update

stylegan_human/openpose/src/model.py

+import torch
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+
+def make_layers(block, no_relu_layers):
+    layers = []
+    for layer_name, v in block.items():
+        if 'pool' in layer_name:
+            layer = nn.MaxPool2d(kernel_size=v[0], stride=v[1],
+                                    padding=v[2])
+            layers.append((layer_name, layer))
+        else:
+            conv2d = nn.Conv2d(in_channels=v[0], out_channels=v[1],
+                               kernel_size=v[2], stride=v[3],
+                               padding=v[4])
+            layers.append((layer_name, conv2d))
+            if layer_name not in no_relu_layers:
+                layers.append(('relu_'+layer_name, nn.ReLU(inplace=True)))
+
+    return nn.Sequential(OrderedDict(layers))
+
+class bodypose_model(nn.Module):
+    def __init__(self):
+        super(bodypose_model, self).__init__()
+
+        # these layers have no relu layer
+        no_relu_layers = ['conv5_5_CPM_L1', 'conv5_5_CPM_L2', 'Mconv7_stage2_L1',\
+                          'Mconv7_stage2_L2', 'Mconv7_stage3_L1', 'Mconv7_stage3_L2',\
+                          'Mconv7_stage4_L1', 'Mconv7_stage4_L2', 'Mconv7_stage5_L1',\
+                          'Mconv7_stage5_L2', 'Mconv7_stage6_L1', 'Mconv7_stage6_L1']
+        blocks = {}
+        block0 = OrderedDict([
+                      ('conv1_1', [3, 64, 3, 1, 1]),
+                      ('conv1_2', [64, 64, 3, 1, 1]),
+                      ('pool1_stage1', [2, 2, 0]),
+                      ('conv2_1', [64, 128, 3, 1, 1]),
+                      ('conv2_2', [128, 128, 3, 1, 1]),
+                      ('pool2_stage1', [2, 2, 0]),
+                      ('conv3_1', [128, 256, 3, 1, 1]),
+                      ('conv3_2', [256, 256, 3, 1, 1]),
+                      ('conv3_3', [256, 256, 3, 1, 1]),
+                      ('conv3_4', [256, 256, 3, 1, 1]),
+                      ('pool3_stage1', [2, 2, 0]),
+                      ('conv4_1', [256, 512, 3, 1, 1]),
+                      ('conv4_2', [512, 512, 3, 1, 1]),
+                      ('conv4_3_CPM', [512, 256, 3, 1, 1]),
+                      ('conv4_4_CPM', [256, 128, 3, 1, 1])
+                  ])
+
+
+        # Stage 1
+        block1_1 = OrderedDict([
+                        ('conv5_1_CPM_L1', [128, 128, 3, 1, 1]),
+                        ('conv5_2_CPM_L1', [128, 128, 3, 1, 1]),
+                        ('conv5_3_CPM_L1', [128, 128, 3, 1, 1]),
+                        ('conv5_4_CPM_L1', [128, 512, 1, 1, 0]),
+                        ('conv5_5_CPM_L1', [512, 38, 1, 1, 0])
+                    ])
+
+        block1_2 = OrderedDict([
+                        ('conv5_1_CPM_L2', [128, 128, 3, 1, 1]),
+                        ('conv5_2_CPM_L2', [128, 128, 3, 1, 1]),
+                        ('conv5_3_CPM_L2', [128, 128, 3, 1, 1]),
+                        ('conv5_4_CPM_L2', [128, 512, 1, 1, 0]),
+                        ('conv5_5_CPM_L2', [512, 19, 1, 1, 0])
+                    ])
+        blocks['block1_1'] = block1_1
+        blocks['block1_2'] = block1_2
+
+        self.model0 = make_layers(block0, no_relu_layers)
+
+        # Stages 2 - 6
+        for i in range(2, 7):
+            blocks['block%d_1' % i] = OrderedDict([
+                    ('Mconv1_stage%d_L1' % i, [185, 128, 7, 1, 3]),
+                    ('Mconv2_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv3_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv4_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv5_stage%d_L1' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv6_stage%d_L1' % i, [128, 128, 1, 1, 0]),
+                    ('Mconv7_stage%d_L1' % i, [128, 38, 1, 1, 0])
+                ])
+
+            blocks['block%d_2' % i] = OrderedDict([
+                    ('Mconv1_stage%d_L2' % i, [185, 128, 7, 1, 3]),
+                    ('Mconv2_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv3_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv4_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv5_stage%d_L2' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv6_stage%d_L2' % i, [128, 128, 1, 1, 0]),
+                    ('Mconv7_stage%d_L2' % i, [128, 19, 1, 1, 0])
+                ])
+
+        for k in blocks.keys():
+            blocks[k] = make_layers(blocks[k], no_relu_layers)
+
+        self.model1_1 = blocks['block1_1']
+        self.model2_1 = blocks['block2_1']
+        self.model3_1 = blocks['block3_1']
+        self.model4_1 = blocks['block4_1']
+        self.model5_1 = blocks['block5_1']
+        self.model6_1 = blocks['block6_1']
+
+        self.model1_2 = blocks['block1_2']
+        self.model2_2 = blocks['block2_2']
+        self.model3_2 = blocks['block3_2']
+        self.model4_2 = blocks['block4_2']
+        self.model5_2 = blocks['block5_2']
+        self.model6_2 = blocks['block6_2']
+
+
+    def forward(self, x):
+
+        out1 = self.model0(x)
+
+        out1_1 = self.model1_1(out1)
+        out1_2 = self.model1_2(out1)
+        out2 = torch.cat([out1_1, out1_2, out1], 1)
+
+        out2_1 = self.model2_1(out2)
+        out2_2 = self.model2_2(out2)
+        out3 = torch.cat([out2_1, out2_2, out1], 1)
+
+        out3_1 = self.model3_1(out3)
+        out3_2 = self.model3_2(out3)
+        out4 = torch.cat([out3_1, out3_2, out1], 1)
+
+        out4_1 = self.model4_1(out4)
+        out4_2 = self.model4_2(out4)
+        out5 = torch.cat([out4_1, out4_2, out1], 1)
+
+        out5_1 = self.model5_1(out5)
+        out5_2 = self.model5_2(out5)
+        out6 = torch.cat([out5_1, out5_2, out1], 1)
+
+        out6_1 = self.model6_1(out6)
+        out6_2 = self.model6_2(out6)
+
+        return out6_1, out6_2
+
+class handpose_model(nn.Module):
+    def __init__(self):
+        super(handpose_model, self).__init__()
+
+        # these layers have no relu layer
+        no_relu_layers = ['conv6_2_CPM', 'Mconv7_stage2', 'Mconv7_stage3',\
+                          'Mconv7_stage4', 'Mconv7_stage5', 'Mconv7_stage6']
+        # stage 1
+        block1_0 = OrderedDict([
+                ('conv1_1', [3, 64, 3, 1, 1]),
+                ('conv1_2', [64, 64, 3, 1, 1]),
+                ('pool1_stage1', [2, 2, 0]),
+                ('conv2_1', [64, 128, 3, 1, 1]),
+                ('conv2_2', [128, 128, 3, 1, 1]),
+                ('pool2_stage1', [2, 2, 0]),
+                ('conv3_1', [128, 256, 3, 1, 1]),
+                ('conv3_2', [256, 256, 3, 1, 1]),
+                ('conv3_3', [256, 256, 3, 1, 1]),
+                ('conv3_4', [256, 256, 3, 1, 1]),
+                ('pool3_stage1', [2, 2, 0]),
+                ('conv4_1', [256, 512, 3, 1, 1]),
+                ('conv4_2', [512, 512, 3, 1, 1]),
+                ('conv4_3', [512, 512, 3, 1, 1]),
+                ('conv4_4', [512, 512, 3, 1, 1]),
+                ('conv5_1', [512, 512, 3, 1, 1]),
+                ('conv5_2', [512, 512, 3, 1, 1]),
+                ('conv5_3_CPM', [512, 128, 3, 1, 1])
+            ])
+
+        block1_1 = OrderedDict([
+            ('conv6_1_CPM', [128, 512, 1, 1, 0]),
+            ('conv6_2_CPM', [512, 22, 1, 1, 0])
+        ])
+
+        blocks = {}
+        blocks['block1_0'] = block1_0
+        blocks['block1_1'] = block1_1
+
+        # stage 2-6
+        for i in range(2, 7):
+            blocks['block%d' % i] = OrderedDict([
+                    ('Mconv1_stage%d' % i, [150, 128, 7, 1, 3]),
+                    ('Mconv2_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv3_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv4_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv5_stage%d' % i, [128, 128, 7, 1, 3]),
+                    ('Mconv6_stage%d' % i, [128, 128, 1, 1, 0]),
+                    ('Mconv7_stage%d' % i, [128, 22, 1, 1, 0])
+                ])
+
+        for k in blocks.keys():
+            blocks[k] = make_layers(blocks[k], no_relu_layers)
+
+        self.model1_0 = blocks['block1_0']
+        self.model1_1 = blocks['block1_1']
+        self.model2 = blocks['block2']
+        self.model3 = blocks['block3']
+        self.model4 = blocks['block4']
+        self.model5 = blocks['block5']
+        self.model6 = blocks['block6']
+
+    def forward(self, x):
+        out1_0 = self.model1_0(x)
+        out1_1 = self.model1_1(out1_0)
+        concat_stage2 = torch.cat([out1_1, out1_0], 1)
+        out_stage2 = self.model2(concat_stage2)
+        concat_stage3 = torch.cat([out_stage2, out1_0], 1)
+        out_stage3 = self.model3(concat_stage3)
+        concat_stage4 = torch.cat([out_stage3, out1_0], 1)
+        out_stage4 = self.model4(concat_stage4)
+        concat_stage5 = torch.cat([out_stage4, out1_0], 1)
+        out_stage5 = self.model5(concat_stage5)
+        concat_stage6 = torch.cat([out_stage5, out1_0], 1)
+        out_stage6 = self.model6(concat_stage6)
+        return out_stage6
+
+

