Merge branch 'master' of https://github.com/deepinsight/insightface

2026-05-18 06:38:19 +00:00 · 2018-02-19 19:13:27 +08:00
parent c180f5d1b9 fbf69f1e78
commit fa6a84a59a
11 changed files with 1969 additions and 476 deletions
--- a/README.md
+++ b/README.md
@@ -5,7 +5,11 @@
  
 ### Recent Update

-  **`2018.02.07`**: We evaluate LFW,CFP,AgeDB-30 again after removing training set overlaps, the results almost stay the same. See [Results](#results) for detail.
+  **`2018.02.16`**: We put the MegaFace noise list in this repo. Please refer to [https://github.com/deepinsight/insightface/blob/master/src/megaface] for detail.
+
+  **`2018.02.13`**: We achieved state-of-the-art performance on [MegaFace-Challenge-1](http://megaface.cs.washington.edu/results/facescrub.html), at `98.06%`. Also note that the training set we used has NO identities overlap with facescrub, please check our paper for detail.
+
+  **`2018.02.07`**: We evaluate LFW,CFP,AgeDB-30 again after removing training set overlaps, the results almost stay the same. See [Results](#results) for detail.

  **`2018.01.30`**: We provide a *LResNet50E-IR* model which can achieve **`99.80@LFW`** and **`97.64%`** at MegaFace 1M Acc. See [Pretrained-Models](#pretrained-models) for detail.
  
--- a/src/common/noise_sgd.py
+++ b/src/common/noise_sgd.py
@@ -0,0 +1,38 @@
+import mxnet.optimizer as optimizer
+from mxnet import ndarray as nd
+
+class NoiseSGD(optimizer.SGD):
+    """Noise SGD.
+
+
+    This optimizer accepts the same arguments as :class:`.SGD`.
+    """
+    def __init__(self, scale, **kwargs):
+        super(NoiseSGD, self).__init__(**kwargs)
+        print('init noise sgd with', scale)
+        self.scale = scale
+
+    def update(self, index, weight, grad, state):
+        assert(isinstance(weight, NDArray))
+        assert(isinstance(grad, NDArray))
+        self._update_count(index)
+        lr = self._get_lr(index)
+        wd = self._get_wd(index)
+
+        grad = grad * self.rescale_grad
+        if self.clip_gradient is not None:
+            grad = clip(grad, -self.clip_gradient, self.clip_gradient)
+        noise = nd.random.normal(scale = self.scale, shape = grad.shape, dtype=grad.dtype, ctx = grad.context)
+        grad += noise
+
+        if state is not None:
+            mom = state
+            mom[:] *= self.momentum
+            grad += wd * weight
+            mom[:] += grad
+            grad[:] += self.momentum * mom
+            weight[:] += -lr * grad
+        else:
+            assert self.momentum == 0.0
+            weight[:] += -lr * (grad + wd * weight)
+
--- a/src/eval/ytf.py
+++ b/src/eval/ytf.py
@@ -0,0 +1,302 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import cv2
+import argparse
+import sys
+import numpy as np
+from scipy import misc
+from sklearn.model_selection import KFold
+from scipy import interpolate
+import sklearn
+from sklearn.metrics.pairwise import euclidean_distances
+import datetime
+import pickle
+from sklearn.decomposition import PCA
+import mxnet as mx
+from mxnet import ndarray as nd
+from verification import evaluate
+from verification import calculate_accuracy
+sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
+import face_image
+
+
+model = None
+feature_cache = {}
+image_size = [112,112]
+
+def get_feature(name, vid, args):
+  global feature_cache
+  key = (name,vid)
+  if key in feature_cache:
+    return feature_cache[key]
+
+  input_dir = os.path.join(args.image_dir, name, str(vid))
+  data = nd.zeros( (1 ,3, image_size[0], image_size[1]) )
+  F = []
+  for img in os.listdir(input_dir):
+    img = os.path.join(input_dir, img)
+    img = cv2.imread(img)
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img = np.transpose(img, (2,0,1))
+    data[0][:] = img
+    db = mx.io.DataBatch(data=(data,))
+    model.forward(db, is_train=False)
+    net_out = model.get_outputs()[0].asnumpy().flatten()
+    F.append(net_out)
+  F = np.array(F)
+  F = sklearn.preprocessing.normalize(F)
+  feature = np.mean(F, axis=0, keepdims=True)
+  feature = sklearn.preprocessing.normalize(feature).flatten()
+
+  feature_cache[key] = feature
+  return feature
+
+def get_feature_set(name, vid, args):
+  global feature_cache
+  key = (name,vid)
+  if key in feature_cache:
+    return feature_cache[key]
+
+  input_dir = os.path.join(args.image_dir, name, str(vid))
+  data = nd.zeros( (1 ,3, image_size[0], image_size[1]) )
+  F = []
+  for img in os.listdir(input_dir):
+    img = os.path.join(input_dir, img)
+    img = cv2.imread(img)
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img = np.transpose(img, (2,0,1))
+    data[0][:] = img
+    db = mx.io.DataBatch(data=(data,))
+    model.forward(db, is_train=False)
+    net_out = model.get_outputs()[0].asnumpy().flatten()
+    F.append(net_out)
+  F = np.array(F)
+  F = sklearn.preprocessing.normalize(F)
+
+  feature_cache[key] = F
+  return F
+
+def main(args):
+  global model
+  ctx = mx.gpu(args.gpu)
+  args.ctx_num = 1
+  print('image_size', image_size)
+  vec = args.model.split(',')
+  prefix = vec[0]
+  epoch = int(vec[1])
+  print('loading',prefix, epoch)
+  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+  all_layers = sym.get_internals()
+  sym = all_layers['fc1_output']
+  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
+  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
+  model.set_params(arg_params, aux_params)
+  args.image_dir = os.path.join(args.data_dir, 'images')
+  pairs_file = os.path.join(args.data_dir, 'splits2.txt')
+  embeddings = []
+  issame_list = []
+  data = []
+  pp = 0
+  for line in open(pairs_file, 'r'):
+    line = line.strip()
+    if line.startswith('split'):
+      continue
+    pp+=1
+    if pp%10==0:
+      print('processing', pp)
+    vec = line.split(',')
+    assert len(vec)>=5
+    issame_list.append(int(vec[-1]))
+    for i in [2,3]:
+      _str = vec[i].strip()
+      _vec = _str.split('/')
+      assert len(_vec)==2
+      name = _vec[0]
+      vid = int(_vec[1])
+      feature = get_feature(name, vid, args)
+      print('feature', feature.shape)
+      embeddings.append(feature)
+      data.append( (name, vid) )
+    #if len(issame_list)==20:
+    #  break
+  embeddings = np.array(embeddings)
+  print(embeddings.shape)
+  thresholds = np.arange(0, 4, 0.01)
+  actual_issame = np.asarray(issame_list)
+  nrof_folds = 10
+  embeddings1 = embeddings[0::2]
+  embeddings2 = embeddings[1::2]
+  assert(embeddings1.shape[0] == embeddings2.shape[0])
+  assert(embeddings1.shape[1] == embeddings2.shape[1])
+  nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
+  nrof_thresholds = len(thresholds)
+  k_fold = KFold(n_splits=nrof_folds, shuffle=False)
+  
+  tprs = np.zeros((nrof_folds,nrof_thresholds))
+  fprs = np.zeros((nrof_folds,nrof_thresholds))
+  accuracy = np.zeros((nrof_folds))
+  indices = np.arange(nrof_pairs)
+  
+  diff = np.subtract(embeddings1, embeddings2)
+  dist = np.sum(np.square(diff),1)
+  pouts = []
+  nouts = []
+  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+      # Find the best threshold for the fold
+      acc_train = np.zeros((nrof_thresholds))
+      #print(train_set)
+      #print(train_set.__class__)
+      for threshold_idx, threshold in enumerate(thresholds):
+          p2 = dist[train_set]
+          p3 = actual_issame[train_set]
+          _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, p2, p3)
+      best_threshold_index = np.argmax(acc_train)
+      for threshold_idx, threshold in enumerate(thresholds):
+          tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
+      _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
+      best_threshold = thresholds[best_threshold_index]
+      for iid in test_set:
+        ida = iid*2
+        idb = ida+1
+        asame = actual_issame[iid]
+        _dist = dist[iid]
+        violate = _dist - best_threshold
+        if not asame:
+          violate *= -1.0
+        if violate>0.0:
+          dataa = data[ida]
+          datab = data[idb]
+          #print(imga.shape, imgb.shape, violate, asame, _dist)
+          if asame:
+            pouts.append( (dataa, datab, _dist, best_threshold, ida) )
+          else:
+            nouts.append( (dataa, datab, _dist, best_threshold, ida) )
+
+        
+  tpr = np.mean(tprs,0)
+  fpr = np.mean(fprs,0)
+  acc = np.mean(accuracy)
+  pouts = sorted(pouts, key = lambda x: x[2], reverse=True)
+  nouts = sorted(nouts, key = lambda x: x[2], reverse=False)
+  print(len(pouts), len(nouts))
+  print('acc', acc)
+  if len(nouts)>0:
+    threshold = nouts[0][3]
+  else:
+    threshold = pouts[-1][3]
+  #print('threshold', threshold)
+  print('positive(false negative):')
+  for out in pouts:
+    print("\t%s\t%s\t(distance:%f, threshold:%f)"%(out[0], out[1], out[2], out[3]))
+  print('negative(false positive):')
+  for out in nouts:
+    print("\t%s\t%s\t(distance:%f, threshold:%f)"%(out[0], out[1], out[2], out[3]))
+
+
+
+
+
+  #_, _, accuracy, val, val_std, far = evaluate(embeddings, issame_list, nrof_folds=10)
+  #acc2, std2 = np.mean(accuracy), np.std(accuracy)
+  #print('acc', acc2)
+
+def main2(args):
+  global model
+  ctx = mx.gpu(args.gpu)
+  args.ctx_num = 1
+  print('image_size', image_size)
+  vec = args.model.split(',')
+  prefix = vec[0]
+  epoch = int(vec[1])
+  print('loading',prefix, epoch)
+  sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
+  all_layers = sym.get_internals()
+  sym = all_layers['fc1_output']
+  model = mx.mod.Module(symbol=sym, context=ctx, label_names = None)
+  model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))])
+  model.set_params(arg_params, aux_params)
+  args.image_dir = os.path.join(args.data_dir, 'images')
+  pairs_file = os.path.join(args.data_dir, 'splits2.txt')
+  issame_list = []
+  dist = []
+  pp = 0
+  for line in open(pairs_file, 'r'):
+    line = line.strip()
+    if line.startswith('split'):
+      continue
+    pp+=1
+    if pp%10==0:
+      print('processing', pp)
+    vec = line.split(',')
+    assert len(vec)>=5
+    issame_list.append(int(vec[-1]))
+    feature_sets = []
+    for i in [2,3]:
+      _str = vec[i].strip()
+      _vec = _str.split('/')
+      assert len(_vec)==2
+      name = _vec[0]
+      vid = int(_vec[1])
+      feature = get_feature_set(name, vid, args)
+      print('feature', len(feature))
+      feature_sets.append(feature)
+    X = feature_sets[0]
+    Y = feature_sets[1]
+    _dist = euclidean_distances(X, Y)
+    _dist = _dist*_dist
+    #_tmp = np.eye(_dist.shape[0], dtype=np.float32)
+    #_dist += _tmp
+    if args.mode==2:
+      _dist = np.amin(_dist)
+    elif args.mode==3:
+      _dist = np.mean(_dist)
+    else:
+      _dist = np.amax(_dist)
+    print(_dist)
+    dist.append(_dist)
+    #if len(dist)==10:
+    #  break
+
+  dist = np.array(dist)
+  nrof_folds = 10
+  thresholds = np.arange(0, 4, 0.01)
+  actual_issame = np.array(issame_list)
+  nrof_pairs = len(actual_issame)
+  nrof_thresholds = len(thresholds)
+  k_fold = KFold(n_splits=nrof_folds, shuffle=False)
+  
+  tprs = np.zeros((nrof_folds,nrof_thresholds))
+  fprs = np.zeros((nrof_folds,nrof_thresholds))
+  accuracy = np.zeros((nrof_folds))
+  indices = np.arange(nrof_pairs)
+  for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
+      
+    # Find the best threshold for the fold
+    acc_train = np.zeros((nrof_thresholds))
+    for threshold_idx, threshold in enumerate(thresholds):
+        _, _, acc_train[threshold_idx] = calculate_accuracy(threshold, dist[train_set], actual_issame[train_set])
+    best_threshold_index = np.argmax(acc_train)
+    for threshold_idx, threshold in enumerate(thresholds):
+        tprs[fold_idx,threshold_idx], fprs[fold_idx,threshold_idx], _ = calculate_accuracy(threshold, dist[test_set], actual_issame[test_set])
+    _, _, accuracy[fold_idx] = calculate_accuracy(thresholds[best_threshold_index], dist[test_set], actual_issame[test_set])
+  acc2, std2 = np.mean(accuracy), np.std(accuracy)
+  print('acc', acc2)
+
+if __name__ == '__main__':
+
+  parser = argparse.ArgumentParser(description='do verification')
+  # general
+  parser.add_argument('--data-dir', default='/raid5data/dplearn/YTF', help='')
+  parser.add_argument('--model', default='../model/softmax,50', help='path to load model.')
