Bundle StyleFeatureEditor code packages in Space to fix ModuleNotFoundError

arguments/inference_arguments.py

@@ -0,0 +1,80 @@
+import os
+
+from pathlib import Path
+from typing import Optional, List, Tuple, Dict
+from dataclasses import dataclass, field
+from omegaconf import OmegaConf, MISSING
+from utils.class_registry import ClassRegistry
+from models.methods import methods_registry
+from metrics.metrics import metrics_registry
+
+
+
+args = ClassRegistry()
+
+
+@args.add_to_registry("exp")
+@dataclass
+class ExperimentArgs:
+    config_dir: str = str(Path(__file__).resolve().parent / "configs")
+    config: str = MISSING
+    output_dir: str = "results_dir"
+    seed: int = 1
+    root: str = os.getenv("EXP_ROOT", ".")
+    domain: str = "human_faces"
+    wandb: bool = False
+
+
+@args.add_to_registry("data")
+@dataclass
+class DataArgs:
+    inference_dir: str = ""
+    transform: str = "face_1024"
+
+
+@args.add_to_registry("inference")
+@dataclass
+class InferenceArgs:
+    inference_runner: str = "base_inference_runner"
+    editings_data: Dict = field(default_factory=lambda: {})
+
+
+@args.add_to_registry("model")
+@dataclass
+class ModelArgs:
+    method: str = "fse_full"
+    device: str = "0"
+    batch_size: int = 4
+    workers: int = 4
+    checkpoint_path: str = ""
+
+
+
+MethodsArgs = methods_registry.make_dataclass_from_args("MethodsArgs")
+args.add_to_registry("methods_args")(MethodsArgs)
+
+MetricsArgs = metrics_registry.make_dataclass_from_args("MetricsArgs")
+args.add_to_registry("metrics")(MetricsArgs)
+
+
+
+Args = args.make_dataclass_from_classes("Args")
+
+
+def load_config():
+    config = OmegaConf.structured(Args)
+
+    conf_cli = OmegaConf.from_cli()
+    config.exp.config = conf_cli.exp.config
+    config.exp.config_dir = conf_cli.exp.config_dir
+
+    config_path = os.path.join(config.exp.config_dir, config.exp.config)
+    conf_file = OmegaConf.load(config_path)
+    config = OmegaConf.merge(config, conf_file)
+    for method in list(config.methods_args.keys()):
+        if method != config.model.method:
+            config.methods_args.__delattr__(method)
+
+    config = OmegaConf.merge(config, conf_cli)
+
+    return config

arguments/training_arguments.py

@@ -0,0 +1,119 @@
+import os
+from training.losses import disc_losses
+from training.optimizers import optimizers
+from pathlib import Path
+from typing import Optional, List, Tuple, Dict
+from dataclasses import dataclass, field
+from omegaconf import OmegaConf, MISSING
+from utils.class_registry import ClassRegistry
+from models.methods import methods_registry
+from metrics.metrics import metrics_registry
+
+
+args = ClassRegistry()
+
+
+@args.add_to_registry("exp")
+@dataclass
+class ExperimentArgs:
+    config_dir: str = str(Path(__file__).resolve().parent / "configs")
+    config: str = MISSING
+    exp_dir: str = "experiments"
+    name: str = MISSING
+    seed: int = 1
+    root: str = os.getenv("EXP_ROOT", ".")
+    wandb: bool = True
+    wandb_project: str = "sfe"
+    domain: str = "human_faces"
+
+
+@args.add_to_registry("data")
+@dataclass
+class DataArgs:
+    special_dir: str = MISSING
+    transform: str = "face_1024"
+    input_train_dir: str = MISSING
+    input_val_dir: str = MISSING
+
+
+@args.add_to_registry("train")
+@dataclass
+class TrainingArgs:
+    train_runner: str = "base_training_runner"
+    encoder_optimizer: str = "ranger"
+    disc_optimizer: str = "adam"
+    resume_path: str = ""
+    val_metrics: List[str] = field(
+        default_factory=lambda: ["msssim", "lpips", "l2", "fid"]
+    )
+    start_step: int = 0
+    steps: int = 300000
+    log_step: int = 500
+    checkpoint_step: int = 15000
+    val_step: int = 15000
+    train_dis: bool = False
+    dis_train_start_step: int = 150000
+    bs_used_before_adv_loss: int = 8
+    disc_edits: List[str] = field(
+        default_factory=lambda: []
+    )
+
+@args.add_to_registry("model")
+@dataclass
+class ModelArgs:
+    method: str = "fse_full"
+    device: str = "0"
+    batch_size: int = 4
+    workers: int = 4
+    checkpoint_path: str = ""
+
+
+@args.add_to_registry("encoder_losses")
+@dataclass
+class EncoderLossesArgs:
+    l2: float = 0.0
+    lpips: float = 0.0
+    lpips_scale: float = 0.0
+    id: float = 0.0
+    moco: float = 0.0
+    adv: float = 0.0
+    feat_rec: float = 0.0
+    feat_rec_l1: float = 0.0
+    l2_latent: float = 0.0
+    id_vit: float = 0.0
+
+
+MethodsArgs = methods_registry.make_dataclass_from_args("MethodsArgs")
+args.add_to_registry("methods_args")(MethodsArgs)
+
+DiscLossesArgs = disc_losses.make_dataclass_from_args("DiscLossesArgs")
+args.add_to_registry("disc_losses")(DiscLossesArgs)
+
+OptimizersArgs = optimizers.make_dataclass_from_args("OptimizersArgs")
+args.add_to_registry("optimizers")(OptimizersArgs)
+
+MetricsArgs = metrics_registry.make_dataclass_from_args("MetricsArgs")
+args.add_to_registry("metrics")(MetricsArgs)
+
+
+Args = args.make_dataclass_from_classes("Args")
+
+
+def load_config():
+    config = OmegaConf.structured(Args)
+
+    conf_cli = OmegaConf.from_cli()
+    config.exp.config = conf_cli.exp.config
+    config.exp.config_dir = conf_cli.exp.config_dir
+
+    config_path = os.path.join(config.exp.config_dir, config.exp.config)
+    conf_file = OmegaConf.load(config_path)
+    config = OmegaConf.merge(config, conf_file)
+
+    for method in list(config.methods_args.keys()):
+        if method != config.model.method:
+            config.methods_args.__delattr__(method)
+
+    config = OmegaConf.merge(config, conf_cli)
+
+    return config

available_directions.txt

@@ -0,0 +1,63 @@
+direction method\name   |    direction effect   | approximate direction |  additional comments
+                        |  during positive edit |     power range to    |       
+                        | (during negative edit |   editing works well  |
+                        |  effect is reversed)  |  (found empirically)  |
+
+
+fs_directions:
+    fs_glasses          |      add glasses      |   [-20; 30]   | may open mouth 
+    fs_smiling          |       add smile       |   [-10; 10]
+    fs_makeup           |      add make-up      |   [-10; 15]   | bad works with men
+
+ganspace_directions:
+    eye_openness        |      close  eyes      |   [-30; 45]
+    trimmed_beard       |      remove beard     |   [-30; 30]   | bad works with women
+    lipstick            |      add lipstick     |   [-30; 30]   | bad works with men
+    face_roundness      |   make face rounder   |   [-20; 15]
+    nose_length         | decreace nose length  |   [-30; 30]   | may open mouth 
+    eyebrow_thickness   |   decreace eyebrow    |   [-20; 20]   
+                        |       thickness       |
+    displeased          |      add sadness      |   [-10; 10]    
+
+interfacegan_directions:
+    age                 |      increase age     |   [-10; 10]    
+    smile               |       add smile       |   [-5 ; 5 ]   | may cause hair artefacts, better use fs_smiling
+    rotation            |    turn  face right   |   [-7 ; 7 ]      
+
+styleclip_directions:
+    afro                |     afro hairstyle    |   [0; 0.14]
+    angry               |      make  angrier    |   [0; 0.14]   | cause background artefacts
+    bobcut              |    bobcut hairstyle   |   [0; 0.18]   | cause background artefacts
+    bowlcut             |    bowlcut h. style   |   [0; 0.14]   | cause background artefacts
+    mohawk              |    mohawk hairstyle   |   [0; 0.10]   | cause background artefacts
+    curly_hair          |        add curls      |   [0; 0.12]
+    purple_hair         |    dye hair purple    |   [0; 0.12]
+    surprised           |  make  more surprised |   [0; 0.10]
+    beyonce             |  make similar to ...  |   [0; 0.12]
+    hilary_clinton      |  make similar to ...  |   [0; 0.10]
+    depp                |  make similar to ...  |   [0; 0.12]   
+    taylor_swift        |  make similar to ...  |   [0; 0.10]
+    trump               |  make similar to ...  |   [0; 0.10]
+    zuckerberg          |  make similar to ...  |   [0; 0.10]
+
+stylespace_directions:
+    black hair          |      darken  hair     |   [-7; 10]
+    blond hair          |      darken  hair     |   [-10; 7]    | yes, positive power darkens hair
+    grey hair           |   make hair colored   |   [-7 ; 7]    | negative means make hair grey
+    wavy hair           |    add hair length    |   [-7 ; 7]
+    receding hairline   |       remove bald     |   [-10; 10]
+    smiling             |      remove smile     |   [-10; 10]   | better use fs_smiling
+    sideburns           |    remove sideburns   |   [-7 ; 7]    | bad works with women
+    goatee              |     remove goatee     |   [-7 ; 7]    | bad works with women
+    earrings            |      add earrings     |   [0 ; 15]    | bad works with men
+    glasses             |     remove glasses    |   [-10; 10]   
+    gender              |     add femininity    |   [-10; 7]    | better use global mapper
+
+You can alse use directions from text prompts via StyleClip Global Mapper (https://arxiv.org/abs/2103.17249).
+Such directions look as follows: "styleclip_global_{neutral prompt}_{target prompt}_{disentanglement}" where
+neutral prompt -- some neutral description of the original image (e.g. "a face")
+target prompt -- text that contains the desired edit (e.g. "a smilling face")
+disentanglement -- positive number, the more this attribute - the more related attributes will also be changed (e.g. 
+for grey hair editing, wrinkle, skin colour and glasses may also be edited)
+
+Example: "styleclip_global_a face_a face with black hair_0.18"

configs/fse_editor_train.yaml

@@ -0,0 +1,30 @@
+train:
+  train_runner: fse_editor
+  train_dis: True
+  dis_train_start_step: 45000
+  steps: 500000
+
+
+model:
+  method: fse_full
+  device: "0"
+  batch_size: 8
+
+
+encoder_losses:
+  l2: 1.0
+  lpips: 0.8
+  id: 0.1
+  adv: 0.01
+
+
+disc_losses:
+  main:
+    coef: 1
+  r1:
+    coef: 10
+
+
+optimizers:
+    ranger:
+      lr: 0.0001

configs/fse_inference.yaml

@@ -0,0 +1,15 @@
+inference:
+  inference_runner: fse_inference_runner
+     
+  editings_data: {
+              "age": [ 2, 3, 4, 5, 6, 7, 8],
+              "fs_glasses": [5, 10, 15, 20, 25, 30],
+              "fs_smiling": [-10, -9, -8, -7, -6, -5, -4,-3],
+              "styleclip_global_face with hair_face with fire hair_0.10": [5, 9]
+            }
+
+
+model:
+  method: fse_full
+  device: "0"
+  batch_size: 8

configs/fse_inverter_inference.yaml

@@ -0,0 +1,23 @@
+inference:
+  inference_runner: fse_inverter_inference_runner
+  editings_data: {
+              "age": [-7, -5, -3, 3, 5, 7, 10],
+              "fs_makeup": [5, 10, 15],
+              "afro": [0.03, 0.07, 0.085, 0.1],
+              "angry": [0.07, 0.1, 0.12],
+              "purple_hair": [0.07, 0.1, 0.12],
+              "glasses": [-7, 5],
+              "face_roundness": [-17, -12, -8, 8, 12, 17], 
+              "rotation": [-5.0, -3.0, -1.0, 1.0, 3.0, 5.0],
+              "bobcut": [0.07, 0.12, 0.18],
+              "bowlcut": [0.07, 0.14],
+              "mohawk": [0.07, 0.10],
+              "blond hair ": [-8, -4, 4, 8],
+              "fs_smiling": [-9, -6, -3, 3, 6, 9]
+            }
+
+
+model:
+  method: fse_inverter
+  device: "0"
+  batch_size: 8

configs/fse_inverter_train.yaml

@@ -0,0 +1,30 @@
+train:
+  train_runner: fse_inverter
+  train_dis: True
+  dis_train_start_step: 45000
+
+
+model:
+  method: fse_inverter
+  device: "0"
+  batch_size: 8
+
+
+encoder_losses:
+  l2: 1.0
+  lpips: 0.8
+  id: 0.1
+  adv: 0.01
+  feat_rec: 0.01
+
+
+disc_losses:
+  main:
+    coef: 1
+  r1:
+    coef: 10
+
+
+optimizers:
+    ranger:
+      lr: 0.0001

configs/paths.py

@@ -0,0 +1,28 @@
+from dataclasses import dataclass, asdict, fields
+
+
+models_dir = "pretrained_models/"
+
+
+@dataclass
+class DefaultPathsClass:
+    psp_path: str = models_dir + "psp_ffhq_encode.pt"
+    e4e_path: str = models_dir + "e4e_ffhq_encode.pt"
+    farl_path:str = models_dir + "face_parsing.farl.lapa.main_ema_136500_jit191.pt"
+    mobile_net_pth: str = models_dir + "mobilenet0.25_Final.pth"
+    ir_se50_path: str = models_dir + "model_ir_se50.pth"
+    stylegan_weights: str = models_dir + "stylegan2-ffhq-config-f.pt"
+    stylegan_car_weights: str = models_dir + "stylegan2-car-config-f-new.pkl"
+    stylegan_weights_pkl: str = models_dir + "stylegan2-ffhq-config-f.pkl"
+    arcface_model_path: str = models_dir + "iresnet50-7f187506.pth"
+    moco: str = models_dir + "moco_v2_800ep_pretrain.pt"
+    curricular_face_path: str = models_dir + "CurricularFace_Backbone.pth"
+    mtcnn: str = models_dir + "mtcnn"
+    landmark: str = models_dir + "79999_iter.pth"
+
+    def __iter__(self):
+      for field in fields(self):
+          yield field.name, getattr(self, field.name)
+
+
+DefaultPaths = DefaultPathsClass()

configs/simple_inference.yaml

@@ -0,0 +1,12 @@
+exp:
+  domain: human_faces
+
+model:
+  method: fse_full
+  device: "0"
+  batch_size: 1
+  workers: 1
+
+methods_args:
+  fse_full:
+    

criteria/id_loss.py

@@ -0,0 +1,38 @@
+import torch
+from torch import nn
+from configs.paths import DefaultPaths
+from models.psp.encoders.model_irse import Backbone
+
+
+class IDLoss(nn.Module):
+    def __init__(self):
+        super(IDLoss, self).__init__()
+        print("Loading ResNet ArcFace")
+        self.facenet = Backbone(
+            input_size=112, num_layers=50, drop_ratio=0.6, mode="ir_se"
+        )
+        self.facenet.load_state_dict(torch.load(DefaultPaths.ir_se50_path))
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
+        self.facenet = self.facenet.cuda().eval()
+
+    def extract_feats(self, x):
+        x = x[:, :, 35:223, 32:220]  # Crop interesting region
+        x = self.face_pool(x)
+        x_feats = self.facenet(x.cuda())
+        return x_feats
+
+
+    def forward(self, y_hat, y):
+        n_samples = y.shape[0]
+        y_feats = self.extract_feats(y)
+        y_hat_feats = self.extract_feats(y_hat)
+
+        y_feats = y_feats.detach()
+        loss = 0
+        count = 0
+        for i in range(n_samples):
+            diff_target = y_hat_feats[i].dot(y_feats[i])
+            loss += 1 - diff_target
+            count += 1
+
+        return loss / count

