（附代码）实战 | PyTorch+OpenCV进行人脸识别

人脸检测

pip install split-folders

import splitfolders
splitfolders.ratio('dataset', output="/data", seed=1337, ratio=(.8, 0.2))

import fnmatch
import os
from matplotlib import pyplot as plt
import cv2

# Load the cascade
face_cascade = cv2.CascadeClassifier('/haarcascade_frontalface_default.xml')

paths="/data/"


for root,_,files in os.walk(paths):
    for filename in files: 
        file = os.path.join(root,filename)
        if fnmatch.fnmatch(file,'*.jpg'):
            
            img = cv2.imread(file)        
            gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
            # Detect faces
            faces = face_cascade.detectMultiScale(gray, 1.1, 4)
            # Draw rectangle around the faces
            for (x, y, w, h) in faces:
              crop_face = img[y:y+h, x:x+w]
            path = os.path.join(root,filename)
            cv2.imwrite(path,crop_face)

from torch import nn, optim, as_tensor
from torch.utils.data import Dataset, DataLoader
import torch.nn.functional as F
from torch.optim import lr_scheduler
from torch.nn.init import *
from torchvision import transforms, utils, datasets, models
import cv2
from PIL import Image
from pdb import set_trace
import time
import copy
from pathlib import Path
import os
import sys
import matplotlib.pyplot as plt
import matplotlib.animation as animation
from skimage import io, transform
from tqdm import trange, tqdm
import csv
import glob
import dlib
import pandas as pd
import numpy as np

data_transforms = {
    'train': transforms.Compose([
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Scale((224,224)),
        transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.4),
        transforms.RandomRotation(5, resample=False,expand=False, center=None),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.ToTensor(),
        transforms.Scale((224,224)),
       transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.4),
        transforms.RandomRotation(5, resample=False,expand=False, center=None),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
    ]),
}
data_dir = '/content/drive/MyDrive/AttendanceCapturingSystem/data/'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
                                          data_transforms[x])
                  for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x],
                                              batch_size=8, 
                                             shuffle=True)
              for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train','val']}
class_names = image_datasets['train'].classes
class_names

def imshow(inp, title=None):
    """Imshow for Tensor."""
    inp = inp.numpy().transpose((1, 2, 0))
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])
    inp = std * inp + mean
    inp = np.clip(inp, 0, 1)
    plt.imshow(inp)
    if title is not None:
        plt.title(title)
    plt.pause(0.001)  # pause a bit so that plots are updated
# Get a batch of training data
inputs, classes = next(iter(dataloaders['train']))
# Make a grid from batch
out = utils.make_grid(inputs)
imshow(out, title=[class_names[x] for x in classes])

from models.inception_resnet_v1 import InceptionResnetV1
print('Running on device: {}'.format(device))

model_ft = InceptionResnetV1(pretrained='vggface2', classify=False, num_classes = len(class_names))

list(model_ft.children())[-6:]

layer_list = list(model_ft.children())[-5:] # all final layers

model_ft = nn.Sequential(*list(model_ft.children())[:-5])

for param in model_ft.parameters():
    param.requires_grad = False
    
class Flatten(nn.Module):
    def __init__(self):
        super(Flatten, self).__init__()
        
    def forward(self, x):
        x = x.view(x.size(0), -1)
        return x
      
class normalize(nn.Module):
    def __init__(self):
        super(normalize, self).__init__()
        
    def forward(self, x):
        x = F.normalize(x, p=2, dim=1)
        return x    
model_ft.avgpool_1a = nn.AdaptiveAvgPool2d(output_size=1)

model_ft.last_linear = nn.Sequential(
    Flatten(),
    nn.Linear(in_features=1792, out_features=512, bias=False),
    normalize()
)

model_ft.logits = nn.Linear(layer_list[2].out_features, len(class_names))
model_ft.softmax = nn.Softmax(dim=1)
model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()

# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=1e-2, momentum=0.9)

# Decay LR by a factor of *gamma* every *step_size* epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)  

model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=1e-2, momentum=0.9)
# Decay LR by a factor of *gamma* every *step_size* epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)

model_ft

def train_model(model, criterion, optimizer, scheduler,
                num_epochs=25):
    since = time.time()
    FT_losses = []
    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0
    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)
    # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # Set model to training mode
            else:
                model.eval()   # Set model to evaluate mode
            running_loss = 0.0
            running_corrects = 0
            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)
                # zero the parameter gradients
                optimizer.zero_grad()
                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)
                    # backward + optimize only if in training phase
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()
                        scheduler.step()
                
                FT_losses.append(loss.item())
                # statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]
            print('{} Loss: {:.4f} Acc: {:.4f}'.format(
                phase, epoch_loss, epoch_acc))
            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())
    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}'.format(best_acc))
    # load best model weights
    model.load_state_dict(best_model_wts)
    return model, FT_losses

model_ft, FT_losses = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=200)
plt.figure(figsize=(10,5))
plt.title("FRT Loss During Training")
plt.plot(FT_losses, label="FT loss")
plt.xlabel("iterations")
plt.ylabel("Loss")
plt.legend()
plt.show()

torch.save(model, "/model.pt")

import fnmatch
import os
from matplotlib import pyplot as plt
import cv2
from facenet_pytorch import MTCNN, InceptionResnetV1
resnet = InceptionResnetV1(pretrained='vggface2').eval()
# Load the cascade
face_cascade = cv2.CascadeClassifier('/haarcascade_frontalface_default.xml')

def face_match(img_path, data_path): # img_path= location of photo, data_path= location of data.pt 
    # getting embedding matrix of the given img
    img = cv2.imread(img_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    # Detect faces
    faces = face_cascade.detectMultiScale(gray, 1.1, 4)
    # Draw rectangle around the faces
    for (x, y, w, h) in faces:
        crop_face = img[y:y+h, x:x+w]
            
    img = cv2.imwrite(img_path,crop_face)
    emb = resnet(img.unsqueeze(0)).detach() # detech is to make required gradient false
    
    saved_data = torch.load('model.pt') # loading data.pt file
    embedding_list = saved_data[0] # getting embedding data
    name_list = saved_data[1] # getting list of names
    dist_list = [] # list of matched distances, minimum distance is used to identify the person
    
    for idx, emb_db in enumerate(embedding_list):
        dist = torch.dist(emb, emb_db).item()
        dist_list.append(dist)
        
    idx_min = dist_list.index(min(dist_list))
    return (name_list[idx_min], min(dist_list))


result = face_match('trainset/0006/0006_0000546/0006_0000546_script.jpg', '/model.pt')

print('Face matched with: ',result[0], 'With distance: ',result[1])