stylegan_human/openpose/src/util.py

+import numpy as np
+import math
+import cv2
+import matplotlib
+from matplotlib.backends.backend_agg import FigureCanvasAgg as FigureCanvas
+from matplotlib.figure import Figure
+import numpy as np
+import matplotlib.pyplot as plt
+import cv2
+
+
+def padRightDownCorner(img, stride, padValue):
+    h = img.shape[0]
+    w = img.shape[1]
+
+    pad = 4 * [None]
+    pad[0] = 0 # up
+    pad[1] = 0 # left
+    pad[2] = 0 if (h % stride == 0) else stride - (h % stride) # down
+    pad[3] = 0 if (w % stride == 0) else stride - (w % stride) # right
+
+    img_padded = img
+    pad_up = np.tile(img_padded[0:1, :, :]*0 + padValue, (pad[0], 1, 1))
+    img_padded = np.concatenate((pad_up, img_padded), axis=0)
+    pad_left = np.tile(img_padded[:, 0:1, :]*0 + padValue, (1, pad[1], 1))
+    img_padded = np.concatenate((pad_left, img_padded), axis=1)
+    pad_down = np.tile(img_padded[-2:-1, :, :]*0 + padValue, (pad[2], 1, 1))
+    img_padded = np.concatenate((img_padded, pad_down), axis=0)
+    pad_right = np.tile(img_padded[:, -2:-1, :]*0 + padValue, (1, pad[3], 1))
+    img_padded = np.concatenate((img_padded, pad_right), axis=1)
+
+    return img_padded, pad
+
+# transfer caffe model to pytorch which will match the layer name
+def transfer(model, model_weights):
+    transfered_model_weights = {}
+    for weights_name in model.state_dict().keys():
+        transfered_model_weights[weights_name] = model_weights['.'.join(weights_name.split('.')[1:])]
+    return transfered_model_weights
+
+# draw the body keypoint and lims
+def draw_bodypose(canvas, candidate, subset,show_number=False):
+    stickwidth = 4
+    limbSeq = [[2, 3], [2, 6], [3, 4], [4, 5], [6, 7], [7, 8], [2, 9], [9, 10], \
+               [10, 11], [2, 12], [12, 13], [13, 14], [2, 1], [1, 15], [15, 17], \
+               [1, 16], [16, 18], [3, 17], [6, 18]]
+
+    colors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0], \
+              [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255], \
+              [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
+    for i in range(18):
+        for n in range(len(subset)):
+            index = int(subset[n][i])
+            if index == -1:
+                continue
+            x, y = candidate[index][0:2]
+            cv2.circle(canvas, (int(x), int(y)), 4, colors[i], thickness=-1)
+            if show_number:
+                cv2.putText(canvas, f'{index}', (int(x), int(y)),cv2.FONT_HERSHEY_SIMPLEX, 0.6, 
+                    (255,255,0), 1, cv2.LINE_AA)
+                ## calc and print average
+    for i in range(17):
+        for n in range(len(subset)):
+            index = subset[n][np.array(limbSeq[i]) - 1]
+            if -1 in index:
+                continue
+            cur_canvas = canvas.copy()
+            Y = candidate[index.astype(int), 0]
+            X = candidate[index.astype(int), 1]
+            mX = np.mean(X)
+            mY = np.mean(Y)
+            length = ((X[0] - X[1]) ** 2 + (Y[0] - Y[1]) ** 2) ** 0.5
+            angle = math.degrees(math.atan2(X[0] - X[1], Y[0] - Y[1]))
+            polygon = cv2.ellipse2Poly((int(mY), int(mX)), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
+            cv2.fillConvexPoly(cur_canvas, polygon, colors[i])
+            canvas = cv2.addWeighted(canvas, 0.4, cur_canvas, 0.6, 0)
+
+    return canvas
+
+# get max index of 2d array
+def npmax(array):
+    arrayindex = array.argmax(1)
+    arrayvalue = array.max(1)
+    i = arrayvalue.argmax()
+    j = arrayindex[i]
+    return i, j
+
+# get max index of 2d array
+def npmax_with_score(array):
+    arrayindex = array.argmax(1)
+    arrayvalue = array.max(1)
+    i = arrayvalue.argmax()
+    j = arrayindex[i]
+    score =array[i][j]
+    return i, j,score

stylegan_human/pti/pti_configs/global_config.py

+## Device
+cuda_visible_devices = '0'
+device = 'cuda:0'
+
+## Logs
+training_step = 1
+image_rec_result_log_snapshot = 100
+pivotal_training_steps = 0
+model_snapshot_interval = 400
+
+## Run name to be updated during PTI
+run_name = 'exp'