+  parser.add_argument('--gpu', default=0, type=int, help='gpu id')
+  parser.add_argument('--batch-size', default=32, type=int, help='')
+  parser.add_argument('--mode', default=1, type=int, help='')
+  args = parser.parse_args()
+  if args.mode>=2:
+    main2(args)
+  else:
+    main(args)
+
--- a/src/eval/ytf_badcases.py
+++ b/src/eval/ytf_badcases.py
@@ -0,0 +1,75 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sys
+import os
+import numpy as np
+import cv2
+
+pairs_file = '/raid5data/dplearn/YTF/splits2.txt'
+stat = [0,0]
+for line in open(pairs_file, 'r'):
+  line = line.strip()
+  if line.startswith('split'):
+    continue
+  vec = line.split(',')
+  issame = int(vec[-1])
+  if issame:
+    stat[0]+=1
+  else:
+    stat[1]+=1
+print('stat', stat)
+
+image_dir = '/raid5data/dplearn/YTF/images'
+
+def get_img(name, vid):
+  input_dir = os.path.join(image_dir, name, str(vid))
+  paths = []
+  for img in os.listdir(input_dir):
+    path = os.path.join(input_dir, img)
+    paths.append(path)
+  paths = sorted(paths)
+  parts = 8
+  assert len(paths)>=parts
+  gap = len(paths)//parts
+  img = None
+  for i in xrange(parts):
+    idx = gap*i
+    path = paths[idx]
+    _img = cv2.imread(path)
+    #print(_img.shape)
+    if img is None:
+      img = _img
+    else:
+      img = np.concatenate( (img, _img), axis=1)
+  return img
+
+
+text_color = (153,255,51)
+for input in ['ytf_false_positive', 'ytf_false_negative']:
+  all_img = None
+  pp = 0
+  for line in open(input+".log", 'r'):
+    if line.startswith("\t"):
+      break
+    vec = line.strip().split(',')
+    img1 = get_img(vec[0], int(vec[1]))
+    img2 = get_img(vec[2], int(vec[3]))
+    img = np.concatenate( (img1, img2), axis=0)
+    if all_img is None:
+      all_img = img
+    else:
+      all_img = np.concatenate( (all_img, img), axis=0)
+    blank_img = np.zeros( (20, 112*8,3), dtype=np.uint8)
+    blank_img[:,:,:] = 255
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    k = "centre-distance:%.3f"%(float(vec[4]))
+    #print(k)
+    cv2.putText(blank_img,k,(350,blank_img.shape[0]-4), font, 0.6, text_color, 2)
+    all_img = np.concatenate( (all_img, blank_img), axis=0)
+    pp+=1
+
+  filename = os.path.join('badcases', input+".png")
+  cv2.imwrite(filename, all_img)
+
--- a/src/megaface/README.md
+++ b/src/megaface/README.md
@@ -0,0 +1,6 @@
+Please strictly follow these rules if you want to use our MegaFace noises list.
+
+* Please cite our paper and git repo if you want to use this list in your paper.
+* Please include the information like `We used the noises list proposed by InsightFace, at https://github.com/deepinsight/insightface/tree/master/src/megaface` if you want to submit the result to MegaFace challenge.
+* To be fair, if you want to submit MegaFace result, please ensure there's no training set overlaps with FaceScrub identities. You can do this by removing identities from your training set whose cosine similarity is larger than 0.4 with any FaceScrub identity by comparing their centre feature vectors. 
+* If you find more overlaps noise, please open an issue at InsightFace.
--- a/src/megaface/facescrub_noises.txt
+++ b/src/megaface/facescrub_noises.txt
@@ -0,0 +1,606 @@
+#Please strictly follow the rules in https://github.com/deepinsight/insightface/tree/master/src/megaface/README.md if you want to use this list.
+Aaron_Eckhart_136.png
+Aaron_Eckhart_221.png
+Adam_McKay_468.png
+Adam_McKay_478.png
+Adam_Sandler_558.png
+Adrienne_Barbeau_4259.png
+Adrienne_Frantz_27683.png
+Adrienne_Frantz_27693.png
+Adrienne_Frantz_27699.png
+Adrienne_Frantz_27701.png
+Adrienne_Frantz_27829.png
+Aisha_Hinds_35768.png
+Alec_Baldwin_2121.png
+Alice_Krige_40717.png
+Alley_Mills_52029.png
+Alley_Mills_52039.png
+Alley_Mills_52044.png
+Allison_Janney_37357.png
+Allison_Janney_37447.png
+Alyson_Hannigan_32288.png
+Alyssa_Milano_51641.png
+America_Ferrera_25859.png
+America_Ferrera_25870.png
+Amy_Davidson_18553.png
+Andrea_Bogart_8427.png
+Andrea_Bowen_8992.png
+Andrea_Bowen_9090.png
+Andy_Richter_3141.png
+Angell_Conwell_15508.png
+Anne_Hathaway_33648.png
+Anne_Hathaway_33685.png
+Annie_Ilonzeh_37123.png
+Anthony_Hopkins_3562.png
+Anthony_Hopkins_3595.png
+Anthony_Hopkins_3694.png
+Antonio_Banderas_4016.png
+Arnold_Vosloo_4287.png
+Ashley_Benson_4144.png
+Ashley_Johnson_37631.png
+Ashley_Johnson_37649.png
+Audra_McDonald_49657.png
+Audrey_Landers_42201.png
+Audrey_Landers_42246.png
+Barbara_Carrera_11961.png
+Ben_Affleck_4700.png
+Ben_Affleck_4706.png
+Ben_Affleck_4762.png
+Ben_Affleck_4885.png
+Ben_Kingsley_5146.png
+Ben_McKenzie_5219.png
+Ben_Stiller_5419.png
+Ben_Stiller_5432.png
+Bernard_Hill_5869.png
+Bernard_Hill_5969.png
+Bernie_Mac_6083.png
+Billy_Bob_Thornton_7010.png
+Billy_Bob_Thornton_7016.png
+Billy_Bob_Thornton_7028.png
+Billy_Bob_Thornton_7047.png
+Billy_Bob_Thornton_7069.png
+Billy_Bob_Thornton_7093.png
+Billy_Bob_Thornton_7106.png
+Billy_Bob_Thornton_7140.png
+Billy_Bob_Thornton_7161.png
+Billy_Boyd_7281.png
+Billy_Boyd_7334.png
+Billy_Burke_7418.png
+Billy_Burke_7556.png
+Billy_Burke_7558.png
+Billy_Zane_7799.png
+Billy_Zane_7821.png
+Bobbie_Eakes_23690.png
+Bobbie_Eakes_23702.png
+Brad_Pitt_8285.png
+Bradley_Cooper_8416.png
+Brianna_Brown_10455.png
+Brianna_Brown_10482.png
+Brooke_Langton_42515.png
+Brooke_Langton_42603.png
+Brooke_Langton_42604.png
+Brooke_Langton_42628.png
+Brooke_Langton_42634.png
+Brooke_Langton_42643.png
+Bruce_Greenwood_8831.png
+Candice_Bergen_6028.png
+Candice_Bergen_6055.png
+Candice_Bergen_6080.png
+Carla_Gallo_28085.png
+Carla_Gallo_28139.png
+Carmen_Electra_24315.png
+Caroline_Dhavernas_21720.png
+Cary_Elwes_9746.png
+Casey_Affleck_9925.png
+Cathy_Lee_Crosby_16548.png
+Cathy_Lee_Crosby_16564.png
+Chase_Masterson_48256.png
+Chazz_Palminteri_10758.png
+Cheryl_Hines_35834.png
+Chris_Evans_10942.png
+Chris_Evans_10943.png
+Chris_Evans_10990.png
+Chris_Evans_11000.png
+Chris_Evans_11039.png
+Chris_Kattan_11206.png
+Chris_Klein_11329.png
+Chris_Klein_11346.png
+Chris_Klein_11471.png
+Christa_Miller_51872.png
+Christian_Bale_11961.png
+Christina_Applegate_2776.png
+Christopher_Lloyd_12504.png
+Christopher_Reeve_12692.png
+Christopher_Reeve_12737.png
+Chyler_Leigh_43366.png
+Chyler_Leigh_43374.png
+Ciara_Bravo_10094.png