criteria/id_vit_loss.py

@@ -0,0 +1,453 @@
+import torch
+from torch import nn
+from configs.paths import DefaultPaths
+from models.psp.encoders.model_irse import Backbone
+
+
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from torch import nn
+
+from torch.nn import Parameter
+#from IPython import embed
+
+MIN_NUM_PATCHES = 16
+
+class Softmax(nn.Module):
+    r"""Implement of Softmax (normal classification head):
+        Args:
+            in_features: size of each input sample
+            out_features: size of each output sample
+            device_id: the ID of GPU where the model will be trained by model parallel.
+                       if device_id=None, it will be trained on CPU without model parallel.
+        """
+    def __init__(self, in_features, out_features, device_id):
+        super(Softmax, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.device_id = device_id
+
+        self.weight = Parameter(torch.FloatTensor(out_features, in_features))
+        self.bias = Parameter(torch.FloatTensor(out_features))
+        nn.init.xavier_uniform_(self.weight)
+        nn.init.zeros_(self.bias)
+
+    def forward(self, input, label):
+        if self.device_id == None:
+            out = F.linear(x, self.weight, self.bias)
+        else:
+            x = input
+            sub_weights = torch.chunk(self.weight, len(self.device_id), dim=0)
+            sub_biases = torch.chunk(self.bias, len(self.device_id), dim=0)
+            temp_x = x.cuda(self.device_id[0])
+            weight = sub_weights[0].cuda(self.device_id[0])
+            bias = sub_biases[0].cuda(self.device_id[0])
+            out = F.linear(temp_x, weight, bias)
+            for i in range(1, len(self.device_id)):
+                temp_x = x.cuda(self.device_id[i])
+                weight = sub_weights[i].cuda(self.device_id[i])
+                bias = sub_biases[i].cuda(self.device_id[i])
+                out = torch.cat((out, F.linear(temp_x, weight, bias).cuda(self.device_id[0])), dim=1)
+        return out
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.xavier_uniform_(m.weight.data)
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+            elif isinstance(m, nn.BatchNorm1d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+            elif isinstance(m, nn.Linear):
+                nn.init.xavier_uniform_(m.weight.data)
+                if m.bias is not None:
+                    m.bias.data.zero_()
+
+
+
+class ArcFace(nn.Module):
+    r"""Implement of ArcFace (https://arxiv.org/pdf/1801.07698v1.pdf):
+        Args:
+            in_features: size of each input sample
+            out_features: size of each output sample
+            device_id: the ID of GPU where the model will be trained by model parallel.
+                       if device_id=None, it will be trained on CPU without model parallel.
+            s: norm of input feature
+            m: margin
+            cos(theta+m)
+        """
+
+    def __init__(self, in_features, out_features, device_id, s=64.0, m=0.50, easy_margin=False):
+        super(ArcFace, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.device_id = device_id
+
+        self.s = s
+        self.m = m
+
+        self.weight = Parameter(torch.FloatTensor(out_features, in_features))
+        nn.init.xavier_uniform_(self.weight)
+
+        self.easy_margin = easy_margin
+        self.cos_m = math.cos(m)
+        self.sin_m = math.sin(m)
+        self.th = math.cos(math.pi - m)
+        self.mm = math.sin(math.pi - m) * m
+
+    def forward(self, input, label):
+        # --------------------------- cos(theta) & phi(theta) ---------------------------
+        if self.device_id == None:
+            cosine = F.linear(F.normalize(input), F.normalize(self.weight))
+        else:
+            x = input
+            sub_weights = torch.chunk(self.weight, len(self.device_id), dim=0)
+            temp_x = x.cuda(self.device_id[0])
+            weight = sub_weights[0].cuda(self.device_id[0])
+            cosine = F.linear(F.normalize(temp_x), F.normalize(weight))
+            for i in range(1, len(self.device_id)):
+                temp_x = x.cuda(self.device_id[i])
+                weight = sub_weights[i].cuda(self.device_id[i])
+                cosine = torch.cat((cosine, F.linear(F.normalize(temp_x), F.normalize(weight)).cuda(self.device_id[0])),
+                                   dim=1)
+        sine = torch.sqrt(1.0 - torch.pow(cosine, 2))
+        phi = cosine * self.cos_m - sine * self.sin_m
+        if self.easy_margin:
+            phi = torch.where(cosine > 0, phi, cosine)
+        else:
+            phi = torch.where(cosine > self.th, phi, cosine - self.mm)
+        # --------------------------- convert label to one-hot ---------------------------
+        one_hot = torch.zeros(cosine.size())
+        if self.device_id != None:
+            one_hot = one_hot.cuda(self.device_id[0])
+        one_hot.scatter_(1, label.view(-1, 1).long(), 1)
+        # -------------torch.where(out_i = {x_i if condition_i else y_i) -------------
+        output = (one_hot * phi) + (
+                    (1.0 - one_hot) * cosine)  # you can use torch.where if your torch.__version__ is 0.4
+        output *= self.s
+
+        return output
+
+
+class CosFace(nn.Module):
+    r"""Implement of CosFace (https://arxiv.org/pdf/1801.09414.pdf):
+    Args:
+        in_features: size of each input sample
+        out_features: size of each output sample
+        device_id: the ID of GPU where the model will be trained by model parallel.
+                       if device_id=None, it will be trained on CPU without model parallel.
+        s: norm of input feature
+        m: margin
+        cos(theta)-m
+    """
+
+    def __init__(self, in_features, out_features, device_id, s=64.0, m=0.35):
+        super(CosFace, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.device_id = device_id
+        self.s = s
+        self.m = m
+        print("self.device_id", self.device_id)
+        self.weight = Parameter(torch.FloatTensor(out_features, in_features))
+        nn.init.xavier_uniform_(self.weight)
+
+    def forward(self, input, label):
+        # --------------------------- cos(theta) & phi(theta) ---------------------------
+
+        if self.device_id == None:
+            cosine = F.linear(F.normalize(input), F.normalize(self.weight))
+        else:
+            x = input
+            sub_weights = torch.chunk(self.weight, len(self.device_id), dim=0)
+            temp_x = x.cuda(self.device_id[0])
+            weight = sub_weights[0].cuda(self.device_id[0])
+            cosine = F.linear(F.normalize(temp_x), F.normalize(weight))
+            for i in range(1, len(self.device_id)):
+                temp_x = x.cuda(self.device_id[i])
+                weight = sub_weights[i].cuda(self.device_id[i])
+                cosine = torch.cat((cosine, F.linear(F.normalize(temp_x), F.normalize(weight)).cuda(self.device_id[0])),
+                                   dim=1)
+        phi = cosine - self.m
+        # --------------------------- convert label to one-hot ---------------------------
+        one_hot = torch.zeros(cosine.size())
+        if self.device_id != None:
+            one_hot = one_hot.cuda(self.device_id[0])
+        # one_hot = one_hot.cuda() if cosine.is_cuda else one_hot
+
+        one_hot.scatter_(1, label.cuda(self.device_id[0]).view(-1, 1).long(), 1)
+        # -------------torch.where(out_i = {x_i if condition_i else y_i) -------------
+        output = (one_hot * phi) + (
+                    (1.0 - one_hot) * cosine)  # you can use torch.where if your torch.__version__ is 0.4
+        output *= self.s
+
+        return output
+
+    def __repr__(self):
+        return self.__class__.__name__ + '(' \
+               + 'in_features = ' + str(self.in_features) \
+               + ', out_features = ' + str(self.out_features) \
+               + ', s = ' + str(self.s) \
+               + ', m = ' + str(self.m) + ')'
+
+class SFaceLoss(nn.Module):
+
+    def __init__(self, in_features, out_features, device_id, s = 64.0, k = 80.0, a = 0.90, b = 1.2):
+        super(SFaceLoss, self).__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.device_id = device_id
+        self.s = s
+        self.k = k
+        self.a = a
+        self.b = b
+        self.weight = Parameter(torch.FloatTensor(out_features, in_features))
+        #nn.init.xavier_uniform_(self.weight)
+        xavier_normal_(self.weight, gain=2, mode='out')
+
+    def forward(self, input, label):
+        # --------------------------- cos(theta) & phi(theta) ---------------------------
+        if self.device_id == None:
+            cosine = F.linear(F.normalize(input), F.normalize(self.weight))
+        else:
+            x = input
+            sub_weights = torch.chunk(self.weight, len(self.device_id), dim=0)
+            temp_x = x.cuda(self.device_id[0])
+            weight = sub_weights[0].cuda(self.device_id[0])
+            cosine = F.linear(F.normalize(temp_x), F.normalize(weight))
+
+            for i in range(1, len(self.device_id)):
+                temp_x = x.cuda(self.device_id[i])
+                weight = sub_weights[i].cuda(self.device_id[i])
+
+                cosine = torch.cat((cosine, F.linear(F.normalize(temp_x), F.normalize(weight)).cuda(self.device_id[0])), dim=1)
+        # --------------------------- s*cos(theta) ---------------------------
+        output = cosine * self.s
+        # --------------------------- sface loss ---------------------------
+
+        one_hot = torch.zeros(cosine.size())
+        if self.device_id != None:
+            one_hot = one_hot.cuda(self.device_id[0])
+        one_hot.scatter_(1, label.view(-1, 1), 1)
+
+        zero_hot = torch.ones(cosine.size())
+        if self.device_id != None:
+            zero_hot = zero_hot.cuda(self.device_id[0])
+        zero_hot.scatter_(1, label.view(-1, 1), 0)
+
+
+        WyiX = torch.sum(one_hot * output, 1)
+        with torch.no_grad():
+            theta_yi = torch.acos(WyiX / self.s)
+            weight_yi = 1.0 / (1.0 + torch.exp(-self.k * (theta_yi - self.a)))
+        intra_loss = - weight_yi * WyiX
+
+        Wj = zero_hot * output
+        with torch.no_grad():
+            # theta_j = torch.acos(Wj)
+            theta_j = torch.acos(Wj / self.s)
+            weight_j = 1.0 / (1.0 + torch.exp(self.k * (theta_j - self.b)))
+        inter_loss = torch.sum(weight_j * Wj, 1)
+
+        loss = intra_loss.mean() + inter_loss.mean()
+        Wyi_s = WyiX / self.s
+        Wj_s = Wj / self.s
+        return output, loss, intra_loss.mean(), inter_loss.mean(), Wyi_s.mean(), Wj_s.mean()
+
+
+
+class Residual(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+    def forward(self, x, **kwargs):
+        return self.fn(x, **kwargs) + x
+
+class PreNorm(nn.Module):
+    def __init__(self, dim, fn):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.fn = fn
+    def forward(self, x, **kwargs):
+        return self.fn(self.norm(x), **kwargs)
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout = 0.):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        self.heads = heads
+        self.scale = dim ** -0.5
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, mask = None):
+        b, n, _, h = *x.shape, self.heads
+        qkv = self.to_qkv(x).chunk(3, dim = -1)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
+        dots = torch.einsum('bhid,bhjd->bhij', q, k) * self.scale
+        mask_value = -torch.finfo(dots.dtype).max
+        #embed()
+        if mask is not None:
+            mask = F.pad(mask.flatten(1), (1, 0), value = True)
+            assert mask.shape[-1] == dots.shape[-1], 'mask has incorrect dimensions'
+            mask = mask[:, None, :] * mask[:, :, None]
+            dots.masked_fill_(~mask, mask_value)
+            del mask
+
+        attn = dots.softmax(dim=-1)
+
+        out = torch.einsum('bhij,bhjd->bhid', attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        out =  self.to_out(out)
+
+        return out
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
+                Residual(PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)))
+            ]))
+    def forward(self, x, mask = None):
+        for attn, ff in self.layers:
+            x = attn(x, mask = mask)
+            #embed()
+            x = ff(x)
+        return x
+
+
+class ViTs_face(nn.Module):
+    def __init__(self, *, loss_type, GPU_ID, num_class, image_size, patch_size, ac_patch_size,
+                         pad, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.):
+        super().__init__()
+        assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
+        num_patches = (image_size // patch_size) ** 2
+        patch_dim = channels * ac_patch_size ** 2
+        assert num_patches > MIN_NUM_PATCHES, f'your number of patches ({num_patches}) is way too small for attention to be effective (at least 16). Try decreasing your patch size'
+        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'
+
+        self.patch_size = patch_size
+        self.soft_split = nn.Unfold(kernel_size=(ac_patch_size, ac_patch_size), stride=(self.patch_size, self.patch_size), padding=(pad, pad))
+
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
+        self.patch_to_embedding = nn.Linear(patch_dim, dim)
+        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
+        self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)
+
+        self.pool = pool
+        self.to_latent = nn.Identity()
+
+        self.mlp_head = nn.Sequential(
+            nn.LayerNorm(dim),
+        )
+        self.loss_type = loss_type
+        self.GPU_ID = GPU_ID
+        if self.loss_type == 'None':
+            print("no loss for vit_face")
+        else:
+            if self.loss_type == 'Softmax':
+                self.loss = Softmax(in_features=dim, out_features=num_class, device_id=self.GPU_ID)
+            elif self.loss_type == 'CosFace':
+                self.loss = CosFace(in_features=dim, out_features=num_class, device_id=self.GPU_ID)
+            elif self.loss_type == 'ArcFace':
+                self.loss = ArcFace(in_features=dim, out_features=num_class, device_id=self.GPU_ID)
+            elif self.loss_type == 'SFace':
+                self.loss = SFaceLoss(in_features=dim, out_features=num_class, device_id=self.GPU_ID)
+
+    def forward(self, img, label= None , mask = None):
+        p = self.patch_size
+        x = self.soft_split(img).transpose(1, 2)
+        x = self.patch_to_embedding(x)
+        b, n, _ = x.shape
+
+        cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
+        x = torch.cat((cls_tokens, x), dim=1)
+        x += self.pos_embedding[:, :(n + 1)]
+        x = self.dropout(x)
+        x = self.transformer(x, mask)
+
+        x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]
+
+        x = self.to_latent(x)
+        emb = self.mlp_head(x)
+        if label is not None:
+            x = self.loss(emb, label)
+            return x, emb
+        else:
+            return emb
+
+def l2_norm(input, axis=1):
+    norm = torch.norm(input, 2, axis, True)
+    output = torch.div(input, norm)
+    return output
+
+
+class IDVitLoss(nn.Module):
+    def __init__(self):
+        super(IDVitLoss, self).__init__()
+        print("Loading Vit ArcFace")
+        DEVICE = torch.device("cuda:0")
+        NUM_CLASS = 93431
+        self.facenet = ViTs_face(
+                    loss_type='CosFace',
+                    GPU_ID=DEVICE,
+                    num_class=NUM_CLASS,
+                    image_size=112,
+                    patch_size=8,
+                    ac_patch_size=12,
+                    pad=4,
+                    dim=512,
+                    depth=20,
+                    heads=8,
+                    mlp_dim=2048,
+                    dropout=0.1,
+                    emb_dropout=0.1
+                )
+        self.facenet.load_state_dict(torch.load("pretrained_models/Backbone_VITs_Epoch_2_Batch_12000_Time_2021-03-17-04-05_checkpoint.pth"))
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((112, 112))
+        self.facenet = self.facenet.cuda().eval()
+
+    def extract_feats(self, x):
+        #x = x[:, :, 35:223, 32:220]  # Crop interesting region
+        x = self.face_pool(x)
+        x_feats = self.facenet(x.cuda())
+        return x_feats
+
+    def forward(self, y_hat, y):
+        n_samples = y.shape[0]
+        y_feats = self.extract_feats(y)
+        y_hat_feats = self.extract_feats(y_hat)
+        y_feats = y_feats.detach()
+        loss = torch.mean((y_hat_feats - y_feats)**2)
+
+        return loss * 10000

criteria/lpips/lpips.py

@@ -0,0 +1,35 @@
+import torch
+import torch.nn as nn
+
+from criteria.lpips.networks import get_network, LinLayers
+from criteria.lpips.utils import get_state_dict
+
+
+class LPIPS(nn.Module):
+    r"""Creates a criterion that measures
+    Learned Perceptual Image Patch Similarity (LPIPS).
+    Arguments:
+        net_type (str): the network type to compare the features:
+                        'alex' | 'squeeze' | 'vgg'. Default: 'alex'.
+        version (str): the version of LPIPS. Default: 0.1.
+    """
+
+    def __init__(self, net_type: str = "vgg", version: str = "0.1"):
+        assert version in ["0.1"], "v0.1 is only supported now"
+
+        super(LPIPS, self).__init__()
+
+        # pretrained network
+        self.net = get_network(net_type).to("cuda")
+
+        # linear layers
+        self.lin = LinLayers(self.net.n_channels_list).to("cuda")
+        self.lin.load_state_dict(get_state_dict(net_type, version))
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor):
+        feat_x, feat_y = self.net(x), self.net(y)
+
+        diff = [(fx - fy) ** 2 for fx, fy in zip(feat_x, feat_y)]
+        res = [l(d).mean((2, 3), True) for d, l in zip(diff, self.lin)]
+
+        return torch.sum(torch.cat(res, 0)) / x.shape[0]

criteria/lpips/networks.py

@@ -0,0 +1,98 @@
+from typing import Sequence
+
+from itertools import chain
+
+import torch
+import torch.nn as nn
+from torchvision import models
+
+from criteria.lpips.utils import normalize_activation
+
+
+def get_network(net_type: str):
+    if net_type == "alex":
+        return AlexNet()
+    elif net_type == "squeeze":
+        return SqueezeNet()
+    elif net_type == "vgg":
+        return VGG16()
+    else:
+        raise NotImplementedError("choose net_type from [alex, squeeze, vgg].")
+
+
+class LinLayers(nn.ModuleList):
+    def __init__(self, n_channels_list: Sequence[int]):
+        super(LinLayers, self).__init__(
+            [
+                nn.Sequential(nn.Identity(), nn.Conv2d(nc, 1, 1, 1, 0, bias=False))
+                for nc in n_channels_list
+            ]
+        )
+
+        for param in self.parameters():
+            param.requires_grad = False
+
+
+class BaseNet(nn.Module):
+    def __init__(self):
+        super(BaseNet, self).__init__()
+
+        # register buffer
+        self.register_buffer(
+            "mean", torch.Tensor([-0.030, -0.088, -0.188])[None, :, None, None]
+        )
+        self.register_buffer(
+            "std", torch.Tensor([0.458, 0.448, 0.450])[None, :, None, None]
+        )
+
+    def set_requires_grad(self, state: bool):
+        for param in chain(self.parameters(), self.buffers()):
+            param.requires_grad = state
+
+    def z_score(self, x: torch.Tensor):
+        return (x - self.mean) / self.std
+
+    def forward(self, x: torch.Tensor):
+        x = self.z_score(x)
+
+        output = []
+        for i, (_, layer) in enumerate(self.layers._modules.items(), 1):
+            x = layer(x)
+            if i in self.target_layers:
+                output.append(normalize_activation(x))
+            if len(output) == len(self.target_layers):
+                break
+        return output
+
+
+class SqueezeNet(BaseNet):
+    def __init__(self):
+        super(SqueezeNet, self).__init__()
+
+        self.layers = models.squeezenet1_1(True).features
+        self.target_layers = [2, 5, 8, 10, 11, 12, 13]
+        self.n_channels_list = [64, 128, 256, 384, 384, 512, 512]
+
+        self.set_requires_grad(False)
+
+
+class AlexNet(BaseNet):
+    def __init__(self):
+        super(AlexNet, self).__init__()
+
+        self.layers = models.alexnet(True).features
+        self.target_layers = [2, 5, 8, 10, 12]
+        self.n_channels_list = [64, 192, 384, 256, 256]
+
+        self.set_requires_grad(False)
+
+
+class VGG16(BaseNet):
+    def __init__(self):
+        super(VGG16, self).__init__()
+
+        self.layers = models.vgg16(True).features
+        self.target_layers = [4, 9, 16, 23, 30]
+        self.n_channels_list = [64, 128, 256, 512, 512]
+
+        self.set_requires_grad(False)

criteria/lpips/utils.py

@@ -0,0 +1,33 @@
+from collections import OrderedDict
+
+import torch
+
+
+def normalize_activation(x, eps=1e-10):
+    norm_factor = torch.sqrt(torch.sum(x**2, dim=1, keepdim=True))
+    return x / (norm_factor + eps)
+
+
+def get_state_dict(net_type: str = "alex", version: str = "0.1"):
+    # build url
+    url = (
+        "https://raw.githubusercontent.com/richzhang/PerceptualSimilarity/"
+        + f"master/lpips/weights/v{version}/{net_type}.pth"
+    )
+
+    # download
+    old_state_dict = torch.hub.load_state_dict_from_url(
+        url,
+        progress=True,
+        map_location=None if torch.cuda.is_available() else torch.device("cpu"),
+    )
+
+    # rename keys
+    new_state_dict = OrderedDict()
+    for key, val in old_state_dict.items():
+        new_key = key
+        new_key = new_key.replace("lin", "")
+        new_key = new_key.replace("model.", "")
+        new_state_dict[new_key] = val
+
+    return new_state_dict

criteria/moco_loss.py

@@ -0,0 +1,62 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from configs.paths import DefaultPaths
+
+
+class MocoLoss(nn.Module):
+    def __init__(self):
+        super(MocoLoss, self).__init__()
+        print("Loading MOCO model from path: {}".format(DefaultPaths.moco))
+        self.model = self.__load_model()
+        self.model.cuda()
+        self.model.eval()
+
+    @staticmethod
+    def __load_model():
+        import torchvision.models as models
+
+        model = models.__dict__["resnet50"]()
+        # freeze all layers but the last fc
+        for name, param in model.named_parameters():
+            if name not in ["fc.weight", "fc.bias"]:
+                param.requires_grad = False
+        checkpoint = torch.load(DefaultPaths.moco, map_location="cpu")
+        state_dict = checkpoint["state_dict"]
+        # rename moco pre-trained keys
+        for k in list(state_dict.keys()):
+            # retain only encoder_q up to before the embedding layer
+            if k.startswith("module.encoder_q") and not k.startswith(
+                "module.encoder_q.fc"
+            ):
+                # remove prefix
+                state_dict[k[len("module.encoder_q.") :]] = state_dict[k]
+            # delete renamed or unused k
+            del state_dict[k]
+        msg = model.load_state_dict(state_dict, strict=False)
+        assert set(msg.missing_keys) == {"fc.weight", "fc.bias"}
+        # remove output layer
+        model = nn.Sequential(*list(model.children())[:-1]).cuda()
+        return model
+
+    def extract_feats(self, x):
+        x = F.interpolate(x, size=224)
+        x_feats = self.model(x)
+        x_feats = nn.functional.normalize(x_feats, dim=1)
+        x_feats = x_feats.squeeze()
+        return x_feats
+        
+
+    def forward(self, y_hat, y):
+        n_samples = y.shape[0]
+        y_feats = self.extract_feats(y)
+        y_hat_feats = self.extract_feats(y_hat)
+        y_feats = y_feats.detach()
+        loss = 0
+        count = 0
+        for i in range(n_samples):
+            diff_target = y_hat_feats[i].dot(y_feats[i])
+            loss += 1 - diff_target
+            count += 1
+
+        return loss / count

criteria/ms_ssim.py

@@ -0,0 +1,180 @@
+import torch
+import torch.nn.functional as F
+from math import exp
+import numpy as np
+
+"""
+Taken from https://github.com/jorge-pessoa/pytorch-msssim
+"""
+
+
+def gaussian(window_size, sigma):
+    gauss = torch.Tensor(
+        [
+            exp(-((x - window_size // 2) ** 2) / float(2 * sigma**2))
+            for x in range(window_size)
+        ]
+    )
+    return gauss / gauss.sum()
+
+
+def create_window(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
+    window = _2D_window.expand(channel, 1, window_size, window_size).contiguous()
+    return window
+
+
+def ssim(
+    img1,
+    img2,
+    window_size=11,
+    window=None,
+    size_average=True,
+    full=False,
+    val_range=None,
+):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, channel, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window(real_size, channel=channel).to(img1.device)
+
+    mu1 = F.conv2d(img1, window, padding=padd, groups=channel)
+    mu2 = F.conv2d(img2, window, padding=padd, groups=channel)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv2d(img1 * img1, window, padding=padd, groups=channel) - mu1_sq
+    sigma2_sq = F.conv2d(img2 * img2, window, padding=padd, groups=channel) - mu2_sq
+    sigma12 = F.conv2d(img1 * img2, window, padding=padd, groups=channel) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = v1 / v2  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        cs = cs.mean()
+        ret = ssim_map.mean()
+    else:
+        cs = cs.mean(1).mean(1).mean(1)
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def msssim(
+    img1, img2, window_size=11, size_average=True, val_range=None, normalize=None
+):
+    device = img1.device
+    weights = torch.FloatTensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333]).to(device)
+    levels = weights.size()[0]
+    ssims = []
+    mcs = []
+    for _ in range(levels):
+        sim, cs = ssim(
+            img1,
+            img2,
+            window_size=window_size,
+            size_average=size_average,
+            full=True,
+            val_range=val_range,
+        )
+
+        # Relu normalize (not compliant with original definition)
+        if normalize == "relu":
+            ssims.append(torch.relu(sim))
+            mcs.append(torch.relu(cs))
+        else:
+            ssims.append(sim)
+            mcs.append(cs)
+
+        img1 = F.avg_pool2d(img1, (2, 2))
+        img2 = F.avg_pool2d(img2, (2, 2))
+
+    ssims = torch.stack(ssims)
+    mcs = torch.stack(mcs)
+
+    # Simple normalize (not compliant with original definition)
+    if normalize == "simple" or normalize == True:
+        ssims = (ssims + 1) / 2
+        mcs = (mcs + 1) / 2
+
+    pow1 = mcs**weights
+    pow2 = ssims**weights
+
+    # From Matlab implementation https://ece.uwaterloo.ca/~z70wang/research/iwssim/
+    output = torch.prod(pow1[:-1]) * pow2[-1]
+    return output
+
+
+# Classes to re-use window
+class SSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, val_range=None):
+        super(SSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.val_range = val_range
+
+        # Assume 1 channel for SSIM
+        self.channel = 1
+        self.window = create_window(window_size)
+
+    def forward(self, img1, img2):
+        (_, channel, _, _) = img1.size()
+
+        if channel == self.channel and self.window.dtype == img1.dtype:
+            window = self.window
+        else:
+            window = (
+                create_window(self.window_size, channel)
+                .to(img1.device)
+                .type(img1.dtype)
+            )
+            self.window = window
+            self.channel = channel
+
+        return ssim(
+            img1,
+            img2,
+            window=window,
+            window_size=self.window_size,
+            size_average=self.size_average,
+        )
+
+
+class MSSSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, channel=3):
+        super(MSSSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.channel = channel
+
+    def forward(self, img1, img2):
+        return msssim(
+            img1, img2, window_size=self.window_size, size_average=self.size_average
+        )

criteria/resnet.py

@@ -0,0 +1,100 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.model_zoo as modelzoo
+
+import os
+
+
+# from modules.bn import InPlaceABNSync as BatchNorm2d
+
+resnet18_url = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
+                     padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    def __init__(self, in_chan, out_chan, stride=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(in_chan, out_chan, stride)
+        self.bn1 = nn.BatchNorm2d(out_chan)
+        self.conv2 = conv3x3(out_chan, out_chan)
+        self.bn2 = nn.BatchNorm2d(out_chan)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        if in_chan != out_chan or stride != 1:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_chan, out_chan,
+                          kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(out_chan),
+                )
+
+    def forward(self, x):
+        residual = self.conv1(x)
+        residual = F.relu(self.bn1(residual))
+        residual = self.conv2(residual)
+        residual = self.bn2(residual)
+
+        shortcut = x
+        if self.downsample is not None:
+            shortcut = self.downsample(x)
+
+        out = shortcut + residual
+        out = self.relu(out)
+        return out
+
+
+def create_layer_basic(in_chan, out_chan, bnum, stride=1):
+    layers = [BasicBlock(in_chan, out_chan, stride=stride)]
+    for i in range(bnum-1):
+        layers.append(BasicBlock(out_chan, out_chan, stride=1))
+    return nn.Sequential(*layers)
+
+
+class Resnet18(nn.Module):
+    def __init__(self):
+        super(Resnet18, self).__init__()
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
+                               bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = create_layer_basic(64, 64, bnum=2, stride=1)
+        self.layer2 = create_layer_basic(64, 128, bnum=2, stride=2)
+        self.layer3 = create_layer_basic(128, 256, bnum=2, stride=2)
+        self.layer4 = create_layer_basic(256, 512, bnum=2, stride=2)
+        self.init_weight()
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = F.relu(self.bn1(x))
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        feat8 = self.layer2(x) # 1/8
+        feat16 = self.layer3(feat8) # 1/16
+        feat32 = self.layer4(feat16) # 1/32
+        return feat8, feat16, feat32
+
+    def init_weight(self):
+        #state_dict = modelzoo.load_url(resnet18_url)
+        state_dict = torch.load('pretrained_models/resnet18-5c106cde.pth')
+        self_state_dict = self.state_dict()
+        for k, v in state_dict.items():
+            if 'fc' in k: continue
+            self_state_dict.update({k: v})
+        self.load_state_dict(self_state_dict)
+
+    def get_params(self):
+        wd_params, nowd_params = [], []
+        for name, module in self.named_modules():
+            if isinstance(module, (nn.Linear, nn.Conv2d)):
+                wd_params.append(module.weight)
+                if not module.bias is None:
+                    nowd_params.append(module.bias)
+            elif isinstance(module,  nn.BatchNorm2d):
+                nowd_params += list(module.parameters())
+        return wd_params, nowd_params

criteria/w_norm.py

@@ -0,0 +1,13 @@
+import torch
+from torch import nn
+
+
+class WNormLoss(nn.Module):
+    def __init__(self, start_from_latent_avg=True):
+        super(WNormLoss, self).__init__()
+        self.start_from_latent_avg = start_from_latent_avg
+
+    def forward(self, latent, latent_avg=None):
+        if self.start_from_latent_avg:
+            latent = latent - latent_avg
+        return torch.sum(latent.norm(2, dim=(1, 2))) / latent.shape[0]

