/face.evoLVe.PyTorch

🔥🔥High-Performance Face Recognition Library on PyTorch🔥🔥

Primary LanguagePythonMIT LicenseMIT

face.evoLVe: High-Performance Face Recognition Library based on PyTorch

  • Evolve to be more comprehensive, effective and efficient for face related analytics & applications!
  • About the name:
    • "face" means this repo is dedicated for face related analytics & applications.
    • "evolve" means unleash your greatness to be better and better. "LV" are capitalized to acknowledge the nurturing of Learning and Vision (LV) group, Nation University of Singapore (NUS).
  • This work was done during Jian Zhao served as a short-term "Texpert" Research Scientist at Tencent FiT DeepSea AI Lab, Shenzhen, China.
Author Jian Zhao
Homepage https://zhaoj9014.github.io

License

The code of face.evoLVe is released under the MIT License.

News

🚩 OPEN 17 Feb 2019: We are training ResNet-50, IR-SE-50 and IR-SE-152 models on MS-Celeb-1M_Align_112x112, and will release them soon.

🚩 OPEN 01 Feb 2019: We are training a IR-152 model on MS-Celeb-1M_Align_112x112, and will release it soon.

✅ CLOSED 23 Jan 2019: The current distributed training schema with multi-GPUs under PyTorch and other mainstream platforms parallels the backbone across multi-GPUs while relying on a single master to compute the final bottleneck (fully-connected/softmax) layer. This is not an issue for conventional face recognition with moderate number of identities. However, it struggles with large-scale face recognition, which requires recognizing millions of identities in the real world. The master can hardly hold the oversized final layer while the slaves still have redundant computation resource, leading to small-batch training or even failed training. To address this problem, we are developing a highly-elegant, effective and efficient distributed training schema with multi-GPUs under PyTorch, supporting not only the backbone, but also the head with the fully-connected (softmax) layer, to facilitate high-performance large-scale face recognition. We will added this support into our repo.

🚩 OPEN 23 Jan 2019: We will release MS-Celeb-1M_Align_224x224 soon to facilitate larger-resolution face recongition model training.

✅ CLOSED 22 Jan 2019: We have released two feature extraction APIs for extracting features from pre-trained models, implemented with PyTorch build-in functions and OpenCV, respectively. Please check ./util/extract_feature_v1.py and ./util/extract_feature_v2.py.

✅ CLOSED 22 Jan 2019: We are fine-tuning our released IR-50 model on our private Asia face data, which will be released soon to facilitate high-performance Asia face recognition.

✅ CLOSED 21 Jan 2019: We are training a better-performing IR-50 model on MS-Celeb-1M_Align_112x112, and will replace the current model soon.

Contents

face.evoLVe for High-Performance Face Recognition

Introduction

💁

  • This repo provides a comprehensive face recognition library for face related analytics & applications, including face alignment (detection, landmark localization, affine transformation, etc.), data processing (e.g., augmentation, data balancing, normalization, etc.), various backbones (e.g., ResNet, IR, IR-SE, ResNeXt, SE-ResNeXt, DenseNet, LightCNN, MobileNet, ShuffleNet, DPN, etc.), various losses (e.g., Softmax, Focal, Center, SphereFace, CosFace, AmSoftmax, ArcFace, Triplet, etc.) and bags of tricks for improving performance (e.g., training refinements, model tweaks, knowledge distillation, etc.).
  • The current distributed training schema with multi-GPUs under PyTorch and other mainstream platforms parallels the backbone across multi-GPUs while relying on a single master to compute the final bottleneck (fully-connected/softmax) layer. This is not an issue for conventional face recognition with moderate number of identities. However, it struggles with large-scale face recognition, which requires recognizing millions of identities in the real world. The master can hardly hold the oversized final layer while the slaves still have redundant computation resource, leading to small-batch training or even failed training. To address this problem, this repo provides a highly-elegant, effective and efficient distributed training schema with multi-GPUs under PyTorch, supporting not only the backbone, but also the head with the fully-connected (softmax) layer, to facilitate high-performance large-scale face recognition.
  • All data before & after alignment, source codes and trained models are provided.
  • This repo can help researchers/engineers develop high-performance deep face recognition models and algorithms quickly for practical use and deployment.

Pre-Requisites

🍰

  • Linux or macOS
  • Python 3.7 (for training & validation) and Python 2.7 (for visualization w/ tensorboardX)
  • PyTorch 1.0 (for traininig & validation, install w/ pip install torch torchvision)
  • MXNet 1.3.1 (optinal, for data processing, install w/ pip install mxnet-cu90)
  • TensorFlow 1.12 (optinal, for visualization, install w/ pip install tensorflow-gpu)
  • tensorboardX 1.6 (optinal, for visualization, install w/ pip install tensorboardX)
  • OpenCV 3.4.5 (install w/ pip install opencv-python)
  • bcolz 1.2.0 (install w/ pip install bcolz)

While not required, for optimal performance it is highly recommended to run the code using a CUDA enabled GPU. We used 4-8 NVIDIA Tesla P40 in parallel.