+Ciara_Bravo_10161.png
+Ciara_Bravo_10214.png
+Clint_Eastwood_12918.png
+Colin_Farrell_13196.png
+Colin_Farrell_13268.png
+Colin_Farrell_13282.png
+Colin_Firth_13455.png
+Colin_Firth_13566.png
+Courteney_Cox_16360.png
+Crystal_Chappell_13859.png
+Crystal_Chappell_13921.png
+Crystal_Chappell_13932.png
+Crystal_Chappell_13949.png
+Crystal_Chappell_13961.png
+Dan_Lauria_13992.png
+Dan_Lauria_14073.png
+Dana_Delany_20425.png
+Dana_Delany_20561.png
+Dana_Delany_20621.png
+Dana_Delany_20632.png
+Daniel_Day-Lewis_14356.png
+Daniel_Day-Lewis_14361.png
+Daniel_Day-Lewis_14397.png
+Daniel_Day-Lewis_14545.png
+Daniel_Radcliffe_14770.png
+David_Schwimmer_15942.png
+David_Wenham_16210.png
+David_Wenham_16221.png
+David_Wenham_16224.png
+David_Wenham_16235.png
+David_Wenham_16241.png
+David_Wenham_16261.png
+David_Wenham_16295.png
+David_Wenham_16300.png
+David_Wenham_16327.png
+Dean_Cain_16435.png
+Debra_Messing_50555.png
+Debra_Messing_50597.png
+Debra_Messing_50703.png
+Debra_Messing_50746.png
+Delta_Burke_11111.png
+Delta_Burke_11176.png
+Denzel_Washington_16698.png
+Desmond_Harrington_17178.png
+Diahann_Carroll_12351.png
+Diahann_Carroll_12397.png
+Dianna_Agron_269.png
+Dianna_Agron_363.png
+Dianna_Agron_375.png
+Dianna_Agron_396.png
+Dustin_Hoffman_17932.png
+Elizabeth_Berkley_6214.png
+Elizabeth_Berkley_6259.png
+Elizabeth_Hendrickson_34413.png
+Ellen_Greene_30978.png
+Ellen_Greene_31003.png
+Ellen_Greene_31005.png
+Emily_Deschanel_20666.png
+Erin_Cummings_17368.png
+Erin_Cummings_17400.png
+Erin_Cummings_17457.png
+Ethan_Hawke_19548.png
+Eva_Longoria_45027.png
+Farah_Fath_25299.png
+Farah_Fath_25343.png
+Farah_Fath_25365.png
+Farah_Fath_25401.png
+Farrah_Fawcett_25594.png
+Florencia_Lozano_45592.png
+Florencia_Lozano_45594.png
+Florencia_Lozano_45606.png
+Florencia_Lozano_45632.png
+Florencia_Lozano_45640.png
+Fran_Drescher_22955.png
+Fran_Drescher_23030.png
+Gabrielle_Carteris_12657.png
+Gates_McFadden_49864.png
+Gates_McFadden_49868.png
+Glenn_Close_14672.png
+Glenn_Close_14739.png
+Harrison_Ford_22656.png
+Hayden_Christensen_23089.png
+Heather_Locklear_44410.png
+Heather_Locklear_44498.png
+Heather_Locklear_44504.png
+Heather_Locklear_44527.png
+Holly_Marie_Combs_15225.png
+Ian_Holm_24214.png
+Jackee_Harry_32964.png
+Jackee_Harry_33011.png
+James_Brolin_26065.png
+James_Frain_26352.png
+James_Franco_26574.png
+James_Marsden_26731.png
+James_Marsden_26756.png
+James_McAvoy_26932.png
+Jamie_Lee_Curtis_17869.png
+Jamie_Lee_Curtis_18018.png
+Jamie_Lee_Curtis_18056.png
+Jamie_Luner_46356.png
+Jamie_Luner_46423.png
+Jane_Curtin_17787.png
+Jane_Curtin_17820.png
+Jane_Leeves_42833.png
+Jane_Lynch_46639.png
+Jane_Lynch_46743.png
+January_Jones_38213.png
+January_Jones_38384.png
+January_Jones_38414.png
+Jasmine_Guy_31496.png
+Jasmine_Guy_31562.png
+Jasmine_Guy_31594.png
+Jasmine_Guy_31618.png
+Jasmine_Guy_31678.png
+Jason_Biggs_28339.png
+Jason_Lee_55690.png
+Jason_Lee_55703.png
+Jeanne_Cooper_15551.png
+Jeanne_Cooper_15652.png
+Jeanne_Cooper_15696.png
+Jenilee_Harrison_32833.png
+Jenilee_Harrison_32864.png
+Jenilee_Harrison_32895.png
+Jennette_McCurdy_49314.png
+Jennie_Garth_28672.png
+Jennie_Garth_28686.png
+Jeremy_Sisto_29948.png
+Jessica_Biel_7413.png
+Jessica_Capshaw_11891.png
+Jessica_Capshaw_11909.png
+Jessica_Leccia_42714.png
+Jessica_Leccia_42719.png
+Jill_Eikenberry_24108.png
+Jill_Eikenberry_24159.png
+Jill_Eikenberry_24166.png
+Jill_Eikenberry_24179.png
+Jill_Eikenberry_24188.png
+Jill_Hennessy_35087.png
+Jim_Carrey_30976.png
+Jim_Carrey_31137.png
+Jim_Carrey_31150.png
+Joan_Collins_14963.png
+Joanna_Kerns_40119.png
+Joanna_Kerns_40127.png
+Joanna_Kerns_40145.png
+Joanna_Kerns_40162.png
+Joanna_Kerns_40170.png
+Joanna_Kerns_40177.png
+John_Malkovich_32908.png
+John_Malkovich_32934.png
+Jon_Hamm_33859.png
+Jonathan_Rhys_Meyers_34545.png
+Jonathan_Sadowski_55304.png
+Josh_Brolin_34866.png
+Josh_Duhamel_35073.png
+Josie_Bissett_7155.png
+Josie_Bissett_7222.png
+Josie_Bissett_7243.png
+Joyce_DeWitt_20190.png
+Julia_Louis-Dreyfus_45687.png
+Julia_Louis-Dreyfus_45785.png
+Julie_Marie_Berman_6437.png
+Justin_Long_35979.png
+Justin_Long_35983.png
+Justine_Bateman_4879.png
+Kassie_DePaiva_19841.png
+Katherine_Helmond_34045.png
+Kathy_Baker_3934.png
+Kathy_Baker_3984.png
+Kathy_Griffin_31087.png
+Kathy_Griffin_31113.png
+Kellan_Lutz_36883.png
+Kevin_Costner_37976.png
+Kim_Cattrall_13087.png
+Kim_Delaney_20329.png
+Kim_Delaney_20363.png
+Kim_Fields_26573.png
+Kimberlin_Brown_10564.png
+Kimberly_McCullough_49225.png
+Kirstie_Alley_1565.png
+Kit_Harington_38601.png
+Kit_Harington_38638.png
+Kristen_Alderson_615.png
+Kristen_Alderson_714.png
+Kristen_Johnston_37751.png
+Kristin_Chenoweth_14161.png
+Kristin_Chenoweth_14220.png
+Kristin_Chenoweth_14242.png
+Kristy_McNichol_50183.png
+Kristy_McNichol_50205.png
+Kristy_McNichol_50256.png
+Kristy_McNichol_50272.png
+Kristy_McNichol_50295.png
+Kristy_McNichol_50304.png
+Kristy_McNichol_50314.png
+Lacey_Chabert_13239.png
+Lacey_Chabert_13291.png
+Laura_Innes_37302.png
+Laura_Innes_37329.png
+Laura_Leighton_43101.png
+Laura_Leighton_43169.png
+Lauralee_Bell_5169.png
+Lauralee_Bell_5206.png
+Lauralee_Bell_5272.png
+Lauren_Holly_36179.png
+Lauren_Holly_36191.png
+Lauren_Holly_36211.png
+Lauren_Koslow_40373.png
+Lauren_Koslow_40426.png
+Lauren_Koslow_40428.png
+Lauren_Koslow_40434.png
+Lauren_Koslow_40453.png
+Laurie_Metcalf_50846.png
+Laurie_Metcalf_50878.png
+Laurie_Metcalf_50883.png
+Laurie_Metcalf_50897.png
+Laurie_Metcalf_50916.png
+Laurie_Metcalf_50917.png
+Laurie_Metcalf_50920.png
+Laurie_Metcalf_50946.png
+Laurie_Metcalf_50973.png
+Laurie_Metcalf_50984.png
+Laurie_Metcalf_50986.png
+Laurie_Metcalf_50989.png
+Laurie_Metcalf_51019.png
+Lea_Michele_51327.png
+Lea_Michele_51337.png
+Lea_Michele_51362.png
+Lea_Michele_51363.png
+Lea_Michele_51427.png
+Lecy_Goranson_30511.png
+Lecy_Goranson_30538.png
+Lecy_Goranson_30539.png
+Lesley-Anne_Down_22118.png
+Lesley-Anne_Down_22144.png
+Lesley-Anne_Down_22249.png
+Lexi_Ainsworth_126.png
+Lexi_Ainsworth_129.png
+Lexi_Ainsworth_150.png
+Lexi_Ainsworth_28.png
+Lexi_Ainsworth_47.png
+Lexi_Ainsworth_54.png
+Liev_Schreiber_39673.png
+Linda_Evans_24646.png
+Linda_Evans_24670.png
+Linda_Gray_30789.png
+Linda_Gray_30823.png
+Linda_Gray_30895.png
+Lindsay_Hartley_33089.png