datasets/datasets.py

@@ -0,0 +1,136 @@
+import os
+from torch.utils.data import Dataset
+from PIL import Image
+from utils import data_utils
+from torchvision import transforms
+
+
+class ImageDataset(Dataset):
+    def __init__(self, root, transform=None):
+        self.paths = sorted(data_utils.make_dataset(root))
+        self.transform = transform
+
+    def __len__(self):
+        return len(self.paths)
+
+    def __getitem__(self, index):
+        path = self.paths[index]
+        image = Image.open(path).convert("RGB")
+
+        if self.transform:
+            image = self.transform(image)
+        return image
+
+
+class CelebaAttributeDataset(Dataset):
+    def __init__(self, images_root, attr, transform=None, attributes_root="", use_attr=True):
+        self.paths = data_utils.make_dataset(images_root)
+        self.transform = transform
+        with open(attributes_root, "r") as f:
+            lines = f.readlines()
+
+        attr_num = -1
+        for i, data_attr in enumerate(lines[1].split(" ")):
+            if data_attr.strip() == attr.strip():
+                attr_num = i
+                break
+        assert attr_num > -1, f"Can not find attribute {attr}"
+
+        filtred_paths = []
+        for path in self.paths:
+            pic_num = int(path.split("/")[-1].replace(".jpg", "").replace(".png", "")) 
+            pic_attrs = lines[pic_num + 2].strip().split(" ")
+            pic_attrs = pic_attrs[2:]
+            if use_attr and pic_attrs[attr_num] == "1" or not use_attr and pic_attrs[attr_num] == "-1":
+                filtred_paths.append(path)
+        self.paths = sorted(filtred_paths)
+
+    def __len__(self):
+        return len(self.paths)
+
+    def __getitem__(self, index):
+        from_path = self.paths[index]
+        image = Image.open(from_path).convert("RGB")
+
+        if self.transform:
+            image = self.transform(image)
+        return image
+
+
+class FIDDataset(Dataset):
+    def __init__(self, files, transforms=None):
+        self.files = files
+        self.transforms = transforms
+
+    def __len__(self):
+        return len(self.files)
+
+    def __getitem__(self, i):
+        file = self.files[i]
+        image = file.convert("RGB")
+
+        if self.transforms is not None:
+            image = self.transforms(image)
+
+        return image
+
+
+class MetricsPathsDataset(Dataset):
+    def __init__(self, root_path, gt_dir=None, transform=None, transform_train=None, return_path=False, ignore=[]):
+        self.pairs = []
+        self.paths = []
+        self.names = []
+
+        for f in os.listdir(root_path):
+            if f not in ignore:
+                self.names.append(f)
+                image_path = os.path.join(root_path, f)
+                gt_path = os.path.join(gt_dir, f)
+                if f.endswith(".jpg") or f.endswith(".png"):
+                    self.pairs.append([image_path, gt_path.replace(".png", ".jpg"), None])
+                    self.paths.append(image_path)
+        self.transform = transform
+        self.transform_train = transform_train
+        self.return_path = return_path
+
+    def __len__(self):
+        return len(self.pairs)
+
+    def __getitem__(self, index):
+        from_path, to_path, _ = self.pairs[index]
+        from_im = Image.open(from_path).convert("RGB")
+        to_im = Image.open(to_path).convert("RGB")
+
+        if self.transform:
+            to_im = self.transform(to_im)
+            from_im = self.transform(from_im)
+
+        if not self.return_path:
+            return from_im, to_im
+        else:
+            return from_im, to_im, self.names[index]
+
+
+class MetricsDataDataset(Dataset):
+    def __init__(
+        self, paths, target_data, fake_data, transform=None, transform_train=None
+    ):
+        self.fake_data = fake_data
+        self.target_data = target_data
+        self.paths = paths
+        self.transform = transform
+        self.transform_train = transform_train
+
+    def __len__(self):
+        return len(self.fake_data)
+
+    def __getitem__(self, index):
+
+        target_im = self.target_data[index]
+        fake_im = self.fake_data[index]
+
+        if self.transform:
+            fake_im = self.transform(fake_im)
+            target_im = self.transform(target_im)
+
+        return target_im, fake_im

datasets/loaders.py

@@ -0,0 +1,36 @@
+from torch.utils.data import DataLoader
+
+
+class InfiniteLoader(DataLoader):
+    def __init__(
+        self,
+        *args,
+        num_workers=0,
+        pin_memory=True,
+        is_infinite = True,
+        **kwargs,
+    ):
+        super().__init__(
+            *args,
+            multiprocessing_context="fork" if num_workers > 0 else None,
+            num_workers=num_workers,
+            pin_memory=pin_memory,
+            **kwargs,
+        )
+        self.dataset_iterator = super().__iter__()
+        self.is_infinite = is_infinite
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        try:
+            x = next(self.dataset_iterator)
+        except StopIteration:
+            self.dataset_iterator = super().__iter__()
+            if self.is_infinite:
+                x = next(self.dataset_iterator)
+            else:
+                raise StopIteration
+
+        return x

datasets/transforms.py

@@ -0,0 +1,55 @@
+from abc import abstractmethod
+import torchvision.transforms as transforms
+from utils.class_registry import ClassRegistry
+
+transforms_registry = ClassRegistry()
+
+
+class TransformsConfig(object):
+    def __init__(self):
+        pass
+
+    @abstractmethod
+    def get_transforms(self):
+        pass
+
+class FaceTransforms(TransformsConfig):
+    def __init__(self):
+        super(FaceTransforms, self).__init__()
+        self.image_size = None
+
+    def get_transforms(self):
+        transforms_dict = {
+            "train": transforms.Compose(
+                [
+                    transforms.Resize(self.image_size),
+                    transforms.RandomHorizontalFlip(0.5),
+                    transforms.ToTensor(),
+                    transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
+                ]
+            ),
+            "test": transforms.Compose(
+                [
+                    transforms.Resize(self.image_size),
+                    transforms.ToTensor(),
+                    transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
+                ]
+            )
+        }
+        return transforms_dict
+
+
+@transforms_registry.add_to_registry(name="face_256")
+class Face256Transforms(FaceTransforms):
+    def __init__(self):
+        super(Face256Transforms, self).__init__()
+        self.image_size = (256, 256)
+
+
+@transforms_registry.add_to_registry(name="face_1024")
+class Face1024Transforms(FaceTransforms):
+    def __init__(self):
+        super(Face1024Transforms, self).__init__()
+        self.image_size = (1024, 1024)
+
+

dnnlib/__init__.py

@@ -0,0 +1,12 @@
+# 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
+
+
+__all__ = ["EasyDict", "make_cache_dir_path"]

dnnlib/tflib/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+from . import autosummary
+from . import network
+from . import optimizer
+from . import tfutil
+from . import custom_ops
+
+from .tfutil import *
+from .network import Network
+
+from .optimizer import Optimizer
+
+from .custom_ops import get_plugin

dnnlib/tflib/autosummary.py

@@ -0,0 +1,191 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helper for adding automatically tracked values to Tensorboard.
+
+Autosummary creates an identity op that internally keeps track of the input
+values and automatically shows up in TensorBoard. The reported value
+represents an average over input components. The average is accumulated
+constantly over time and flushed when save_summaries() is called.
+
+Notes:
+- The output tensor must be used as an input for something else in the
+  graph. Otherwise, the autosummary op will not get executed, and the average
+  value will not get accumulated.
+- It is perfectly fine to include autosummaries with the same name in
+  several places throughout the graph, even if they are executed concurrently.
+- It is ok to also pass in a python scalar or numpy array. In this case, it
+  is added to the average immediately.
+"""
+
+from collections import OrderedDict
+import numpy as np
+import tensorflow as tf
+from tensorboard import summary as summary_lib
+from tensorboard.plugins.custom_scalar import layout_pb2
+
+from . import tfutil
+from .tfutil import TfExpression
+from .tfutil import TfExpressionEx
+
+# Enable "Custom scalars" tab in TensorBoard for advanced formatting.
+# Disabled by default to reduce tfevents file size.
+enable_custom_scalars = False
+
+_dtype = tf.float64
+_vars = OrderedDict()  # name => [var, ...]
+_immediate = OrderedDict()  # name => update_op, update_value
+_finalized = False
+_merge_op = None
+
+
+def _create_var(name: str, value_expr: TfExpression) -> TfExpression:
+    """Internal helper for creating autosummary accumulators."""
+    assert not _finalized
+    name_id = name.replace("/", "_")
+    v = tf.cast(value_expr, _dtype)
+
+    if v.shape.is_fully_defined():
+        size = np.prod(v.shape.as_list())
+        size_expr = tf.constant(size, dtype=_dtype)
+    else:
+        size = None
+        size_expr = tf.reduce_prod(tf.cast(tf.shape(v), _dtype))
+
+    if size == 1:
+        if v.shape.ndims != 0:
+            v = tf.reshape(v, [])
+        v = [size_expr, v, tf.square(v)]
+    else:
+        v = [size_expr, tf.reduce_sum(v), tf.reduce_sum(tf.square(v))]
+    v = tf.cond(tf.is_finite(v[1]), lambda: tf.stack(v), lambda: tf.zeros(3, dtype=_dtype))
+
+    with tfutil.absolute_name_scope("Autosummary/" + name_id), tf.control_dependencies(None):
+        var = tf.Variable(tf.zeros(3, dtype=_dtype), trainable=False)  # [sum(1), sum(x), sum(x**2)]
+    update_op = tf.cond(tf.is_variable_initialized(var), lambda: tf.assign_add(var, v), lambda: tf.assign(var, v))
+
+    if name in _vars:
+        _vars[name].append(var)
+    else:
+        _vars[name] = [var]
+    return update_op
+
+
+def autosummary(name: str, value: TfExpressionEx, passthru: TfExpressionEx = None, condition: TfExpressionEx = True) -> TfExpressionEx:
+    """Create a new autosummary.
+
+    Args:
+        name:     Name to use in TensorBoard
+        value:    TensorFlow expression or python value to track
+        passthru: Optionally return this TF node without modifications but tack an autosummary update side-effect to this node.
+
+    Example use of the passthru mechanism:
+
+    n = autosummary('l2loss', loss, passthru=n)
+
+    This is a shorthand for the following code:
+
+    with tf.control_dependencies([autosummary('l2loss', loss)]):
+        n = tf.identity(n)
+    """
+    tfutil.assert_tf_initialized()
+    name_id = name.replace("/", "_")
+
+    if tfutil.is_tf_expression(value):
+        with tf.name_scope("summary_" + name_id), tf.device(value.device):
+            condition = tf.convert_to_tensor(condition, name='condition')
+            update_op = tf.cond(condition, lambda: tf.group(_create_var(name, value)), tf.no_op)
+            with tf.control_dependencies([update_op]):
+                return tf.identity(value if passthru is None else passthru)
+
+    else:  # python scalar or numpy array
+        assert not tfutil.is_tf_expression(passthru)
+        assert not tfutil.is_tf_expression(condition)
+        if condition:
+            if name not in _immediate:
+                with tfutil.absolute_name_scope("Autosummary/" + name_id), tf.device(None), tf.control_dependencies(None):
+                    update_value = tf.placeholder(_dtype)
+                    update_op = _create_var(name, update_value)
+                    _immediate[name] = update_op, update_value
+            update_op, update_value = _immediate[name]
+            tfutil.run(update_op, {update_value: value})
+        return value if passthru is None else passthru
+
+
+def finalize_autosummaries() -> None:
+    """Create the necessary ops to include autosummaries in TensorBoard report.
+    Note: This should be done only once per graph.
+    """
+    global _finalized
+    tfutil.assert_tf_initialized()
+
+    if _finalized:
+        return None
+
+    _finalized = True
+    tfutil.init_uninitialized_vars([var for vars_list in _vars.values() for var in vars_list])
+
+    # Create summary ops.
+    with tf.device(None), tf.control_dependencies(None):
+        for name, vars_list in _vars.items():
+            name_id = name.replace("/", "_")
+            with tfutil.absolute_name_scope("Autosummary/" + name_id):
+                moments = tf.add_n(vars_list)
+                moments /= moments[0]
+                with tf.control_dependencies([moments]):  # read before resetting
+                    reset_ops = [tf.assign(var, tf.zeros(3, dtype=_dtype)) for var in vars_list]
+                    with tf.name_scope(None), tf.control_dependencies(reset_ops):  # reset before reporting
+                        mean = moments[1]
+                        std = tf.sqrt(moments[2] - tf.square(moments[1]))
+                        tf.summary.scalar(name, mean)
+                        if enable_custom_scalars:
+                            tf.summary.scalar("xCustomScalars/" + name + "/margin_lo", mean - std)
+                            tf.summary.scalar("xCustomScalars/" + name + "/margin_hi", mean + std)
+
+    # Setup layout for custom scalars.
+    layout = None
+    if enable_custom_scalars:
+        cat_dict = OrderedDict()
+        for series_name in sorted(_vars.keys()):
+            p = series_name.split("/")
+            cat = p[0] if len(p) >= 2 else ""
+            chart = "/".join(p[1:-1]) if len(p) >= 3 else p[-1]
+            if cat not in cat_dict:
+                cat_dict[cat] = OrderedDict()
+            if chart not in cat_dict[cat]:
+                cat_dict[cat][chart] = []
+            cat_dict[cat][chart].append(series_name)
+        categories = []
+        for cat_name, chart_dict in cat_dict.items():
+            charts = []
+            for chart_name, series_names in chart_dict.items():
+                series = []
+                for series_name in series_names:
+                    series.append(layout_pb2.MarginChartContent.Series(
+                        value=series_name,
+                        lower="xCustomScalars/" + series_name + "/margin_lo",
+                        upper="xCustomScalars/" + series_name + "/margin_hi"))
+                margin = layout_pb2.MarginChartContent(series=series)
+                charts.append(layout_pb2.Chart(title=chart_name, margin=margin))
+            categories.append(layout_pb2.Category(title=cat_name, chart=charts))
+        layout = summary_lib.custom_scalar_pb(layout_pb2.Layout(category=categories))
+    return layout
+
+def save_summaries(file_writer, global_step=None):
+    """Call FileWriter.add_summary() with all summaries in the default graph,
+    automatically finalizing and merging them on the first call.
+    """
+    global _merge_op
+    tfutil.assert_tf_initialized()
+
+    if _merge_op is None:
+        layout = finalize_autosummaries()
+        if layout is not None:
+            file_writer.add_summary(layout)
+        with tf.device(None), tf.control_dependencies(None):
+            _merge_op = tf.summary.merge_all()
+
+    file_writer.add_summary(_merge_op.eval(), global_step)

dnnlib/tflib/custom_ops.py

@@ -0,0 +1,169 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""TensorFlow custom ops builder.
+"""
+
+import os
+import re
+import uuid
+import hashlib
+import tempfile
+import shutil
+import tensorflow as tf
+from tensorflow.python.client import device_lib # pylint: disable=no-name-in-module
+
+#----------------------------------------------------------------------------
+# Global options.
+
+cuda_cache_path = os.path.join(os.path.dirname(__file__), '_cudacache')
+cuda_cache_version_tag = 'v1'
+do_not_hash_included_headers = False # Speed up compilation by assuming that headers included by the CUDA code never change. Unsafe!
+verbose = True # Print status messages to stdout.
+
+compiler_bindir_search_path = [
+    'C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.14.26428/bin/Hostx64/x64',
+    'C:/Program Files (x86)/Microsoft Visual Studio/2019/Community/VC/Tools/MSVC/14.23.28105/bin/Hostx64/x64',
+    'C:/Program Files (x86)/Microsoft Visual Studio 14.0/vc/bin',
+]
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _find_compiler_bindir():
+    for compiler_path in compiler_bindir_search_path:
+        if os.path.isdir(compiler_path):
+            return compiler_path
+    return None
+
+def _get_compute_cap(device):
+    caps_str = device.physical_device_desc
+    m = re.search('compute capability: (\\d+).(\\d+)', caps_str)
+    major = m.group(1)
+    minor = m.group(2)
+    return (major, minor)
+
+def _get_cuda_gpu_arch_string():
+    gpus = [x for x in device_lib.list_local_devices() if x.device_type == 'GPU']
+    if len(gpus) == 0:
+        raise RuntimeError('No GPU devices found')
+    (major, minor) = _get_compute_cap(gpus[0])
+    return 'sm_%s%s' % (major, minor)
+
+def _run_cmd(cmd):
+    with os.popen(cmd) as pipe:
+        output = pipe.read()
+        status = pipe.close()
+    if status is not None:
+        raise RuntimeError('NVCC returned an error. See below for full command line and output log:\n\n%s\n\n%s' % (cmd, output))
+
+def _prepare_nvcc_cli(opts):
+    cmd = 'nvcc ' + opts.strip()
+    cmd += ' --disable-warnings'
+    cmd += ' --include-path "%s"' % tf.sysconfig.get_include()
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'protobuf_archive', 'src')
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'com_google_absl')
+    cmd += ' --include-path "%s"' % os.path.join(tf.sysconfig.get_include(), 'external', 'eigen_archive')
+
+    compiler_bindir = _find_compiler_bindir()
+    if compiler_bindir is None:
+        # Require that _find_compiler_bindir succeeds on Windows.  Allow
+        # nvcc to use whatever is the default on Linux.
+        if os.name == 'nt':
+            raise RuntimeError('Could not find MSVC/GCC/CLANG installation on this computer. Check compiler_bindir_search_path list in "%s".' % __file__)
+    else:
+        cmd += ' --compiler-bindir "%s"' % compiler_bindir
+    cmd += ' 2>&1'
+    return cmd
+
+#----------------------------------------------------------------------------
+# Main entry point.
+
+_plugin_cache = dict()
+
+def get_plugin(cuda_file):
+    cuda_file_base = os.path.basename(cuda_file)
+    cuda_file_name, cuda_file_ext = os.path.splitext(cuda_file_base)
+
+    # Already in cache?
+    if cuda_file in _plugin_cache:
+        return _plugin_cache[cuda_file]
+
+    # Setup plugin.
+    if verbose:
+        print('Setting up TensorFlow plugin "%s": ' % cuda_file_base, end='', flush=True)
+    try:
+        # Hash CUDA source.
+        md5 = hashlib.md5()
+        with open(cuda_file, 'rb') as f:
+            md5.update(f.read())
+        md5.update(b'\n')
+
+        # Hash headers included by the CUDA code by running it through the preprocessor.
+        if not do_not_hash_included_headers:
+            if verbose:
+                print('Preprocessing... ', end='', flush=True)
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                tmp_file = os.path.join(tmp_dir, cuda_file_name + '_tmp' + cuda_file_ext)
+                _run_cmd(_prepare_nvcc_cli('"%s" --preprocess -o "%s" --keep --keep-dir "%s"' % (cuda_file, tmp_file, tmp_dir)))
+                with open(tmp_file, 'rb') as f:
+                    bad_file_str = ('"' + cuda_file.replace('\\', '/') + '"').encode('utf-8') # __FILE__ in error check macros
+                    good_file_str = ('"' + cuda_file_base + '"').encode('utf-8')
+                    for ln in f:
+                        if not ln.startswith(b'# ') and not ln.startswith(b'#line '): # ignore line number pragmas
+                            ln = ln.replace(bad_file_str, good_file_str)
+                            md5.update(ln)
+                    md5.update(b'\n')
+
+        # Select compiler options.
+        compile_opts = ''
+        if os.name == 'nt':
+            compile_opts += '"%s"' % os.path.join(tf.sysconfig.get_lib(), 'python', '_pywrap_tensorflow_internal.lib')
+        elif os.name == 'posix':
+            compile_opts += '"%s"' % os.path.join(tf.sysconfig.get_lib(), 'python', '_pywrap_tensorflow_internal.so')
+            compile_opts += ' --compiler-options \'-fPIC -D_GLIBCXX_USE_CXX11_ABI=0\''
+        else:
+            assert False # not Windows or Linux, w00t?
+        compile_opts += ' --gpu-architecture=%s' % _get_cuda_gpu_arch_string()
+        compile_opts += ' --use_fast_math'
+        nvcc_cmd = _prepare_nvcc_cli(compile_opts)
+
+        # Hash build configuration.
+        md5.update(('nvcc_cmd: ' + nvcc_cmd).encode('utf-8') + b'\n')
+        md5.update(('tf.VERSION: ' + tf.VERSION).encode('utf-8') + b'\n')
+        md5.update(('cuda_cache_version_tag: ' + cuda_cache_version_tag).encode('utf-8') + b'\n')
+
+        # Compile if not already compiled.
+        bin_file_ext = '.dll' if os.name == 'nt' else '.so'
+        bin_file = os.path.join(cuda_cache_path, cuda_file_name + '_' + md5.hexdigest() + bin_file_ext)
+        if not os.path.isfile(bin_file):
+            if verbose:
+                print('Compiling... ', end='', flush=True)
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                tmp_file = os.path.join(tmp_dir, cuda_file_name + '_tmp' + bin_file_ext)
+                _run_cmd(nvcc_cmd + ' "%s" --shared -o "%s" --keep --keep-dir "%s"' % (cuda_file, tmp_file, tmp_dir))
+                os.makedirs(cuda_cache_path, exist_ok=True)
+                intermediate_file = os.path.join(cuda_cache_path, cuda_file_name + '_' + uuid.uuid4().hex + '_tmp' + bin_file_ext)
+                shutil.copyfile(tmp_file, intermediate_file)
+                os.rename(intermediate_file, bin_file) # atomic
+
+        # Load.
+        if verbose:
+            print('Loading... ', end='', flush=True)
+        plugin = tf.load_op_library(bin_file)
+
+        # Add to cache.
+        _plugin_cache[cuda_file] = plugin
+        if verbose:
+            print('Done.', flush=True)
+        return plugin
+
+    except:
+        if verbose:
+            print('Failed!', flush=True)
+        raise
+
+#----------------------------------------------------------------------------