stylegan_human/pti/pti_configs/hyperparameters.py

+## Architechture
+lpips_type = 'alex'
+first_inv_type = 'w+'#'w+'
+optim_type = 'adam'
+
+## Locality regularization
+latent_ball_num_of_samples = 1
+locality_regularization_interval = 1
+use_locality_regularization = False
+regulizer_l2_lambda = 0.1
+regulizer_lpips_lambda = 0.1
+regulizer_alpha = 30
+
+## Loss
+pt_l2_lambda = 1
+pt_lpips_lambda = 1
+
+## Steps
+LPIPS_value_threshold = 0.04
+max_pti_steps = 350
+first_inv_steps = 450
+max_images_to_invert = 30
+
+## Optimization
+pti_learning_rate = 5e-4
+first_inv_lr = 8e-3
+train_batch_size = 1
+use_last_w_pivots = False

stylegan_human/pti/pti_configs/paths_config.py

+import os
+
+## Pretrained models paths
+e4e = './pti/e4e_w+.pt'
+stylegan2_ada_shhq = './pretrained_models/stylegan_human_v2_1024.pkl'
+ir_se50 =  '' #'./model_ir_se50.pth' 
+
+## Dirs for output files
+checkpoints_dir = './outputs/pti/checkpoints/'
+embedding_base_dir = './outputs/pti/embeddings'
+experiments_output_dir = './outputs/pti/'
+
+## Input info
+### Input dir, where the images reside
+input_data_path = 'aligned_image/'
+### Inversion identifier, used to keeping track of the inversion results. Both the latent code and the generator
+input_data_id = 'test'
+
+## Keywords
+pti_results_keyword = 'PTI'
+e4e_results_keyword = 'e4e'
+sg2_results_keyword = 'SG2'
+sg2_plus_results_keyword = 'SG2_Plus'
+multi_id_model_type = 'multi_id'

stylegan_human/pti/pti_models/e4e/encoders/helpers.py

+from collections import namedtuple
+import torch
+import torch.nn.functional as F
+from torch.nn import Conv2d, BatchNorm2d, PReLU, ReLU, Sigmoid, MaxPool2d, AdaptiveAvgPool2d, Sequential, Module
+
+"""
+ArcFace implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
+"""
+
+
+class Flatten(Module):
+    def forward(self, input):
+        return input.view(input.size(0), -1)
+
+
+def l2_norm(input, axis=1):
+    norm = torch.norm(input, 2, axis, True)
+    output = torch.div(input, norm)
+    return output
+
+
+class Bottleneck(namedtuple('Block', ['in_channel', 'depth', 'stride'])):
+    """ A named tuple describing a ResNet block. """
+
+
+def get_block(in_channel, depth, num_units, stride=2):
+    return [Bottleneck(in_channel, depth, stride)] + [Bottleneck(depth, depth, 1) for i in range(num_units - 1)]
+
+
+def get_blocks(num_layers):
+    if num_layers == 50:
+        blocks = [
+            get_block(in_channel=64, depth=64, num_units=3),
+            get_block(in_channel=64, depth=128, num_units=4),
+            get_block(in_channel=128, depth=256, num_units=14),
+            get_block(in_channel=256, depth=512, num_units=3)
+        ]
+    elif num_layers == 100:
+        blocks = [
+            get_block(in_channel=64, depth=64, num_units=3),
+            get_block(in_channel=64, depth=128, num_units=13),
+            get_block(in_channel=128, depth=256, num_units=30),
+            get_block(in_channel=256, depth=512, num_units=3)
+        ]
+    elif num_layers == 152:
+        blocks = [
+            get_block(in_channel=64, depth=64, num_units=3),
+            get_block(in_channel=64, depth=128, num_units=8),
+            get_block(in_channel=128, depth=256, num_units=36),
+            get_block(in_channel=256, depth=512, num_units=3)
+        ]
+    else:
+        raise ValueError("Invalid number of layers: {}. Must be one of [50, 100, 152]".format(num_layers))
+    return blocks
+
+
+class SEModule(Module):
+    def __init__(self, channels, reduction):
+        super(SEModule, self).__init__()
+        self.avg_pool = AdaptiveAvgPool2d(1)
+        self.fc1 = Conv2d(channels, channels // reduction, kernel_size=1, padding=0, bias=False)
+        self.relu = ReLU(inplace=True)
+        self.fc2 = Conv2d(channels // reduction, channels, kernel_size=1, padding=0, bias=False)
+        self.sigmoid = Sigmoid()
+
+    def forward(self, x):
+        module_input = x
+        x = self.avg_pool(x)
+        x = self.fc1(x)
+        x = self.relu(x)
+        x = self.fc2(x)
+        x = self.sigmoid(x)
+        return module_input * x
+
+
+class bottleneck_IR(Module):
+    def __init__(self, in_channel, depth, stride):
+        super(bottleneck_IR, self).__init__()
+        if in_channel == depth:
+            self.shortcut_layer = MaxPool2d(1, stride)
+        else:
+            self.shortcut_layer = Sequential(
+                Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+                BatchNorm2d(depth)
+            )
+        self.res_layer = Sequential(
+            BatchNorm2d(in_channel),
+            Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False), PReLU(depth),
+            Conv2d(depth, depth, (3, 3), stride, 1, bias=False), BatchNorm2d(depth)
+        )
+
+    def forward(self, x):
+        shortcut = self.shortcut_layer(x)
+        res = self.res_layer(x)
+        return res + shortcut
+
+
+class bottleneck_IR_SE(Module):
+    def __init__(self, in_channel, depth, stride):
+        super(bottleneck_IR_SE, self).__init__()
+        if in_channel == depth:
+            self.shortcut_layer = MaxPool2d(1, stride)
+        else:
+            self.shortcut_layer = Sequential(
+                Conv2d(in_channel, depth, (1, 1), stride, bias=False),
+                BatchNorm2d(depth)
+            )
+        self.res_layer = Sequential(
+            BatchNorm2d(in_channel),
+            Conv2d(in_channel, depth, (3, 3), (1, 1), 1, bias=False),
+            PReLU(depth),
+            Conv2d(depth, depth, (3, 3), stride, 1, bias=False),
+            BatchNorm2d(depth),
+            SEModule(depth, 16)
+        )
+
+    def forward(self, x):
+        shortcut = self.shortcut_layer(x)
+        res = self.res_layer(x)
+        return res + shortcut
+
+
+def _upsample_add(x, y):
+    """Upsample and add two feature maps.
+    Args:
+      x: (Variable) top feature map to be upsampled.
+      y: (Variable) lateral feature map.
+    Returns:
+      (Variable) added feature map.
+    Note in PyTorch, when input size is odd, the upsampled feature map
+    with `F.upsample(..., scale_factor=2, mode='nearest')`
+    maybe not equal to the lateral feature map size.
+    e.g.
+    original input size: [N,_,15,15] ->
+    conv2d feature map size: [N,_,8,8] ->
+    upsampled feature map size: [N,_,16,16]
+    So we choose bilinear upsample which supports arbitrary output sizes.
+    """
+    _, _, H, W = y.size()
+    return F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True) + y