Usage

📙

  • Clone the repo: git clone https://github.com/ZhaoJ9014/face.evoLVe.PyTorch.git.
  • mkdir data checkpoint log at appropriate directory to store your train/val/test data, checkpoints and training logs.
  • Prepare your train/val/test data (refer to Sec. Data Zoo for publicly available face related databases), and ensure each database folder has the following structure: 
    ./data/db_name/
        -> id1/
            -> 1.jpg
            -> ...
        -> id2/
            -> 1.jpg
            -> ...
        -> ...
            -> ...
            -> ...
  • Refer to the codes of corresponding sections for specific purposes.

Face Alignment

📐

  • This section is based on the work of MTCNN.
  • Folder: ./align
  • Face detection, landmark localization APIs and visualization toy example with ipython notebook: 
    from PIL import Image
from detector import detect_faces
from visualization_utils import show_results

img = Image.open('some_img.jpg') # modify the image path to yours
bounding_boxes, landmarks = detect_faces(img) # detect bboxes and landmarks for all faces in the image
show_results(img, bounding_boxes, landmarks) # visualize the results
  • Face alignment API (perform face detection, landmark localization and alignment with affine transformations on a whole database folder source_root with the directory structure as demonstrated in Sec. Usage, and store the aligned results to a new folder dest_root with the same directory structure): 
    python face_align.py -source_root [source_root] -dest_root [dest_root] -crop_size [crop_size]

# python face_align.py -source_root './data/test' -dest_root './data/test_Aligned' -crop_size 112
  • For macOS users, there is no need to worry about *.DS_Store files which may ruin your data, since they will be automatically removed when you run the scripts.
  • Keynotes for customed use: 1) specify the arguments of source_root, dest_root and crop_size to your own values when you run face_align.py; 2) pass your customed min_face_size, thresholds and nms_thresholds values to the detect_faces function of detector.py to match your practical requirements; 3) if you find the speed using face alignment API is a bit slow, you can call face resize API to firstly resize the image whose smaller size is larger than a threshold (specify the arguments of source_root, dest_root and min_side to your own values) before calling the face alignment API: 
    python face_resize.py

Data Processing

📊

  • Folder: ./balance
  • Remove low-shot data API (remove the low-shot classes with less than min_num samples in the training set root with the directory structure as demonstrated in Sec. Usage for data balance and effective model training): 
    python remove_lowshot.py -root [root] -min_num [min_num]

# python remove_lowshot.py -root './data/train' -min_num 10
  • Keynotes for customed use: specify the arguments of root and min_num to your own values when you run remove_lowshot.py.
  • We prefer to include other data processing tricks, e.g., augmentation (flip horizontally, scale hue/satuation/brightness with coefficients uniformly drawn from [0.6,1.4], add PCA noise with a coefficient sampled from a normal distribution N(0,0.1), etc.), weighted random sampling, normalization, etc. to the main training script in Sec. Training and Validation to be self-contained.

Training and Validation

☕

  • Folder: ./

  • Configuration API (configurate your overall settings for training & validation) config.py:

    import torch

configurations = {
    1: dict(
        SEED = 1337, # random seed for reproduce results

        DATA_ROOT = '/media/pc/6T/jasonjzhao/data/faces_emore', # the parent root where your train/val/test data are stored
        MODEL_ROOT = '/media/pc/6T/jasonjzhao/buffer/model', # the root to buffer your checkpoints
        LOG_ROOT = '/media/pc/6T/jasonjzhao/buffer/log', # the root to log your train/val status
        BACKBONE_RESUME_ROOT = './', # the root to resume training from a saved checkpoint
        HEAD_RESUME_ROOT = './', # the root to resume training from a saved checkpoint

        BACKBONE_NAME = 'IR_SE_50', # support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152']
        HEAD_NAME = 'ArcFace', # support:  ['Softmax', 'ArcFace', 'CosFace', 'SphereFace', 'Am_softmax']
        LOSS_NAME = 'Focal', # support: ['Focal', 'Softmax']