dnnlib/tflib/network.py

@@ -0,0 +1,590 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helper for managing networks."""
+
+import types
+import inspect
+import re
+import uuid
+import sys
+import numpy as np
+import tensorflow as tf
+
+from collections import OrderedDict
+from typing import Any, List, Tuple, Union
+
+from . import tfutil
+from .. import util
+
+from .tfutil import TfExpression, TfExpressionEx
+
+_import_handlers = []  # Custom import handlers for dealing with legacy data in pickle import.
+_import_module_src = dict()  # Source code for temporary modules created during pickle import.
+
+
+def import_handler(handler_func):
+    """Function decorator for declaring custom import handlers."""
+    _import_handlers.append(handler_func)
+    return handler_func
+
+
+class Network:
+    """Generic network abstraction.
+
+    Acts as a convenience wrapper for a parameterized network construction
+    function, providing several utility methods and convenient access to
+    the inputs/outputs/weights.
+
+    Network objects can be safely pickled and unpickled for long-term
+    archival purposes. The pickling works reliably as long as the underlying
+    network construction function is defined in a standalone Python module
+    that has no side effects or application-specific imports.
+
+    Args:
+        name: Network name. Used to select TensorFlow name and variable scopes.
+        func_name: Fully qualified name of the underlying network construction function, or a top-level function object.
+        static_kwargs: Keyword arguments to be passed in to the network construction function.
+
+    Attributes:
+        name: User-specified name, defaults to build func name if None.
+        scope: Unique TensorFlow scope containing template graph and variables, derived from the user-specified name.
+        static_kwargs: Arguments passed to the user-supplied build func.
+        components: Container for sub-networks. Passed to the build func, and retained between calls.
+        num_inputs: Number of input tensors.
+        num_outputs: Number of output tensors.
+        input_shapes: Input tensor shapes (NC or NCHW), including minibatch dimension.
+        output_shapes: Output tensor shapes (NC or NCHW), including minibatch dimension.
+        input_shape: Short-hand for input_shapes[0].
+        output_shape: Short-hand for output_shapes[0].
+        input_templates: Input placeholders in the template graph.
+        output_templates: Output tensors in the template graph.
+        input_names: Name string for each input.
+        output_names: Name string for each output.
+        own_vars: Variables defined by this network (local_name => var), excluding sub-networks.
+        vars: All variables (local_name => var).
+        trainables: All trainable variables (local_name => var).
+        var_global_to_local: Mapping from variable global names to local names.
+    """
+
+    def __init__(self, name: str = None, func_name: Any = None, **static_kwargs):
+        tfutil.assert_tf_initialized()
+        assert isinstance(name, str) or name is None
+        assert func_name is not None
+        assert isinstance(func_name, str) or util.is_top_level_function(func_name)
+        assert util.is_pickleable(static_kwargs)
+
+        self._init_fields()
+        self.name = name
+        self.static_kwargs = util.EasyDict(static_kwargs)
+
+        # Locate the user-specified network build function.
+        if util.is_top_level_function(func_name):
+            func_name = util.get_top_level_function_name(func_name)
+        module, self._build_func_name = util.get_module_from_obj_name(func_name)
+        self._build_func = util.get_obj_from_module(module, self._build_func_name)
+        assert callable(self._build_func)
+
+        # Dig up source code for the module containing the build function.
+        self._build_module_src = _import_module_src.get(module, None)
+        if self._build_module_src is None:
+            self._build_module_src = inspect.getsource(module)
+
+        # Init TensorFlow graph.
+        self._init_graph()
+        self.reset_own_vars()
+
+    def _init_fields(self) -> None:
+        self.name = None
+        self.scope = None
+        self.static_kwargs = util.EasyDict()
+        self.components = util.EasyDict()
+        self.num_inputs = 0
+        self.num_outputs = 0
+        self.input_shapes = [[]]
+        self.output_shapes = [[]]
+        self.input_shape = []
+        self.output_shape = []
+        self.input_templates = []
+        self.output_templates = []
+        self.input_names = []
+        self.output_names = []
+        self.own_vars = OrderedDict()
+        self.vars = OrderedDict()
+        self.trainables = OrderedDict()
+        self.var_global_to_local = OrderedDict()
+
+        self._build_func = None  # User-supplied build function that constructs the network.
+        self._build_func_name = None  # Name of the build function.
+        self._build_module_src = None  # Full source code of the module containing the build function.
+        self._run_cache = dict()  # Cached graph data for Network.run().
+
+    def _init_graph(self) -> None:
+        # Collect inputs.
+        self.input_names = []
+
+        for param in inspect.signature(self._build_func).parameters.values():
+            if param.kind == param.POSITIONAL_OR_KEYWORD and param.default is param.empty:
+                self.input_names.append(param.name)
+
+        self.num_inputs = len(self.input_names)
+        assert self.num_inputs >= 1
+
+        # Choose name and scope.
+        if self.name is None:
+            self.name = self._build_func_name
+        assert re.match("^[A-Za-z0-9_.\\-]*$", self.name)
+        with tf.name_scope(None):
+            self.scope = tf.get_default_graph().unique_name(self.name, mark_as_used=True)
+
+        # Finalize build func kwargs.
+        build_kwargs = dict(self.static_kwargs)
+        build_kwargs["is_template_graph"] = True
+        build_kwargs["components"] = self.components
+
+        # Build template graph.
+        with tfutil.absolute_variable_scope(self.scope, reuse=False), tfutil.absolute_name_scope(self.scope):  # ignore surrounding scopes
+            assert tf.get_variable_scope().name == self.scope
+            assert tf.get_default_graph().get_name_scope() == self.scope
+            with tf.control_dependencies(None):  # ignore surrounding control dependencies
+                self.input_templates = [tf.placeholder(tf.float32, name=name) for name in self.input_names]
+                out_expr = self._build_func(*self.input_templates, **build_kwargs)
+
+        # Collect outputs.
+        assert tfutil.is_tf_expression(out_expr) or isinstance(out_expr, tuple)
+        self.output_templates = [out_expr] if tfutil.is_tf_expression(out_expr) else list(out_expr)
+        self.num_outputs = len(self.output_templates)
+        assert self.num_outputs >= 1
+        assert all(tfutil.is_tf_expression(t) for t in self.output_templates)
+
+        # Perform sanity checks.
+        if any(t.shape.ndims is None for t in self.input_templates):
+            raise ValueError("Network input shapes not defined. Please call x.set_shape() for each input.")
+        if any(t.shape.ndims is None for t in self.output_templates):
+            raise ValueError("Network output shapes not defined. Please call x.set_shape() where applicable.")
+        if any(not isinstance(comp, Network) for comp in self.components.values()):
+            raise ValueError("Components of a Network must be Networks themselves.")
+        if len(self.components) != len(set(comp.name for comp in self.components.values())):
+            raise ValueError("Components of a Network must have unique names.")
+
+        # List inputs and outputs.
+        self.input_shapes = [t.shape.as_list() for t in self.input_templates]
+        self.output_shapes = [t.shape.as_list() for t in self.output_templates]
+        self.input_shape = self.input_shapes[0]
+        self.output_shape = self.output_shapes[0]
+        self.output_names = [t.name.split("/")[-1].split(":")[0] for t in self.output_templates]
+
+        # List variables.
+        self.own_vars = OrderedDict((var.name[len(self.scope) + 1:].split(":")[0], var) for var in tf.global_variables(self.scope + "/"))
+        self.vars = OrderedDict(self.own_vars)
+        self.vars.update((comp.name + "/" + name, var) for comp in self.components.values() for name, var in comp.vars.items())
+        self.trainables = OrderedDict((name, var) for name, var in self.vars.items() if var.trainable)
+        self.var_global_to_local = OrderedDict((var.name.split(":")[0], name) for name, var in self.vars.items())
+
+    def reset_own_vars(self) -> None:
+        """Re-initialize all variables of this network, excluding sub-networks."""
+        tfutil.run([var.initializer for var in self.own_vars.values()])
+
+    def reset_vars(self) -> None:
+        """Re-initialize all variables of this network, including sub-networks."""
+        tfutil.run([var.initializer for var in self.vars.values()])
+
+    def reset_trainables(self) -> None:
+        """Re-initialize all trainable variables of this network, including sub-networks."""
+        tfutil.run([var.initializer for var in self.trainables.values()])
+
+    def get_output_for(self, *in_expr: TfExpression, return_as_list: bool = False, **dynamic_kwargs) -> Union[TfExpression, List[TfExpression]]:
+        """Construct TensorFlow expression(s) for the output(s) of this network, given the input expression(s)."""
+        assert len(in_expr) == self.num_inputs
+        assert not all(expr is None for expr in in_expr)
+
+        # Finalize build func kwargs.
+        build_kwargs = dict(self.static_kwargs)
+        build_kwargs.update(dynamic_kwargs)
+        build_kwargs["is_template_graph"] = False
+        build_kwargs["components"] = self.components
+
+        # Build TensorFlow graph to evaluate the network.
+        with tfutil.absolute_variable_scope(self.scope, reuse=True), tf.name_scope(self.name):
+            assert tf.get_variable_scope().name == self.scope
+            valid_inputs = [expr for expr in in_expr if expr is not None]
+            final_inputs = []
+            for expr, name, shape in zip(in_expr, self.input_names, self.input_shapes):
+                if expr is not None:
+                    expr = tf.identity(expr, name=name)
+                else:
+                    expr = tf.zeros([tf.shape(valid_inputs[0])[0]] + shape[1:], name=name)
+                final_inputs.append(expr)
+            out_expr = self._build_func(*final_inputs, **build_kwargs)
+
+        # Propagate input shapes back to the user-specified expressions.
+        for expr, final in zip(in_expr, final_inputs):
+            if isinstance(expr, tf.Tensor):
+                expr.set_shape(final.shape)
+
+        # Express outputs in the desired format.
+        assert tfutil.is_tf_expression(out_expr) or isinstance(out_expr, tuple)
+        if return_as_list:
+            out_expr = [out_expr] if tfutil.is_tf_expression(out_expr) else list(out_expr)
+        return out_expr
+
+    def get_var_local_name(self, var_or_global_name: Union[TfExpression, str]) -> str:
+        """Get the local name of a given variable, without any surrounding name scopes."""
+        assert tfutil.is_tf_expression(var_or_global_name) or isinstance(var_or_global_name, str)
+        global_name = var_or_global_name if isinstance(var_or_global_name, str) else var_or_global_name.name
+        return self.var_global_to_local[global_name]
+
+    def find_var(self, var_or_local_name: Union[TfExpression, str]) -> TfExpression:
+        """Find variable by local or global name."""
+        assert tfutil.is_tf_expression(var_or_local_name) or isinstance(var_or_local_name, str)
+        return self.vars[var_or_local_name] if isinstance(var_or_local_name, str) else var_or_local_name
+
+    def get_var(self, var_or_local_name: Union[TfExpression, str]) -> np.ndarray:
+        """Get the value of a given variable as NumPy array.
+        Note: This method is very inefficient -- prefer to use tflib.run(list_of_vars) whenever possible."""
+        return self.find_var(var_or_local_name).eval()
+
+    def set_var(self, var_or_local_name: Union[TfExpression, str], new_value: Union[int, float, np.ndarray]) -> None:
+        """Set the value of a given variable based on the given NumPy array.
+        Note: This method is very inefficient -- prefer to use tflib.set_vars() whenever possible."""
+        tfutil.set_vars({self.find_var(var_or_local_name): new_value})
+
+    def __getstate__(self) -> dict:
+        """Pickle export."""
+        state = dict()
+        state["version"]            = 4
+        state["name"]               = self.name
+        state["static_kwargs"]      = dict(self.static_kwargs)
+        state["components"]         = dict(self.components)
+        state["build_module_src"]   = self._build_module_src
+        state["build_func_name"]    = self._build_func_name
+        state["variables"]          = list(zip(self.own_vars.keys(), tfutil.run(list(self.own_vars.values()))))
+        return state
+
+    def __setstate__(self, state: dict) -> None:
+        """Pickle import."""
+        # pylint: disable=attribute-defined-outside-init
+        tfutil.assert_tf_initialized()
+        self._init_fields()
+
+        # Execute custom import handlers.
+        for handler in _import_handlers:
+            state = handler(state)
+
+        # Set basic fields.
+        assert state["version"] in [2, 3, 4]
+        self.name = state["name"]
+        self.static_kwargs = util.EasyDict(state["static_kwargs"])
+        self.components = util.EasyDict(state.get("components", {}))
+        self._build_module_src = state["build_module_src"]
+        self._build_func_name = state["build_func_name"]
+
+        # Create temporary module from the imported source code.
+        module_name = "_tflib_network_import_" + uuid.uuid4().hex
+        module = types.ModuleType(module_name)
+        sys.modules[module_name] = module
+        _import_module_src[module] = self._build_module_src
+        exec(self._build_module_src, module.__dict__) # pylint: disable=exec-used
+
+        # Locate network build function in the temporary module.
+        self._build_func = util.get_obj_from_module(module, self._build_func_name)
+        assert callable(self._build_func)
+
+        # Init TensorFlow graph.
+        self._init_graph()
+        self.reset_own_vars()
+        tfutil.set_vars({self.find_var(name): value for name, value in state["variables"]})
+
+    def clone(self, name: str = None, **new_static_kwargs) -> "Network":
+        """Create a clone of this network with its own copy of the variables."""
+        # pylint: disable=protected-access
+        net = object.__new__(Network)
+        net._init_fields()
+        net.name = name if name is not None else self.name
+        net.static_kwargs = util.EasyDict(self.static_kwargs)
+        net.static_kwargs.update(new_static_kwargs)
+        net._build_module_src = self._build_module_src
+        net._build_func_name = self._build_func_name
+        net._build_func = self._build_func
+        net._init_graph()
+        net.copy_vars_from(self)
+        return net
+
+    def copy_own_vars_from(self, src_net: "Network") -> None:
+        """Copy the values of all variables from the given network, excluding sub-networks."""
+        names = [name for name in self.own_vars.keys() if name in src_net.own_vars]
+        tfutil.set_vars(tfutil.run({self.vars[name]: src_net.vars[name] for name in names}))
+
+    def copy_vars_from(self, src_net: "Network") -> None:
+        """Copy the values of all variables from the given network, including sub-networks."""
+        names = [name for name in self.vars.keys() if name in src_net.vars]
+        tfutil.set_vars(tfutil.run({self.vars[name]: src_net.vars[name] for name in names}))
+
+    def copy_trainables_from(self, src_net: "Network") -> None:
+        """Copy the values of all trainable variables from the given network, including sub-networks."""
+        names = [name for name in self.trainables.keys() if name in src_net.trainables]
+        tfutil.set_vars(tfutil.run({self.vars[name]: src_net.vars[name] for name in names}))
+
+    def convert(self, new_func_name: str, new_name: str = None, **new_static_kwargs) -> "Network":
+        """Create new network with the given parameters, and copy all variables from this network."""
+        if new_name is None:
+            new_name = self.name
+        static_kwargs = dict(self.static_kwargs)
+        static_kwargs.update(new_static_kwargs)
+        net = Network(name=new_name, func_name=new_func_name, **static_kwargs)
+        net.copy_vars_from(self)
+        return net
+
+    def setup_as_moving_average_of(self, src_net: "Network", beta: TfExpressionEx = 0.99, beta_nontrainable: TfExpressionEx = 0.0) -> tf.Operation:
+        """Construct a TensorFlow op that updates the variables of this network
+        to be slightly closer to those of the given network."""
+        with tfutil.absolute_name_scope(self.scope + "/_MovingAvg"):
+            ops = []
+            for name, var in self.vars.items():
+                if name in src_net.vars:
+                    cur_beta = beta if name in self.trainables else beta_nontrainable
+                    new_value = tfutil.lerp(src_net.vars[name], var, cur_beta)
+                    ops.append(var.assign(new_value))
+            return tf.group(*ops)
+
+    def run(self,
+            *in_arrays: Tuple[Union[np.ndarray, None], ...],
+            input_transform: dict = None,
+            output_transform: dict = None,
+            return_as_list: bool = False,
+            print_progress: bool = False,
+            minibatch_size: int = None,
+            num_gpus: int = 1,
+            assume_frozen: bool = False,
+            **dynamic_kwargs) -> Union[np.ndarray, Tuple[np.ndarray, ...], List[np.ndarray]]:
+        """Run this network for the given NumPy array(s), and return the output(s) as NumPy array(s).
+
+        Args:
+            input_transform:    A dict specifying a custom transformation to be applied to the input tensor(s) before evaluating the network.
+                                The dict must contain a 'func' field that points to a top-level function. The function is called with the input
+                                TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
+            output_transform:   A dict specifying a custom transformation to be applied to the output tensor(s) after evaluating the network.
+                                The dict must contain a 'func' field that points to a top-level function. The function is called with the output
+                                TensorFlow expression(s) as positional arguments. Any remaining fields of the dict will be passed in as kwargs.
+            return_as_list:     True = return a list of NumPy arrays, False = return a single NumPy array, or a tuple if there are multiple outputs.
+            print_progress:     Print progress to the console? Useful for very large input arrays.
+            minibatch_size:     Maximum minibatch size to use, None = disable batching.
+            num_gpus:           Number of GPUs to use.
+            assume_frozen:      Improve multi-GPU performance by assuming that the trainable parameters will remain changed between calls.
+            dynamic_kwargs:     Additional keyword arguments to be passed into the network build function.
+        """
+        assert len(in_arrays) == self.num_inputs
+        assert not all(arr is None for arr in in_arrays)
+        assert input_transform is None or util.is_top_level_function(input_transform["func"])
+        assert output_transform is None or util.is_top_level_function(output_transform["func"])
+        output_transform, dynamic_kwargs = _handle_legacy_output_transforms(output_transform, dynamic_kwargs)
+        num_items = in_arrays[0].shape[0]
+        if minibatch_size is None:
+            minibatch_size = num_items
+
+        # Construct unique hash key from all arguments that affect the TensorFlow graph.
+        key = dict(input_transform=input_transform, output_transform=output_transform, num_gpus=num_gpus, assume_frozen=assume_frozen, dynamic_kwargs=dynamic_kwargs)
+        def unwind_key(obj):
+            if isinstance(obj, dict):
+                return [(key, unwind_key(value)) for key, value in sorted(obj.items())]
+            if callable(obj):
+                return util.get_top_level_function_name(obj)
+            return obj
+        key = repr(unwind_key(key))
+
+        # Build graph.
+        if key not in self._run_cache:
+            with tfutil.absolute_name_scope(self.scope + "/_Run"), tf.control_dependencies(None):
+                with tf.device("/cpu:0"):
+                    in_expr = [tf.placeholder(tf.float32, name=name) for name in self.input_names]
+                    in_split = list(zip(*[tf.split(x, num_gpus) for x in in_expr]))
+
+                out_split = []
+                for gpu in range(num_gpus):
+                    with tf.device("/gpu:%d" % gpu):
+                        net_gpu = self.clone() if assume_frozen else self
+                        in_gpu = in_split[gpu]
+
+                        if input_transform is not None:
+                            in_kwargs = dict(input_transform)
+                            in_gpu = in_kwargs.pop("func")(*in_gpu, **in_kwargs)
+                            in_gpu = [in_gpu] if tfutil.is_tf_expression(in_gpu) else list(in_gpu)
+
+                        assert len(in_gpu) == self.num_inputs
+                        out_gpu = net_gpu.get_output_for(*in_gpu, return_as_list=True, **dynamic_kwargs)
+
+                        if output_transform is not None:
+                            out_kwargs = dict(output_transform)
+                            out_gpu = out_kwargs.pop("func")(*out_gpu, **out_kwargs)
+                            out_gpu = [out_gpu] if tfutil.is_tf_expression(out_gpu) else list(out_gpu)
+
+                        assert len(out_gpu) == self.num_outputs
+                        out_split.append(out_gpu)
+
+                with tf.device("/cpu:0"):
+                    out_expr = [tf.concat(outputs, axis=0) for outputs in zip(*out_split)]
+                    self._run_cache[key] = in_expr, out_expr
+
+        # Run minibatches.
+        in_expr, out_expr = self._run_cache[key]
+        out_arrays = [np.empty([num_items] + expr.shape.as_list()[1:], expr.dtype.name) for expr in out_expr]
+
+        for mb_begin in range(0, num_items, minibatch_size):
+            if print_progress:
+                print("\r%d / %d" % (mb_begin, num_items), end="")
+
+            mb_end = min(mb_begin + minibatch_size, num_items)
+            mb_num = mb_end - mb_begin
+            mb_in = [src[mb_begin : mb_end] if src is not None else np.zeros([mb_num] + shape[1:]) for src, shape in zip(in_arrays, self.input_shapes)]
+            mb_out = tf.get_default_session().run(out_expr, dict(zip(in_expr, mb_in)))
+
+            for dst, src in zip(out_arrays, mb_out):
+                dst[mb_begin: mb_end] = src
+
+        # Done.
+        if print_progress:
+            print("\r%d / %d" % (num_items, num_items))
+
+        if not return_as_list:
+            out_arrays = out_arrays[0] if len(out_arrays) == 1 else tuple(out_arrays)
+        return out_arrays
+
+    def list_ops(self) -> List[TfExpression]:
+        include_prefix = self.scope + "/"
+        exclude_prefix = include_prefix + "_"
+        ops = tf.get_default_graph().get_operations()
+        ops = [op for op in ops if op.name.startswith(include_prefix)]
+        ops = [op for op in ops if not op.name.startswith(exclude_prefix)]
+        return ops
+
+    def list_layers(self) -> List[Tuple[str, TfExpression, List[TfExpression]]]:
+        """Returns a list of (layer_name, output_expr, trainable_vars) tuples corresponding to
+        individual layers of the network. Mainly intended to be used for reporting."""
+        layers = []
+
+        def recurse(scope, parent_ops, parent_vars, level):
+            # Ignore specific patterns.
+            if any(p in scope for p in ["/Shape", "/strided_slice", "/Cast", "/concat", "/Assign"]):
+                return
+
+            # Filter ops and vars by scope.
+            global_prefix = scope + "/"
+            local_prefix = global_prefix[len(self.scope) + 1:]
+            cur_ops = [op for op in parent_ops if op.name.startswith(global_prefix) or op.name == global_prefix[:-1]]
+            cur_vars = [(name, var) for name, var in parent_vars if name.startswith(local_prefix) or name == local_prefix[:-1]]
+            if not cur_ops and not cur_vars:
+                return
+
+            # Filter out all ops related to variables.
+            for var in [op for op in cur_ops if op.type.startswith("Variable")]:
+                var_prefix = var.name + "/"
+                cur_ops = [op for op in cur_ops if not op.name.startswith(var_prefix)]
+
+            # Scope does not contain ops as immediate children => recurse deeper.
+            contains_direct_ops = any("/" not in op.name[len(global_prefix):] and op.type not in ["Identity", "Cast", "Transpose"] for op in cur_ops)
+            if (level == 0 or not contains_direct_ops) and (len(cur_ops) + len(cur_vars)) > 1:
+                visited = set()
+                for rel_name in [op.name[len(global_prefix):] for op in cur_ops] + [name[len(local_prefix):] for name, _var in cur_vars]:
+                    token = rel_name.split("/")[0]
+                    if token not in visited:
+                        recurse(global_prefix + token, cur_ops, cur_vars, level + 1)
+                        visited.add(token)
+                return
+
+            # Report layer.
+            layer_name = scope[len(self.scope) + 1:]
+            layer_output = cur_ops[-1].outputs[0] if cur_ops else cur_vars[-1][1]
+            layer_trainables = [var for _name, var in cur_vars if var.trainable]
+            layers.append((layer_name, layer_output, layer_trainables))
+
+        recurse(self.scope, self.list_ops(), list(self.vars.items()), 0)
+        return layers
+
+    def print_layers(self, title: str = None, hide_layers_with_no_params: bool = False) -> None:
+        """Print a summary table of the network structure."""
+        rows = [[title if title is not None else self.name, "Params", "OutputShape", "WeightShape"]]
+        rows += [["---"] * 4]
+        total_params = 0
+
+        for layer_name, layer_output, layer_trainables in self.list_layers():
+            num_params = sum(int(np.prod(var.shape.as_list())) for var in layer_trainables)
+            weights = [var for var in layer_trainables if var.name.endswith("/weight:0")]
+            weights.sort(key=lambda x: len(x.name))
+            if len(weights) == 0 and len(layer_trainables) == 1:
+                weights = layer_trainables
+            total_params += num_params
+
+            if not hide_layers_with_no_params or num_params != 0:
+                num_params_str = str(num_params) if num_params > 0 else "-"
+                output_shape_str = str(layer_output.shape)
+                weight_shape_str = str(weights[0].shape) if len(weights) >= 1 else "-"
+                rows += [[layer_name, num_params_str, output_shape_str, weight_shape_str]]
+
+        rows += [["---"] * 4]
+        rows += [["Total", str(total_params), "", ""]]
+
+        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()
+
+    def setup_weight_histograms(self, title: str = None) -> None:
+        """Construct summary ops to include histograms of all trainable parameters in TensorBoard."""
+        if title is None:
+            title = self.name
+
+        with tf.name_scope(None), tf.device(None), tf.control_dependencies(None):
+            for local_name, var in self.trainables.items():
+                if "/" in local_name:
+                    p = local_name.split("/")
+                    name = title + "_" + p[-1] + "/" + "_".join(p[:-1])
+                else:
+                    name = title + "_toplevel/" + local_name
+
+                tf.summary.histogram(name, var)
+
+#----------------------------------------------------------------------------
+# Backwards-compatible emulation of legacy output transformation in Network.run().
+
+_print_legacy_warning = True
+
+def _handle_legacy_output_transforms(output_transform, dynamic_kwargs):
+    global _print_legacy_warning
+    legacy_kwargs = ["out_mul", "out_add", "out_shrink", "out_dtype"]
+    if not any(kwarg in dynamic_kwargs for kwarg in legacy_kwargs):
+        return output_transform, dynamic_kwargs
+
+    if _print_legacy_warning:
+        _print_legacy_warning = False
+        print()
+        print("WARNING: Old-style output transformations in Network.run() are deprecated.")
+        print("Consider using 'output_transform=dict(func=tflib.convert_images_to_uint8)'")
+        print("instead of 'out_mul=127.5, out_add=127.5, out_dtype=np.uint8'.")
+        print()
+    assert output_transform is None
+
+    new_kwargs = dict(dynamic_kwargs)
+    new_transform = {kwarg: new_kwargs.pop(kwarg) for kwarg in legacy_kwargs if kwarg in dynamic_kwargs}
+    new_transform["func"] = _legacy_output_transform_func
+    return new_transform, new_kwargs
+
+def _legacy_output_transform_func(*expr, out_mul=1.0, out_add=0.0, out_shrink=1, out_dtype=None):
+    if out_mul != 1.0:
+        expr = [x * out_mul for x in expr]
+
+    if out_add != 0.0:
+        expr = [x + out_add for x in expr]
+
+    if out_shrink > 1:
+        ksize = [1, 1, out_shrink, out_shrink]
+        expr = [tf.nn.avg_pool(x, ksize=ksize, strides=ksize, padding="VALID", data_format="NCHW") for x in expr]
+
+    if out_dtype is not None:
+        if tf.as_dtype(out_dtype).is_integer:
+            expr = [tf.round(x) for x in expr]
+        expr = [tf.saturate_cast(x, out_dtype) for x in expr]
+    return expr