stylegan_human/pti/pti_models/e4e/encoders/model_irse.py

+from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Dropout, Sequential, Module
+from encoder4editing.models.encoders.helpers import get_blocks, Flatten, bottleneck_IR, bottleneck_IR_SE, l2_norm
+
+"""
+Modified Backbone implementation from [TreB1eN](https://github.com/TreB1eN/InsightFace_Pytorch)
+"""
+
+
+class Backbone(Module):
+    def __init__(self, input_size, num_layers, mode='ir', drop_ratio=0.4, affine=True):
+        super(Backbone, self).__init__()
+        assert input_size in [112, 224], "input_size should be 112 or 224"
+        assert num_layers in [50, 100, 152], "num_layers should be 50, 100 or 152"
+        assert mode in ['ir', 'ir_se'], "mode should be ir or ir_se"
+        blocks = get_blocks(num_layers)
+        if mode == 'ir':
+            unit_module = bottleneck_IR
+        elif mode == 'ir_se':
+            unit_module = bottleneck_IR_SE
+        self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
+                                      BatchNorm2d(64),
+                                      PReLU(64))
+        if input_size == 112:
+            self.output_layer = Sequential(BatchNorm2d(512),
+                                           Dropout(drop_ratio),
+                                           Flatten(),
+                                           Linear(512 * 7 * 7, 512),
+                                           BatchNorm1d(512, affine=affine))
+        else:
+            self.output_layer = Sequential(BatchNorm2d(512),
+                                           Dropout(drop_ratio),
+                                           Flatten(),
+                                           Linear(512 * 14 * 14, 512),
+                                           BatchNorm1d(512, affine=affine))
+
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = Sequential(*modules)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+        x = self.body(x)
+        x = self.output_layer(x)
+        return l2_norm(x)
+
+
+def IR_50(input_size):
+    """Constructs a ir-50 model."""
+    model = Backbone(input_size, num_layers=50, mode='ir', drop_ratio=0.4, affine=False)
+    return model
+
+
+def IR_101(input_size):
+    """Constructs a ir-101 model."""
+    model = Backbone(input_size, num_layers=100, mode='ir', drop_ratio=0.4, affine=False)
+    return model
+
+
+def IR_152(input_size):
+    """Constructs a ir-152 model."""
+    model = Backbone(input_size, num_layers=152, mode='ir', drop_ratio=0.4, affine=False)
+    return model
+
+
+def IR_SE_50(input_size):
+    """Constructs a ir_se-50 model."""
+    model = Backbone(input_size, num_layers=50, mode='ir_se', drop_ratio=0.4, affine=False)
+    return model
+
+
+def IR_SE_101(input_size):
+    """Constructs a ir_se-101 model."""
+    model = Backbone(input_size, num_layers=100, mode='ir_se', drop_ratio=0.4, affine=False)
+    return model
+
+
+def IR_SE_152(input_size):
+    """Constructs a ir_se-152 model."""
+    model = Backbone(input_size, num_layers=152, mode='ir_se', drop_ratio=0.4, affine=False)
+    return model

stylegan_human/pti/pti_models/e4e/encoders/psp_encoders.py

+from enum import Enum
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import Conv2d, BatchNorm2d, PReLU, Sequential, Module
+
+from pti.pti_models.e4e.encoders.helpers import get_blocks, bottleneck_IR, bottleneck_IR_SE, _upsample_add
+from pti.pti_models.e4e.stylegan2.model import EqualLinear
+
+
+class ProgressiveStage(Enum):
+    WTraining = 0
+    Delta1Training = 1
+    Delta2Training = 2
+    Delta3Training = 3
+    Delta4Training = 4
+    Delta5Training = 5
+    Delta6Training = 6
+    Delta7Training = 7
+    Delta8Training = 8
+    Delta9Training = 9
+    Delta10Training = 10
+    Delta11Training = 11
+    Delta12Training = 12
+    Delta13Training = 13
+    Delta14Training = 14
+    Delta15Training = 15
+    Delta16Training = 16
+    Delta17Training = 17
+    Inference = 18
+
+
+class GradualStyleBlock(Module):
+    def __init__(self, in_c, out_c, spatial):
+        super(GradualStyleBlock, self).__init__()
+        self.out_c = out_c
+        self.spatial = spatial
+        num_pools = int(np.log2(spatial))
+        modules = []
+        modules += [Conv2d(in_c, out_c, kernel_size=3, stride=2, padding=1),
+                    nn.LeakyReLU()]
+        for i in range(num_pools - 1):
+            modules += [
+                Conv2d(out_c, out_c, kernel_size=3, stride=2, padding=1),
+                nn.LeakyReLU()
+            ]
+        self.convs = nn.Sequential(*modules)
+        self.linear = EqualLinear(out_c, out_c, lr_mul=1)
+
+    def forward(self, x):
+        x = self.convs(x)
+        x = x.view(-1, self.out_c)
+        x = self.linear(x)
+        return x
+
+
+class GradualStyleEncoder(Module):
+    def __init__(self, num_layers, mode='ir', opts=None):
+        super(GradualStyleEncoder, self).__init__()
+        assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
+        assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
+        blocks = get_blocks(num_layers)
+        if mode == 'ir':
+            unit_module = bottleneck_IR
+        elif mode == 'ir_se':
+            unit_module = bottleneck_IR_SE
+        self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
+                                      BatchNorm2d(64),
+                                      PReLU(64))
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = Sequential(*modules)
+
+        self.styles = nn.ModuleList()
+        log_size = int(math.log(opts.stylegan_size, 2))
+        self.style_count = 2 * log_size - 2
+        self.coarse_ind = 3
+        self.middle_ind = 7
+        for i in range(self.style_count):
+            if i < self.coarse_ind:
+                style = GradualStyleBlock(512, 512, 16)
+            elif i < self.middle_ind:
+                style = GradualStyleBlock(512, 512, 32)
+            else:
+                style = GradualStyleBlock(512, 512, 64)
+            self.styles.append(style)
+        self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
+        self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+
+        latents = []
+        modulelist = list(self.body._modules.values())
+        for i, l in enumerate(modulelist):
+            x = l(x)
+            if i == 6:
+                c1 = x
+            elif i == 20:
+                c2 = x
+            elif i == 23:
+                c3 = x
+
+        for j in range(self.coarse_ind):
+            latents.append(self.styles[j](c3))
+
+        p2 = _upsample_add(c3, self.latlayer1(c2))
+        for j in range(self.coarse_ind, self.middle_ind):
+            latents.append(self.styles[j](p2))
+
+        p1 = _upsample_add(p2, self.latlayer2(c1))
+        for j in range(self.middle_ind, self.style_count):
+            latents.append(self.styles[j](p1))
+
+        out = torch.stack(latents, dim=1)
+        return out
+
+
+class Encoder4Editing(Module):
+    def __init__(self, num_layers, mode='ir', opts=None):
+        super(Encoder4Editing, self).__init__()
+        assert num_layers in [50, 100, 152], 'num_layers should be 50,100, or 152'
+        assert mode in ['ir', 'ir_se'], 'mode should be ir or ir_se'
+        blocks = get_blocks(num_layers)
+        if mode == 'ir':
+            unit_module = bottleneck_IR
+        elif mode == 'ir_se':
+            unit_module = bottleneck_IR_SE
+        self.input_layer = Sequential(Conv2d(3, 64, (3, 3), 1, 1, bias=False),
+                                      BatchNorm2d(64),
+                                      PReLU(64))
+        modules = []
+        for block in blocks:
+            for bottleneck in block:
+                modules.append(unit_module(bottleneck.in_channel,
+                                           bottleneck.depth,
+                                           bottleneck.stride))
+        self.body = Sequential(*modules)
+
+        self.styles = nn.ModuleList()
+        log_size = int(math.log(opts.stylegan_size, 2))
+        self.style_count = 2 * log_size - 2
+        self.coarse_ind = 3
+        self.middle_ind = 7
+
+        for i in range(self.style_count):
+            if i < self.coarse_ind:
+                style = GradualStyleBlock(512, 512, 16)
+            elif i < self.middle_ind:
+                style = GradualStyleBlock(512, 512, 32)
+            else:
+                style = GradualStyleBlock(512, 512, 64)
+            self.styles.append(style)
+
+        self.latlayer1 = nn.Conv2d(256, 512, kernel_size=1, stride=1, padding=0)
+        self.latlayer2 = nn.Conv2d(128, 512, kernel_size=1, stride=1, padding=0)
+
+        self.progressive_stage = ProgressiveStage.Inference
+
+    def get_deltas_starting_dimensions(self):
+        ''' Get a list of the initial dimension of every delta from which it is applied '''
+        return list(range(self.style_count))  # Each dimension has a delta applied to it
+
+    def set_progressive_stage(self, new_stage: ProgressiveStage):
+        self.progressive_stage = new_stage
+        print('Changed progressive stage to: ', new_stage)
+
+    def forward(self, x):
+        x = self.input_layer(x)
+
+        modulelist = list(self.body._modules.values())
+        for i, l in enumerate(modulelist):
+            x = l(x)
+            if i == 6:
+                c1 = x
+            elif i == 20:
+                c2 = x
+            elif i == 23:
+                c3 = x
+
+        # Infer main W and duplicate it
+        w0 = self.styles[0](c3)
+        w = w0.repeat(self.style_count, 1, 1).permute(1, 0, 2)
+        stage = self.progressive_stage.value
+        features = c3
+        for i in range(1, min(stage + 1, self.style_count)):  # Infer additional deltas
+            if i == self.coarse_ind:
+                p2 = _upsample_add(c3, self.latlayer1(c2))  # FPN's middle features
+                features = p2
+            elif i == self.middle_ind:
+                p1 = _upsample_add(p2, self.latlayer2(c1))  # FPN's fine features
+                features = p1
+            delta_i = self.styles[i](features)
+            w[:, i] += delta_i
+        return w