        INPUT_SIZE = [112, 112], # support: [112, 112] and [224, 224]
        RGB_MEAN = [0.5, 0.5, 0.5], # for normalize inputs to [-1, 1]
        RGB_STD = [0.5, 0.5, 0.5],
        EMBEDDING_SIZE = 512, # feature dimension
        BATCH_SIZE = 512,
        DROP_LAST = True, # whether drop the last batch to ensure consistent batch_norm statistics
        LR = 0.1, # initial LR
        NUM_EPOCH = 125, # total epoch number (use the firt 1/25 epochs to warm up)
        WEIGHT_DECAY = 5e-4, # do not apply to batch_norm parameters
        MOMENTUM = 0.9,
        STAGES = [35, 65, 95], # epoch stages to decay learning rate

        DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu"),
        MULTI_GPU = True, # flag to use multiple GPUs; if you choose to train with single GPU, you should first run "export CUDA_VISILE_DEVICES=device_id" to specify the GPU card you want to use
        GPU_ID = [0, 1, 2, 3], # specify your GPU ids
        PIN_MEMORY = True,
        NUM_WORKERS = 0,
),
}

  • Train & validation API (all folks about training & validation, i.e., import package, hyperparameters & data loaders, model & loss & optimizer, train & validation & save checkpoint) train.py. Since MS-Celeb-1M serves as an ImageNet in the filed of face recognition, we pre-train the face.evoLVe models on MS-Celeb-1M and perform validation on LFW, CFP_FF, CFP_FP, AgeDB, CALFW, CPLFW and Vggface2_FP. Let's dive into details together step by step.

    • Import necessary packages: 
      import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import torchvision.datasets as datasets

from config import configurations
from backbone.model_resnet import ResNet_50, ResNet_101, ResNet_152
from backbone.model_irse import IR_50, IR_101, IR_152, IR_SE_50, IR_SE_101, IR_SE_152
from head.metrics import ArcFace, CosFace, SphereFace, Am_softmax
from loss.focal import FocalLoss
from util.utils import make_weights_for_balanced_classes, get_val_data, separate_irse_bn_paras, separate_resnet_bn_paras, warm_up_lr, schedule_lr, perform_val, get_time, buffer_val, AverageMeter, accuracy

from tensorboardX import SummaryWriter
from tqdm import tqdm
import os
    • Initialize hyperparameters: 
      cfg = configurations[1]

SEED = cfg['SEED'] # random seed for reproduce results
torch.manual_seed(SEED)

DATA_ROOT = cfg['DATA_ROOT'] # the parent root where your train/val/test data are stored
MODEL_ROOT = cfg['MODEL_ROOT'] # the root to buffer your checkpoints
LOG_ROOT = cfg['LOG_ROOT'] # the root to log your train/val status
BACKBONE_RESUME_ROOT = cfg['BACKBONE_RESUME_ROOT'] # the root to resume training from a saved checkpoint
HEAD_RESUME_ROOT = cfg['HEAD_RESUME_ROOT']  # the root to resume training from a saved checkpoint

BACKBONE_NAME = cfg['BACKBONE_NAME'] # support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152']
HEAD_NAME = cfg['HEAD_NAME'] # support:  ['Softmax', 'ArcFace', 'CosFace', 'SphereFace', 'Am_softmax']
LOSS_NAME = cfg['LOSS_NAME'] # support: ['Focal', 'Softmax']

INPUT_SIZE = cfg['INPUT_SIZE']
RGB_MEAN = cfg['RGB_MEAN'] # for normalize inputs
RGB_STD = cfg['RGB_STD']
EMBEDDING_SIZE = cfg['EMBEDDING_SIZE'] # feature dimension
BATCH_SIZE = cfg['BATCH_SIZE']
DROP_LAST = cfg['DROP_LAST'] # whether drop the last batch to ensure consistent batch_norm statistics
LR = cfg['LR'] # initial LR
NUM_EPOCH = cfg['NUM_EPOCH']
WEIGHT_DECAY = cfg['WEIGHT_DECAY']
MOMENTUM = cfg['MOMENTUM']
STAGES = cfg['STAGES'] # epoch stages to decay learning rate

DEVICE = cfg['DEVICE']
MULTI_GPU = cfg['MULTI_GPU'] # flag to use multiple GPUs
GPU_ID = cfg['GPU_ID'] # specify your GPU ids
PIN_MEMORY = cfg['PIN_MEMORY']
NUM_WORKERS = cfg['NUM_WORKERS']
print("=" * 60)
print("Overall Configurations:")
print(cfg)
print("=" * 60)

writer = SummaryWriter(LOG_ROOT) # writer for buffering intermedium results
    • Train & validation data loaders: 
      train_transform = transforms.Compose([ # refer to https://pytorch.org/docs/stable/torchvision/transforms.html for more build-in online data augmentation
    transforms.Resize([int(128 * INPUT_SIZE[0] / 112), int(128 * INPUT_SIZE[0] / 112)]), # smaller side resized
    transforms.RandomCrop([INPUT_SIZE[0], INPUT_SIZE[1]]),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(mean = RGB_MEAN,
                         std = RGB_STD),
])

dataset_train = datasets.ImageFolder(os.path.join(DATA_ROOT, 'imgs'), train_transform)