dnnlib/tflib/ops/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+# empty

dnnlib/tflib/ops/fused_bias_act.cu

@@ -0,0 +1,188 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#define EIGEN_USE_GPU
+#define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/shape_inference.h"
+#include <stdio.h>
+
+using namespace tensorflow;
+using namespace tensorflow::shape_inference;
+
+#define OP_CHECK_CUDA_ERROR(CTX, CUDA_CALL) do { cudaError_t err = CUDA_CALL; OP_REQUIRES(CTX, err == cudaSuccess, errors::Internal(cudaGetErrorName(err))); } while (false)
+
+//------------------------------------------------------------------------
+// CUDA kernel.
+
+template <class T>
+struct FusedBiasActKernelParams
+{
+    const T*    x;      // [sizeX]
+    const T*    b;      // [sizeB] or NULL
+    const T*    ref;    // [sizeX] or NULL
+    T*          y;      // [sizeX]
+
+    int         grad;
+    int         axis;
+    int         act;
+    float       alpha;
+    float       gain;
+
+    int         sizeX;
+    int         sizeB;
+    int         stepB;
+    int         loopX;
+};
+
+template <class T>
+static __global__ void FusedBiasActKernel(const FusedBiasActKernelParams<T> p)
+{
+    const float expRange        = 80.0f;
+    const float halfExpRange    = 40.0f;
+    const float seluScale       = 1.0507009873554804934193349852946f;
+    const float seluAlpha       = 1.6732632423543772848170429916717f;
+
+    // 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 and apply bias.
+        float x = (float)p.x[xi];
+        if (p.b)
+            x += (float)p.b[(xi / p.stepB) % p.sizeB];
+        float ref = (p.ref) ? (float)p.ref[xi] : 0.0f;
+        if (p.gain != 0.0f & p.act != 9)
+            ref /= p.gain;
+
+        // Evaluate activation func.
+        float y;
+        switch (p.act * 10 + p.grad)
+        {
+            // linear
+            default:
+            case 10: y = x; break;
+            case 11: y = x; break;
+            case 12: y = 0.0f; break;
+
+            // relu
+            case 20: y = (x > 0.0f) ? x : 0.0f; break;
+            case 21: y = (ref > 0.0f) ? x : 0.0f; break;
+            case 22: y = 0.0f; break;
+
+            // lrelu
+            case 30: y = (x > 0.0f) ? x : x * p.alpha; break;
+            case 31: y = (ref > 0.0f) ? x : x * p.alpha; break;
+            case 32: y = 0.0f; break;
+
+            // tanh
+            case 40: { float c = expf(x); float d = 1.0f / c; y = (x < -expRange) ? -1.0f : (x > expRange) ? 1.0f : (c - d) / (c + d); } break;
+            case 41: y = x * (1.0f - ref * ref); break;
+            case 42: y = x * (1.0f - ref * ref) * (-2.0f * ref); break;
+
+            // sigmoid
+            case 50: y = (x < -expRange) ? 0.0f : 1.0f / (expf(-x) + 1.0f); break;
+            case 51: y = x * ref * (1.0f - ref); break;
+            case 52: y = x * ref * (1.0f - ref) * (1.0f - 2.0f * ref); break;
+
+            // elu
+            case 60: y = (x >= 0.0f) ? x : expf(x) - 1.0f; break;
+            case 61: y = (ref >= 0.0f) ? x : x * (ref + 1.0f); break;
+            case 62: y = (ref >= 0.0f) ? 0.0f : x * (ref + 1.0f); break;
+
+            // selu
+            case 70: y = (x >= 0.0f) ? seluScale * x : (seluScale * seluAlpha) * (expf(x) - 1.0f); break;
+            case 71: y = (ref >= 0.0f) ? x * seluScale : x * (ref + seluScale * seluAlpha); break;
+            case 72: y = (ref >= 0.0f) ? 0.0f : x * (ref + seluScale * seluAlpha); break;
+
+            // softplus
+            case 80: y = (x > expRange) ? x : logf(expf(x) + 1.0f); break;
+            case 81: y = x * (1.0f - expf(-ref)); break;
+            case 82: { float c = expf(-ref); y = x * c * (1.0f - c); } break;
+
+            // swish
+            case 90: y = (x < -expRange) ? 0.0f : x / (expf(-x) + 1.0f); break;
+            case 91: { float c = expf(ref); float d = c + 1.0f; y = (ref > halfExpRange) ? x : x * c * (ref + d) / (d * d); } break;
+            case 92: { float c = expf(ref); float d = c + 1.0f; y = (ref > halfExpRange) ? 0.0f : x * c * (ref * (2.0f - d) + 2.0f * d) / (d * d * d); } break;
+        }
+
+        // Apply gain and store.
+        p.y[xi] = (T)(y * p.gain);
+    }
+}
+
+//------------------------------------------------------------------------
+// TensorFlow op.
+
+template <class T>
+struct FusedBiasActOp : public OpKernel
+{
+    FusedBiasActKernelParams<T> m_attribs;
+
+    FusedBiasActOp(OpKernelConstruction* ctx) : OpKernel(ctx)
+    {
+        memset(&m_attribs, 0, sizeof(m_attribs));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("grad", &m_attribs.grad));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("axis", &m_attribs.axis));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("act", &m_attribs.act));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha", &m_attribs.alpha));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("gain", &m_attribs.gain));
+        OP_REQUIRES(ctx, m_attribs.grad >= 0, errors::InvalidArgument("grad must be non-negative"));
+        OP_REQUIRES(ctx, m_attribs.axis >= 0, errors::InvalidArgument("axis must be non-negative"));
+        OP_REQUIRES(ctx, m_attribs.act >= 0, errors::InvalidArgument("act must be non-negative"));
+    }
+
+    void Compute(OpKernelContext* ctx)
+    {
+        FusedBiasActKernelParams<T> p = m_attribs;
+        cudaStream_t stream = ctx->eigen_device<Eigen::GpuDevice>().stream();
+
+        const Tensor& x     = ctx->input(0); // [...]
+        const Tensor& b     = ctx->input(1); // [sizeB] or [0]
+        const Tensor& ref   = ctx->input(2); // x.shape or [0]
+        p.x = x.flat<T>().data();
+        p.b = (b.NumElements()) ? b.flat<T>().data() : NULL;
+        p.ref = (ref.NumElements()) ? ref.flat<T>().data() : NULL;
+        OP_REQUIRES(ctx, b.NumElements() == 0 || m_attribs.axis < x.dims(), errors::InvalidArgument("axis out of bounds"));
+        OP_REQUIRES(ctx, b.dims() == 1, errors::InvalidArgument("b must have rank 1"));
+        OP_REQUIRES(ctx, b.NumElements() == 0 || b.NumElements() == x.dim_size(m_attribs.axis), errors::InvalidArgument("b has wrong number of elements"));
+        OP_REQUIRES(ctx, ref.NumElements() == ((p.grad == 0) ? 0 : x.NumElements()), errors::InvalidArgument("ref has wrong number of elements"));
+        OP_REQUIRES(ctx, x.NumElements() <= kint32max, errors::InvalidArgument("x is too large"));
+
+        p.sizeX = (int)x.NumElements();
+        p.sizeB = (int)b.NumElements();
+        p.stepB = 1;
+        for (int i = m_attribs.axis + 1; i < x.dims(); i++)
+            p.stepB *= (int)x.dim_size(i);
+
+        Tensor* y = NULL; // x.shape
+        OP_REQUIRES_OK(ctx, ctx->allocate_output(0, x.shape(), &y));
+        p.y = y->flat<T>().data();
+
+        p.loopX = 4;
+        int blockSize = 4 * 32;
+        int gridSize = (p.sizeX - 1) / (p.loopX * blockSize) + 1;
+        void* args[] = {&p};
+        OP_CHECK_CUDA_ERROR(ctx, cudaLaunchKernel((void*)FusedBiasActKernel<T>, gridSize, blockSize, args, 0, stream));
+    }
+};
+
+REGISTER_OP("FusedBiasAct")
+    .Input      ("x: T")
+    .Input      ("b: T")
+    .Input      ("ref: T")
+    .Output     ("y: T")
+    .Attr       ("T: {float, half}")
+    .Attr       ("grad: int = 0")
+    .Attr       ("axis: int = 1")
+    .Attr       ("act: int = 0")
+    .Attr       ("alpha: float = 0.0")
+    .Attr       ("gain: float = 1.0");
+REGISTER_KERNEL_BUILDER(Name("FusedBiasAct").Device(DEVICE_GPU).TypeConstraint<float>("T"), FusedBiasActOp<float>);
+REGISTER_KERNEL_BUILDER(Name("FusedBiasAct").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"), FusedBiasActOp<Eigen::half>);
+
+//------------------------------------------------------------------------

dnnlib/tflib/ops/fused_bias_act.py

@@ -0,0 +1,196 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Custom TensorFlow ops for efficient bias and activation."""
+
+import os
+import numpy as np
+import tensorflow as tf
+from .. import custom_ops
+from ...util import EasyDict
+
+def _get_plugin():
+    return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
+
+#----------------------------------------------------------------------------
+
+activation_funcs = {
+    'linear':   EasyDict(func=lambda x, **_:        x,                          def_alpha=None, def_gain=1.0,           cuda_idx=1, ref='y', zero_2nd_grad=True),
+    'relu':     EasyDict(func=lambda x, **_:        tf.nn.relu(x),              def_alpha=None, def_gain=np.sqrt(2),    cuda_idx=2, ref='y', zero_2nd_grad=True),
+    'lrelu':    EasyDict(func=lambda x, alpha, **_: tf.nn.leaky_relu(x, alpha), def_alpha=0.2,  def_gain=np.sqrt(2),    cuda_idx=3, ref='y', zero_2nd_grad=True),
+    'tanh':     EasyDict(func=lambda x, **_:        tf.nn.tanh(x),              def_alpha=None, def_gain=1.0,           cuda_idx=4, ref='y', zero_2nd_grad=False),
+    'sigmoid':  EasyDict(func=lambda x, **_:        tf.nn.sigmoid(x),           def_alpha=None, def_gain=1.0,           cuda_idx=5, ref='y', zero_2nd_grad=False),
+    'elu':      EasyDict(func=lambda x, **_:        tf.nn.elu(x),               def_alpha=None, def_gain=1.0,           cuda_idx=6, ref='y', zero_2nd_grad=False),
+    'selu':     EasyDict(func=lambda x, **_:        tf.nn.selu(x),              def_alpha=None, def_gain=1.0,           cuda_idx=7, ref='y', zero_2nd_grad=False),
+    'softplus': EasyDict(func=lambda x, **_:        tf.nn.softplus(x),          def_alpha=None, def_gain=1.0,           cuda_idx=8, ref='y', zero_2nd_grad=False),
+    'swish':    EasyDict(func=lambda x, **_:        tf.nn.sigmoid(x) * x,       def_alpha=None, def_gain=np.sqrt(2),    cuda_idx=9, ref='x', zero_2nd_grad=False),
+}
+
+#----------------------------------------------------------------------------
+
+def fused_bias_act(x, b=None, axis=1, act='linear', alpha=None, gain=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 TensorFlow ops. It supports first and second order gradients,
+    but not third order gradients.
+
+    Args:
+        x:      Input activation tensor. Can have any shape, but if `b` is defined, the
+                dimension corresponding to `axis`, as well as the rank, must be known.
+        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 `axis`.
+        axis:   The dimension in `x` corresponding to the elements of `b`.
+                The value of `axis` 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 `1.0`.
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+
+    impl_dict = {
+        'ref':  _fused_bias_act_ref,
+        'cuda': _fused_bias_act_cuda,
+    }
+    return impl_dict[impl](x=x, b=b, axis=axis, act=act, alpha=alpha, gain=gain)
+
+#----------------------------------------------------------------------------
+
+def _fused_bias_act_ref(x, b, axis, act, alpha, gain):
+    """Slow reference implementation of `fused_bias_act()` using standard TensorFlow ops."""
+
+    # Validate arguments.
+    x = tf.convert_to_tensor(x)
+    b = tf.convert_to_tensor(b) if b is not None else tf.constant([], dtype=x.dtype)
+    act_spec = activation_funcs[act]
+    assert b.shape.rank == 1 and (b.shape[0] == 0 or b.shape[0] == x.shape[axis])
+    assert b.shape[0] == 0 or 0 <= axis < x.shape.rank
+    if alpha is None:
+        alpha = act_spec.def_alpha
+    if gain is None:
+        gain = act_spec.def_gain
+
+    # Add bias.
+    if b.shape[0] != 0:
+        x += tf.reshape(b, [-1 if i == axis else 1 for i in range(x.shape.rank)])
+
+    # Evaluate activation function.
+    x = act_spec.func(x, alpha=alpha)
+
+    # Scale by gain.
+    if gain != 1:
+        x *= gain
+    return x
+
+#----------------------------------------------------------------------------
+
+def _fused_bias_act_cuda(x, b, axis, act, alpha, gain):
+    """Fast CUDA implementation of `fused_bias_act()` using custom ops."""
+
+    # Validate arguments.
+    x = tf.convert_to_tensor(x)
+    empty_tensor = tf.constant([], dtype=x.dtype)
+    b = tf.convert_to_tensor(b) if b is not None else empty_tensor
+    act_spec = activation_funcs[act]
+    assert b.shape.rank == 1 and (b.shape[0] == 0 or b.shape[0] == x.shape[axis])
+    assert b.shape[0] == 0 or 0 <= axis < x.shape.rank
+    if alpha is None:
+        alpha = act_spec.def_alpha
+    if gain is None:
+        gain = act_spec.def_gain
+
+    # Special cases.
+    if act == 'linear' and b is None and gain == 1.0:
+        return x
+    if act_spec.cuda_idx is None:
+        return _fused_bias_act_ref(x=x, b=b, axis=axis, act=act, alpha=alpha, gain=gain)
+
+    # CUDA kernel.
+    cuda_kernel = _get_plugin().fused_bias_act
+    cuda_kwargs = dict(axis=axis, act=act_spec.cuda_idx, alpha=alpha, gain=gain)
+
+    # Forward pass: y = func(x, b).
+    def func_y(x, b):
+        y = cuda_kernel(x=x, b=b, ref=empty_tensor, grad=0, **cuda_kwargs)
+        y.set_shape(x.shape)
+        return y
+
+    # Backward pass: dx, db = grad(dy, x, y)
+    def grad_dx(dy, x, y):
+        ref = {'x': x, 'y': y}[act_spec.ref]
+        dx = cuda_kernel(x=dy, b=empty_tensor, ref=ref, grad=1, **cuda_kwargs)
+        dx.set_shape(x.shape)
+        return dx
+    def grad_db(dx):
+        if b.shape[0] == 0:
+            return empty_tensor
+        db = dx
+        if axis < x.shape.rank - 1:
+            db = tf.reduce_sum(db, list(range(axis + 1, x.shape.rank)))
+        if axis > 0:
+            db = tf.reduce_sum(db, list(range(axis)))
+        db.set_shape(b.shape)
+        return db
+
+    # Second order gradients: d_dy, d_x = grad2(d_dx, d_db, x, y)
+    def grad2_d_dy(d_dx, d_db, x, y):
+        ref = {'x': x, 'y': y}[act_spec.ref]
+        d_dy = cuda_kernel(x=d_dx, b=d_db, ref=ref, grad=1, **cuda_kwargs)
+        d_dy.set_shape(x.shape)
+        return d_dy
+    def grad2_d_x(d_dx, d_db, x, y):
+        ref = {'x': x, 'y': y}[act_spec.ref]
+        d_x = cuda_kernel(x=d_dx, b=d_db, ref=ref, grad=2, **cuda_kwargs)
+        d_x.set_shape(x.shape)
+        return d_x
+
+    # Fast version for piecewise-linear activation funcs.
+    @tf.custom_gradient
+    def func_zero_2nd_grad(x, b):
+        y = func_y(x, b)
+        @tf.custom_gradient
+        def grad(dy):
+            dx = grad_dx(dy, x, y)
+            db = grad_db(dx)
+            def grad2(d_dx, d_db):
+                d_dy = grad2_d_dy(d_dx, d_db, x, y)
+                return d_dy
+            return (dx, db), grad2
+        return y, grad
+
+    # Slow version for general activation funcs.
+    @tf.custom_gradient
+    def func_nonzero_2nd_grad(x, b):
+        y = func_y(x, b)
+        def grad_wrap(dy):
+            @tf.custom_gradient
+            def grad_impl(dy, x):
+                dx = grad_dx(dy, x, y)
+                db = grad_db(dx)
+                def grad2(d_dx, d_db):
+                    d_dy = grad2_d_dy(d_dx, d_db, x, y)
+                    d_x = grad2_d_x(d_dx, d_db, x, y)
+                    return d_dy, d_x
+                return (dx, db), grad2
+            return grad_impl(dy, x)
+        return y, grad_wrap
+
+    # Which version to use?
+    if act_spec.zero_2nd_grad:
+        return func_zero_2nd_grad(x, b)
+    return func_nonzero_2nd_grad(x, b)
+
+#----------------------------------------------------------------------------

