AI face editor makes Lena smile

1. Get into AI Face editing page case page , And complete the basic configuration

AI Face editing  (huaweicloud.com)

Click on  

Run in ModelArts

, Get into  JupyterLab  page .

Wait for initialization

​ Switch specifications , And select  [ Time limited free ]GPU: 1*V100|CPU: 8 nucleus  64GB 

​

​ Resource switching is complete , Click ok .

Click on the top right corner Select Kernel： choice

PyTorch-1.4

​

​2. Download code and data and install dependencies

!pip install ninja

!pip install dlib

!pip uninstall -y torch
!pip uninstall -y torchvision
!pip install torch==1.6.0
!pip install torchvision==0.7.0

After installation , Need to restart kernel, Click on the above Restart the kernel

Get into HFGI Under the path ：

%cd HFGI

3. Start running code

#@title Setup Repository
import os
from argparse import Namespace
import time
import os
import sys
import numpy as np
from PIL import Image
import torch
import torchvision.transforms as transforms
 
 
# from utils.common import tensor2im
from models.psp import pSp # we use the pSp framework to load the e4e encoder.
 
%load_ext autoreload
%autoreload 2
def tensor2im(var):
 # var shape: (3, H, W)
 var = var.cpu().detach().transpose(0, 2).transpose(0, 1).numpy()
 var = ((var + 1) / 2)
 var[var < 0] = 0
 var[var > 1] = 1
 var = var * 255
 return Image.fromarray(var.astype('uint8'))

Load pre training model

model_path = "checkpoint/ckpt.pt"
ckpt = torch.load(model_path, map_location='cpu')
opts = ckpt['opts']
opts['is_train'] = False
opts['checkpoint_path'] = model_path
opts= Namespace(**opts)
net = pSp(opts)
net.eval()
net.cuda()
print('Model successfully loaded!')

Set the input image

# Setup required image transformations
EXPERIMENT_ARGS = {
 "image_path": "test_imgs/Lina.jpg"
 }
 
EXPERIMENT_ARGS['transform'] = transforms.Compose([
 transforms.Resize((256, 256)),
 transforms.ToTensor(),
 transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])
resize_dims = (256, 256)

Drag photos here to upload

image_path = EXPERIMENT_ARGS["image_path"]
original_image = Image.open(image_path)
original_image = original_image.convert("RGB")
 
run_align = True

image alignment

import numpy as np
import PIL
import PIL.Image
import scipy
import scipy.ndimage
import dlib
 
 
def get_landmark(filepath, predictor):
 """get landmark with dlib
 :return: np.array shape=(68, 2)
 """
 detector = dlib.get_frontal_face_detector()
 
 img = dlib.load_rgb_image(filepath)
 dets = detector(img, 1)
 
 for k, d in enumerate(dets):
 shape = predictor(img, d)
 
 t = list(shape.parts())
 a = []
 for tt in t:
 a.append([tt.x, tt.y])
 lm = np.array(a)
 return lm
 
 
def align_face(filepath, predictor):
 """
 :param filepath: str
 :return: PIL Image
 """
 
 lm = get_landmark(filepath, predictor)
 
 lm_chin = lm[0: 17] # left-right
 lm_eyebrow_left = lm[17: 22] # left-right
 lm_eyebrow_right = lm[22: 27] # left-right
 lm_nose = lm[27: 31] # top-down
 lm_nostrils = lm[31: 36] # top-down
 lm_eye_left = lm[36: 42] # left-clockwise
 lm_eye_right = lm[42: 48] # left-clockwise
 lm_mouth_outer = lm[48: 60] # left-clockwise
 lm_mouth_inner = lm[60: 68] # left-clockwise
 
 # Calculate auxiliary vectors.
 eye_left = np.mean(lm_eye_left, axis=0)
 eye_right = np.mean(lm_eye_right, axis=0)
 eye_avg = (eye_left + eye_right) * 0.5
 eye_to_eye = eye_right - eye_left
 mouth_left = lm_mouth_outer[0]
 mouth_right = lm_mouth_outer[6]
 mouth_avg = (mouth_left + mouth_right) * 0.5
 eye_to_mouth = mouth_avg - eye_avg
 
 # Choose oriented crop rectangle.
 x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
 x /= np.hypot(*x)
 x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
 y = np.flipud(x) * [-1, 1]
 c = eye_avg + eye_to_mouth * 0.1
 quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
 qsize = np.hypot(*x) * 2
 
 # read image
 img = PIL.Image.open(filepath)
 
 output_size = 256
 transform_size = 256
 enable_padding = True
 
 # Shrink.
 shrink = int(np.floor(qsize / output_size * 0.5))
 if shrink > 1:
 rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
 img = img.resize(rsize, PIL.Image.ANTIALIAS)
 quad /= shrink
 qsize /= shrink
 