# create a weighted random sampler to process imbalanced data
weights = make_weights_for_balanced_classes(dataset_train.imgs, len(dataset_train.classes))
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights))

train_loader = torch.utils.data.DataLoader(
    dataset_train, batch_size = BATCH_SIZE, sampler = sampler, pin_memory = PIN_MEMORY,
    num_workers = NUM_WORKERS, drop_last = DROP_LAST
)

NUM_CLASS = len(train_loader.dataset.classes)
print("Number of Training Classes: {}".format(NUM_CLASS))

lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(DATA_ROOT)
    • Define and initialize model (backbone & head): 
      BACKBONE_DICT = {'ResNet_50': ResNet_50(INPUT_SIZE), 
                 'ResNet_101': ResNet_101(INPUT_SIZE), 
                 'ResNet_152': ResNet_152(INPUT_SIZE),
                 'IR_50': IR_50(INPUT_SIZE), 
                 'IR_101': IR_101(INPUT_SIZE), 
                 'IR_152': IR_152(INPUT_SIZE),
                 'IR_SE_50': IR_SE_50(INPUT_SIZE), 
                 'IR_SE_101': IR_SE_101(INPUT_SIZE), 
                 'IR_SE_152': IR_SE_152(INPUT_SIZE)}
BACKBONE = BACKBONE_DICT[BACKBONE_NAME]
print("=" * 60)
print(BACKBONE)
print("{} Backbone Generated".format(BACKBONE_NAME))
print("=" * 60)

HEAD_DICT = {'ArcFace': ArcFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID),
             'CosFace': CosFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID),
             'SphereFace': SphereFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID),
             'Am_softmax': Am_softmax(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID)}
HEAD = HEAD_DICT[HEAD_NAME]
print("=" * 60)
print(HEAD)
print("{} Head Generated".format(HEAD_NAME))
print("=" * 60)
    • Define and initialize loss function: 
      LOSS_DICT = {'Focal': FocalLoss(), 
             'Softmax': nn.CrossEntropyLoss()}
LOSS = LOSS_DICT[LOSS_NAME]
print("=" * 60)
print(LOSS)
print("{} Loss Generated".format(LOSS_NAME))
print("=" * 60)
    • Define and initialize optimizer: 
      if BACKBONE_NAME.find("IR") >= 0:
    backbone_paras_only_bn, backbone_paras_wo_bn = separate_irse_bn_paras(BACKBONE) # separate batch_norm parameters from others; do not do weight decay for batch_norm parameters to improve the generalizability
    _, head_paras_wo_bn = separate_irse_bn_paras(HEAD)
else:
    backbone_paras_only_bn, backbone_paras_wo_bn = separate_resnet_bn_paras(BACKBONE) # separate batch_norm parameters from others; do not do weight decay for batch_norm parameters to improve the generalizability
    _, head_paras_wo_bn = separate_resnet_bn_paras(HEAD)
OPTIMIZER = optim.SGD([{'params': backbone_paras_wo_bn + head_paras_wo_bn, 'weight_decay': WEIGHT_DECAY}, {'params': backbone_paras_only_bn}], lr = LR, momentum = MOMENTUM)
print("=" * 60)
print(OPTIMIZER)
print("Optimizer Generated")
print("=" * 60)
    • Whether resume from a checkpoint or not: 
      if BACKBONE_RESUME_ROOT and HEAD_RESUME_ROOT:
    print("=" * 60)
    if os.path.isfile(BACKBONE_RESUME_ROOT) and os.path.isfile(HEAD_RESUME_ROOT):
        print("Loading Backbone Checkpoint '{}'".format(BACKBONE_RESUME_ROOT))
        BACKBONE.load_state_dict(torch.load(BACKBONE_RESUME_ROOT))
        print("Loading Head Checkpoint '{}'".format(HEAD_RESUME_ROOT))
        HEAD.load_state_dict(torch.load(HEAD_RESUME_ROOT))
    else:
        print("No Checkpoint Found at '{}' and '{}'. Please Have a Check or Continue to Train from Scratch".format(BACKBONE_RESUME_ROOT, HEAD_RESUME_ROOT))
    print("=" * 60)
    • Whether use multi-GPU or not: 
      if MULTI_GPU:
    # multi-GPU setting
    BACKBONE = nn.DataParallel(BACKBONE, device_ids = GPU_ID)
    BACKBONE = BACKBONE.to(DEVICE)
else:
    # single-GPU setting
    BACKBONE = BACKBONE.to(DEVICE)
    • Minor settings prior to training: 
      DISP_FREQ = len(train_loader) // 100 # frequency to display training loss & acc