dnnlib/tflib/ops/upfirdn_2d.cu

@@ -0,0 +1,326 @@
+// Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+//
+// This work is made available under the Nvidia Source Code License-NC.
+// To view a copy of this license, visit
+// https://nvlabs.github.io/stylegan2/license.html
+
+#define EIGEN_USE_GPU
+#define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
+#include "tensorflow/core/framework/op.h"
+#include "tensorflow/core/framework/op_kernel.h"
+#include "tensorflow/core/framework/shape_inference.h"
+#include <stdio.h>
+
+using namespace tensorflow;
+using namespace tensorflow::shape_inference;
+
+//------------------------------------------------------------------------
+// Helpers.
+
+#define OP_CHECK_CUDA_ERROR(CTX, CUDA_CALL) do { cudaError_t err = CUDA_CALL; OP_REQUIRES(CTX, err == cudaSuccess, errors::Internal(cudaGetErrorName(err))); } while (false)
+
+static __host__ __device__ __forceinline__ int floorDiv(int a, int b)
+{
+    int c = a / b;
+    if (c * b > a)
+        c--;
+    return c;
+}
+
+//------------------------------------------------------------------------
+// CUDA kernel params.
+
+template <class T>
+struct UpFirDn2DKernelParams
+{
+    const T*    x;          // [majorDim, inH, inW, minorDim]
+    const T*    k;          // [kernelH, kernelW]
+    T*          y;          // [majorDim, outH, outW, minorDim]
+
+    int         upx;
+    int         upy;
+    int         downx;
+    int         downy;
+    int         padx0;
+    int         padx1;
+    int         pady0;
+    int         pady1;
+
+    int         majorDim;
+    int         inH;
+    int         inW;
+    int         minorDim;
+    int         kernelH;
+    int         kernelW;
+    int         outH;
+    int         outW;
+    int         loopMajor;
+    int         loopX;
+};
+
+//------------------------------------------------------------------------
+// General CUDA implementation for large filter kernels.
+
+template <class T>
+static __global__ void UpFirDn2DKernel_large(const UpFirDn2DKernelParams<T> p)
+{
+    // Calculate thread index.
+    int minorIdx = blockIdx.x * blockDim.x + threadIdx.x;
+    int outY = minorIdx / p.minorDim;
+    minorIdx -= outY * p.minorDim;
+    int outXBase = blockIdx.y * p.loopX * blockDim.y + threadIdx.y;
+    int majorIdxBase = blockIdx.z * p.loopMajor;
+    if (outXBase >= p.outW || outY >= p.outH || majorIdxBase >= p.majorDim)
+        return;
+
+    // Setup Y receptive field.
+    int midY = outY * p.downy + p.upy - 1 - p.pady0;
+    int inY = min(max(floorDiv(midY, p.upy), 0), p.inH);
+    int h = min(max(floorDiv(midY + p.kernelH, p.upy), 0), p.inH) - inY;
+    int kernelY = midY + p.kernelH - (inY + 1) * p.upy;
+
+    // Loop over majorDim and outX.
+    for (int loopMajor = 0, majorIdx = majorIdxBase; loopMajor < p.loopMajor && majorIdx < p.majorDim; loopMajor++, majorIdx++)
+    for (int loopX = 0, outX = outXBase; loopX < p.loopX && outX < p.outW; loopX++, outX += blockDim.y)
+    {
+        // Setup X receptive field.
+        int midX = outX * p.downx + p.upx - 1 - p.padx0;
+        int inX = min(max(floorDiv(midX, p.upx), 0), p.inW);
+        int w = min(max(floorDiv(midX + p.kernelW, p.upx), 0), p.inW) - inX;
+        int kernelX = midX + p.kernelW - (inX + 1) * p.upx;
+
+        // Initialize pointers.
+        const T* xp = &p.x[((majorIdx * p.inH + inY) * p.inW + inX) * p.minorDim + minorIdx];
+        const T* kp = &p.k[kernelY * p.kernelW + kernelX];
+        int xpx = p.minorDim;
+        int kpx = -p.upx;
+        int xpy = p.inW * p.minorDim;
+        int kpy = -p.upy * p.kernelW;
+
+        // Inner loop.
+        float v = 0.0f;
+        for (int y = 0; y < h; y++)
+        {
+            for (int x = 0; x < w; x++)
+            {
+                v += (float)(*xp) * (float)(*kp);
+                xp += xpx;
+                kp += kpx;
+            }
+            xp += xpy - w * xpx;
+            kp += kpy - w * kpx;
+        }
+
+        // Store result.
+        p.y[((majorIdx * p.outH + outY) * p.outW + outX) * p.minorDim + minorIdx] = (T)v;
+    }
+}
+
+//------------------------------------------------------------------------
+// Specialized CUDA implementation for small filter kernels.
+
+template <class T, int upx, int upy, int downx, int downy, int kernelW, int kernelH, int tileOutW, int tileOutH>
+static __global__ void UpFirDn2DKernel_small(const UpFirDn2DKernelParams<T> p)
+{
+    //assert(kernelW % upx == 0);
+    //assert(kernelH % upy == 0);
+    const int tileInW = ((tileOutW - 1) * downx + kernelW - 1) / upx + 1;
+    const int tileInH = ((tileOutH - 1) * downy + kernelH - 1) / upy + 1;
+    __shared__ volatile float sk[kernelH][kernelW];
+    __shared__ volatile float sx[tileInH][tileInW];
+
+    // Calculate tile index.
+    int minorIdx = blockIdx.x;
+    int tileOutY = minorIdx / p.minorDim;
+    minorIdx -= tileOutY * p.minorDim;
+    tileOutY *= tileOutH;
+    int tileOutXBase = blockIdx.y * p.loopX * tileOutW;
+    int majorIdxBase = blockIdx.z * p.loopMajor;
+    if (tileOutXBase >= p.outW | tileOutY >= p.outH | majorIdxBase >= p.majorDim)
+        return;
+
+    // Load filter kernel (flipped).
+    for (int tapIdx = threadIdx.x; tapIdx < kernelH * kernelW; tapIdx += blockDim.x)
+    {
+        int ky = tapIdx / kernelW;
+        int kx = tapIdx - ky * kernelW;
+        float v = 0.0f;
+        if (kx < p.kernelW & ky < p.kernelH)
+            v = (float)p.k[(p.kernelH - 1 - ky) * p.kernelW + (p.kernelW - 1 - kx)];
+        sk[ky][kx] = v;
+    }
+
+    // Loop over majorDim and outX.
+    for (int loopMajor = 0, majorIdx = majorIdxBase; loopMajor < p.loopMajor & majorIdx < p.majorDim; loopMajor++, majorIdx++)
+    for (int loopX = 0, tileOutX = tileOutXBase; loopX < p.loopX & tileOutX < p.outW; loopX++, tileOutX += tileOutW)
+    {
+        // Load input pixels.
+        int tileMidX = tileOutX * downx + upx - 1 - p.padx0;
+        int tileMidY = tileOutY * downy + upy - 1 - p.pady0;
+        int tileInX = floorDiv(tileMidX, upx);
+        int tileInY = floorDiv(tileMidY, upy);
+        __syncthreads();
+        for (int inIdx = threadIdx.x; inIdx < tileInH * tileInW; inIdx += blockDim.x)
+        {
+            int relInY = inIdx / tileInW;
+            int relInX = inIdx - relInY * tileInW;
+            int inX = relInX + tileInX;
+            int inY = relInY + tileInY;
+            float v = 0.0f;
+            if (inX >= 0 & inY >= 0 & inX < p.inW & inY < p.inH)
+                v = (float)p.x[((majorIdx * p.inH + inY) * p.inW + inX) * p.minorDim + minorIdx];
+            sx[relInY][relInX] = v;
+        }
+
+        // Loop over output pixels.
+        __syncthreads();
+        for (int outIdx = threadIdx.x; outIdx < tileOutH * tileOutW; outIdx += blockDim.x)
+        {
+            int relOutY = outIdx / tileOutW;
+            int relOutX = outIdx - relOutY * tileOutW;
+            int outX = relOutX + tileOutX;
+            int outY = relOutY + tileOutY;
+
+            // Setup receptive field.
+            int midX = tileMidX + relOutX * downx;
+            int midY = tileMidY + relOutY * downy;
+            int inX = floorDiv(midX, upx);
+            int inY = floorDiv(midY, upy);
+            int relInX = inX - tileInX;
+            int relInY = inY - tileInY;
+            int kernelX = (inX + 1) * upx - midX - 1; // flipped
+            int kernelY = (inY + 1) * upy - midY - 1; // flipped
+
+            // Inner loop.
+            float v = 0.0f;
+            #pragma unroll
+            for (int y = 0; y < kernelH / upy; y++)
+                #pragma unroll
+                for (int x = 0; x < kernelW / upx; x++)
+                    v += sx[relInY + y][relInX + x] * sk[kernelY + y * upy][kernelX + x * upx];
+
+            // Store result.
+            if (outX < p.outW & outY < p.outH)
+                p.y[((majorIdx * p.outH + outY) * p.outW + outX) * p.minorDim + minorIdx] = (T)v;
+        }
+    }
+}
+
+//------------------------------------------------------------------------
+// TensorFlow op.
+
+template <class T>
+struct UpFirDn2DOp : public OpKernel
+{
+    UpFirDn2DKernelParams<T> m_attribs;
+
+    UpFirDn2DOp(OpKernelConstruction* ctx) : OpKernel(ctx)
+    {
+        memset(&m_attribs, 0, sizeof(m_attribs));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("upx", &m_attribs.upx));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("upy", &m_attribs.upy));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("downx", &m_attribs.downx));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("downy", &m_attribs.downy));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("padx0", &m_attribs.padx0));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("padx1", &m_attribs.padx1));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("pady0", &m_attribs.pady0));
+        OP_REQUIRES_OK(ctx, ctx->GetAttr("pady1", &m_attribs.pady1));
+        OP_REQUIRES(ctx, m_attribs.upx >= 1 && m_attribs.upy >= 1, errors::InvalidArgument("upx and upy must be at least 1x1"));
+        OP_REQUIRES(ctx, m_attribs.downx >= 1 && m_attribs.downy >= 1, errors::InvalidArgument("downx and downy must be at least 1x1"));
+    }
+
+    void Compute(OpKernelContext* ctx)
+    {
+        UpFirDn2DKernelParams<T> p = m_attribs;
+        cudaStream_t stream = ctx->eigen_device<Eigen::GpuDevice>().stream();
+
+        const Tensor& x = ctx->input(0); // [majorDim, inH, inW, minorDim]
+        const Tensor& k = ctx->input(1); // [kernelH, kernelW]
+        p.x = x.flat<T>().data();
+        p.k = k.flat<T>().data();
+        OP_REQUIRES(ctx, x.dims() == 4, errors::InvalidArgument("input must have rank 4"));
+        OP_REQUIRES(ctx, k.dims() == 2, errors::InvalidArgument("kernel must have rank 2"));
+        OP_REQUIRES(ctx, x.NumElements() <= kint32max, errors::InvalidArgument("input too large"));
+        OP_REQUIRES(ctx, k.NumElements() <= kint32max, errors::InvalidArgument("kernel too large"));
+
+        p.majorDim  = (int)x.dim_size(0);
+        p.inH       = (int)x.dim_size(1);
+        p.inW       = (int)x.dim_size(2);
+        p.minorDim  = (int)x.dim_size(3);
+        p.kernelH   = (int)k.dim_size(0);
+        p.kernelW   = (int)k.dim_size(1);
+        OP_REQUIRES(ctx, p.kernelW >= 1 && p.kernelH >= 1, errors::InvalidArgument("kernel must be at least 1x1"));
+
+        p.outW = (p.inW * p.upx + p.padx0 + p.padx1 - p.kernelW + p.downx) / p.downx;
+        p.outH = (p.inH * p.upy + p.pady0 + p.pady1 - p.kernelH + p.downy) / p.downy;
+        OP_REQUIRES(ctx, p.outW >= 1 && p.outH >= 1, errors::InvalidArgument("output must be at least 1x1"));
+
+        Tensor* y = NULL; // [majorDim, outH, outW, minorDim]
+        TensorShape ys;
+        ys.AddDim(p.majorDim);
+        ys.AddDim(p.outH);
+        ys.AddDim(p.outW);
+        ys.AddDim(p.minorDim);
+        OP_REQUIRES_OK(ctx, ctx->allocate_output(0, ys, &y));
+        p.y = y->flat<T>().data();
+        OP_REQUIRES(ctx, y->NumElements() <= kint32max, errors::InvalidArgument("output too large"));
+
+        // Choose CUDA kernel to use.
+        void* cudaKernel = (void*)UpFirDn2DKernel_large<T>;
+        int tileOutW = -1;
+        int tileOutH = -1;
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 7 && p.kernelH <= 7) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 7,7, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 6 && p.kernelH <= 6) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 6,6, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 5 && p.kernelH <= 5) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 5,5, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 4 && p.kernelH <= 4) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 4,4, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 1 && p.downy == 1 && p.kernelW <= 3 && p.kernelH <= 3) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 1,1, 3,3, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 8 && p.kernelH <= 8) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 8,8, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 6 && p.kernelH <= 6) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 6,6, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 4 && p.kernelH <= 4) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 4,4, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 2 && p.upy == 2 && p.downx == 1 && p.downy == 1 && p.kernelW <= 2 && p.kernelH <= 2) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 2,2, 1,1, 2,2, 64,16>; tileOutW = 64; tileOutH = 16; }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 8 && p.kernelH <= 8) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 8,8, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 6 && p.kernelH <= 6) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 6,6, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 4 && p.kernelH <= 4) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 4,4, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+        if (p.upx == 1 && p.upy == 1 && p.downx == 2 && p.downy == 2 && p.kernelW <= 2 && p.kernelH <= 2) { cudaKernel = (void*)UpFirDn2DKernel_small<T, 1,1, 2,2, 2,2, 32,8>;  tileOutW = 32; tileOutH = 8;  }
+
+        // Choose launch params.
+        dim3 blockSize;
+        dim3 gridSize;
+        if (tileOutW > 0 && tileOutH > 0) // small
+        {
+            p.loopMajor = (p.majorDim - 1) / 16384 + 1;
+            p.loopX = 1;
+            blockSize = dim3(32 * 8, 1, 1);
+            gridSize = dim3(((p.outH - 1) / tileOutH + 1) * p.minorDim, (p.outW - 1) / (p.loopX * tileOutW) + 1, (p.majorDim - 1) / p.loopMajor + 1);
+        }
+        else // large
+        {
+            p.loopMajor = (p.majorDim - 1) / 16384 + 1;
+            p.loopX = 4;
+            blockSize = dim3(4, 32, 1);
+            gridSize = dim3((p.outH * p.minorDim - 1) / blockSize.x + 1, (p.outW - 1) / (p.loopX * blockSize.y) + 1, (p.majorDim - 1) / p.loopMajor + 1);
+        }
+
+        // Launch CUDA kernel.
+        void* args[] = {&p};
+        OP_CHECK_CUDA_ERROR(ctx, cudaLaunchKernel(cudaKernel, gridSize, blockSize, args, 0, stream));
+    }
+};
+
+REGISTER_OP("UpFirDn2D")
+    .Input      ("x: T")
+    .Input      ("k: T")
+    .Output     ("y: T")
+    .Attr       ("T: {float, half}")
+    .Attr       ("upx: int = 1")
+    .Attr       ("upy: int = 1")
+    .Attr       ("downx: int = 1")
+    .Attr       ("downy: int = 1")
+    .Attr       ("padx0: int = 0")
+    .Attr       ("padx1: int = 0")
+    .Attr       ("pady0: int = 0")
+    .Attr       ("pady1: int = 0");
+REGISTER_KERNEL_BUILDER(Name("UpFirDn2D").Device(DEVICE_GPU).TypeConstraint<float>("T"), UpFirDn2DOp<float>);
+REGISTER_KERNEL_BUILDER(Name("UpFirDn2D").Device(DEVICE_GPU).TypeConstraint<Eigen::half>("T"), UpFirDn2DOp<Eigen::half>);
+
+//------------------------------------------------------------------------

dnnlib/tflib/ops/upfirdn_2d.py

@@ -0,0 +1,364 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Custom TensorFlow ops for efficient resampling of 2D images."""
+
+import os
+import numpy as np
+import tensorflow as tf
+from .. import custom_ops
+
+def _get_plugin():
+    return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
+
+#----------------------------------------------------------------------------
+
+def upfirdn_2d(x, k, upx=1, upy=1, downx=1, downy=1, padx0=0, padx1=0, pady0=0, pady1=0, impl='cuda'):
+    r"""Pad, upsample, FIR filter, and downsample a batch of 2D images.
+
+    Accepts a batch of 2D images of the shape `[majorDim, inH, inW, minorDim]`
+    and performs the following operations for each image, batched across
+    `majorDim` and `minorDim`:
+
+    1. Pad the image with zeros by the specified number of pixels on each side
+       (`padx0`, `padx1`, `pady0`, `pady1`). Specifying a negative value
+       corresponds to cropping the image.
+
+    2. Upsample the image by inserting the zeros after each pixel (`upx`, `upy`).
+
+    3. Convolve the image with the specified 2D FIR filter (`k`), shrinking the
+       image so that the footprint of all output pixels lies within the input image.
+
+    4. Downsample the image by throwing away pixels (`downx`, `downy`).
+
+    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 TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:      Input tensor of the shape `[majorDim, inH, inW, minorDim]`.
+        k:      2D FIR filter of the shape `[firH, firW]`.
+        upx:    Integer upsampling factor along the X-axis (default: 1).
+        upy:    Integer upsampling factor along the Y-axis (default: 1).
+        downx:  Integer downsampling factor along the X-axis (default: 1).
+        downy:  Integer downsampling factor along the Y-axis (default: 1).
+        padx0:  Number of pixels to pad on the left side (default: 0).
+        padx1:  Number of pixels to pad on the right side (default: 0).
+        pady0:  Number of pixels to pad on the top side (default: 0).
+        pady1:  Number of pixels to pad on the bottom side (default: 0).
+        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[majorDim, outH, outW, minorDim]`, and same datatype as `x`.
+    """
+
+    impl_dict = {
+        'ref':  _upfirdn_2d_ref,
+        'cuda': _upfirdn_2d_cuda,
+    }
+    return impl_dict[impl](x=x, k=k, upx=upx, upy=upy, downx=downx, downy=downy, padx0=padx0, padx1=padx1, pady0=pady0, pady1=pady1)
+
+#----------------------------------------------------------------------------
+
+def _upfirdn_2d_ref(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
+    """Slow reference implementation of `upfirdn_2d()` using standard TensorFlow ops."""
+
+    x = tf.convert_to_tensor(x)
+    k = np.asarray(k, dtype=np.float32)
+    assert x.shape.rank == 4
+    inH = x.shape[1].value
+    inW = x.shape[2].value
+    minorDim = _shape(x, 3)
+    kernelH, kernelW = k.shape
+    assert inW >= 1 and inH >= 1
+    assert kernelW >= 1 and kernelH >= 1
+    assert isinstance(upx, int) and isinstance(upy, int)
+    assert isinstance(downx, int) and isinstance(downy, int)
+    assert isinstance(padx0, int) and isinstance(padx1, int)
+    assert isinstance(pady0, int) and isinstance(pady1, int)
+
+    # Upsample (insert zeros).
+    x = tf.reshape(x, [-1, inH, 1, inW, 1, minorDim])
+    x = tf.pad(x, [[0, 0], [0, 0], [0, upy - 1], [0, 0], [0, upx - 1], [0, 0]])
+    x = tf.reshape(x, [-1, inH * upy, inW * upx, minorDim])
+
+    # Pad (crop if negative).
+    x = tf.pad(x, [[0, 0], [max(pady0, 0), max(pady1, 0)], [max(padx0, 0), max(padx1, 0)], [0, 0]])
+    x = x[:, max(-pady0, 0) : x.shape[1].value - max(-pady1, 0), max(-padx0, 0) : x.shape[2].value - max(-padx1, 0), :]
+
+    # Convolve with filter.
+    x = tf.transpose(x, [0, 3, 1, 2])
+    x = tf.reshape(x, [-1, 1, inH * upy + pady0 + pady1, inW * upx + padx0 + padx1])
+    w = tf.constant(k[::-1, ::-1, np.newaxis, np.newaxis], dtype=x.dtype)
+    x = tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='VALID', data_format='NCHW')
+    x = tf.reshape(x, [-1, minorDim, inH * upy + pady0 + pady1 - kernelH + 1, inW * upx + padx0 + padx1 - kernelW + 1])
+    x = tf.transpose(x, [0, 2, 3, 1])
+
+    # Downsample (throw away pixels).
+    return x[:, ::downy, ::downx, :]
+
+#----------------------------------------------------------------------------
+
+def _upfirdn_2d_cuda(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
+    """Fast CUDA implementation of `upfirdn_2d()` using custom ops."""
+
+    x = tf.convert_to_tensor(x)
+    k = np.asarray(k, dtype=np.float32)
+    majorDim, inH, inW, minorDim = x.shape.as_list()
+    kernelH, kernelW = k.shape
+    assert inW >= 1 and inH >= 1
+    assert kernelW >= 1 and kernelH >= 1
+    assert isinstance(upx, int) and isinstance(upy, int)
+    assert isinstance(downx, int) and isinstance(downy, int)
+    assert isinstance(padx0, int) and isinstance(padx1, int)
+    assert isinstance(pady0, int) and isinstance(pady1, int)
+
+    outW = (inW * upx + padx0 + padx1 - kernelW) // downx + 1
+    outH = (inH * upy + pady0 + pady1 - kernelH) // downy + 1
+    assert outW >= 1 and outH >= 1
+
+    kc = tf.constant(k, dtype=x.dtype)
+    gkc = tf.constant(k[::-1, ::-1], dtype=x.dtype)
+    gpadx0 = kernelW - padx0 - 1
+    gpady0 = kernelH - pady0 - 1
+    gpadx1 = inW * upx - outW * downx + padx0 - upx + 1
+    gpady1 = inH * upy - outH * downy + pady0 - upy + 1
+
+    @tf.custom_gradient
+    def func(x):
+        y = _get_plugin().up_fir_dn2d(x=x, k=kc, upx=upx, upy=upy, downx=downx, downy=downy, padx0=padx0, padx1=padx1, pady0=pady0, pady1=pady1)
+        y.set_shape([majorDim, outH, outW, minorDim])
+        @tf.custom_gradient
+        def grad(dy):
+            dx = _get_plugin().up_fir_dn2d(x=dy, k=gkc, upx=downx, upy=downy, downx=upx, downy=upy, padx0=gpadx0, padx1=gpadx1, pady0=gpady0, pady1=gpady1)
+            dx.set_shape([majorDim, inH, inW, minorDim])
+            return dx, func
+        return y, grad
+    return func(x)
+
+#----------------------------------------------------------------------------
+
+def filter_2d(x, k, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Filter a batch of 2D images with the given FIR filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and filters each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the same shape and datatype as `x`.
+    """
+
+    k = _setup_kernel(k) * gain
+    p = k.shape[0] - 1
+    return _simple_upfirdn_2d(x, k, pad0=(p+1)//2, pad1=p//2, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample_2d(x, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Upsample a batch of 2D images with the given filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and upsamples each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero, and the filter is padded with
+    zeros so that its shape is a multiple of the upsampling factor.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to nearest-neighbor
+                      upsampling.
+        factor:       Integer upsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H * factor, W * factor]` or
+        `[N, H * factor, W * factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * (gain * (factor ** 2))
+    p = k.shape[0] - factor
+    return _simple_upfirdn_2d(x, k, up=factor, pad0=(p+1)//2+factor-1, pad1=p//2, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def downsample_2d(x, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Downsample a batch of 2D images with the given filter.
+
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
+    and downsamples each image with the given filter. The filter is normalized so that
+    if the input pixels are constant, they will be scaled by the specified `gain`.
+    Pixels outside the image are assumed to be zero, and the filter is padded with
+    zeros so that its shape is a multiple of the downsampling factor.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to average pooling.
+        factor:       Integer downsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H // factor, W // factor]` or
+        `[N, H // factor, W // factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * gain
+    p = k.shape[0] - factor
+    return _simple_upfirdn_2d(x, k, down=factor, pad0=(p+1)//2, pad1=p//2, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def upsample_conv_2d(x, w, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Fused `upsample_2d()` followed by `tf.nn.conv2d()`.
+
+    Padding is performed only once at the beginning, not between the operations.
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        w:            Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
+                      Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to nearest-neighbor
+                      upsampling.
+        factor:       Integer upsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H * factor, W * factor]` or
+        `[N, H * factor, W * factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+
+    # Check weight shape.
+    w = tf.convert_to_tensor(w)
+    assert w.shape.rank == 4
+    convH = w.shape[0].value
+    convW = w.shape[1].value
+    inC = _shape(w, 2)
+    outC = _shape(w, 3)
+    assert convW == convH
+
+    # Setup filter kernel.
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * (gain * (factor ** 2))
+    p = (k.shape[0] - factor) - (convW - 1)
+
+    # Determine data dimensions.
+    if data_format == 'NCHW':
+        stride = [1, 1, factor, factor]
+        output_shape = [_shape(x, 0), outC, (_shape(x, 2) - 1) * factor + convH, (_shape(x, 3) - 1) * factor + convW]
+        num_groups = _shape(x, 1) // inC
+    else:
+        stride = [1, factor, factor, 1]
+        output_shape = [_shape(x, 0), (_shape(x, 1) - 1) * factor + convH, (_shape(x, 2) - 1) * factor + convW, outC]
+        num_groups = _shape(x, 3) // inC
+
+    # Transpose weights.
+    w = tf.reshape(w, [convH, convW, inC, num_groups, -1])
+    w = tf.transpose(w[::-1, ::-1], [0, 1, 4, 3, 2])
+    w = tf.reshape(w, [convH, convW, -1, num_groups * inC])
+
+    # Execute.
+    x = tf.nn.conv2d_transpose(x, w, output_shape=output_shape, strides=stride, padding='VALID', data_format=data_format)
+    return _simple_upfirdn_2d(x, k, pad0=(p+1)//2+factor-1, pad1=p//2+1, data_format=data_format, impl=impl)
+
+#----------------------------------------------------------------------------
+
+def conv_downsample_2d(x, w, k=None, factor=2, gain=1, data_format='NCHW', impl='cuda'):
+    r"""Fused `tf.nn.conv2d()` followed by `downsample_2d()`.
+
+    Padding is performed only once at the beginning, not between the operations.
+    The fused op is considerably more efficient than performing the same calculation
+    using standard TensorFlow ops. It supports gradients of arbitrary order.
+
+    Args:
+        x:            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        w:            Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
+                      Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
+        k:            FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
+                      The default is `[1] * factor`, which corresponds to average pooling.
+        factor:       Integer downsampling factor (default: 2).
+        gain:         Scaling factor for signal magnitude (default: 1.0).
+        data_format:  `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
+        impl:         Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
+
+    Returns:
+        Tensor of the shape `[N, C, H // factor, W // factor]` or
+        `[N, H // factor, W // factor, C]`, and same datatype as `x`.
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    w = tf.convert_to_tensor(w)
+    convH, convW, _inC, _outC = w.shape.as_list()
+    assert convW == convH
+    if k is None:
+        k = [1] * factor
+    k = _setup_kernel(k) * gain
+    p = (k.shape[0] - factor) + (convW - 1)
+    if data_format == 'NCHW':
+        s = [1, 1, factor, factor]
+    else:
+        s = [1, factor, factor, 1]
+    x = _simple_upfirdn_2d(x, k, pad0=(p+1)//2, pad1=p//2, data_format=data_format, impl=impl)
+    return tf.nn.conv2d(x, w, strides=s, padding='VALID', data_format=data_format)
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _shape(tf_expr, dim_idx):
+    if tf_expr.shape.rank is not None:
+        dim = tf_expr.shape[dim_idx].value
+        if dim is not None:
+            return dim
+    return tf.shape(tf_expr)[dim_idx]
+
+def _setup_kernel(k):
+    k = np.asarray(k, dtype=np.float32)
+    if k.ndim == 1:
+        k = np.outer(k, k)
+    k /= np.sum(k)
+    assert k.ndim == 2
+    assert k.shape[0] == k.shape[1]
+    return k
+
+def _simple_upfirdn_2d(x, k, up=1, down=1, pad0=0, pad1=0, data_format='NCHW', impl='cuda'):
+    assert data_format in ['NCHW', 'NHWC']
+    assert x.shape.rank == 4
+    y = x
+    if data_format == 'NCHW':
+        y = tf.reshape(y, [-1, _shape(y, 2), _shape(y, 3), 1])
+    y = upfirdn_2d(y, k, upx=up, upy=up, downx=down, downy=down, padx0=pad0, padx1=pad1, pady0=pad0, pady1=pad1, impl=impl)
+    if data_format == 'NCHW':
+        y = tf.reshape(y, [-1, _shape(x, 1), _shape(y, 1), _shape(y, 2)])
+    return y
+
+#----------------------------------------------------------------------------