+Lindsay_Hartley_33091.png
+Lindsay_Hartley_33182.png
+Lindsay_Hartley_33188.png
+Lindsay_Hartley_33192.png
+Lisa_Bonet_8489.png
+Lisa_Bonet_8547.png
+Lisa_Kudrow_40922.png
+Lisa_LoCicero_44328.png
+Lisa_LoCicero_44334.png
+Loni_Anderson_2152.png
+Loni_Anderson_2190.png
+Lorraine_Bracco_9632.png
+Lorraine_Bracco_9746.png
+Lourdes_Benedicto_5324.png
+Lourdes_Benedicto_5360.png
+Mary_Beth_Evans_24742.png
+Mary_Crosby_16814.png
+Mary_Crosby_16826.png
+Mary_Crosby_16827.png
+Mary_Crosby_16835.png
+Mary_Crosby_16856.png
+Mary_Crosby_16867.png
+Mary_Crosby_16876.png
+Matt_Czuchry_41100.png
+Matt_Dillon_41441.png
+Matt_Dillon_41584.png
+Melina_Kanakaredes_39464.png
+Melissa_Archer_3127.png
+Melissa_Archer_3169.png
+Melissa_Benoist_5425.png
+Melissa_Benoist_5490.png
+Melissa_Benoist_5507.png
+Melissa_Benoist_5523.png
+Melissa_Benoist_5538.png
+Michael_Douglas_43514.png
+Michael_Douglas_43520.png
+Michael_Douglas_43593.png
+Michael_Landes_43643.png
+Michael_Landes_43712.png
+Michael_Vartan_43752.png
+Michael_Vartan_43873.png
+Mila_Kunis_41190.png
+Miranda_Cosgrove_15892.png
+Molly_Burnett_11287.png
+Molly_Burnett_11323.png
+Morena_Baccarin_3662.png
+Natalia_Livingston_44183.png
+Natalie_Hall_31866.png
+Natalie_Hall_31875.png
+Natalie_Hall_31898.png
+Natalie_Hall_31914.png
+Natalie_Martinez_48106.png
+Natalie_Martinez_48117.png
+Natalie_Martinez_48191.png
+Neve_Campbell_11584.png
+Nicole_Eggert_23928.png
+Nicole_Eggert_24088.png
+Nicole_de_Boer_19422.png
+Nicole_de_Boer_19448.png
+Olivia_d'Abo_18077.png
+Olivia_d'Abo_18126.png
+Olivia_d'Abo_18133.png
+Olivia_d'Abo_18166.png
+Olivia_d'Abo_18188.png
+Olivia_d'Abo_18202.png
+Olivia_d'Abo_18224.png
+Olivia_d'Abo_18228.png
+Olivia_d'Abo_18229.png
+Pamela_Sue_Martin_47876.png
+Pamela_Sue_Martin_48018.png
+Pamela_Sue_Martin_48027.png
+Patricia_Arquette_3360.png
+Patricia_Kalember_39346.png
+Patricia_Kalember_39361.png
+Patricia_Kalember_39365.png
+Patricia_Kalember_39366.png
+Patricia_Kalember_39367.png
+Patricia_Kalember_39369.png
+Peggy_Lipton_43900.png
+Peggy_Lipton_43935.png
+Peggy_Lipton_43962.png
+Peggy_Lipton_44032.png
+Peggy_Lipton_44070.png
+Peggy_Lipton_44071.png
+Peggy_McCay_48867.png
+Peggy_McCay_48870.png
+Peggy_McCay_48883.png
+Peggy_McCay_48897.png
+Peri_Gilpin_29456.png
+Philip_Seymour_Hoffman_47737.png
+Philip_Seymour_Hoffman_47778.png
+Portia_Doubleday_22036.png
+Portia_Doubleday_22101.png
+Portia_de_Rossi_19903.png
+Portia_de_Rossi_19908.png
+Portia_de_Rossi_19957.png
+Portia_de_Rossi_19963.png
+Portia_de_Rossi_19997.png
+Portia_de_Rossi_20036.png
+Portia_de_Rossi_20069.png
+Rachel_Dratch_22646.png
+Rachel_Dratch_22751.png
+Rebecca_Budig_10668.png
+Rebecca_Budig_10714.png
+Rebecca_Budig_10756.png
+Rebecca_Budig_10785.png
+Rebecca_Budig_10792.png
+Rebecca_Budig_10805.png
+Rebecca_Herbst_35381.png
+Rebecca_Herbst_35391.png
+Rebecca_Herbst_35407.png
+Rebecca_Herbst_35454.png
+Richard_E._Grant_48348.png
+Richard_Gere_48555.png
+Richard_Gere_48559.png
+Richard_Madden_48806.png
+Robert_Di_Niro_49242.png
+Robert_Downey_Jr._49395.png
+Robert_Downey_Jr._49485.png
+Robert_Duvall_49683.png
+Robert_Duvall_49684.png
+Robert_Duvall_49744.png
+Robert_Knepper_49865.png
+Roma_Downey_22357.png
+Roma_Downey_22378.png
+Roseanne_Barr_4520.png
+Roseanne_Barr_4546.png
+Roseanne_Barr_4622.png
+Rue_McClanahan_48968.png
+Rue_McClanahan_49092.png
+Rue_McClanahan_49151.png
+Rupert_Friend_50825.png
+Ryan_Gosling_51576.png
+S._Epatha_Merkerson_50436.png
+S._Epatha_Merkerson_50464.png
+Sara_Gilbert_29260.png
+Sara_Gilbert_29333.png
+Sarah_Drew_23164.png
+Sarah_Drew_23166.png
+Sarah_Drew_23174.png
+Sarah_Drew_23209.png
+Sarah_Drew_23245.png
+Sarah_Drew_23272.png
+Sarah_Drew_23286.png
+Sarah_Hyland_36867.png
+Sean_Bean_52769.png
+Selena_Gomez_30458.png
+Selena_Gomez_30471.png
+Seth_Rogen_52915.png
+Seth_Rogen_53117.png
+Seth_Rogen_53123.png
+Shannen_Doherty_21826.png
+Shannon_Kane_39638.png
+Shannon_Kane_39729.png
+Shannon_Kane_39732.png
+Sharon_Case_12719.png
+Sharon_Case_12779.png
+Sharon_Case_12889.png
+Sharon_Gless_29936.png
+Sharon_Gless_29976.png
+Sharon_Gless_29987.png
+Sharon_Gless_29989.png
+Shelley_Hack_31728.png
+Shelley_Hack_31736.png
+Shelley_Hack_31826.png
+Sherilyn_Fenn_25694.png
+Shirley_Jones_38512.png
+Shirley_Jones_38556.png
+Staci_Keanan_39993.png
+Staci_Keanan_40018.png
+Staci_Keanan_40038.png
+Staci_Keanan_40068.png
+Stana_Katic_39947.png
+Susan_Flannery_27075.png
+Susan_Flannery_27095.png
+Susan_Flannery_27128.png
+Swoosie_Kurtz_41518.png
+Swoosie_Kurtz_41605.png
+Tamala_Jones_38648.png
+Tamara_Braun_9874.png
+Tatyana_M._Ali_1228.png
+Tatyana_M._Ali_1266.png
+Tatyana_M._Ali_1290.png
+Tatyana_M._Ali_1349.png
+Taylor_Atelian_3507.png
+Taylor_Atelian_3517.png
+Taylor_Atelian_3526.png
+Taylor_Atelian_3532.png
+Taylor_Atelian_3533.png
+Taylor_Atelian_3553.png
+Taylor_Atelian_3559.png
+Taylor_Atelian_3566.png
+Taylor_Atelian_3581.png
+Taylor_Lautner_54170.png
+Taylor_Lautner_54238.png
+Taylor_Lautner_54258.png
+Taylor_Lautner_54279.png
+Taylor_Lautner_54292.png
+Taylor_Lautner_54296.png
+Taylor_Lautner_54328.png
+Tempestt_Bledsoe_8014.png
+Tempestt_Bledsoe_8062.png
+Tempestt_Bledsoe_8063.png
+Teri_Hatcher_33393.png
+Teri_Hatcher_33395.png
+Teri_Hatcher_33401.png
+Terry_Farrell_25136.png
+Terry_Farrell_25203.png
+Tia_Carrere_12136.png
+Tia_Carrere_12240.png
+Tobey_Maguire_54398.png
+Tom_Hanks_54730.png
+Tyne_Daly_18276.png
+Tyne_Daly_18293.png
+Tyne_Daly_18322.png
+Tyne_Daly_18362.png
+Tyne_Daly_18370.png
+Tyne_Daly_18399.png
+Tyne_Daly_18420.png
+Valerie_Cruz_17158.png
+Valerie_Harper_32638.png
+Valerie_Harper_32705.png
+Valerie_Harper_32771.png
+Vanessa_Marcil_47131.png
+Vanessa_Marcil_47158.png
+Vanessa_Marcil_47182.png
+Veronica_Hamel_32010.png
+Veronica_Hamel_32016.png
+Veronica_Hamel_32020.png
+Veronica_Hamel_32030.png
+Victor_Garber_54809.png
+Victor_Garber_54894.png
+Victoria_Justice_39219.png
+Wanda_De_Jesus_19687.png
+Wendie_Malick_46848.png
+Yasmine_Bleeth_8219.png
+Yasmine_Bleeth_8238.png
+Yasmine_Bleeth_8259.png
+Yasmine_Bleeth_8350.png
+Zooey_Deschanel_20937.png
--- a/src/megaface/megaface_noises.txt
+++ b/src/megaface/megaface_noises.txt
@@ -0,0 +1,720 @@
+#Please strictly follow the rules in https://github.com/deepinsight/insightface/tree/master/src/megaface/README.md if you want to use this list.