stylegan_human/pti/pti_models/e4e/latent_codes_pool.py

+import random
+import torch
+
+
+class LatentCodesPool:
+    """This class implements latent codes buffer that stores previously generated w latent codes.
+    This buffer enables us to update discriminators using a history of generated w's
+    rather than the ones produced by the latest encoder.
+    """
+
+    def __init__(self, pool_size):
+        """Initialize the ImagePool class
+        Parameters:
+            pool_size (int) -- the size of image buffer, if pool_size=0, no buffer will be created
+        """
+        self.pool_size = pool_size
+        if self.pool_size > 0:  # create an empty pool
+            self.num_ws = 0
+            self.ws = []
+
+    def query(self, ws):
+        """Return w's from the pool.
+        Parameters:
+            ws: the latest generated w's from the generator
+        Returns w's from the buffer.
+        By 50/100, the buffer will return input w's.
+        By 50/100, the buffer will return w's previously stored in the buffer,
+        and insert the current w's to the buffer.
+        """
+        if self.pool_size == 0:  # if the buffer size is 0, do nothing
+            return ws
+        return_ws = []
+        for w in ws:  # ws.shape: (batch, 512) or (batch, n_latent, 512)
+            # w = torch.unsqueeze(image.data, 0)
+            if w.ndim == 2:
+                i = random.randint(0, len(w) - 1)  # apply a random latent index as a candidate
+                w = w[i]
+            self.handle_w(w, return_ws)
+        return_ws = torch.stack(return_ws, 0)   # collect all the images and return
+        return return_ws
+
+    def handle_w(self, w, return_ws):
+        if self.num_ws < self.pool_size:  # if the buffer is not full; keep inserting current codes to the buffer
+            self.num_ws = self.num_ws + 1
+            self.ws.append(w)
+            return_ws.append(w)
+        else:
+            p = random.uniform(0, 1)
+            if p > 0.5:  # by 50% chance, the buffer will return a previously stored latent code, and insert the current code into the buffer
+                random_id = random.randint(0, self.pool_size - 1)  # randint is inclusive
+                tmp = self.ws[random_id].clone()
+                self.ws[random_id] = w
+                return_ws.append(tmp)
+            else:  # by another 50% chance, the buffer will return the current image
+                return_ws.append(w)

stylegan_human/pti/pti_models/e4e/psp.py

+import matplotlib
+from pti.pti_configs import paths_config
+matplotlib.use('Agg')
+import torch
+from torch import nn
+from pti.pti_models.e4e.encoders import psp_encoders
+from pti.pti_models.e4e.stylegan2.model import Generator
+
+
+def get_keys(d, name):
+    if 'state_dict' in d:
+        d = d['state_dict']
+    d_filt = {k[len(name) + 1:]: v for k, v in d.items() if k[:len(name)] == name}
+    return d_filt
+
+
+class pSp(nn.Module):
+
+    def __init__(self, opts):
+        super(pSp, self).__init__()
+        self.opts = opts
+        # Define architecture
+        self.encoder = self.set_encoder()
+        self.decoder = Generator(opts.stylegan_size, 512, 8, channel_multiplier=2)
+        self.face_pool = torch.nn.AdaptiveAvgPool2d((256, 256 // 2))
+        # Load weights if needed
+        self.load_weights()
+
+    def set_encoder(self):
+        if self.opts.encoder_type == 'GradualStyleEncoder':
+            encoder = psp_encoders.GradualStyleEncoder(50, 'ir_se', self.opts)
+        elif self.opts.encoder_type == 'Encoder4Editing':
+            encoder = psp_encoders.Encoder4Editing(50, 'ir_se', self.opts)
+        elif self.opts.encoder_type == 'SingleStyleCodeEncoder':
+            encoder = psp_encoders.BackboneEncoderUsingLastLayerIntoW(50, 'ir_se', self.opts)
+        else:
+            raise Exception('{} is not a valid encoders'.format(self.opts.encoder_type))
+        return encoder
+
+    def load_weights(self):
+        if self.opts.checkpoint_path is not None:
+            print('Loading e4e over the pSp framework from checkpoint: {}'.format(self.opts.checkpoint_path))
+            ckpt = torch.load(self.opts.checkpoint_path, map_location='cpu')
+            self.encoder.load_state_dict(get_keys(ckpt, 'encoder'), strict=True)
+            self.decoder.load_state_dict(get_keys(ckpt, 'decoder'), strict=True)
+            self.__load_latent_avg(ckpt)
+        else:
+            print('Loading encoders weights from irse50!')
+            encoder_ckpt = torch.load(model_paths['ir_se50'])
+            self.encoder.load_state_dict(encoder_ckpt, strict=False)
+            print('Loading decoder weights from pretrained!')
+            ckpt = torch.load(self.opts.stylegan_weights)
+            self.decoder.load_state_dict(ckpt['g_ema'], strict=False)
+            self.__load_latent_avg(ckpt, repeat=self.encoder.style_count)
+
+    def forward(self, x, resize=True, latent_mask=None, input_code=False, randomize_noise=True,
+                inject_latent=None, return_latents=False, alpha=None):
+        if input_code:
+            codes = x
+        else:
+            codes = self.encoder(x)
+            # normalize with respect to the center of an average face
+            if self.opts.start_from_latent_avg:
+                if codes.ndim == 2:
+                    codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1)[:, 0, :]
+                else:
+                    codes = codes + self.latent_avg.repeat(codes.shape[0], 1, 1)
+
+        if latent_mask is not None:
+            for i in latent_mask:
+                if inject_latent is not None:
+                    if alpha is not None:
+                        codes[:, i] = alpha * inject_latent[:, i] + (1 - alpha) * codes[:, i]
+                    else:
+                        codes[:, i] = inject_latent[:, i]
+                else:
+                    codes[:, i] = 0
+
+        input_is_latent = not input_code
+        images, result_latent = self.decoder([codes],
+                                             input_is_latent=input_is_latent,
+                                             randomize_noise=randomize_noise,
+                                             return_latents=return_latents)
+
+        if resize:
+            images = self.face_pool(images)
+
+        if return_latents:
+            return images, result_latent
+        else:
+            return images
+
+    def __load_latent_avg(self, ckpt, repeat=None):
+        if 'latent_avg' in ckpt:
+            self.latent_avg = ckpt['latent_avg'].to(self.opts.device)
+            if repeat is not None:
+                self.latent_avg = self.latent_avg.repeat(repeat, 1)
+        else:
+            self.latent_avg = None