dnnlib/tflib/optimizer.py

@@ -0,0 +1,370 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Helper wrapper for a Tensorflow optimizer."""
+
+import platform
+import numpy as np
+import tensorflow as tf
+
+from collections import OrderedDict
+from typing import List, Union
+
+from . import autosummary
+from . import tfutil
+from .. import util
+
+from .tfutil import TfExpression, TfExpressionEx
+
+_collective_ops_warning_printed = False
+_collective_ops_group_key       = 831766147
+_collective_ops_instance_key    = 436340067
+
+class Optimizer:
+    """A Wrapper for tf.train.Optimizer.
+
+    Automatically takes care of:
+    - Gradient averaging for multi-GPU training.
+    - Gradient accumulation for arbitrarily large minibatches.
+    - Dynamic loss scaling and typecasts for FP16 training.
+    - Ignoring corrupted gradients that contain NaNs/Infs.
+    - Reporting statistics.
+    - Well-chosen default settings.
+    """
+
+    def __init__(self,
+        name:                   str             = "Train",                  # Name string that will appear in TensorFlow graph.
+        tf_optimizer:           str             = "tf.train.AdamOptimizer", # Underlying optimizer class.
+        learning_rate:          TfExpressionEx  = 0.001,                    # Learning rate. Can vary over time.
+        minibatch_multiplier:   TfExpressionEx  = None,                     # Treat N consecutive minibatches as one by accumulating gradients.
+        share:                  "Optimizer"     = None,                     # Share internal state with a previously created optimizer?
+        use_loss_scaling:       bool            = False,                    # Enable dynamic loss scaling for robust mixed-precision training?
+        loss_scaling_init:      float           = 64.0,                     # Log2 of initial loss scaling factor.
+        loss_scaling_inc:       float           = 0.0005,                   # Log2 of per-minibatch loss scaling increment when there is no overflow.
+        loss_scaling_dec:       float           = 1.0,                      # Log2 of per-minibatch loss scaling decrement when there is an overflow.
+        report_mem_usage:       bool            = False,                    # Report fine-grained memory usage statistics in TensorBoard?
+        **kwargs):
+
+        # Public fields.
+        self.name                   = name
+        self.learning_rate          = learning_rate
+        self.minibatch_multiplier   = minibatch_multiplier
+        self.id                     = self.name.replace("/", ".")
+        self.scope                  = tf.get_default_graph().unique_name(self.id)
+        self.optimizer_class        = util.get_obj_by_name(tf_optimizer)
+        self.optimizer_kwargs       = dict(kwargs)
+        self.use_loss_scaling       = use_loss_scaling
+        self.loss_scaling_init      = loss_scaling_init
+        self.loss_scaling_inc       = loss_scaling_inc
+        self.loss_scaling_dec       = loss_scaling_dec
+
+        # Private fields.
+        self._updates_applied       = False
+        self._devices               = OrderedDict() # device_name => EasyDict()
+        self._shared_optimizers     = OrderedDict() # device_name => optimizer_class
+        self._gradient_shapes       = None          # [shape, ...]
+        self._report_mem_usage      = report_mem_usage
+
+        # Validate arguments.
+        assert callable(self.optimizer_class)
+
+        # Share internal state if requested.
+        if share is not None:
+            assert isinstance(share, Optimizer)
+            assert self.optimizer_class is share.optimizer_class
+            assert self.learning_rate is share.learning_rate
+            assert self.optimizer_kwargs == share.optimizer_kwargs
+            self._shared_optimizers = share._shared_optimizers # pylint: disable=protected-access
+
+    def _get_device(self, device_name: str):
+        """Get internal state for the given TensorFlow device."""
+        tfutil.assert_tf_initialized()
+        if device_name in self._devices:
+            return self._devices[device_name]
+
+        # Initialize fields.
+        device = util.EasyDict()
+        device.name             = device_name
+        device.optimizer        = None          # Underlying optimizer:     optimizer_class
+        device.loss_scaling_var = None          # Log2 of loss scaling:     tf.Variable
+        device.grad_raw         = OrderedDict() # Raw gradients:            var => [grad, ...]
+        device.grad_clean       = OrderedDict() # Clean gradients:          var => grad
+        device.grad_acc_vars    = OrderedDict() # Accumulation sums:        var => tf.Variable
+        device.grad_acc_count   = None          # Accumulation counter:     tf.Variable
+        device.grad_acc         = OrderedDict() # Accumulated gradients:    var => grad
+
+        # Setup TensorFlow objects.
+        with tfutil.absolute_name_scope(self.scope + "/Devices"), tf.device(device_name), tf.control_dependencies(None):
+            if device_name not in self._shared_optimizers:
+                optimizer_name = self.scope.replace("/", "_") + "_opt%d" % len(self._shared_optimizers)
+                self._shared_optimizers[device_name] = self.optimizer_class(name=optimizer_name, learning_rate=self.learning_rate, **self.optimizer_kwargs)
+            device.optimizer = self._shared_optimizers[device_name]
+            if self.use_loss_scaling:
+                device.loss_scaling_var = tf.Variable(np.float32(self.loss_scaling_init), trainable=False, name="loss_scaling_var")
+
+        # Register device.
+        self._devices[device_name] = device
+        return device
+
+    def register_gradients(self, loss: TfExpression, trainable_vars: Union[List, dict]) -> None:
+        """Register the gradients of the given loss function with respect to the given variables.
+        Intended to be called once per GPU."""
+        tfutil.assert_tf_initialized()
+        assert not self._updates_applied
+        device = self._get_device(loss.device)
+
+        # Validate trainables.
+        if isinstance(trainable_vars, dict):
+            trainable_vars = list(trainable_vars.values())  # allow passing in Network.trainables as vars
+        assert isinstance(trainable_vars, list) and len(trainable_vars) >= 1
+        assert all(tfutil.is_tf_expression(expr) for expr in trainable_vars + [loss])
+        assert all(var.device == device.name for var in trainable_vars)
+
+        # Validate shapes.
+        if self._gradient_shapes is None:
+            self._gradient_shapes = [var.shape.as_list() for var in trainable_vars]
+        assert len(trainable_vars) == len(self._gradient_shapes)
+        assert all(var.shape.as_list() == var_shape for var, var_shape in zip(trainable_vars, self._gradient_shapes))
+
+        # Report memory usage if requested.
+        deps = []
+        if self._report_mem_usage:
+            self._report_mem_usage = False
+            try:
+                with tf.name_scope(self.id + '_mem'), tf.device(device.name), tf.control_dependencies([loss]):
+                    deps.append(autosummary.autosummary(self.id + "/mem_usage_gb", tf.contrib.memory_stats.BytesInUse() / 2**30))
+            except tf.errors.NotFoundError:
+                pass
+
+        # Compute gradients.
+        with tf.name_scope(self.id + "_grad"), tf.device(device.name), tf.control_dependencies(deps):
+            loss = self.apply_loss_scaling(tf.cast(loss, tf.float32))
+            gate = tf.train.Optimizer.GATE_NONE  # disable gating to reduce memory usage
+            grad_list = device.optimizer.compute_gradients(loss=loss, var_list=trainable_vars, gate_gradients=gate)
+
+        # Register gradients.
+        for grad, var in grad_list:
+            if var not in device.grad_raw:
+                device.grad_raw[var] = []
+            device.grad_raw[var].append(grad)
+
+    def apply_updates(self, allow_no_op: bool = False) -> tf.Operation:
+        """Construct training op to update the registered variables based on their gradients."""
+        tfutil.assert_tf_initialized()
+        assert not self._updates_applied
+        self._updates_applied = True
+        all_ops = []
+
+        # Check for no-op.
+        if allow_no_op and len(self._devices) == 0:
+            with tfutil.absolute_name_scope(self.scope):
+                return tf.no_op(name='TrainingOp')
+
+        # Clean up gradients.
+        for device_idx, device in enumerate(self._devices.values()):
+            with tfutil.absolute_name_scope(self.scope + "/Clean%d" % device_idx), tf.device(device.name):
+                for var, grad in device.grad_raw.items():
+
+                    # Filter out disconnected gradients and convert to float32.
+                    grad = [g for g in grad if g is not None]
+                    grad = [tf.cast(g, tf.float32) for g in grad]
+
+                    # Sum within the device.
+                    if len(grad) == 0:
+                        grad = tf.zeros(var.shape)  # No gradients => zero.
+                    elif len(grad) == 1:
+                        grad = grad[0]              # Single gradient => use as is.
+                    else:
+                        grad = tf.add_n(grad)       # Multiple gradients => sum.
+
+                    # Scale as needed.
+                    scale = 1.0 / len(device.grad_raw[var]) / len(self._devices)
+                    scale = tf.constant(scale, dtype=tf.float32, name="scale")
+                    if self.minibatch_multiplier is not None:
+                        scale /= tf.cast(self.minibatch_multiplier, tf.float32)
+                    scale = self.undo_loss_scaling(scale)
+                    device.grad_clean[var] = grad * scale
+
+        # Sum gradients across devices.
+        if len(self._devices) > 1:
+            with tfutil.absolute_name_scope(self.scope + "/Broadcast"), tf.device(None):
+                if platform.system() == "Windows":    # Windows => NCCL ops are not available.
+                    self._broadcast_fallback()
+                elif tf.VERSION.startswith("1.15."):  # TF 1.15 => NCCL ops are broken: https://github.com/tensorflow/tensorflow/issues/41539
+                    self._broadcast_fallback()
+                else:                                 # Otherwise => NCCL ops are safe to use.
+                    self._broadcast_nccl()
+
+        # Apply updates separately on each device.
+        for device_idx, device in enumerate(self._devices.values()):
+            with tfutil.absolute_name_scope(self.scope + "/Apply%d" % device_idx), tf.device(device.name):
+                # pylint: disable=cell-var-from-loop
+
+                # Accumulate gradients over time.
+                if self.minibatch_multiplier is None:
+                    acc_ok = tf.constant(True, name='acc_ok')
+                    device.grad_acc = OrderedDict(device.grad_clean)
+                else:
+                    # Create variables.
+                    with tf.control_dependencies(None):
+                        for var in device.grad_clean.keys():
+                            device.grad_acc_vars[var] = tf.Variable(tf.zeros(var.shape), trainable=False, name="grad_acc_var")
+                        device.grad_acc_count = tf.Variable(tf.zeros([]), trainable=False, name="grad_acc_count")
+
+                    # Track counter.
+                    count_cur = device.grad_acc_count + 1.0
+                    count_inc_op = lambda: tf.assign(device.grad_acc_count, count_cur)
+                    count_reset_op = lambda: tf.assign(device.grad_acc_count, tf.zeros([]))
+                    acc_ok = (count_cur >= tf.cast(self.minibatch_multiplier, tf.float32))
+                    all_ops.append(tf.cond(acc_ok, count_reset_op, count_inc_op))
+
+                    # Track gradients.
+                    for var, grad in device.grad_clean.items():
+                        acc_var = device.grad_acc_vars[var]
+                        acc_cur = acc_var + grad
+                        device.grad_acc[var] = acc_cur
+                        with tf.control_dependencies([acc_cur]):
+                            acc_inc_op = lambda: tf.assign(acc_var, acc_cur)
+                            acc_reset_op = lambda: tf.assign(acc_var, tf.zeros(var.shape))
+                            all_ops.append(tf.cond(acc_ok, acc_reset_op, acc_inc_op))
+
+                # No overflow => apply gradients.
+                all_ok = tf.reduce_all(tf.stack([acc_ok] + [tf.reduce_all(tf.is_finite(g)) for g in device.grad_acc.values()]))
+                apply_op = lambda: device.optimizer.apply_gradients([(tf.cast(grad, var.dtype), var) for var, grad in device.grad_acc.items()])
+                all_ops.append(tf.cond(all_ok, apply_op, tf.no_op))
+
+                # Adjust loss scaling.
+                if self.use_loss_scaling:
+                    ls_inc_op = lambda: tf.assign_add(device.loss_scaling_var, self.loss_scaling_inc)
+                    ls_dec_op = lambda: tf.assign_sub(device.loss_scaling_var, self.loss_scaling_dec)
+                    ls_update_op = lambda: tf.group(tf.cond(all_ok, ls_inc_op, ls_dec_op))
+                    all_ops.append(tf.cond(acc_ok, ls_update_op, tf.no_op))
+
+                # Last device => report statistics.
+                if device_idx == len(self._devices) - 1:
+                    all_ops.append(autosummary.autosummary(self.id + "/learning_rate", tf.convert_to_tensor(self.learning_rate)))
+                    all_ops.append(autosummary.autosummary(self.id + "/overflow_frequency", tf.where(all_ok, 0, 1), condition=acc_ok))
+                    if self.use_loss_scaling:
+                        all_ops.append(autosummary.autosummary(self.id + "/loss_scaling_log2", device.loss_scaling_var))
+
+        # Initialize variables.
+        self.reset_optimizer_state()
+        if self.use_loss_scaling:
+            tfutil.init_uninitialized_vars([device.loss_scaling_var for device in self._devices.values()])
+        if self.minibatch_multiplier is not None:
+            tfutil.run([var.initializer for device in self._devices.values() for var in list(device.grad_acc_vars.values()) + [device.grad_acc_count]])
+
+        # Group everything into a single op.
+        with tfutil.absolute_name_scope(self.scope):
+            return tf.group(*all_ops, name="TrainingOp")
+
+    def reset_optimizer_state(self) -> None:
+        """Reset internal state of the underlying optimizer."""
+        tfutil.assert_tf_initialized()
+        tfutil.run([var.initializer for device in self._devices.values() for var in device.optimizer.variables()])
+
+    def get_loss_scaling_var(self, device: str) -> Union[tf.Variable, None]:
+        """Get or create variable representing log2 of the current dynamic loss scaling factor."""
+        return self._get_device(device).loss_scaling_var
+
+    def apply_loss_scaling(self, value: TfExpression) -> TfExpression:
+        """Apply dynamic loss scaling for the given expression."""
+        assert tfutil.is_tf_expression(value)
+        if not self.use_loss_scaling:
+            return value
+        return value * tfutil.exp2(self.get_loss_scaling_var(value.device))
+
+    def undo_loss_scaling(self, value: TfExpression) -> TfExpression:
+        """Undo the effect of dynamic loss scaling for the given expression."""
+        assert tfutil.is_tf_expression(value)
+        if not self.use_loss_scaling:
+            return value
+        return value * tfutil.exp2(-self.get_loss_scaling_var(value.device)) # pylint: disable=invalid-unary-operand-type
+
+    def _broadcast_nccl(self):
+        """Sum gradients across devices using NCCL ops (fast path)."""
+        from tensorflow.python.ops import nccl_ops # pylint: disable=no-name-in-module
+        for all_vars in zip(*[device.grad_clean.keys() for device in self._devices.values()]):
+            if any(x.shape.num_elements() > 0 for x in all_vars):
+                all_grads = [device.grad_clean[var] for device, var in zip(self._devices.values(), all_vars)]
+                all_grads = nccl_ops.all_sum(all_grads)
+                for device, var, grad in zip(self._devices.values(), all_vars, all_grads):
+                    device.grad_clean[var] = grad
+
+    def _broadcast_fallback(self):
+        """Sum gradients across devices using TensorFlow collective ops (slow fallback path)."""
+        from tensorflow.python.ops import collective_ops # pylint: disable=no-name-in-module
+        global _collective_ops_warning_printed, _collective_ops_group_key, _collective_ops_instance_key
+        if all(x.shape.num_elements() == 0 for device in self._devices.values() for x in device.grad_clean.values()):
+            return
+        if not _collective_ops_warning_printed:
+            print("------------------------------------------------------------------------")
+            print("WARNING: Using slow fallback implementation for inter-GPU communication.")
+            print("Please use TensorFlow 1.14 on Linux for optimal training performance.")
+            print("------------------------------------------------------------------------")
+            _collective_ops_warning_printed = True
+        for device in self._devices.values():
+            with tf.device(device.name):
+                combo = [tf.reshape(x, [x.shape.num_elements()]) for x in device.grad_clean.values()]
+                combo = tf.concat(combo, axis=0)
+                combo = collective_ops.all_reduce(combo, merge_op='Add', final_op='Id',
+                    group_size=len(self._devices), group_key=_collective_ops_group_key,
+                    instance_key=_collective_ops_instance_key)
+                cur_ofs = 0
+                for var, grad_old in device.grad_clean.items():
+                    grad_new = tf.reshape(combo[cur_ofs : cur_ofs + grad_old.shape.num_elements()], grad_old.shape)
+                    cur_ofs += grad_old.shape.num_elements()
+                    device.grad_clean[var] = grad_new
+        _collective_ops_instance_key += 1
+
+
+class SimpleAdam:
+    """Simplified version of tf.train.AdamOptimizer that behaves identically when used with dnnlib.tflib.Optimizer."""
+
+    def __init__(self, name="Adam", learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-8):
+        self.name = name
+        self.learning_rate = learning_rate
+        self.beta1 = beta1
+        self.beta2 = beta2
+        self.epsilon = epsilon
+        self.all_state_vars = []
+
+    def variables(self):
+        return self.all_state_vars
+
+    def compute_gradients(self, loss, var_list, gate_gradients=tf.train.Optimizer.GATE_NONE):
+        assert gate_gradients == tf.train.Optimizer.GATE_NONE
+        return list(zip(tf.gradients(loss, var_list), var_list))
+
+    def apply_gradients(self, grads_and_vars):
+        with tf.name_scope(self.name):
+            state_vars = []
+            update_ops = []
+
+            # Adjust learning rate to deal with startup bias.
+            with tf.control_dependencies(None):
+                b1pow_var = tf.Variable(dtype=tf.float32, initial_value=1, trainable=False)
+                b2pow_var = tf.Variable(dtype=tf.float32, initial_value=1, trainable=False)
+                state_vars += [b1pow_var, b2pow_var]
+            b1pow_new = b1pow_var * self.beta1
+            b2pow_new = b2pow_var * self.beta2
+            update_ops += [tf.assign(b1pow_var, b1pow_new), tf.assign(b2pow_var, b2pow_new)]
+            lr_new = self.learning_rate * tf.sqrt(1 - b2pow_new) / (1 - b1pow_new)
+
+            # Construct ops to update each variable.
+            for grad, var in grads_and_vars:
+                with tf.control_dependencies(None):
+                    m_var = tf.Variable(dtype=tf.float32, initial_value=tf.zeros_like(var), trainable=False)
+                    v_var = tf.Variable(dtype=tf.float32, initial_value=tf.zeros_like(var), trainable=False)
+                    state_vars += [m_var, v_var]
+                m_new = self.beta1 * m_var + (1 - self.beta1) * grad
+                v_new = self.beta2 * v_var + (1 - self.beta2) * tf.square(grad)
+                var_delta = lr_new * m_new / (tf.sqrt(v_new) + self.epsilon)
+                update_ops += [tf.assign(m_var, m_new), tf.assign(v_var, v_new), tf.assign_sub(var, var_delta)]
+
+            # Group everything together.
+            self.all_state_vars += state_vars
+            return tf.group(*update_ops)