NUM_EPOCH_WARM_UP = NUM_EPOCH // 25  # use the first 1/25 epochs to warm up
NUM_BATCH_WARM_UP = len(train_loader) * NUM_EPOCH_WARM_UP  # use the first 1/25 epochs to warm up
batch = 0  # batch index
    • Training & validation & save checkpoint (use the first 1/25 epochs to warm up -- gradually increase LR to the initial value to ensure stable convergence): 
      for epoch in range(NUM_EPOCH): # start training process
    
    if epoch == STAGES[0]: # adjust LR for each training stage after warm up, you can also choose to adjust LR manually (with slight modification) once plaueau observed
        schedule_lr(OPTIMIZER)
    if epoch == STAGES[1]:
        schedule_lr(OPTIMIZER)
    if epoch == STAGES[2]:
        schedule_lr(OPTIMIZER)

    BACKBONE.train()  # set to training mode
    HEAD.train()

    losses = AverageMeter()
    top1 = AverageMeter()
    top5 = AverageMeter()

    for inputs, labels in tqdm(iter(train_loader)):

        if (epoch + 1 <= NUM_EPOCH_WARM_UP) and (batch + 1 <= NUM_BATCH_WARM_UP): # adjust LR for each training batch during warm up
            warm_up_lr(batch + 1, NUM_BATCH_WARM_UP, LR, OPTIMIZER)

        # compute output
        inputs = inputs.to(DEVICE)
        labels = labels.to(DEVICE).long()
        features = BACKBONE(inputs)
        outputs = HEAD(features, labels)
        loss = LOSS(outputs, labels)

        # measure accuracy and record loss
        prec1, prec5 = accuracy(outputs.data, labels, topk = (1, 5))
        losses.update(loss.data.item(), inputs.size(0))
        top1.update(prec1.data.item(), inputs.size(0))
        top5.update(prec5.data.item(), inputs.size(0))

        # compute gradient and do SGD step
        OPTIMIZER.zero_grad()
        loss.backward()
        OPTIMIZER.step()
        
        # dispaly training loss & acc every DISP_FREQ
        if ((batch + 1) % DISP_FREQ == 0) and batch != 0:
            print("=" * 60)
            print('Epoch {}/{} Batch {}/{}\t'
                  'Training Loss {loss.val:.4f} ({loss.avg:.4f})\t'
                  'Training Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
                  'Training Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
                epoch + 1, NUM_EPOCH, batch + 1, len(train_loader) * NUM_EPOCH, loss = losses, top1 = top1, top5 = top5))
            print("=" * 60)

        batch += 1 # batch index

    # training statistics per epoch (buffer for visualization)
    epoch_loss = losses.avg
    epoch_acc = top1.avg
    writer.add_scalar("Training_Loss", epoch_loss, epoch + 1)
    writer.add_scalar("Training_Accuracy", epoch_acc, epoch + 1)
    print("=" * 60)
    print('Epoch: {}/{}\t'
          'Training Loss {loss.val:.4f} ({loss.avg:.4f})\t'
          'Training Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
          'Training Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
        epoch + 1, NUM_EPOCH, loss = losses, top1 = top1, top5 = top5))
    print("=" * 60)

    # perform validation & save checkpoints per epoch
    # validation statistics per epoch (buffer for visualization)
    print("=" * 60)
    print("Perform Evaluation on LFW, CFP_FF, CFP_FP, AgeDB, CALFW, CPLFW and VGG2_FP, and Save Checkpoints...")
    accuracy_lfw, best_threshold_lfw, roc_curve_lfw = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, lfw, lfw_issame)
    buffer_val(writer, "LFW", accuracy_lfw, best_threshold_lfw, roc_curve_lfw, epoch + 1)
    accuracy_cfp_ff, best_threshold_cfp_ff, roc_curve_cfp_ff = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, cfp_ff, cfp_ff_issame)
    buffer_val(writer, "CFP_FF", accuracy_cfp_ff, best_threshold_cfp_ff, roc_curve_cfp_ff, epoch + 1)
    accuracy_cfp_fp, best_threshold_cfp_fp, roc_curve_cfp_fp = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, cfp_fp, cfp_fp_issame)
    buffer_val(writer, "CFP_FP", accuracy_cfp_fp, best_threshold_cfp_fp, roc_curve_cfp_fp, epoch + 1)
    accuracy_agedb, best_threshold_agedb, roc_curve_agedb = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, agedb, agedb_issame)
    buffer_val(writer, "AgeDB", accuracy_agedb, best_threshold_agedb, roc_curve_agedb, epoch + 1)
    accuracy_calfw, best_threshold_calfw, roc_curve_calfw = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, calfw, calfw_issame)
    buffer_val(writer, "CALFW", accuracy_calfw, best_threshold_calfw, roc_curve_calfw, epoch + 1)
    accuracy_cplfw, best_threshold_cplfw, roc_curve_cplfw = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, cplfw, cplfw_issame)
    buffer_val(writer, "CPLFW", accuracy_cplfw, best_threshold_cplfw, roc_curve_cplfw, epoch + 1)
    accuracy_vgg2_fp, best_threshold_vgg2_fp, roc_curve_vgg2_fp = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, vgg2_fp, vgg2_fp_issame)
    buffer_val(writer, "VGGFace2_FP", accuracy_vgg2_fp, best_threshold_vgg2_fp, roc_curve_vgg2_fp, epoch + 1)
    print("Epoch {}/{}, Evaluation: LFW Acc: {}, CFP_FF Acc: {}, CFP_FP Acc: {}, AgeDB Acc: {}, CALFW Acc: {}, CPLFW Acc: {}, VGG2_FP Acc: {}".format(epoch + 1, NUM_EPOCH, accuracy_lfw, accuracy_cfp_ff, accuracy_cfp_fp, accuracy_agedb, accuracy_calfw, accuracy_cplfw, accuracy_vgg2_fp))
    print("=" * 60)