stylegan_human/pti/pti_models/e4e/stylegan2/model.py

+import math
+import random
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from .op.fused_act import FusedLeakyReLU, fused_leaky_relu
+from .op.upfirdn2d import upfirdn2d
+
+
+class PixelNorm(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, input):
+        return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
+
+
+def make_kernel(k):
+    k = torch.tensor(k, dtype=torch.float32)
+
+    if k.ndim == 1:
+        k = k[None, :] * k[:, None]
+
+    k /= k.sum()
+
+    return k
+
+
+class Upsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel) * (factor ** 2)
+        self.register_buffer('kernel', kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2 + factor - 1
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=self.factor, down=1, pad=self.pad)
+
+        return out
+
+
+class Downsample(nn.Module):
+    def __init__(self, kernel, factor=2):
+        super().__init__()
+
+        self.factor = factor
+        kernel = make_kernel(kernel)
+        self.register_buffer('kernel', kernel)
+
+        p = kernel.shape[0] - factor
+
+        pad0 = (p + 1) // 2
+        pad1 = p // 2
+
+        self.pad = (pad0, pad1)
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, up=1, down=self.factor, pad=self.pad)
+
+        return out
+
+
+class Blur(nn.Module):
+    def __init__(self, kernel, pad, upsample_factor=1):
+        super().__init__()
+
+        kernel = make_kernel(kernel)
+
+        if upsample_factor > 1:
+            kernel = kernel * (upsample_factor ** 2)
+
+        self.register_buffer('kernel', kernel)
+
+        self.pad = pad
+
+    def forward(self, input):
+        out = upfirdn2d(input, self.kernel, pad=self.pad)
+
+        return out
+
+
+class EqualConv2d(nn.Module):
+    def __init__(
+            self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(
+            torch.randn(out_channel, in_channel, kernel_size, kernel_size)
+        )
+        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
+
+        self.stride = stride
+        self.padding = padding
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channel))
+
+        else:
+            self.bias = None
+
+    def forward(self, input):
+        out = F.conv2d(
+            input,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},'
+            f' {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})'
+        )
+
+
+class EqualLinear(nn.Module):
+    def __init__(
+            self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None
+    ):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
+
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
+
+        else:
+            self.bias = None
+
+        self.activation = activation
+
+        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
+        self.lr_mul = lr_mul
+
+    def forward(self, input):
+        if self.activation:
+            out = F.linear(input, self.weight * self.scale)
+            out = fused_leaky_relu(out, self.bias * self.lr_mul)
+
+        else:
+            out = F.linear(
+                input, self.weight * self.scale, bias=self.bias * self.lr_mul
+            )
+
+        return out
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})'
+        )
+
+
+class ScaledLeakyReLU(nn.Module):
+    def __init__(self, negative_slope=0.2):
+        super().__init__()
+
+        self.negative_slope = negative_slope
+
+    def forward(self, input):
+        out = F.leaky_relu(input, negative_slope=self.negative_slope)
+
+        return out * math.sqrt(2)
+
+
+class ModulatedConv2d(nn.Module):
+    def __init__(
+            self,
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            demodulate=True,
+            upsample=False,
+            downsample=False,
+            blur_kernel=[1, 3, 3, 1],
+    ):
+        super().__init__()
+
+        self.eps = 1e-8
+        self.kernel_size = kernel_size
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+        self.upsample = upsample
+        self.downsample = downsample
+
+        if upsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) - (kernel_size - 1)
+            pad0 = (p + 1) // 2 + factor - 1
+            pad1 = p // 2 + 1
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1), upsample_factor=factor)
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            self.blur = Blur(blur_kernel, pad=(pad0, pad1))
+
+        fan_in = in_channel * kernel_size ** 2
+        self.scale = 1 / math.sqrt(fan_in)
+        self.padding = kernel_size // 2
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channel, in_channel, kernel_size, kernel_size)
+        )
+
+        self.modulation = EqualLinear(style_dim, in_channel, bias_init=1)
+
+        self.demodulate = demodulate
+
+    def __repr__(self):
+        return (
+            f'{self.__class__.__name__}({self.in_channel}, {self.out_channel}, {self.kernel_size}, '
+            f'upsample={self.upsample}, downsample={self.downsample})'
+        )
+
+    def forward(self, input, style):
+        batch, in_channel, height, width = input.shape
+
+        style = self.modulation(style).view(batch, 1, in_channel, 1, 1)
+        weight = self.scale * self.weight * style
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + 1e-8)
+            weight = weight * demod.view(batch, self.out_channel, 1, 1, 1)
+
+        weight = weight.view(
+            batch * self.out_channel, in_channel, self.kernel_size, self.kernel_size
+        )
+
+        if self.upsample:
+            input = input.view(1, batch * in_channel, height, width)
+            weight = weight.view(
+                batch, self.out_channel, in_channel, self.kernel_size, self.kernel_size
+            )
+            weight = weight.transpose(1, 2).reshape(
+                batch * in_channel, self.out_channel, self.kernel_size, self.kernel_size
+            )
+            out = F.conv_transpose2d(input, weight, padding=0, stride=2, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+            out = self.blur(out)
+
+        elif self.downsample:
+            input = self.blur(input)
+            _, _, height, width = input.shape
+            input = input.view(1, batch * in_channel, height, width)
+            out = F.conv2d(input, weight, padding=0, stride=2, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        else:
+            input = input.view(1, batch * in_channel, height, width)
+            out = F.conv2d(input, weight, padding=self.padding, groups=batch)
+            _, _, height, width = out.shape
+            out = out.view(batch, self.out_channel, height, width)
+
+        return out
+
+
+class NoiseInjection(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.zeros(1))
+
+    def forward(self, image, noise=None):
+        if noise is None:
+            batch, _, height, width = image.shape
+            noise = image.new_empty(batch, 1, height, width).normal_()
+
+        return image + self.weight * noise
+
+
+class ConstantInput(nn.Module):
+    def __init__(self, channel, size=4):
+        super().__init__()
+
+        self.input = nn.Parameter(torch.randn(1, channel, size, size // 2))
+
+    def forward(self, input):
+        batch = input.shape[0]
+        out = self.input.repeat(batch, 1, 1, 1)
+
+        return out
+
+
+class StyledConv(nn.Module):
+    def __init__(
+            self,
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            upsample=False,
+            blur_kernel=[1, 3, 3, 1],
+            demodulate=True,
+    ):
+        super().__init__()
+
+        self.conv = ModulatedConv2d(
+            in_channel,
+            out_channel,
+            kernel_size,
+            style_dim,
+            upsample=upsample,
+            blur_kernel=blur_kernel,
+            demodulate=demodulate,
+        )
+
+        self.noise = NoiseInjection()
+        # self.bias = nn.Parameter(torch.zeros(1, out_channel, 1, 1))
+        # self.activate = ScaledLeakyReLU(0.2)
+        self.activate = FusedLeakyReLU(out_channel)
+
+    def forward(self, input, style, noise=None):
+        out = self.conv(input, style)
+        out = self.noise(out, noise=noise)
+        # out = out + self.bias
+        out = self.activate(out)
+
+        return out
+
+
+class ToRGB(nn.Module):
+    def __init__(self, in_channel, style_dim, upsample=True, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        if upsample:
+            self.upsample = Upsample(blur_kernel)
+
+        self.conv = ModulatedConv2d(in_channel, 3, 1, style_dim, demodulate=False)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, input, style, skip=None):
+        out = self.conv(input, style)
+        out = out + self.bias