dnnlib/tflib/tfutil.py

@@ -0,0 +1,252 @@
+# Copyright (c) 2019, NVIDIA Corporation. All rights reserved.
+#
+# This work is made available under the Nvidia Source Code License-NC.
+# To view a copy of this license, visit
+# https://nvlabs.github.io/stylegan2/license.html
+
+"""Miscellaneous helper utils for Tensorflow."""
+
+import os
+import numpy as np
+import tensorflow as tf
+
+# Silence deprecation warnings from TensorFlow 1.13 onwards
+import logging
+logging.getLogger('tensorflow').setLevel(logging.ERROR)
+import tensorflow.contrib   # requires TensorFlow 1.x!
+tf.contrib = tensorflow.contrib
+
+from typing import Any, Iterable, List, Union
+
+TfExpression = Union[tf.Tensor, tf.Variable, tf.Operation]
+"""A type that represents a valid Tensorflow expression."""
+
+TfExpressionEx = Union[TfExpression, int, float, np.ndarray]
+"""A type that can be converted to a valid Tensorflow expression."""
+
+
+def run(*args, **kwargs) -> Any:
+    """Run the specified ops in the default session."""
+    assert_tf_initialized()
+    return tf.get_default_session().run(*args, **kwargs)
+
+
+def is_tf_expression(x: Any) -> bool:
+    """Check whether the input is a valid Tensorflow expression, i.e., Tensorflow Tensor, Variable, or Operation."""
+    return isinstance(x, (tf.Tensor, tf.Variable, tf.Operation))
+
+
+def shape_to_list(shape: Iterable[tf.Dimension]) -> List[Union[int, None]]:
+    """Convert a Tensorflow shape to a list of ints. Retained for backwards compatibility -- use TensorShape.as_list() in new code."""
+    return [dim.value for dim in shape]
+
+
+def flatten(x: TfExpressionEx) -> TfExpression:
+    """Shortcut function for flattening a tensor."""
+    with tf.name_scope("Flatten"):
+        return tf.reshape(x, [-1])
+
+
+def log2(x: TfExpressionEx) -> TfExpression:
+    """Logarithm in base 2."""
+    with tf.name_scope("Log2"):
+        return tf.log(x) * np.float32(1.0 / np.log(2.0))
+
+
+def exp2(x: TfExpressionEx) -> TfExpression:
+    """Exponent in base 2."""
+    with tf.name_scope("Exp2"):
+        return tf.exp(x * np.float32(np.log(2.0)))
+
+
+def lerp(a: TfExpressionEx, b: TfExpressionEx, t: TfExpressionEx) -> TfExpressionEx:
+    """Linear interpolation."""
+    with tf.name_scope("Lerp"):
+        return a + (b - a) * t
+
+
+def lerp_clip(a: TfExpressionEx, b: TfExpressionEx, t: TfExpressionEx) -> TfExpression:
+    """Linear interpolation with clip."""
+    with tf.name_scope("LerpClip"):
+        return a + (b - a) * tf.clip_by_value(t, 0.0, 1.0)
+
+
+def absolute_name_scope(scope: str) -> tf.name_scope:
+    """Forcefully enter the specified name scope, ignoring any surrounding scopes."""
+    return tf.name_scope(scope + "/")
+
+
+def absolute_variable_scope(scope: str, **kwargs) -> tf.variable_scope:
+    """Forcefully enter the specified variable scope, ignoring any surrounding scopes."""
+    return tf.variable_scope(tf.VariableScope(name=scope, **kwargs), auxiliary_name_scope=False)
+
+
+def _sanitize_tf_config(config_dict: dict = None) -> dict:
+    # Defaults.
+    cfg = dict()
+    cfg["rnd.np_random_seed"]               = None      # Random seed for NumPy. None = keep as is.
+    cfg["rnd.tf_random_seed"]               = "auto"    # Random seed for TensorFlow. 'auto' = derive from NumPy random state. None = keep as is.
+    cfg["env.TF_CPP_MIN_LOG_LEVEL"]         = "1"       # 0 = Print all available debug info from TensorFlow. 1 = Print warnings and errors, but disable debug info.
+    cfg["graph_options.place_pruned_graph"] = True      # False = Check that all ops are available on the designated device. True = Skip the check for ops that are not used.
+    cfg["gpu_options.allow_growth"]         = True      # False = Allocate all GPU memory at the beginning. True = Allocate only as much GPU memory as needed.
+
+    # Remove defaults for environment variables that are already set.
+    for key in list(cfg):
+        fields = key.split(".")
+        if fields[0] == "env":
+            assert len(fields) == 2
+            if fields[1] in os.environ:
+                del cfg[key]
+
+    # User overrides.
+    if config_dict is not None:
+        cfg.update(config_dict)
+    return cfg
+
+
+def init_tf(config_dict: dict = None) -> None:
+    """Initialize TensorFlow session using good default settings."""
+    # Skip if already initialized.
+    if tf.get_default_session() is not None:
+        return
+
+    # Setup config dict and random seeds.
+    cfg = _sanitize_tf_config(config_dict)
+    np_random_seed = cfg["rnd.np_random_seed"]
+    if np_random_seed is not None:
+        np.random.seed(np_random_seed)
+    tf_random_seed = cfg["rnd.tf_random_seed"]
+    if tf_random_seed == "auto":
+        tf_random_seed = np.random.randint(1 << 31)
+    if tf_random_seed is not None:
+        tf.set_random_seed(tf_random_seed)
+
+    # Setup environment variables.
+    for key, value in cfg.items():
+        fields = key.split(".")
+        if fields[0] == "env":
+            assert len(fields) == 2
+            os.environ[fields[1]] = str(value)
+
+    # Create default TensorFlow session.
+    create_session(cfg, force_as_default=True)
+
+
+def assert_tf_initialized():
+    """Check that TensorFlow session has been initialized."""
+    if tf.get_default_session() is None:
+        raise RuntimeError("No default TensorFlow session found. Please call dnnlib.tflib.init_tf().")
+
+
+def create_session(config_dict: dict = None, force_as_default: bool = False) -> tf.Session:
+    """Create tf.Session based on config dict."""
+    # Setup TensorFlow config proto.
+    cfg = _sanitize_tf_config(config_dict)
+    config_proto = tf.ConfigProto()
+    for key, value in cfg.items():
+        fields = key.split(".")
+        if fields[0] not in ["rnd", "env"]:
+            obj = config_proto
+            for field in fields[:-1]:
+                obj = getattr(obj, field)
+            setattr(obj, fields[-1], value)
+
+    # Create session.
+    session = tf.Session(config=config_proto)
+    if force_as_default:
+        # pylint: disable=protected-access
+        session._default_session = session.as_default()
+        session._default_session.enforce_nesting = False
+        session._default_session.__enter__()
+    return session
+
+
+def init_uninitialized_vars(target_vars: List[tf.Variable] = None) -> None:
+    """Initialize all tf.Variables that have not already been initialized.
+
+    Equivalent to the following, but more efficient and does not bloat the tf graph:
+    tf.variables_initializer(tf.report_uninitialized_variables()).run()
+    """
+    assert_tf_initialized()
+    if target_vars is None:
+        target_vars = tf.global_variables()
+
+    test_vars = []
+    test_ops = []
+
+    with tf.control_dependencies(None):  # ignore surrounding control_dependencies
+        for var in target_vars:
+            assert is_tf_expression(var)
+
+            try:
+                tf.get_default_graph().get_tensor_by_name(var.name.replace(":0", "/IsVariableInitialized:0"))
+            except KeyError:
+                # Op does not exist => variable may be uninitialized.
+                test_vars.append(var)
+
+                with absolute_name_scope(var.name.split(":")[0]):
+                    test_ops.append(tf.is_variable_initialized(var))
+
+    init_vars = [var for var, inited in zip(test_vars, run(test_ops)) if not inited]
+    run([var.initializer for var in init_vars])
+
+
+def set_vars(var_to_value_dict: dict) -> None:
+    """Set the values of given tf.Variables.
+
+    Equivalent to the following, but more efficient and does not bloat the tf graph:
+    tflib.run([tf.assign(var, value) for var, value in var_to_value_dict.items()]
+    """
+    assert_tf_initialized()
+    ops = []
+    feed_dict = {}
+
+    for var, value in var_to_value_dict.items():
+        assert is_tf_expression(var)
+
+        try:
+            setter = tf.get_default_graph().get_tensor_by_name(var.name.replace(":0", "/setter:0"))  # look for existing op
+        except KeyError:
+            with absolute_name_scope(var.name.split(":")[0]):
+                with tf.control_dependencies(None):  # ignore surrounding control_dependencies
+                    setter = tf.assign(var, tf.placeholder(var.dtype, var.shape, "new_value"), name="setter")  # create new setter
+
+        ops.append(setter)
+        feed_dict[setter.op.inputs[1]] = value
+
+    run(ops, feed_dict)
+
+
+def create_var_with_large_initial_value(initial_value: np.ndarray, *args, **kwargs):
+    """Create tf.Variable with large initial value without bloating the tf graph."""
+    assert_tf_initialized()
+    assert isinstance(initial_value, np.ndarray)
+    zeros = tf.zeros(initial_value.shape, initial_value.dtype)
+    var = tf.Variable(zeros, *args, **kwargs)
+    set_vars({var: initial_value})
+    return var
+
+
+def convert_images_from_uint8(images, drange=[-1,1], nhwc_to_nchw=False):
+    """Convert a minibatch of images from uint8 to float32 with configurable dynamic range.
+    Can be used as an input transformation for Network.run().
+    """
+    images = tf.cast(images, tf.float32)
+    if nhwc_to_nchw:
+        images = tf.transpose(images, [0, 3, 1, 2])
+    return images * ((drange[1] - drange[0]) / 255) + drange[0]
+
+
+def convert_images_to_uint8(images, drange=[-1,1], nchw_to_nhwc=False, shrink=1):
+    """Convert a minibatch of images from float32 to uint8 with configurable dynamic range.
+    Can be used as an output transformation for Network.run().
+    """
+    images = tf.cast(images, tf.float32)
+    if shrink > 1:
+        ksize = [1, 1, shrink, shrink]
+        images = tf.nn.avg_pool(images, ksize=ksize, strides=ksize, padding="VALID", data_format="NCHW")
+    if nchw_to_nhwc:
+        images = tf.transpose(images, [0, 2, 3, 1])
+    scale = 255 / (drange[1] - drange[0])
+    images = images * scale + (0.5 - drange[0] * scale)
+    return tf.saturate_cast(images, tf.uint8)

dnnlib/util.py

@@ -0,0 +1,515 @@
+# 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 glob
+import hashlib
+import html
+import importlib
+import inspect
+import io
+import os
+import pickle
+import re
+import shutil
+import sys
+import tempfile
+import types
+import urllib
+import urllib.request
+import uuid
+from distutils.util import strtobool
+from typing import Any, List, Tuple, Union
+
+import numpy as np
+import requests
+
+
+# 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 Exception:
+        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 Exception:
+            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 Exception:
+        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")
+
+    print(cache_dir)
+
+    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 Exception:
+                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)

editings/bound/Eyeglasses_boundary.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:227c8edae45f8cf06a0578e8422ef2dc59de6fb741d55280746a63fd40d2936c
+size 36992

editings/bound/Heavy_Makeup_boundary.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbd6336240825cf1a8077c5e40bde544ebd6b7c300ee76e28991e8e2d892030d
+size 36992

editings/bound/Smiling_boundary.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5782190b00340d58f8418885dc9ad9ab53462874d7a9a1b91241e76099c9fc22
+size 36992

editings/deltaedit/delta_mapper.py

@@ -0,0 +1,73 @@
+import math
+
+import torch
+from torch import nn
+from torch.nn import Module
+import torch.nn.functional as F
+
+from models.psp.stylegan2.model import EqualLinear, PixelNorm
+
+class Mapper(Module):
+
+    def __init__(self, in_channel=512, out_channel=512, norm=True, num_layers=4):
+        super(Mapper, self).__init__()
+
+        layers = [PixelNorm()] if norm else []
+        
+        layers.append(EqualLinear(in_channel, out_channel, lr_mul=0.01, activation='fused_lrelu'))
+        for _ in range(num_layers-1):
+            layers.append(EqualLinear(out_channel, out_channel, lr_mul=0.01, activation='fused_lrelu'))
+        self.mapping = nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.mapping(x)
+        return x
+
+class DeltaMapper(Module):
+
+    def __init__(self):
+        super(DeltaMapper, self).__init__()
+
+        #Style Module(sm)
+        self.sm_coarse = Mapper(512,  512)
+        self.sm_medium = Mapper(512,  512)
+        self.sm_fine   = Mapper(2464, 2464)
+
+        #Condition Module(cm)
+        self.cm_coarse = Mapper(1024, 512)
+        self.cm_medium = Mapper(1024, 512)
+        self.cm_fine   = Mapper(1024, 2464)
+
+        #Fusion Module(fm)
+        self.fm_coarse = Mapper(512*2,  512,  norm=False)
+        self.fm_medium = Mapper(512*2,  512,  norm=False)
+        self.fm_fine   = Mapper(2464*2, 2464, norm=False)
+        
+    def forward(self, sspace_feat, clip_feat):
+
+        s_coarse = sspace_feat[:, :3*512].view(-1,3,512)
+        s_medium = sspace_feat[:, 3*512:7*512].view(-1,4,512)
+        s_fine   = sspace_feat[:, 7*512:] #channels:2464
+
+        s_coarse = self.sm_coarse(s_coarse)
+        s_medium = self.sm_medium(s_medium)
+        s_fine   = self.sm_fine(s_fine)
+
+        c_coarse = self.cm_coarse(clip_feat)
+        c_medium = self.cm_medium(clip_feat)
+        c_fine   = self.cm_fine(clip_feat)
+
+        x_coarse = torch.cat([s_coarse, torch.stack([c_coarse]*3, dim=1)], dim=2) #[b,3,1024]
+        x_medium = torch.cat([s_medium, torch.stack([c_medium]*4, dim=1)], dim=2) #[b,4,1024]
+        x_fine   = torch.cat([s_fine, c_fine], dim=1) #[b,2464*2]
+
+        x_coarse = self.fm_coarse(x_coarse)
+        x_coarse = x_coarse.view(-1,3*512)
+
+        x_medium = self.fm_medium(x_medium)
+        x_medium = x_medium.view(-1,4*512)
+
+        x_fine   = self.fm_fine(x_fine)
+
+        out = torch.cat([x_coarse, x_medium, x_fine], dim=1)
+        return out

editings/deltaedit/editor.py

@@ -0,0 +1,59 @@
+import torch
+import clip
+import copy
+import numpy as np
+import torch.nn.functional as F
+
+from editings.deltaedit import map_tool
+from editings.deltaedit.delta_mapper import DeltaMapper
+
+
+STYLE_DIM = [512] * 10 + [256, 256, 128, 128, 64, 64, 32]
+
+
+def GetBoundary(fs3, dt, threshold):
+    tmp = np.dot(fs3, dt)
+    select = np.abs(tmp) < threshold
+    return select
+
+def improved_ds(ds, select):
+    ds_imp = copy.copy(ds)
+    ds_imp[select] = 0
+    ds_imp = ds_imp.unsqueeze(0)
+    return ds_imp
+
+
+class DeltaEditor:
+    def __init__(self):
+        device = "cuda"
+        self.fs3 = np.load("pretrained_models/fs3.npy")
+        np.set_printoptions(suppress=True)
+
+        self.net = DeltaMapper()
+        net_ckpt = torch.load("pretrained_models/delta_mapper.pt")
+        self.net.load_state_dict(net_ckpt)
+        self.net = self.net.to(device).eval()
+
+        self.clip_model, self.preprocess = clip.load("ViT-B/32", device=device)
+        self.avg_pool = torch.nn.AdaptiveAvgPool2d((224, 224))
+        self.upsample = torch.nn.Upsample(scale_factor=7)
+
+    def get_delta_s(self, neutral, target, trash, orig_image, start_s):
+        with torch.no_grad():
+            classnames = [target, neutral]
+            dt = map_tool.GetDt(classnames, self.clip_model)
+            select = GetBoundary(self.fs3, dt, trash)
+            dt = torch.Tensor(dt).cuda()
+            dt = dt / dt.norm(dim=-1, keepdim=True).float().clamp(min=1e-5)
+
+            img_gen_for_clip = self.avg_pool(orig_image)
+            c_latents = self.clip_model.encode_image(img_gen_for_clip.cuda())
+            c_latents = c_latents / c_latents.norm(dim=-1, keepdim=True).float()
+
+            delta_c = torch.cat((c_latents, dt.unsqueeze(0)), dim=1)
+            fake_delta_s = self.net(torch.cat(start_s, dim=-1), delta_c)
+            improved_fake_delta_s = improved_ds(fake_delta_s[0], select)
+            return torch.split(improved_fake_delta_s, STYLE_DIM, dim=-1)
+
+
+

editings/deltaedit/map_tool.py

@@ -0,0 +1,109 @@
+import torch
+import clip
+import os
+import numpy as np
+
+imagenet_templates = [
+    'a bad photo of a {}.',
+#    'a photo of many {}.',
+    'a sculpture of a {}.',
+    'a photo of the hard to see {}.',
+    'a low resolution photo of the {}.',
+    'a rendering of a {}.',
+    'graffiti of a {}.',
+    'a bad photo of the {}.',
+    'a cropped photo of the {}.',
+    'a tattoo of a {}.',
+    'the embroidered {}.',
+    'a photo of a hard to see {}.',
+    'a bright photo of a {}.',
+    'a photo of a clean {}.',
+    'a photo of a dirty {}.',
+    'a dark photo of the {}.',
+    'a drawing of a {}.',
+    'a photo of my {}.',
+    'the plastic {}.',
+    'a photo of the cool {}.',
+    'a close-up photo of a {}.',
+    'a black and white photo of the {}.',
+    'a painting of the {}.',
+    'a painting of a {}.',
+    'a pixelated photo of the {}.',
+    'a sculpture of the {}.',
+    'a bright photo of the {}.',
+    'a cropped photo of a {}.',
+    'a plastic {}.',
+    'a photo of the dirty {}.',
+    'a jpeg corrupted photo of a {}.',
+    'a blurry photo of the {}.',
+    'a photo of the {}.',
+    'a good photo of the {}.',
+    'a rendering of the {}.',
+    'a {} in a video game.',
+    'a photo of one {}.',
+    'a doodle of a {}.',
+    'a close-up photo of the {}.',
+    'a photo of a {}.',
+    'the origami {}.',
+    'the {} in a video game.',
+    'a sketch of a {}.',
+    'a doodle of the {}.',
+    'a origami {}.',
+    'a low resolution photo of a {}.',
+    'the toy {}.',
+    'a rendition of the {}.',
+    'a photo of the clean {}.',
+    'a photo of a large {}.',
+    'a rendition of a {}.',
+    'a photo of a nice {}.',
+    'a photo of a weird {}.',
+    'a blurry photo of a {}.',
+    'a cartoon {}.',
+    'art of a {}.',
+    'a sketch of the {}.',
+    'a embroidered {}.',
+    'a pixelated photo of a {}.',
+    'itap of the {}.',
+    'a jpeg corrupted photo of the {}.',
+    'a good photo of a {}.',
+    'a plushie {}.',
+    'a photo of the nice {}.',
+    'a photo of the small {}.',
+    'a photo of the weird {}.',
+    'the cartoon {}.',
+    'art of the {}.',
+    'a drawing of the {}.',
+    'a photo of the large {}.',
+    'a black and white photo of a {}.',
+    'the plushie {}.',
+    'a dark photo of a {}.',
+    'itap of a {}.',
+    'graffiti of the {}.',
+    'a toy {}.',
+    'itap of my {}.',
+    'a photo of a cool {}.',
+    'a photo of a small {}.',
+    'a tattoo of the {}.',
+]
+
+def zeroshot_classifier(classnames, templates,model):
+    with torch.no_grad():
+        zeroshot_weights = []
+        for classname in classnames:
+            texts = [template.format(classname) for template in templates] #format with class
+            texts = clip.tokenize(texts).cuda() #tokenize
+            class_embeddings = model.encode_text(texts) #embed with text encoder
+            class_embeddings /= class_embeddings.norm(dim=-1, keepdim=True)
+            class_embedding = class_embeddings.mean(dim=0)
+            class_embedding /= class_embedding.norm()
+            zeroshot_weights.append(class_embedding)
+        zeroshot_weights = torch.stack(zeroshot_weights, dim=1).cuda()
+    return zeroshot_weights
+
+def GetDt(classnames,model):
+    text_features=zeroshot_classifier(classnames, imagenet_templates,model).t()
+    
+    dt=text_features[0]-text_features[1]
+    dt=dt.cpu().numpy()
+    
+    return dt

editings/ganspace.py

@@ -0,0 +1,32 @@
+import torch
+
+
+def edit(latents, pca, edit_directions):
+    edit_latents = []
+    for latent in latents:
+        for pca_idx, start, end, strength in edit_directions:
+            delta = get_delta(pca, latent, pca_idx, strength)
+            delta_padded = torch.zeros(latent.shape).to("cuda")
+            delta_padded[start:end] += delta.repeat(end - start, 1)
+            edit_latents.append(latent + delta_padded)
+    return torch.stack(edit_latents)
+
+
+def get_delta(pca, latent, idx, strength):
+    w_centered = latent - pca["mean"].to("cuda")
+    lat_comp = pca["comp"].to("cuda")
+    lat_std = pca["std"].to("cuda")
+    w_coord = (
+        torch.sum(w_centered[0].reshape(-1) * lat_comp[idx].reshape(-1)) / lat_std[idx]
+    )
+    delta = (strength - w_coord) * lat_comp[idx] * lat_std[idx]
+    return delta
+
+
+def edit_latent(latent, pca, edit_direction):
+    pca_idx, start, end, strength = edit_direction
+    delta = get_delta(pca, latent, pca_idx, strength)
+    delta_padded = torch.zeros(latent.shape).to("cuda")
+    delta_padded[start:end] += delta.repeat(end - start, 1)
+    edit_latent = latent + delta_padded
+    return edit_latent

editings/ganspace_pca/cars_pca.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5c3bae61ecd85de077fbbf103f5f30cf4b7676fe23a8508166eaf2ce73c8392
+size 167562

editings/ganspace_pca/church_pca.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b848bc301a7a01a4303797768fe95bc88fea2e335a8e818a5c5575383f0496a5
+size 166790

editings/ganspace_pca/ffhq_pca.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d7f9df1c96180d9026b9cb8d04753579fbf385f321a9d0e263641601c5e5d36
+size 167562

editings/interfacegan_directions/age.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50074516b1629707d89b5e43d6b8abd1792212fa3b961a87a11323d6a5222ae0
+size 2808

editings/interfacegan_directions/rotation.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ef9787ac6b7b37940e47fef22e78cbbe5aeae2e566c0b391bf279f3264b1d1c
+size 2808

editings/interfacegan_directions/smile.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:817a7e732b59dee9eba862bec8bd7e8373568443bc9f9731a21cf9b0356f0653
+size 2808