFlow kwarg', unknown_kwargs[0])
+
+    # Collect params.
+    tf_params = _collect_tf_params(tf_G)
+    for name, value in list(tf_params.items()):
+        match = re.fullmatch(r'ToRGB_lod(\d+)/(.*)', name)
+        if match:
+            r = kwargs.img_resolution // (2 ** int(match.group(1)))
+            tf_params[f'{r}x{r}/ToRGB/{match.group(2)}'] = value
+            kwargs.synthesis.kwargs.architecture = 'orig'
+    #for name, value in tf_params.items(): print(f'{name:<50s}{list(value.shape)}')
+
+    # Convert params.
+    from training import networks
+    G = networks.Generator(**kwargs).eval().requires_grad_(False)
+    # pylint: disable=unnecessary-lambda
+    _populate_module_params(G,
+        r'mapping\.w_avg',                                  lambda:     tf_params[f'dlatent_avg'],
+        r'mapping\.embed\.weight',                          lambda:     tf_params[f'mapping/LabelEmbed/weight'].transpose(),
+        r'mapping\.embed\.bias',                            lambda:     tf_params[f'mapping/LabelEmbed/bias'],
+        r'mapping\.fc(\d+)\.weight',                        lambda i:   tf_params[f'mapping/Dense{i}/weight'].transpose(),
+        r'mapping\.fc(\d+)\.bias',                          lambda i:   tf_params[f'mapping/Dense{i}/bias'],
+        r'synthesis\.b4\.const',                            lambda:     tf_params[f'synthesis/4x4/Const/const'][0],
+        r'synthesis\.b4\.conv1\.weight',                    lambda:     tf_params[f'synthesis/4x4/Conv/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b4\.conv1\.bias',                      lambda:     tf_params[f'synthesis/4x4/Conv/bias'],
+        r'synthesis\.b4\.conv1\.noise_const',               lambda:     tf_params[f'synthesis/noise0'][0, 0],
+        r'synthesis\.b4\.conv1\.noise_strength',            lambda:     tf_params[f'synthesis/4x4/Conv/noise_strength'],
+        r'synthesis\.b4\.conv1\.affine\.weight',            lambda:     tf_params[f'synthesis/4x4/Conv/mod_weight'].transpose(),
+        r'synthesis\.b4\.conv1\.affine\.bias',              lambda:     tf_params[f'synthesis/4x4/Conv/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.conv0\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/weight'][::-1, ::-1].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.conv0\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/bias'],
+        r'synthesis\.b(\d+)\.conv0\.noise_const',           lambda r:   tf_params[f'synthesis/noise{int(np.log2(int(r)))*2-5}'][0, 0],
+        r'synthesis\.b(\d+)\.conv0\.noise_strength',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/noise_strength'],
+        r'synthesis\.b(\d+)\.conv0\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.conv0\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.conv1\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.conv1\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/bias'],
+        r'synthesis\.b(\d+)\.conv1\.noise_const',           lambda r:   tf_params[f'synthesis/noise{int(np.log2(int(r)))*2-4}'][0, 0],
+        r'synthesis\.b(\d+)\.conv1\.noise_strength',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/noise_strength'],
+        r'synthesis\.b(\d+)\.conv1\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.conv1\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.torgb\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.torgb\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/bias'],
+        r'synthesis\.b(\d+)\.torgb\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.torgb\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.skip\.weight',                 lambda r:   tf_params[f'synthesis/{r}x{r}/Skip/weight'][::-1, ::-1].transpose(3, 2, 0, 1),
+        r'.*\.resample_filter',                             None,
+    )
+    return G
+
+#----------------------------------------------------------------------------
+
+def convert_tf_discriminator(tf_D):
+    if tf_D.version < 4:
+        raise ValueError('TensorFlow pickle version too low')
+
+    # Collect kwargs.
+    tf_kwargs = tf_D.static_kwargs
+    known_kwargs = set()
+    def kwarg(tf_name, default=None):
+        known_kwargs.add(tf_name)
+        return tf_kwargs.get(tf_name, default)
+
+    # Convert kwargs.
+    kwargs = dnnlib.EasyDict(
+        c_dim                   = kwarg('label_size',           0),
+        img_resolution          = kwarg('resolution',           1024),
+        img_channels            = kwarg('num_channels',         3),
+        architecture            = kwarg('architecture',         'resnet'),
+        channel_base            = kwarg('fmap_base',            16384) * 2,
+        channel_max             = kwarg('fmap_max',             512),
+        num_fp16_res            = kwarg('num_fp16_res',         0),
+        conv_clamp              = kwarg('conv_clamp',           None),