+
+        if skip is not None:
+            skip = self.upsample(skip)
+
+            out = out + skip
+
+        return out
+
+
+class Generator(nn.Module):
+    def __init__(
+            self,
+            size,
+            style_dim,
+            n_mlp,
+            channel_multiplier=2,
+            blur_kernel=[1, 3, 3, 1],
+            lr_mlp=0.01,
+    ):
+        super().__init__()
+
+        self.size = size
+
+        self.style_dim = style_dim
+
+        layers = [PixelNorm()]
+
+        for i in range(n_mlp):
+            layers.append(
+                EqualLinear(
+                    style_dim, style_dim, lr_mul=lr_mlp, activation='fused_lrelu'
+                )
+            )
+
+        self.style = nn.Sequential(*layers)
+
+        self.channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        self.input = ConstantInput(self.channels[4])
+        self.conv1 = StyledConv(
+            self.channels[4], self.channels[4], 3, style_dim, blur_kernel=blur_kernel
+        )
+        self.to_rgb1 = ToRGB(self.channels[4], style_dim, upsample=False)
+
+        self.log_size = int(math.log(size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+
+        self.convs = nn.ModuleList()
+        self.upsamples = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channel = self.channels[4]
+
+        for layer_idx in range(self.num_layers):
+            res = (layer_idx + 5) // 2
+            shape = [1, 1, 2 ** res, 2 ** res // 2]
+            self.noises.register_buffer(
+                "noise_{}".format(layer_idx), torch.randn(*shape)
+            )
+
+        for i in range(3, self.log_size + 1):
+            out_channel = self.channels[2 ** i]
+
+            self.convs.append(
+                StyledConv(
+                    in_channel,
+                    out_channel,
+                    3,
+                    style_dim,
+                    upsample=True,
+                    blur_kernel=blur_kernel,
+                )
+            )
+
+            self.convs.append(
+                StyledConv(
+                    out_channel, out_channel, 3, style_dim, blur_kernel=blur_kernel
+                )
+            )
+
+            self.to_rgbs.append(ToRGB(out_channel, style_dim))
+
+            in_channel = out_channel
+
+        self.n_latent = self.log_size * 2 - 2
+
+    def make_noise(self):
+        device = self.input.input.device
+
+        noises = [torch.randn(1, 1, 2 ** 2, 2 ** 2 // 2, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2 ** i, 2 ** i // 2, device=device))
+
+        return noises
+
+    def mean_latent(self, n_latent):
+        latent_in = torch.randn(
+            n_latent, self.style_dim, device=self.input.input.device
+        )
+        latent = self.style(latent_in).mean(0, keepdim=True)
+
+        return latent
+
+    def get_latent(self, input):
+        return self.style(input)
+
+    def forward(
+            self,
+            styles,
+            return_latents=False,
+            return_features=False,
+            inject_index=None,
+            truncation=1,
+            truncation_latent=None,
+            input_is_latent=False,
+            noise=None,
+            randomize_noise=True,
+    ):
+        if not input_is_latent:
+            styles = [self.style(s) for s in styles]
+
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers
+            else:
+                noise = [
+                    getattr(self.noises, f'noise_{i}') for i in range(self.num_layers)
+                ]
+
+        if truncation < 1:
+            style_t = []
+
+            for style in styles:
+                style_t.append(
+                    truncation_latent + truncation * (style - truncation_latent)
+                )
+
+            styles = style_t
+
+        if len(styles) < 2:
+            inject_index = self.n_latent
+            if styles[0].ndim < 3:
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:
+                latent = styles[0]
+
+        else:
+            if inject_index is None:
+                inject_index = random.randint(1, self.n_latent - 1)
+
+            # latent = styles[0].unsqueeze(0)
+            # if latent.shape[1] == 1:
+            #     latent = latent.repeat(1, inject_index, 1)
+            # else:
+            #     latent = latent[:, :inject_index, :]
+            latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.n_latent - inject_index, 1)
+            # latent = styles[0][:, :inject_index, :]
+            # latent2 = styles[1][:, inject_index:, :]
+            latent = torch.cat([latent, latent2], 1)
+        out = self.input(latent)
+        out = self.conv1(out, latent[:, 0], noise=noise[0])
+
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(
+                self.convs[::2], self.convs[1::2], noise[1::2], noise[2::2], self.to_rgbs
+        ):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)
+
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        elif return_features:
+            return image, out
+        else:
+            return image, None
+
+
+class ConvLayer(nn.Sequential):
+    def __init__(
+            self,
+            in_channel,
+            out_channel,
+            kernel_size,
+            downsample=False,
+            blur_kernel=[1, 3, 3, 1],
+            bias=True,
+            activate=True,
+    ):
+        layers = []
+
+        if downsample:
+            factor = 2
+            p = (len(blur_kernel) - factor) + (kernel_size - 1)
+            pad0 = (p + 1) // 2
+            pad1 = p // 2
+
+            layers.append(Blur(blur_kernel, pad=(pad0, pad1)))
+
+            stride = 2
+            self.padding = 0
+
+        else:
+            stride = 1
+            self.padding = kernel_size // 2
+
+        layers.append(
+            EqualConv2d(
+                in_channel,
+                out_channel,
+                kernel_size,
+                padding=self.padding,
+                stride=stride,
+                bias=bias and not activate,
+            )
+        )
+
+        if activate:
+            if bias:
+                layers.append(FusedLeakyReLU(out_channel))
+
+            else:
+                layers.append(ScaledLeakyReLU(0.2))
+
+        super().__init__(*layers)
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_channel, out_channel, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        self.conv1 = ConvLayer(in_channel, in_channel, 3)
+        self.conv2 = ConvLayer(in_channel, out_channel, 3, downsample=True)
+
+        self.skip = ConvLayer(
+            in_channel, out_channel, 1, downsample=True, activate=False, bias=False
+        )
+
+    def forward(self, input):
+        out = self.conv1(input)
+        out = self.conv2(out)
+
+        skip = self.skip(input)
+        out = (out + skip) / math.sqrt(2)
+
+        return out
+
+
+class Discriminator(nn.Module):
+    def __init__(self, size, channel_multiplier=2, blur_kernel=[1, 3, 3, 1]):
+        super().__init__()
+
+        channels = {
+            4: 512,
+            8: 512,
+            16: 512,
+            32: 512,
+            64: 256 * channel_multiplier,
+            128: 128 * channel_multiplier,
+            256: 64 * channel_multiplier,
+            512: 32 * channel_multiplier,
+            1024: 16 * channel_multiplier,
+        }
+
+        convs = [ConvLayer(3, channels[size], 1)]
+
+        log_size = int(math.log(size, 2))
+
+        in_channel = channels[size]
+
+        for i in range(log_size, 2, -1):
+            out_channel = channels[2 ** (i - 1)]
+
+            convs.append(ResBlock(in_channel, out_channel, blur_kernel))
+
+            in_channel = out_channel
+
+        self.convs = nn.Sequential(*convs)
+
+        self.stddev_group = 4
+        self.stddev_feat = 1
+
+        self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
+        self.final_linear = nn.Sequential(
+            EqualLinear(channels[4] * 4 * 4 // 2, channels[4], activation='fused_lrelu'),
+            EqualLinear(channels[4], 1),
+        )
+
+    def forward(self, input):
+        out = self.convs(input)
+
+        batch, channel, height, width = out.shape
+        group = min(batch, self.stddev_group)
+        stddev = out.view(
+            group, -1, self.stddev_feat, channel // self.stddev_feat, height, width
+        )
+        stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
+        stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
+        stddev = stddev.repeat(group, 1, height, width)
+        out = torch.cat([out, stddev], 1)
+
+        out = self.final_conv(out)
+
+        out = out.view(batch, -1)
+        out = self.final_linear(out)
+
+        return out