    # save checkpoints per epoch
    if MULTI_GPU:
        torch.save(BACKBONE.module.state_dict(), os.path.join(MODEL_ROOT, "Backbone_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(BACKBONE_NAME, epoch + 1, batch, get_time())))
        torch.save(HEAD.state_dict(), os.path.join(MODEL_ROOT, "Head_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(HEAD_NAME, epoch + 1, batch, get_time())))
    else:
        torch.save(BACKBONE.state_dict(), os.path.join(MODEL_ROOT, "Backbone_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(BACKBONE_NAME, epoch + 1, batch, get_time())))
        torch.save(HEAD.state_dict(), os.path.join(MODEL_ROOT, "Head_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(HEAD_NAME, epoch + 1, batch, get_time())))

  • Now, you can start to play with face.evoLVe and run train.py. User friendly information will popped out on your terminal:

    • About overall configuration:

    • About number of training classes:

    • About backbone details:

    • About head details:

    • About loss details:

    • About optimizer details:

    • About resume training:

    • About training status & statistics (when batch index reachs DISP_FREQ or at the end of each epoch):

    • About validation statistics & save checkpoints (at the end of each epoch):

  • Monitor on-the-fly GPU occupancy with watch -d -n 0.01 nvidia-smi.

  • Please refer to Sec. Model Zoo for specific model weights and corresponding performance.

  • Feature extraction API (extract features from pre-trained models) ./util/extract_feature_v1.py (implemented with PyTorch build-in functions) and ./util/extract_feature_v2.py (implemented with OpenCV).

  • Visualize training & validation statistics with tensorboardX (see Sec. Model Zoo):

    tensorboard --logdir /media/pc/6T/jasonjzhao/buffer/log

Data Zoo

🐯

Database Version #Identity #Image #Frame #Video Download Link
LFW Raw 5,749 13,233 - - Google Drive, Baidu Drive
LFW Align_250x250 5,749 13,233 - - Google Drive, Baidu Drive
LFW Align_112x112 5,749 13,233 - - Google Drive, Baidu Drive
CALFW Raw 4,025 12,174 - - Google Drive, Baidu Drive
CALFW Align_112x112 4,025 12,174 - - Google Drive, Baidu Drive
CPLFW Raw 3,884 11,652 - - Google Drive, Baidu Drive
CPLFW Align_112x112 3,884 11,652 - - Google Drive, Baidu Drive
CASIA-WebFace Raw_v1 10,575 494,414 - - Baidu Drive
CASIA-WebFace Raw_v2 10,575 494,414 - - Google Drive, Baidu Drive
CASIA-WebFace Clean 10,575 455,594 - - Google Drive, Baidu Drive
MS-Celeb-1M Clean 100,000 5,084,127 - - Google Drive
MS-Celeb-1M Align_224x224 100,000 5,084,127 - -
MS-Celeb-1M Align_112x112 85,742 5,822,653 - - Google Drive
Vggface2 Clean 8,631 3,086,894 - - Google Drive
Vggface2_FP Align_112x112 - - - - Google Drive, Baidu Drive
AgeDB Raw 570 16,488 - - Google Drive, Baidu Drive
AgeDB Align_112x112 570 16,488 - - Google Drive, Baidu Drive
IJB-A Clean 500 5,396 20,369 2,085 Google Drive, Baidu Drive
IJB-B Raw 1,845 21,798 55,026 7,011 Google Drive
CFP Raw 500 7,000 - - Google Drive, Baidu Drive
CFP Align_112x112 500 7,000 - - Google Drive, Baidu Drive
Umdfaces Align_112x112 8,277 367,888 - - Google Drive, Baidu Drive
CelebA Raw 10,177 202,599 - - Google Drive, Baidu Drive
CACD-VS Raw 2,000 163,446 - - Google Drive, Baidu Drive
YTF Align_344x344 1,595 - 3,425 621,127 Google Drive, Baidu Drive
DeepGlint Align_112x112 180,855 6,753,545 - - Google Drive
UTKFace Align_200x200 - 23,708 - - Google Drive, Baidu Drive
  • Remark: unzip CASIA-WebFace clean version with 
    unzip casia-maxpy-clean.zip    
cd casia-maxpy-clean    
zip -F CASIA-maxpy-clean.zip --out CASIA-maxpy-clean_fix.zip    
unzip CASIA-maxpy-clean_fix.zip
  • Remark: after unzip, get image data & pair ground truths from AgeDB, CFP, LFW and VGGFace2_FP align_112x112 versions with 
    import numpy as np
import bcolz
import os