+        cmap_dim                = kwarg('mapping_fmaps',        None),
+        block_kwargs = dnnlib.EasyDict(
+            activation          = kwarg('nonlinearity',         'lrelu'),
+            resample_filter     = kwarg('resample_kernel',      [1,3,3,1]),
+            freeze_layers       = kwarg('freeze_layers',        0),
+        ),
+        mapping_kwargs = dnnlib.EasyDict(
+            num_layers          = kwarg('mapping_layers',       0),
+            embed_features      = kwarg('mapping_fmaps',        None),
+            layer_features      = kwarg('mapping_fmaps',        None),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+            lr_multiplier       = kwarg('mapping_lrmul',        0.1),
+        ),
+        epilogue_kwargs = dnnlib.EasyDict(
+            mbstd_group_size    = kwarg('mbstd_group_size',     None),
+            mbstd_num_channels  = kwarg('mbstd_num_features',   1),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+        ),
+    )
+
+    # Check for unknown kwargs.
+    kwarg('structure')
+    unknown_kwargs = list(set(tf_kwargs.keys()) - known_kwargs)
+    if len(unknown_kwargs) > 0:
+        raise ValueError('Unknown TensorFlow kwarg', unknown_kwargs[0])
+
+    # Collect params.
+    tf_params = _collect_tf_params(tf_D)
+    for name, value in list(tf_params.items()):
+        match = re.fullmatch(r'FromRGB_lod(\d+)/(.*)', name)
+        if match:
+            r = kwargs.img_resolution // (2 ** int(match.group(1)))
+            tf_params[f'{r}x{r}/FromRGB/{match.group(2)}'] = value
+            kwargs.architecture = 'orig'
+    #for name, value in tf_params.items(): print(f'{name:<50s}{list(value.shape)}')
+
+    # Convert params.
+    from training import networks
+    D = networks.Discriminator(**kwargs).eval().requires_grad_(False)
+    # pylint: disable=unnecessary-lambda
+    _populate_module_params(D,
+        r'b(\d+)\.fromrgb\.weight',     lambda r:       tf_params[f'{r}x{r}/FromRGB/weight'].transpose(3, 2, 0, 1),
+        r'b(\d+)\.fromrgb\.bias',       lambda r:       tf_params[f'{r}x{r}/FromRGB/bias'],
+        r'b(\d+)\.conv(\d+)\.weight',   lambda r, i:    tf_params[f'{r}x{r}/Conv{i}{["","_down"][int(i)]}/weight'].transpose(3, 2, 0, 1),
+        r'b(\d+)\.conv(\d+)\.bias',     lambda r, i:    tf_params[f'{r}x{r}/Conv{i}{["","_down"][int(i)]}/bias'],
+        r'b(\d+)\.skip\.weight',        lambda r:       tf_params[f'{r}x{r}/Skip/weight'].transpose(3, 2, 0, 1),
+        r'mapping\.embed\.weight',      lambda:         tf_params[f'LabelEmbed/weight'].transpose(),
+        r'mapping\.embed\.bias',        lambda:         tf_params[f'LabelEmbed/bias'],
+        r'mapping\.fc(\d+)\.weight',    lambda i:       tf_params[f'Mapping{i}/weight'].transpose(),
+        r'mapping\.fc(\d+)\.bias',      lambda i:       tf_params[f'Mapping{i}/bias'],
+        r'b4\.conv\.weight',            lambda:         tf_params[f'4x4/Conv/weight'].transpose(3, 2, 0, 1),
+        r'b4\.conv\.bias',              lambda:         tf_params[f'4x4/Conv/bias'],
+        r'b4\.fc\.weight',              lambda:         tf_params[f'4x4/Dense0/weight'].transpose(),
+        r'b4\.fc\.bias',                lambda:         tf_params[f'4x4/Dense0/bias'],
+        r'b4\.out\.weight',             lambda:         tf_params[f'Output/weight'].transpose(),
+        r'b4\.out\.bias',               lambda:         tf_params[f'Output/bias'],
+        r'.*\.resample_filter',         None,
+    )
+    return D
+
+#----------------------------------------------------------------------------
+
+@click.command()
+@click.option('--source', help='Input pickle', required=True, metavar='PATH')
+@click.option('--dest', help='Output pickle', required=True, metavar='PATH')
+@click.option('--force-fp16', help='Force the networks to use FP16', type=bool, default=False, metavar='BOOL', show_default=True)
+def convert_network_pickle(source, dest, force_fp16):
+    """Convert legacy network pickle into the native PyTorch format.
+
+    The tool is able to load the main network configurations exported using the TensorFlow version of StyleGAN2 or StyleGAN2-ADA.
+    It does not support e.g. StyleGAN2-ADA comparison methods, StyleGAN2 configs A-D, or StyleGAN1 networks.
+
+    Example:
+
+    \b
+    python legacy.py \\
+        --source=https://nvlabs-fi-cdn.nvidia.com/stylegan2/networks/stylegan2-cat-config-f.pkl \\
+        --dest=stylegan2-cat-config-f.pkl
+    """
+    print(f'Loading "{source}"...')
+    with dnnlib.util.open_url(source) as f:
+        data = load_network_pkl(f, force_fp16=force_fp16)
+    print(f'Saving "{dest}"...')
+    with open(dest, 'wb') as f:
+        pickle.dump(data, f)
+    print('Done.')
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    convert_network_pickle() # pylint: disable=no-value-for-parameter
+
+#----------------------------------------------------------------------------