stylegan_human/pti/pti_models/e4e/stylegan2/op/__init__.py

+from .fused_act import FusedLeakyReLU, fused_leaky_relu
+from .upfirdn2d import upfirdn2d

stylegan_human/pti/pti_models/e4e/stylegan2/op/fused_act.py

+import os
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.autograd import Function
+
+
+module_path = os.path.dirname(__file__)
+
+
+
+class FusedLeakyReLU(nn.Module):
+    def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5):
+        super().__init__()
+
+        self.bias = nn.Parameter(torch.zeros(channel))
+        self.negative_slope = negative_slope
+        self.scale = scale
+
+    def forward(self, input):
+        return fused_leaky_relu(input, self.bias, self.negative_slope, self.scale)
+
+
+def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
+    rest_dim = [1] * (input.ndim - bias.ndim - 1)
+    input = input.cuda()
+    return (
+        F.leaky_relu(
+            input + bias.view(1, bias.shape[0], *rest_dim), negative_slope=negative_slope
+        )
+        * scale
+    )
+

stylegan_human/pti/pti_models/e4e/stylegan2/op/fused_bias_act.cpp

+#include <torch/extension.h>
+
+
+torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale);
+
+#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x)
+
+torch::Tensor fused_bias_act(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale) {
+    CHECK_CUDA(input);
+    CHECK_CUDA(bias);
+
+    return fused_bias_act_op(input, bias, refer, act, grad, alpha, scale);
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("fused_bias_act", &fused_bias_act, "fused bias act (CUDA)");
+}
\ No newline at end of file

stylegan_human/pti/pti_models/e4e/stylegan2/op/fused_bias_act_kernel.cu

+// 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
+
+#include <torch/types.h>
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+
+template <typename scalar_t>
+static __global__ void fused_bias_act_kernel(scalar_t* out, const scalar_t* p_x, const scalar_t* p_b, const scalar_t* p_ref,
+    int act, int grad, scalar_t alpha, scalar_t scale, int loop_x, int size_x, int step_b, int size_b, int use_bias, int use_ref) {
+    int xi = blockIdx.x * loop_x * blockDim.x + threadIdx.x;
+
+    scalar_t zero = 0.0;
+
+    for (int loop_idx = 0; loop_idx < loop_x && xi < size_x; loop_idx++, xi += blockDim.x) {
+        scalar_t x = p_x[xi];
+
+        if (use_bias) {
+            x += p_b[(xi / step_b) % size_b];
+        }
+
+        scalar_t ref = use_ref ? p_ref[xi] : zero;
+
+        scalar_t y;
+
+        switch (act * 10 + grad) {
+            default:
+            case 10: y = x; break;
+            case 11: y = x; break;
+            case 12: y = 0.0; break;
+
+            case 30: y = (x > 0.0) ? x : x * alpha; break;
+            case 31: y = (ref > 0.0) ? x : x * alpha; break;
+            case 32: y = 0.0; break;
+        }
+
+        out[xi] = y * scale;
+    }
+}
+
+
+torch::Tensor fused_bias_act_op(const torch::Tensor& input, const torch::Tensor& bias, const torch::Tensor& refer,
+    int act, int grad, float alpha, float scale) {
+    int curDevice = -1;
+    cudaGetDevice(&curDevice);
+    cudaStream_t stream = at::cuda::getCurrentCUDAStream(curDevice);
+
+    auto x = input.contiguous();
+    auto b = bias.contiguous();
+    auto ref = refer.contiguous();
+
+    int use_bias = b.numel() ? 1 : 0;
+    int use_ref = ref.numel() ? 1 : 0;
+
+    int size_x = x.numel();
+    int size_b = b.numel();
+    int step_b = 1;
+
+    for (int i = 1 + 1; i < x.dim(); i++) {
+        step_b *= x.size(i);
+    }
+
+    int loop_x = 4;
+    int block_size = 4 * 32;
+    int grid_size = (size_x - 1) / (loop_x * block_size) + 1;
+
+    auto y = torch::empty_like(x);
+
+    AT_DISPATCH_FLOATING_TYPES_AND_HALF(x.scalar_type(), "fused_bias_act_kernel", [&] {
+        fused_bias_act_kernel<scalar_t><<<grid_size, block_size, 0, stream>>>(
+            y.data_ptr<scalar_t>(),
+            x.data_ptr<scalar_t>(),
+            b.data_ptr<scalar_t>(),
+            ref.data_ptr<scalar_t>(),
+            act,
+            grad,
+            alpha,
+            scale,
+            loop_x,
+            size_x,
+            step_b,
+            size_b,
+            use_bias,
+            use_ref
+        );
+    });
+
+    return y;
+}
\ No newline at end of file