def get_pair(root, name):
    carray = bcolz.carray(rootdir = os.path.join(root, name), mode='r')
    issame = np.load('{}/{}_list.npy'.format(root, name))
    return carray, issame

def get_data(data_root):
    agedb_30, agedb_30_issame = get_pair(data_root, 'agedb_30')
    cfp_fp, cfp_fp_issame = get_pair(data_root, 'cfp_fp')
    lfw, lfw_issame = get_pair(data_root, 'lfw')
    vgg2_fp, vgg2_fp_issame = get_pair(data_root, 'vgg2_fp')
    return agedb_30, cfp_fp, lfw, vgg2_fp, agedb_30_issame, cfp_fp_issame, lfw_issame, vgg2_fp_issame

agedb_30, cfp_fp, lfw, vgg2_fp, agedb_30_issame, cfp_fp_issame, lfw_issame, vgg2_fp_issame = get_data(DATA_ROOT)
  • Due to release license issue, for other face related databases, please make contact with us in person for more details.

Model Zoo

🐒

  • Model

    Backbone Head Loss Training Data Download Link
    IR-50 ArcFace Focal MS-Celeb-1M_Align_112x112 Google Drive, Baidu Drive

    • Setting

      INPUT_SIZE: [112, 112]; RGB_MEAN: [0.5, 0.5, 0.5]; RGB_STD: [0.5, 0.5, 0.5]; BATCH_SIZE: 512 (drop the last batch to ensure consistent batch_norm statistics); Initial LR: 0.1; NUM_EPOCH: 120; WEIGHT_DECAY: 5e-4 (do not apply to batch_norm parameters); MOMENTUM: 0.9; STAGES: [30, 60, 90]; Augmentation: Random Crop + Horizontal Flip; Imbalanced Data Processing: Weighted Random Sampling; Solver: SGD; GPUs: 4 NVIDIA Tesla P40 in Parallel

    • Training & validation statistics

    • Performance

      LFW CFP_FF CFP_FP AgeDB CALFW CPLFW Vggface2_FP
      99.81 99.68 97.32 97.50 95.75 91.05 94.88

  • Model

    Backbone Head Loss Training Data Download Link
    IR-50 ArcFace Focal Private Asia Face Data Google Drive, Baidu Drive

    • Setting

      INPUT_SIZE: [112, 112]; RGB_MEAN: [0.5, 0.5, 0.5]; RGB_STD: [0.5, 0.5, 0.5]; BATCH_SIZE: 1024 (drop the last batch to ensure consistent batch_norm statistics); Initial LR: 0.01 (initialize weights from the above model pre-trained on MS-Celeb-1M_Align_112x112); NUM_EPOCH: 80; WEIGHT_DECAY: 5e-4 (do not apply to batch_norm parameters); MOMENTUM: 0.9; STAGES: [20, 40, 60]; Augmentation: Random Crop + Horizontal Flip; Imbalanced Data Processing: Weighted Random Sampling; Solver: SGD; GPUs: 8 NVIDIA Tesla P40 in Parallel

    • Performance (please perform evaluation on your own Asia face benchmark dataset)

Achievement

🎊

  • 2017 No.1 on ICCV 2017 MS-Celeb-1M Large-Scale Face Recognition Hard Set/Random Set/Low-Shot Learning Challenges. WeChat News, NUS ECE News, NUS ECE Poster, Award Certificate for Track-1, Award Certificate for Track-2, Award Ceremony.

  • 2017 No.1 on National Institute of Standards and Technology (NIST) IARPA Janus Benchmark A (IJB-A) Unconstrained Face Verification challenge and Identification challenge. WeChat News.