stylegan2/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

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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

stylegan2/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);
+}
+
+//------------------------------------------------------------------------

stylegan2/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);
+
+//------------------------------------------------------------------------

stylegan2/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);
+
+//------------------------------------------------------------------------

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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
+
+#----------------------------------------------------------------------------

stylegan2/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);
+}
+
+//------------------------------------------------------------------------

stylegan2/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);
+
+//------------------------------------------------------------------------

stylegan2/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);
+
+//------------------------------------------------------------------------

stylegan2/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)
+
+#----------------------------------------------------------------------------

stylegan2/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)
+
+#----------------------------------------------------------------------------

stylegan2/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]
+
+#----------------------------------------------------------------------------

utils.py

@@ -0,0 +1,176 @@
+from typing import List, Optional, Tuple
+import numpy as np
+import torch
+import PIL
+import math
+
+
+def tensor_to_PIL(img: torch.Tensor) -> PIL.Image.Image:
+    """
+    Converts a tensor image to a PIL Image.
+
+    Args:
+        img (torch.Tensor): The tensor image of shape [batch_size, num_channels, height, width].
+
+    Returns:
+        A PIL Image object.
+    """
+    img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
+    return PIL.Image.fromarray(img[0].cpu().numpy(), "RGB")
+
+
+def get_ellipse_coords(
+    point: Tuple[int, int], radius: int = 5
+) -> Tuple[int, int, int, int]:
+    """
+    Returns the coordinates of an ellipse centered at the given point.
+
+    Args:
+        point (Tuple[int, int]): The center point of the ellipse.
+        radius (int): The radius of the ellipse.
+
+    Returns:
+        A tuple containing the coordinates of the ellipse in the format (x_min, y_min, x_max, y_max).
+    """
+    center = point
+    return (
+        center[0] - radius,
+        center[1] - radius,
+        center[0] + radius,
+        center[1] + radius,
+    )
+
+
+def draw_handle_target_points(
+        img: PIL.Image.Image, 
+        handle_points: List[Tuple[int, int]],
+        target_points: List[Tuple[int, int]],
+        radius: int = 5):
+    """
+    Draws handle and target points with arrow pointing towards the target point.
+
+    Args:
+        img (PIL.Image.Image): The image to draw on.
+        handle_points (List[Tuple[int, int]]): A list of handle [x,y] points.
+        target_points (List[Tuple[int, int]]): A list of target [x,y] points.
+        radius (int): The radius of the handle and target points.
+    """
+    if len(handle_points) == len(target_points) + 1:
+        target_points.append(None)
+    draw = PIL.ImageDraw.Draw(img)
+    for handle_point, target_point in zip(handle_points, target_points):
+        # Draw the handle point
+        handle_coords = get_ellipse_coords(handle_point, radius)
+        draw.ellipse(handle_coords, fill="red")
+
+        if target_point:
+            # Draw the target point
+            target_coords = get_ellipse_coords(target_point, radius)
+            draw.ellipse(target_coords, fill="blue")
+
+            # Draw arrow head
+            arrow_head_length = 10.0