+375/37531198@N00/4174611825_0.jpg
+375/37531198@N00/4174611825_1.jpg
+925/92596048@N00/3910487430_2.jpg
+587/58736276@N00/3023685783_2.jpg
+905/90553008@N00/824787574_0.jpg
+350/35034345966@N01/3086429616_0.jpg
+388/38847947@N03/4351812234_0.jpg
+788/78865207@N05/8267101710_0.jpg
+495/49503172960@N01/2674784671_0.jpg
+481/48122140@N00/3488870947_0.jpg
+770/77029195@N00/490108118_3.jpg
+566/56692742@N05/8202697602_1.jpg
+546/54613470@N00/6155519468_0.jpg
+486/48600102544@N01/206574155_6.jpg
+857/85705006@N00/2052818739_0.jpg
+345/34531071@N07/3374444062_1.jpg
+714/71490499@N06/6460912773_0.jpg
+790/79045485@N00/8312851597_9.jpg
+497/49703429@N00/487110767_5.jpg
+588/58819758@N00/502852994_0.jpg
+603/60370779@N07/12733436605_1.jpg
+351/35197415@N06/3392955585_0.jpg
+414/41464693@N02/4431128027_1.jpg
+745/74536253@N07/7744893278_0.jpg
+745/74536253@N07/7744893278_1.jpg
+745/74536253@N07/7744893278_2.jpg
+745/74536253@N07/7746445288_0.jpg
+720/72093892@N00/8394540214_0.jpg
+943/94306126@N05/8585087758_0.jpg
+316/31635293@N06/10300923803_0.jpg
+316/31635293@N06/10300923803_1.jpg
+461/46151867@N00/1361048872_0.jpg
+316/31658559@N04/11248071845_2.jpg
+510/51035763730@N01/3487351916_0.jpg
+381/38176611@N04/6845223441_1.jpg
+812/8127814@N07/1169159373_0.jpg
+596/59689442@N00/397368405_0.jpg
+457/45702277@N04/4404748345_2.jpg
+457/45702277@N04/4405512456_0.jpg
+315/31557769@N04/5183029470_0.jpg
+105/105042823@N05/10224905845_1.jpg
+233/23302147@N08/6717536683_0.jpg
+943/94324445@N02/8585299005_5.jpg
+865/86525452@N06/9026905203_1.jpg
+575/57593906@N00/183103326_0.jpg
+575/57593906@N00/183103326_1.jpg
+513/51313580@N08/6077728920_10.jpg
+794/79447530@N00/2646557330_3.jpg
+838/8385182@N06/8791791352_2.jpg
+834/83452322@N00/1379291226_0.jpg
+299/29932833@N06/4605882096_0.jpg
+125/12523672@N02/1368391674_0.jpg
+251/25193498@N00/4015722351_2.jpg
+207/20741443@N00/5045852254_2.jpg
+313/31369133@N04/5050193084_3.jpg
+313/31369133@N04/5050193084_5.jpg
+504/50440126@N03/8061480293_0.jpg
+353/35336901@N00/4353383267_0.jpg
+102/10249456@N08/839447586_7.jpg
+621/62158648@N07/5765099754_0.jpg
+779/77991982@N00/3351774300_0.jpg
+779/77991982@N00/3351774300_2.jpg
+207/20741443@N00/5045852254_0.jpg
+579/57954142@N07/6082189380_0.jpg
+210/21091679@N08/5811404199_1.jpg
+555/55530018@N04/5481680731_1.jpg
+830/8302818@N06/7180492423_4.jpg
+108/108788939@N08/11207124956_0.jpg
+358/35803015@N03/5571012744_5.jpg
+125/12508217@N08/3458308016_3.jpg
+526/52617239@N07/9550573423_3.jpg
+760/7600622@N07/442516245_2.jpg
+838/83855583@N00/139393859_0.jpg
+357/35775623@N00/4387679624_0.jpg
+508/50824803@N04/5101695889_0.jpg
+813/81369265@N00/5962625787_1.jpg
+849/84987970@N00/6503971169_3.jpg
+366/36692718@N00/7508265904_1.jpg
+101/101878765@N05/9792203325_2.jpg
+149/14994959@N04/3611150762_2.jpg
+257/25777683@N04/3781609919_2.jpg
+206/20625597@N07/3362909564_1.jpg
+768/7686538@N06/1370874019_0.jpg
+286/28675521@N05/2677267827_0.jpg
+958/95818633@N00/317066357_0.jpg
+114/11461909@N06/3569039014_2.jpg
+114/11461909@N06/3926753297_2.jpg
+827/8271358@N08/4748146669_0.jpg
+341/34120602@N05/5979478480_5.jpg
+114/11461909@N06/8281364025_1.jpg
+467/46721620@N00/3849799981_0.jpg
+407/40719493@N06/4008300252_1.jpg
+764/76491372@N00/540480752_1.jpg
+330/33019425@N00/275100215_2.jpg
+316/31677763@N00/273039584_0.jpg
+213/21349468@N00/82691007_2.jpg
+425/42568333@N05/4079636121_3.jpg
+972/97205184@N04/10076409423_0.jpg
+827/8279509@N08/2712311284_0.jpg
+311/31150907@N03/4133932870_0.jpg
+412/41259870@N06/5812928114_0.jpg
+332/33213804@N04/7599995760_0.jpg
+375/37552553@N03/8157072811_0.jpg
+737/7376233@N08/8712184839_0.jpg
+665/66598017@N08/9614369266_3.jpg
+352/35237099299@N01/16090400_0.jpg
+395/39545243@N00/1121668561_11.jpg
+529/52968060@N00/4836487807_1.jpg
+466/46648278@N00/68839900_0.jpg
+768/76841197@N00/349452085_0.jpg
+971/97193933@N00/7441014914_1.jpg
+711/71164686@N00/2199159152_3.jpg
+956/95633051@N00/2504698743_0.jpg
+850/85094310@N00/2536647566_0.jpg
+850/85094310@N00/2536647566_1.jpg
+546/54673576@N00/3231373075_1.jpg
+419/41942641@N02/3911619734_0.jpg
+706/70654989@N00/435992217_0.jpg
+665/66588862@N00/4990604401_0.jpg
+665/66588862@N00/4990604401_1.jpg
+175/17514301@N00/504360950_0.jpg
+242/24256726@N00/2284033395_1.jpg
+388/38847947@N03/4278156225_0.jpg
+120/12057709@N00/4817523632_5.jpg
+958/95876452@N07/8757383318_5.jpg
+958/95876452@N07/9136979141_0.jpg
+511/51135741@N00/1197314078_2.jpg
+732/73257494@N00/254360075_0.jpg
+217/21725313@N04/2735944275_4.jpg
+859/85928975@N00/3054044988_0.jpg
+803/8035175@N05/511570673_2.jpg
+169/16922208@N08/5320713367_0.jpg
+370/37016273@N02/8089047614_0.jpg
+108/10807869@N03/9369308884_2.jpg
+859/85928975@N00/3320166109_0.jpg
+365/36521958172@N01/100498068_4.jpg
+724/7243800@N05/2227611248_0.jpg
+724/7243800@N05/2227611248_3.jpg
+117/11738031@N00/2998834476_0.jpg
+795/7959345@N03/5237906433_0.jpg
+236/23634892@N07/8479172294_0.jpg
+334/33484476@N00/1631862642_0.jpg
+793/7934170@N07/4553285915_3.jpg
+603/60370779@N07/12733436605_0.jpg
+943/94366076@N00/3048525915_0.jpg
+277/27721729@N07/3509070727_2.jpg
+759/75933558@N00/3620382660_0.jpg
+460/46041863@N08/4441052512_0.jpg
+129/12989638@N00/6212397379_0.jpg
+337/33757663@N00/7358306448_0.jpg
+943/94306126@N05/8585087758_1.jpg
+953/95388136@N08/8862410806_0.jpg
+223/22320444@N08/5673545350_0.jpg
+434/43411679@N00/262399750_0.jpg
+635/63534101@N00/2646352652_1.jpg
+225/22591084@N00/6196535614_0.jpg
+635/63587775@N03/6196638593_25.jpg
+353/35399223@N08/6016093673_0.jpg
+410/41063459@N02/3864724318_0.jpg
+822/8220615@N05/4552969733_3.jpg
+549/54949957@N00/5233621265_1.jpg
+701/70177760@N04/9477029774_4.jpg
+105/10564470@N04/2648385175_0.jpg
+199/19909714@N00/3830545099_2.jpg
+758/75818171@N02/7146473317_0.jpg
+233/23354880@N03/2452852823_0.jpg
+399/39915396@N00/5193727577_1.jpg
+575/57503924@N07/6296483205_0.jpg
+430/43093552@N00/6705048843_0.jpg
+417/41739894@N04/4830148462_0.jpg
+417/41739894@N04/4830148462_1.jpg
+101/101386609@N04/13523151125_0.jpg
+381/38117284@N00/1590829110_0.jpg
+827/8279509@N08/2712311284_1.jpg
+379/37963250@N06/3588816064_0.jpg
+510/51035607419@N01/3883605563_0.jpg
+311/31150907@N03/4133932870_1.jpg
+311/31150907@N03/4333226349_0.jpg
+409/40989913@N03/5368897603_1.jpg
+332/33213804@N04/7599995760_2.jpg
+519/51973188@N06/8553574872_0.jpg
+665/66598017@N08/9614369266_1.jpg
+131/13117300@N05/2323998041_0.jpg
+489/48949906@N00/3917465141_1.jpg
+489/48949906@N00/3917465141_2.jpg
+251/25117187@N03/9118076515_0.jpg
+225/22547477@N05/10576991153_1.jpg
+169/16926338@N06/3551147251_1.jpg
+120/12057709@N00/4817523632_1.jpg
+327/32721085@N04/5095989423_2.jpg
+404/40406266@N08/7499062102_1.jpg
+227/22766186@N07/5625438347_0.jpg
+227/22766186@N07/5626032050_2.jpg
+302/30237597@N00/3761331400_0.jpg
+412/41259870@N06/5811920381_0.jpg
+369/36934384@N07/7153191901_0.jpg
+369/36934384@N07/7153191901_1.jpg
+332/33213804@N04/7599995760_6.jpg
+526/52604186@N03/9982748374_2.jpg
+125/12508217@N08/3458308016_0.jpg
+793/79383703@N08/8188874944_2.jpg
+328/32820037@N04/12550334735_0.jpg
+554/55456932@N08/5239999140_0.jpg
+193/19323934@N05/5236281857_0.jpg
+300/30035643@N03/6318064786_1.jpg
+260/26071009@N04/8133430835_0.jpg
+395/39545243@N00/1121668561_23.jpg
+927/92723020@N00/198034899_0.jpg
+732/73257494@N00/254360075_1.jpg
+768/76815233@N00/318365457_0.jpg
+768/76815233@N00/318365457_1.jpg
+758/75818171@N02/6811401849_0.jpg
+960/96009072@N00/7213113090_1.jpg
+440/44042276@N00/11078882754_1.jpg
+563/56379629@N00/231943094_0.jpg
+984/98411817@N00/2372250142_0.jpg
+984/98411817@N00/2372250142_4.jpg
+758/75815807@N00/3141877049_8.jpg