  • State-of-the-art performance on

    • MS-Celeb-1M (Challenge1 Hard Set Coverage@P=0.95: 79.10%; Challenge1 Random Set Coverage@P=0.95: 87.50%; Challenge2 Development Set Coverage@P=0.99: 100.00%; Challenge2 Base Set Top 1 Accuracy: 99.74%; Challenge2 Novel Set Coverage@P=0.99: 99.01%).
    • IJB-A (1:1 Veification TAR@FAR=0.1: 99.6%±0.1%; 1:1 Veification TAR@FAR=0.01: 99.1%±0.2%; 1:1 Veification TAR@FAR=0.001: 97.9%±0.4%; 1:N Identification FNIR@FPIR=0.1: 1.3%±0.3%; 1:N Identification FNIR@FPIR=0.01: 5.4%±4.7%; 1:N Identification Rank1 Accuracy: 99.2%±0.1%; 1:N Identification Rank5 Accuracy: 99.7%±0.1%; 1:N Identification Rank10 Accuracy: 99.8%±0.1%).
    • IJB-C (1:1 Veification TAR@FAR=1e-5: 82.6%).
    • Labeled Faces in the Wild (LFW) (Accuracy: 99.85%±0.217%).
    • Celebrities in Frontal-Profile (CFP) (Frontal-Profile Accuracy: 96.01%±0.84%; Frontal-Profile EER: 4.43%±1.04%; Frontal-Profile AUC: 99.00%±0.35%; Frontal-Frontal Accuracy: 99.64%±0.25%; Frontal-Frontal EER: 0.54%±0.37%; Frontal-Frontal AUC: 99.98%±0.03%).
    • CMU Multi-PIE (Rank1 Accuracy Setting-1 under ±90°: 76.12%; Rank1 Accuracy Setting-2 under ±90°: 86.73%).
    • MORPH Album2 (Rank1 Accuracy Setting-1: 99.65%; Rank1 Accuracy Setting-2: 99.26%).
    • CACD-VS (Accuracy: 99.76%).
    • FG-NET (Rank1 Accuracy: 93.20%).

Acknowledgement

👬

Citation

📑

  • Please consult and consider citing the following papers:

    @article{zhao2018look,
title={Look Across Elapse: Disentangled Representation Learning and Photorealistic Cross-Age Face Synthesis for Age-Invariant Face Recognition},
author={Zhao, Jian and Cheng, Yu and Cheng, Yi and Yang, Yang and Lan, Haochong and Zhao, Fang and Xiong, Lin and Xu, Yan and Li, Jianshu and Pranata, Sugiri and others},
journal={AAAI},
year={2019}
}


@article{zhao20183d,
title={3D-Aided Dual-Agent GANs for Unconstrained Face Recognition},
author={Zhao, Jian and Xiong, Lin and Li, Jianshu and Xing, Junliang and Yan, Shuicheng and Feng, Jiashi},
journal={T-PAMI},
year={2018}
}


@inproceedings{zhao2018towards,
title={Towards Pose Invariant Face Recognition in the Wild},
author={Zhao, Jian and Cheng, Yu and Xu, Yan and Xiong, Lin and Li, Jianshu and Zhao, Fang and Jayashree, Karlekar and Pranata,         Sugiri and Shen, Shengmei and Xing, Junliang and others},
booktitle={CVPR},
pages={2207--2216},
year={2018}
}


@inproceedings{zhao2017dual,
title={Dual-agent gans for photorealistic and identity preserving profile face synthesis},
author={Zhao, Jian and Xiong, Lin and Jayashree, Panasonic Karlekar and Li, Jianshu and Zhao, Fang and Wang, Zhecan and Pranata,           Panasonic Sugiri and Shen, Panasonic Shengmei and Yan, Shuicheng and Feng, Jiashi},
booktitle={NIPS},
pages={66--76},
year={2017}
}


@inproceedings{zhao3d,
title={3D-Aided Deep Pose-Invariant Face Recognition},
author={Zhao, Jian and Xiong, Lin and Cheng, Yu and Cheng, Yi and Li, Jianshu and Zhou, Li and Xu, Yan and Karlekar, Jayashree and       Pranata, Sugiri and Shen, Shengmei and others},
booktitle={IJCAI},
pages={1184--1190},
year={2018}
}


@inproceedings{cheng2017know,
title={Know you at one glance: A compact vector representation for low-shot learning},
author={Cheng, Yu and Zhao, Jian and Wang, Zhecan and Xu, Yan and Jayashree, Karlekar and Shen, Shengmei and Feng, Jiashi},
booktitle={ICCVW},
pages={1924--1932},
year={2017}
}