 # Crop.
 border = max(int(np.rint(qsize * 0.1)), 3)
 crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
 int(np.ceil(max(quad[:, 1]))))
 crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
 min(crop[3] + border, img.size[1]))
 if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
 img = img.crop(crop)
 quad -= crop[0:2]
 
 # Pad.
 pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
 int(np.ceil(max(quad[:, 1]))))
 pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
 max(pad[3] - img.size[1] + border, 0))
 if enable_padding and max(pad) > border - 4:
 pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
 img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
 h, w, _ = img.shape
 y, x, _ = np.ogrid[:h, :w, :1]
 mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
 1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
 blur = qsize * 0.02
 img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
 img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
 img = PIL.Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
 quad += pad[:2]
 
 # Transform.
 img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(), PIL.Image.BILINEAR)
 if output_size < transform_size:
 img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
 
 # Return aligned image.
 return img
if 'shape_predictor_68_face_landmarks.dat' not in os.listdir():
# !wget http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
 !bzip2 -dk shape_predictor_68_face_landmarks.dat.bz2
 
def run_alignment(image_path):
 import dlib
 predictor = dlib.shape_predictor(&quot;shape_predictor_68_face_landmarks.dat&quot;)
 aligned_image = align_face(filepath=image_path, predictor=predictor) 
 print(&quot;Aligned image has shape: {}&quot;.format(aligned_image.size))
 return aligned_image 
 
if run_align:
 input_image = run_alignment(image_path)
else:
 input_image = original_image
 
input_image.resize(resize_dims)

High fidelity inverse mapping

def display_alongside_source_image(result_image, source_image):
 res = np.concatenate([np.array(source_image.resize(resize_dims)),
 np.array(result_image.resize(resize_dims))], axis=1)
 return Image.fromarray(res)
 
def get_latents(net, x, is_cars=False):
 codes = net.encoder(x)
 if net.opts.start_from_latent_avg:
 if codes.ndim == 2:
 codes = codes + net.latent_avg.repeat(codes.shape[0], 1, 1)[:, 0, :]
 else:
 codes = codes + net.latent_avg.repeat(codes.shape[0], 1, 1)
 if codes.shape[1] == 18 and is_cars:
 codes = codes[:, :16, :]
 return codes
with torch.no_grad():
 x = transformed_image.unsqueeze(0).cuda()
 
 tic = time.time()
 latent_codes = get_latents(net, x)
 
 # calculate the distortion map
 imgs, _ = net.decoder([latent_codes[0].unsqueeze(0).cuda()],None, input_is_latent=True, randomize_noise=False, return_latents=True)
 res = x - torch.nn.functional.interpolate(torch.clamp(imgs, -1., 1.), size=(256,256) , mode='bilinear')
 
 # ADA
 img_edit = torch.nn.functional.interpolate(torch.clamp(imgs, -1., 1.), size=(256,256) , mode='bilinear')
 res_align = net.grid_align(torch.cat((res, img_edit ), 1))
 
 # consultation fusion
 conditions = net.residue(res_align)
 
 result_image, _ = net.decoder([latent_codes],conditions, input_is_latent=True, randomize_noise=False, return_latents=True)
 toc = time.time()
 print('Inference took {:.4f} seconds.'.format(toc - tic))
 
# Display inversion:
display_alongside_source_image(tensor2im(result_image[0]), input_image)

Running results ：

High fidelity image editing

from editings import latent_editor
editor = latent_editor.LatentEditor(net.decoder)
# interface-GAN
interfacegan_directions = {
 'age': './editings/interfacegan_directions/age.pt',
 'smile': './editings/interfacegan_directions/smile.pt' }
edit_direction = torch.load(interfacegan_directions['smile']).cuda() 
edit_degree = 1.5 #  Set the smile range
img_edit, edit_latents = editor.apply_interfacegan(latent_codes[0].unsqueeze(0).cuda(), edit_direction, factor=edit_degree) #  Set smile
# align the distortion map
img_edit = torch.nn.functional.interpolate(torch.clamp(img_edit, -1., 1.), size=(256,256) , mode='bilinear')
res_align = net.grid_align(torch.cat((res, img_edit ), 1))
 
# fusion
conditions = net.residue(res_align)
result, _ = net.decoder([edit_latents],conditions, input_is_latent=True, randomize_noise=False, return_latents=True)
 
result = torch.nn.functional.interpolate(result, size=(256,256) , mode='bilinear')
display_alongside_source_image(tensor2im(result[0]), input_image)

Running results :