+779/77991982@N00/3351774300_1.jpg
+956/95668756@N00/9137041215_1.jpg
+664/66487272@N00/3175332074_0.jpg
+714/71484472@N08/7749429212_0.jpg
+622/62292236@N03/12280954993_0.jpg
+716/71635685@N00/2171921996_1.jpg
+703/70335407@N00/2629913923_0.jpg
+375/37526864@N03/3575384689_0.jpg
+441/44124482892@N01/4076872185_0.jpg
+760/7600622@N07/442516245_2.jpg
+923/92399379@N00/3745627039_0.jpg
+236/23634892@N07/8479172294_1.jpg
+620/62092813@N00/12460977104_8.jpg
+486/48600102544@N01/206574155_2.jpg
+486/48600102544@N01/206574155_9.jpg
+120/12057709@N00/4817523632_4.jpg
+421/42197860@N05/7964010316_0.jpg
+790/79045485@N00/8312851597_13.jpg
+943/94366076@N00/3048525915_2.jpg
+904/9047144@N06/4039658431_2.jpg
+222/22292214@N00/4243141797_0.jpg
+736/73694732@N00/4431487460_0.jpg
+527/52706816@N04/4862401074_0.jpg
+696/69654695@N04/8112732403_0.jpg
+147/14754516@N00/6746387961_0.jpg
+497/49703429@N00/487110767_7.jpg
+145/14537247@N02/3736211228_8.jpg
+286/28629203@N00/3064397634_3.jpg
+332/33213624@N05/3106373748_0.jpg
+334/33415234@N02/3114165513_1.jpg
+314/31492856@N08/3175678730_1.jpg
+333/33369864@N07/3244523696_4.jpg
+309/30975003@N06/3322588992_1.jpg
+262/26212498@N08/3372483710_1.jpg
+373/37310116@N04/3434757683_0.jpg
+246/24652718@N07/3504274334_0.jpg
+311/31170710@N02/3541648434_0.jpg
+868/8683186@N05/3611020792_0.jpg
+868/8683186@N05/3611020792_6.jpg
+145/14537247@N02/3736211228_1.jpg
+145/14537247@N02/3736211228_10.jpg
+145/14537247@N02/3736211228_3.jpg
+145/14537247@N02/3736211228_7.jpg
+403/40384578@N06/4101631794_0.jpg
+495/49503010002@N01/4159540475_1.jpg
+868/8683186@N05/4808516991_6.jpg
+412/41275121@N06/4903601396_5.jpg
+529/52947003@N08/5062807160_1.jpg
+529/52947003@N08/5062807160_10.jpg
+529/52947003@N08/5062807160_6.jpg
+556/55642376@N02/5907157097_2.jpg
+556/55642376@N02/5907157097_3.jpg
+294/29475291@N05/6520878321_0.jpg
+788/78832981@N04/7439193650_0.jpg
+861/86125876@N07/7890416460_0.jpg
+495/49503124519@N01/8054014324_8.jpg
+170/17040371@N08/8095753999_5.jpg
+170/17040371@N08/8095788948_4.jpg
+336/33672038@N05/8147806935_1.jpg
+340/34085730@N06/8292774555_1.jpg
+268/26881063@N08/13878266544_2.jpg
+309/30932831@N00/2773069134_2.jpg
+276/27674701@N00/306654282_1.jpg
+402/40272616@N07/4800636882_1.jpg
+980/98055082@N00/992566662_1.jpg
+232/23206546@N04/8621794327_0.jpg
+232/23206546@N04/8621794327_1.jpg
+224/22491837@N08/3626048130_0.jpg
+224/22491837@N08/3641245154_2.jpg
+950/95026061@N00/2542285242_0.jpg
+759/75923111@N02/8430338965_0.jpg
+106/10647915@N04/6503063251_1.jpg
+759/7598168@N03/5226476291_0.jpg
+105/105042823@N05/10224905845_2.jpg
+450/45019754@N02/7606879826_2.jpg
+943/94324445@N02/8585299005_4.jpg
+310/31083236@N06/2986649565_0.jpg
+441/44124482892@N01/4076872185_4.jpg
+292/29276771@N03/11641854816_0.jpg
+363/36382552@N04/5425274524_3.jpg
+767/76795962@N03/12297733704_0.jpg
+121/121110687@N08/13323427194_2.jpg
+229/22969849@N00/1349964647_17.jpg
+371/37147296@N03/3720755210_3.jpg
+366/36684003@N06/4437979063_0.jpg
+533/53384351@N06/4978715463_0.jpg
+125/12572929@N05/5221666460_1.jpg
+641/64158255@N00/5489160164_0.jpg
+299/29912007@N05/5921111759_2.jpg
+388/38820321@N06/6281589012_0.jpg
+266/26657772@N02/7323904516_0.jpg
+812/81253095@N08/7563494124_0.jpg
+812/81253095@N08/7563494124_7.jpg
+131/13117300@N05/2323998041_1.jpg
+239/23927487@N05/4499177940_1.jpg
+624/62477948@N02/5691704999_3.jpg
+458/45842803@N00/2702028061_3.jpg
+213/21385420@N00/9351830159_2.jpg
+111/11149280@N02/9354553803_0.jpg
+297/29785808@N04/2869578452_0.jpg
+581/58167807@N00/4216179051_0.jpg
+309/30932831@N00/2773069134_0.jpg
+113/113235388@N08/11701952496_5.jpg
+431/43152922@N00/3293843245_0.jpg
+270/27088911@N08/5601971419_0.jpg
+291/29123483@N00/8613949068_1.jpg
+619/61992454@N00/119165477_2.jpg
+260/26071009@N04/8133430835_1.jpg
+561/56198214@N00/11354524454_0.jpg
+295/29599105@N00/5802624346_6.jpg
+330/33085931@N08/9400799871_0.jpg
+341/34128229@N06/3267947070_0.jpg
+213/21385420@N00/9351830159_0.jpg
+758/75815807@N00/3141877049_7.jpg
+529/52909086@N08/4912165639_0.jpg
+923/92319630@N00/2022738688_0.jpg
+318/31826784@N06/3346271475_0.jpg
+437/43791698@N07/4031782636_4.jpg
+269/26966164@N08/2699439482_3.jpg
+525/52545972@N00/7991923484_0.jpg
+250/25030443@N03/2852010154_6.jpg
+886/88664590@N00/702044241_0.jpg
+658/65873073@N00/8667101819_1.jpg
+274/27433266@N00/5972034863_0.jpg
+292/29233640@N07/3691901778_5.jpg
+122/12254431@N02/5261928850_0.jpg
+765/76562640@N00/2561153799_0.jpg
+664/66487272@N00/3175332074_3.jpg
+349/34946027@N06/3899131678_3.jpg
+914/91409706@N00/7296129156_3.jpg
+914/91409706@N00/7296129156_5.jpg
+329/32996060@N05/7496967290_1.jpg
+241/24112094@N00/7545432624_0.jpg
+588/58871905@N03/8212289784_0.jpg
+284/28426408@N00/4985494194_0.jpg
+315/31594932@N00/8242400525_1.jpg
+315/31594932@N00/8243469114_0.jpg
+500/50016899@N03/4605388504_0.jpg
+263/26357527@N05/7223540266_2.jpg
+498/49814762@N00/8014123645_0.jpg
+498/49814762@N00/8014123645_1.jpg
+875/87504239@N08/10670731404_4.jpg
+875/87504239@N08/10852751085_3.jpg
+372/37244828@N04/13173856803_10.jpg
+371/37147296@N03/3720755210_1.jpg
+125/12572929@N05/5221666460_2.jpg
+641/64158255@N00/5489160164_2.jpg
+356/35658425@N08/5726282580_0.jpg
+766/76681787@N00/5957052480_2.jpg
+812/81253095@N08/7563494124_10.jpg
+812/81253095@N08/7563494124_6.jpg
+699/69958247@N05/13388922034_0.jpg
+748/74896593@N07/6764119963_0.jpg
+846/84655869@N08/7775105362_0.jpg
+181/18155385@N00/239228291_7.jpg
+421/42103613@N08/4040845679_1.jpg
+313/31369133@N04/5050193084_4.jpg
+785/78573292@N00/2256084892_0.jpg
+167/16759096@N00/489481531_0.jpg
+957/95723376@N06/8735477642_1.jpg
+863/8630870@N02/5322350325_0.jpg
+863/8630870@N02/5322350325_2.jpg
+863/8630870@N02/5322350325_3.jpg
+863/8630870@N02/5322350325_4.jpg
+863/8630870@N02/5322350325_5.jpg
+735/73531427@N00/218215108_0.jpg
+309/30932831@N00/2773069134_3.jpg
+727/72707136@N00/33992579_1.jpg
+762/76236359@N00/6234006071_5.jpg
+736/73626930@N00/9428192473_0.jpg
+105/105042823@N05/10224905845_4.jpg
+284/28476480@N04/3975041778_0.jpg
+943/94324445@N02/8585299005_1.jpg
+601/60125017@N00/344049639_2.jpg
+876/87619178@N03/8590731823_0.jpg
+277/27712137@N04/2623505924_0.jpg
+701/70154022@N00/3925362192_1.jpg
+459/45975847@N07/4691684655_2.jpg
+459/45975847@N07/4692316090_0.jpg
+514/51460103@N07/4823559383_0.jpg
+220/22072051@N03/5011260955_0.jpg
+535/53533856@N08/5361207739_0.jpg
+535/53533856@N08/5361207739_1.jpg
+535/53533856@N08/5361207739_4.jpg
+528/52858230@N08/5564688606_0.jpg
+574/57438188@N07/5627487722_0.jpg
+176/17694278@N04/6508529521_0.jpg
+362/36204621@N08/6698329979_0.jpg
+697/69794372@N08/6785690647_2.jpg
+100/100739634@N06/9717775343_0.jpg
+267/26728047@N05/5524462661_5.jpg
+156/15699085@N05/6408711257_0.jpg
+327/32721085@N04/5095989423_0.jpg
+669/66964839@N00/8741217920_1.jpg
+608/60842486@N00/2702417524_0.jpg
+206/20625597@N07/3362909564_5.jpg
+554/55456932@N08/5239999140_1.jpg
+616/61628022@N02/6999143216_0.jpg
+685/68558611@N06/6390155109_1.jpg
+770/77029195@N00/490108118_4.jpg
+258/25827417@N00/2574466319_0.jpg
+878/87855339@N00/3244971644_0.jpg
+878/87855339@N00/3244971644_1.jpg
+654/65497908@N00/6813913952_1.jpg
+868/8683186@N05/3611020792_2.jpg
+868/8683186@N05/4808516991_9.jpg
+294/29475291@N05/6520878321_1.jpg
+716/71678139@N05/6807828131_0.jpg
+716/71678139@N05/6807828131_1.jpg
+716/71678139@N05/6807828131_2.jpg
+716/71678139@N05/6807828131_3.jpg
+716/71678139@N05/6807828131_4.jpg
+716/71678139@N05/6807828131_5.jpg
+716/71678139@N05/6807828131_6.jpg
+170/17040371@N08/8095753999_3.jpg
+170/17040371@N08/8095788948_3.jpg
+336/33672038@N05/8147806935_2.jpg