+
+            # Compute the direction vector of the line
+            dx = target_point[0] - handle_point[0]
+            dy = target_point[1] - handle_point[1]
+            angle = math.atan2(dy, dx)
+
+            # Shorten the target point by the length of the arrowhead
+            shortened_target_point = (
+                target_point[0] - arrow_head_length * math.cos(angle),
+                target_point[1] - arrow_head_length * math.sin(angle),
+            )
+            
+            # Draw the arrow (main line)
+            draw.line([tuple(handle_point), shortened_target_point], fill='green', width=2)
+
+            # Compute the points for the arrowhead
+            arrow_point1 = (
+                target_point[0] - arrow_head_length * math.cos(angle - math.pi / 6),
+                target_point[1] - arrow_head_length * math.sin(angle - math.pi / 6),
+            )
+
+            arrow_point2 = (
+                target_point[0] - arrow_head_length * math.cos(angle + math.pi / 6),
+                target_point[1] - arrow_head_length * math.sin(angle + math.pi / 6),
+            )
+
+            # Draw the arrowhead
+            draw.polygon([tuple(target_point), arrow_point1, arrow_point2], fill='green')
+ 
+    # # Draw shifted coordinates handle + d_i
+    # for points in all_shifted_coordinates:
+    #     if not torch.isnan(points).any():
+    #         coords = utils.get_ellipse_coords(points.mean(0).flip(-1).cpu().long().numpy().tolist(), 7)
+    #         draw.ellipse(coords, fill="orange")
+
+
+
+def create_circular_mask(
+    h: int,
+    w: int,
+    center: Optional[Tuple[int, int]] = None,
+    radius: Optional[int] = None,
+) -> torch.Tensor:
+    """
+    Create a circular mask tensor.
+
+    Args:
+        h (int): The height of the mask tensor.
+        w (int): The width of the mask tensor.
+        center (Optional[Tuple[int, int]]): The center of the circle as a tuple (y, x). If None, the middle of the image is used.
+        radius (Optional[int]): The radius of the circle. If None, the smallest distance between the center and image walls is used.
+
+    Returns:
+        A boolean tensor of shape [h, w] representing the circular mask.
+    """
+    if center is None:  # use the middle of the image
+        center = (int(h / 2), int(w / 2))
+    if radius is None:  # use the smallest distance between the center and image walls
+        radius = min(center[0], center[1], h - center[0], w - center[1])
+
+    Y, X = np.ogrid[:h, :w]
+    dist_from_center = np.sqrt((Y - center[0]) ** 2 + (X - center[1]) ** 2)
+
+    mask = dist_from_center <= radius
+    mask = torch.from_numpy(mask).bool()
+    return mask
+
+
+def create_square_mask(
+    height: int, width: int, center: list, radius: int
+) -> torch.Tensor:
+    """Create a square mask tensor.
+
+    Args:
+        height (int): The height of the mask.
+        width (int): The width of the mask.
+        center (list): The center of the square mask as a list of two integers. Order [y,x]
+        radius (int): The radius of the square mask.
+
+    Returns:
+        torch.Tensor: The square mask tensor of shape (1, 1, height, width).
+
+    Raises:
+        ValueError: If the center or radius is invalid.
+    """
+    if not isinstance(center, list) or len(center) != 2:
+        raise ValueError("center must be a list of two integers")
+    if not isinstance(radius, int) or radius <= 0:
+        raise ValueError("radius must be a positive integer")
+    if (
+        center[0] < radius
+        or center[0] >= height - radius
+        or center[1] < radius
+        or center[1] >= width - radius
+    ):
+        raise ValueError("center and radius must be within the bounds of the mask")
+
+    mask = torch.zeros((height, width), dtype=torch.float32)
+    x1 = int(center[1]) - radius
+    x2 = int(center[1]) + radius
+    y1 = int(center[0]) - radius
+    y2 = int(center[0]) + radius
+    mask[y1 : y2 + 1, x1 : x2 + 1] = 1.0
+    return mask.bool()