+441/44124395142@N01/407637002_0.jpg
+795/79543373@N00/2529893768_1.jpg
+575/57545119@N04/8673142329_0.jpg
+359/35909637@N06/3917967715_0.jpg
+407/40719493@N06/4008300252_0.jpg
+764/76491372@N00/540480752_0.jpg
+221/22132798@N08/8018992056_1.jpg
+402/40245280@N00/1936928371_2.jpg
+114/114772050@N03/12008186175_0.jpg
+636/63601558@N00/199623056_0.jpg
+185/18502090@N00/2524917681_47.jpg
+793/79398354@N00/2622290309_0.jpg
+898/8982863@N07/10257490644_5.jpg
+898/8982863@N07/10257490644_8.jpg
+458/45861060@N00/7900853588_1.jpg
+102/102077743@N07/12190186456_0.jpg
+228/22882274@N04/3632475832_2.jpg
+233/23357263@N03/3953791578_5.jpg
+317/31769130@N03/11011170923_0.jpg
+371/37195519@N02/3424150218_0.jpg
+795/79589933@N00/3352567863_0.jpg
+623/62362697@N00/8577869641_3.jpg
+365/36521958172@N01/100498068_1.jpg
+898/8982863@N07/10257490644_10.jpg
+102/102077743@N07/12190186456_2.jpg
+354/35468148224@N01/12347079_0.jpg
+905/90516586@N00/159075833_0.jpg
+773/77364737@N00/2503183115_0.jpg
+975/97584199@N00/2632241655_1.jpg
+527/52772894@N00/90756428_0.jpg
+527/52772894@N00/90756428_2.jpg
+972/97205184@N04/10076409423_0.jpg
+101/101386609@N04/13523144805_0.jpg
+827/8279509@N08/2712311284_0.jpg
+379/37963250@N06/3588816064_1.jpg
+379/37963250@N06/3588816064_2.jpg
+510/51035607419@N01/3883605563_1.jpg
+311/31150907@N03/4133932870_0.jpg
+409/40989913@N03/5368897603_0.jpg
+412/41259870@N06/5812928114_0.jpg
+332/33213804@N04/7599995760_0.jpg
+332/33213804@N04/7599995760_1.jpg
+375/37552553@N03/8157072811_0.jpg
+665/66598017@N08/9614369266_3.jpg
+241/24112094@N00/3994684107_2.jpg
+388/38847947@N03/4351812234_5.jpg
+510/51035620166@N01/237428859_2.jpg
+450/45019754@N02/7606879826_0.jpg
+664/66474562@N00/19195378_2.jpg
+287/28771658@N03/3905862519_0.jpg
+123/12360228@N04/7166710688_1.jpg
+123/12360228@N04/7166710688_2.jpg
+302/30245869@N07/2843650382_4.jpg
+972/97205184@N04/10079707226_0.jpg
+570/57023246@N00/4623144791_0.jpg
+332/33213804@N04/7599995760_4.jpg
+487/48782814@N07/13847705683_0.jpg
+595/59533494@N04/5479759436_0.jpg
+938/93841400@N00/8327059841_1.jpg
+824/8240241@N06/4114864483_0.jpg
+513/51313580@N08/6077728920_5.jpg
+289/28917877@N00/6333929606_0.jpg
+395/39545243@N00/1121668561_35.jpg
+640/64073015@N00/2831021702_0.jpg
+141/14195956@N04/3948300818_3.jpg
+539/53906287@N00/5101060235_7.jpg
+500/50034633@N05/4837007562_1.jpg
+254/25414047@N00/145668608_3.jpg
+489/48949906@N00/3917465141_0.jpg
+808/80812769@N00/442495313_0.jpg
+459/45964884@N08/5242276755_2.jpg
+103/10394437@N03/4568233297_5.jpg
+169/16989146@N06/7641494140_4.jpg
+597/59751999@N00/3289666452_1.jpg
+437/43709093@N07/4256796727_0.jpg
+445/44521275@N00/506634574_2.jpg
+716/71616417@N05/6509071195_2.jpg
+716/71616417@N05/6509074321_2.jpg
+524/52407821@N00/7984244356_0.jpg
+366/36624962@N03/8288304168_3.jpg
+697/69756126@N00/4662338037_0.jpg
+631/63177605@N08/6792309851_2.jpg
+616/61670399@N07/10102978284_1.jpg
+105/105042823@N05/10225015423_0.jpg
+273/27357514@N00/11343847725_0.jpg
+463/46327697@N00/2124534031_4.jpg
+984/98411817@N00/2372250142_6.jpg
+257/25797798@N07/3211844955_1.jpg
+111/11134149@N02/3317710642_0.jpg
+388/38847947@N03/4178270889_3.jpg
+822/8220615@N05/4552969733_5.jpg
+611/61109191@N00/5543733500_2.jpg
+806/80604586@N07/9070005912_4.jpg
+241/24113869@N07/9379713988_1.jpg
+748/74888386@N00/9692493485_0.jpg
+695/69501650@N00/46407370_1.jpg
+950/95082307@N00/8773335932_0.jpg
+510/51035620166@N01/237428859_2.jpg
+434/43411679@N00/262399750_3.jpg
+634/63465779@N07/6362029047_0.jpg
+987/98736785@N00/4499296_3.jpg
+715/7155702@N03/2302080410_0.jpg
+715/7155702@N03/2302080410_1.jpg
+715/7155702@N03/2302080410_3.jpg
+715/7155702@N03/2302080410_4.jpg
+715/7155702@N03/2302080410_5.jpg
+715/7155702@N03/3631806363_2.jpg
+874/87472210@N00/3105959620_1.jpg
+246/24638567@N00/3321513653_0.jpg
+945/94545836@N00/1443145007_1.jpg
+362/36298222@N08/5935447071_0.jpg
+222/22284790@N06/2176524544_0.jpg
+747/7478277@N04/3155354570_0.jpg
+747/7478277@N04/3155354570_1.jpg
+375/37531198@N00/4174611825_2.jpg
+375/37531198@N00/4174611825_4.jpg
+768/76841197@N00/349452085_2.jpg
+768/76841197@N00/349452085_3.jpg
+468/46868174@N08/6511978997_0.jpg
+100/10058188@N05/8295748234_0.jpg
+108/108285828@N06/10779570026_1.jpg
+108/108285828@N06/10779778303_0.jpg
+168/16873194@N05/1801191486_0.jpg
+857/8572970@N04/2079107678_1.jpg
+472/47253165@N00/360990372_0.jpg
+647/64767764@N00/450008059_0.jpg
+729/72971773@N00/58588039_0.jpg
+911/91148289@N00/7916666268_1.jpg
+416/41608212@N00/8718283094_0.jpg
+911/91113759@N00/2539574881_0.jpg
+293/29322132@N08/3569343093_3.jpg
+876/87671668@N00/9013803280_0.jpg
+381/38176611@N04/8638753913_4.jpg
+976/97644214@N00/6173832483_1.jpg
+232/23206546@N04/8621794327_6.jpg
+563/56379629@N00/231943094_3.jpg
+864/86429182@N00/330808508_1.jpg
+353/35336901@N00/4353383267_1.jpg
+631/63122283@N06/7978644677_2.jpg
+932/93211492@N06/8487757725_206.jpg
+575/57545119@N04/8673142329_3.jpg
+354/35479068@N06/3306075644_0.jpg
+277/27721729@N07/3509070727_0.jpg
+188/18899983@N00/4100299699_2.jpg
+213/21349468@N00/82691007_1.jpg
+110/11021333@N02/2957267195_2.jpg
+925/9259500@N03/5710775169_0.jpg
+486/48600102544@N01/206574155_11.jpg
+486/48600102544@N01/206574155_12.jpg
+486/48600102544@N01/206574155_7.jpg
+339/33959433@N05/8396364213_0.jpg
+997/9975353@N03/12630170935_1.jpg
+247/24733288@N07/5846541456_0.jpg
+872/8729914@N05/5099947628_0.jpg
+889/88976184@N05/8244475370_1.jpg
+889/88976184@N05/8244475370_3.jpg
+630/63095109@N00/1079079769_0.jpg
+664/66422694@N00/182828985_0.jpg
+645/64503524@N00/2480285692_0.jpg
+864/86429182@N00/330808508_0.jpg
+447/44754496@N00/4287694240_1.jpg
+363/36382552@N04/5425274524_6.jpg
+210/21091679@N08/5811404199_2.jpg
+562/56213435@N08/9008148399_0.jpg
+695/69501650@N00/46407370_0.jpg
+889/88976184@N05/8244475370_0.jpg
+889/88976184@N05/8244475370_2.jpg
+889/88976184@N05/8244475370_4.jpg
+889/88976184@N05/8244475370_5.jpg
+527/52706816@N04/4862401074_1.jpg
+794/79428123@N06/7349516920_1.jpg
+790/79024059@N04/7351250586_0.jpg
+693/69318970@N00/9266107252_1.jpg
+943/94324445@N02/8585410655_1.jpg
+599/59947594@N00/5684370261_0.jpg
+849/84986567@N00/177640275_0.jpg
+345/34514184@N03/3562741086_0.jpg
+158/15803691@N00/6914149962_2.jpg
+817/81708918@N03/7855914580_5.jpg
+309/30975003@N06/3322588992_8.jpg
+900/90054085@N00/43987316_0.jpg
+136/13675730@N06/2452368334_0.jpg
+386/38675455@N00/6803844225_0.jpg
+717/71763791@N00/5532357066_0.jpg
+606/60604521@N07/5584865343_0.jpg
+395/39545243@N00/1121668561_29.jpg
+391/39196943@N05/3992501716_2.jpg
+207/20719052@N04/8646466859_0.jpg
+207/20719052@N04/8646466859_1.jpg
+758/75815807@N00/3141877049_0.jpg
+471/47170787@N05/6477792561_2.jpg
+317/31734244@N00/542792072_0.jpg
+931/9319402@N03/3358320101_0.jpg
+362/36223735@N03/3784170734_3.jpg
+534/53400644@N05/5044502598_0.jpg
+513/51313580@N08/6077728920_1.jpg
+137/13774680@N04/12313323523_0.jpg
+636/63608960@N08/13553141815_2.jpg
+434/43463445@N07/5723825501_0.jpg
+237/23740675@N05/6877915376_0.jpg
+868/8683186@N05/3611020792_4.jpg
+216/21653251@N04/5633592631_0.jpg
+216/21653251@N04/5633592631_1.jpg
+170/17040371@N08/8095753999_1.jpg
+170/17040371@N08/8095788948_1.jpg
+620/62092813@N00/12460977104_4.jpg
+236/23686718@N08/10662589695_3.jpg
+264/26451473@N00/1342290947_5.jpg
+636/63608960@N08/13553141815_6.jpg
+409/40926212@N05/13257463775_7.jpg
+449/44915880@N03/4248497494_1.jpg
+235/23588194@N00/4494098639_1.jpg
+532/53201041@N07/4911217327_0.jpg
+532/53201041@N07/4911259327_0.jpg
+606/60625084@N04/6450812495_0.jpg
+650/65013293@N02/7724906252_3.jpg
+560/56021903@N02/8227297222_2.jpg
+359/35904083@N08/4345934789_0.jpg
+593/59319260@N00/8743143610_2.jpg
+271/27125798@N05/2710711437_0.jpg
+402/40210933@N02/6043303646_0.jpg
+122/12280910@N06/6299940378_1.jpg
+207/20775043@N07/3627508299_0.jpg
+595/59592751@N08/9096179270_0.jpg
+395/39545243@N00/1121668561_14.jpg
+296/29625288@N00/12891226005_0.jpg
+548/54858071@N00/205845735_1.jpg
+751/75148156@N00/3485469298_1.jpg
+407/40764492@N08/3797968770_0.jpg
+333/33369864@N07/3244523696_0.jpg
+312/31236383@N03/3259555928_1.jpg
+385/38524878@N04/3563411595_0.jpg
+259/25955358@N06/3921002900_2.jpg
+788/78832981@N04/7447167480_1.jpg
+696/69696568@N05/6872955321_1.jpg
+757/75741170@N02/6829396274_0.jpg
+724/72448202@N00/359801202_1.jpg
+969/96934953@N00/4315376370_0.jpg
+133/13361855@N00/989708629_1.jpg
+664/66478195@N00/11649050243_3.jpg
+122/12276055@N02/4214939371_0.jpg
+463/46327697@N00/2124534031_3.jpg
+534/53400644@N05/5044502598_1.jpg
+669/66944684@N03/6306275408_0.jpg
+710/71006820@N00/4037980088_0.jpg
+766/76656964@N04/9363467181_3.jpg
+112/11262617@N07/12873348265_0.jpg
+409/40926212@N05/13257463775_4.jpg
+409/40926212@N05/13835647984_0.jpg
+528/52858230@N08/5636750038_0.jpg
+528/52858230@N08/5636750038_1.jpg
+203/20335591@N04/9752849313_1.jpg
+350/35034345966@N01/3086429616_1.jpg
+736/73694732@N00/4430689233_0.jpg
+605/60517117@N00/10370054943_0.jpg
+421/42173774@N03/3889203507_0.jpg
+511/51135741@N00/1197314078_3.jpg
+267/26753167@N04/5922076258_2.jpg
+330/33062815@N00/8606375434_1.jpg
+108/10807869@N03/9369308884_0.jpg
+761/76151808@N00/6945480767_0.jpg
+757/75741170@N02/6829396274_1.jpg
+220/22057861@N07/2126995093_0.jpg
+665/66572814@N00/8242014606_0.jpg
+849/84986567@N00/177640275_1.jpg
+444/44460493@N03/4085897885_0.jpg
+330/33085931@N08/9400799871_1.jpg
+533/53314395@N00/5174087013_11.jpg
+864/86429182@N00/330808508_2.jpg
+151/15133426@N08/3646529532_0.jpg
+302/30223854@N00/6535512351_0.jpg
+929/92903356@N04/8497572867_0.jpg
+412/41232325@N03/9045866185_0.jpg
+582/58246614@N00/2842715287_1.jpg
+513/51304493@N05/4932744035_0.jpg
+301/30120696@N02/3068639840_0.jpg
+407/40764492@N08/3797968770_0.jpg
+231/23100621@N04/3986465605_0.jpg
+288/28891066@N00/4210148052_2.jpg
+940/94066910@N00/448278027_1.jpg
+532/53282124@N00/3858373020_0.jpg
+597/59751999@N00/3289666452_2.jpg
+365/36521958172@N01/100498068_4.jpg
+293/29322132@N08/3569343093_3.jpg
--- a/src/megaface/remove_noises.py
+++ b/src/megaface/remove_noises.py
@@ -0,0 +1,175 @@
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+
+import os.path
+import datetime
+from easydict import EasyDict as edict
+import time
+import json
+import shutil
+import sys
+import numpy as np
+import importlib
+import itertools
+import argparse
+import struct
+import cv2
+sys.path.append(os.path.join(os.path.dirname(__file__), '..', 'common'))
+import face_preprocess
+from sklearn.preprocessing import normalize
+import mxnet as mx
+from mxnet import ndarray as nd
+
+
+feature_dim = 512
+feature_ext = 1
+
+def load_bin(path, fill = 0.0):
+  with open(path, 'rb') as f:
+    bb = f.read(4*4)
+    #print(len(bb))
+    v = struct.unpack('4i', bb)
+    #print(v[0])
+    bb = f.read(v[0]*4)
+    v = struct.unpack("%df"%(v[0]), bb)
+    feature = np.full( (feature_dim+feature_ext,), fill, dtype=np.float32)
+    feature[0:feature_dim] = v
+    #feature = np.array( v, dtype=np.float32)
+  #print(feature.shape)
+  #print(np.linalg.norm(feature))
+  return feature
+
+def write_bin(path, feature):
+  feature = list(feature)
+  with open(path, 'wb') as f:
+    f.write(struct.pack('4i', len(feature),1,4,5))
+    f.write(struct.pack("%df"%len(feature), *feature))
+
+def main(args):
+
+  out_algo = args.suffix
+  if len(args.algo)>0:
+    out_algo = args.algo
+
+  fs_noise_map = {}
+  for line in open(args.facescrub_noises, 'r'):
+    if line.startswith('#'):
+      continue
+    line = line.strip()
+    fname = line.split('.')[0]
+    p = fname.rfind('_')
+    fname = fname[0:p]
+    fs_noise_map[line] = fname
+
+  print(len(fs_noise_map))
+
+  i=0
+  fname2center = {}
+  noises = []
+  for line in open(args.facescrub_lst, 'r'):
+    if i%1000==0:
+      print("reading fs",i)
+    i+=1
+    image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
+    assert aligned==True
+    _path = image_path.split('/')
+    a, b = _path[-2], _path[-1]
+    feature_path = os.path.join(args.facescrub_feature_dir, a, "%s_%s.bin"%(b, args.suffix))
+    feature_dir_out = os.path.join(args.facescrub_feature_dir_out, a)
+    if not os.path.exists(feature_dir_out):
+      os.makedirs(feature_dir_out)
+    feature_path_out = os.path.join(args.facescrub_feature_dir_out, a, "%s_%s.bin"%(b, out_algo))
+    #print(b)
+    if not b in fs_noise_map:
+      #shutil.copyfile(feature_path, feature_path_out)
+      feature = load_bin(feature_path)
+      write_bin(feature_path_out, feature)
+      if not a in fname2center:
+        fname2center[a] = np.zeros((feature_dim+feature_ext,), dtype=np.float32)
+      fname2center[a] += feature
+    else:
+      #print('n', b)
+      noises.append( (a,b) )
+  print(len(noises))
+
+  for k in noises:
+    a,b = k
+    assert a in fname2center
+    center = fname2center[a]
+    g = np.zeros( (feature_dim+feature_ext,), dtype=np.float32)
+    g2 = np.random.uniform(-0.001, 0.001, (feature_dim,))
+    g[0:feature_dim] = g2
+    f = center+g
+    _norm=np.linalg.norm(f)
+    f /= _norm
+    feature_path_out = os.path.join(args.facescrub_feature_dir_out, a, "%s_%s.bin"%(b, out_algo))
+    write_bin(feature_path_out, f)
+
+  mf_noise_map = {}
+  for line in open(args.megaface_noises, 'r'):
+    if line.startswith('#'):
+      continue
+    line = line.strip()
+    _vec = line.split("\t")
+    if len(_vec)>1:
+      line = _vec[1]
+    mf_noise_map[line] = 1
+
+  print(len(mf_noise_map))
+
+  i=0
+  nrof_noises = 0
+  for line in open(args.megaface_lst, 'r'):
+    if i%1000==0:
+      print("reading mf",i)
+    i+=1
+    image_path, label, bbox, landmark, aligned = face_preprocess.parse_lst_line(line)
+    assert aligned==True
+    _path = image_path.split('/')
+    a1, a2, b = _path[-3], _path[-2], _path[-1]
+    feature_path = os.path.join(args.megaface_feature_dir, a1, a2, "%s_%s.bin"%(b, args.suffix))
+    feature_dir_out = os.path.join(args.megaface_feature_dir_out, a1, a2)
+    if not os.path.exists(feature_dir_out):
+      os.makedirs(feature_dir_out)
+    feature_path_out = os.path.join(args.megaface_feature_dir_out, a1, a2, "%s_%s.bin"%(b, out_algo))
+    bb = '/'.join([a1, a2, b])
+    #print(b)
+    if not bb in mf_noise_map:
+      feature = load_bin(feature_path)
+      write_bin(feature_path_out, feature)
+      #shutil.copyfile(feature_path, feature_path_out)
+    else:
+      feature = load_bin(feature_path, 100.0)
+      write_bin(feature_path_out, feature)
+      #g = np.random.uniform(-0.001, 0.001, (feature_dim,))
+      #print('n', bb)
+      #write_bin(feature_path_out, g)
+      nrof_noises+=1
+  print(nrof_noises)
+
+
+def parse_arguments(argv):
+  parser = argparse.ArgumentParser()
+  
+  parser.add_argument('--facescrub-noises', type=str, help='', default='./facescrub_noises.txt')
+  parser.add_argument('--megaface-noises', type=str, help='', default='./megaface_noises.txt')
+  parser.add_argument('--suffix', type=str, help='', default='r100_cm_112x112')
+  parser.add_argument('--algo', type=str, help='', default='')
+  parser.add_argument('--megaface-lst', type=str, help='', default='/raid5data/dplearn/megaface/megaface_mtcnn_112x112/lst')
+  parser.add_argument('--facescrub-lst', type=str, help='', default='/raid5data/dplearn/megaface/facescrubr/small_lst')
+  parser.add_argument('--megaface-feature-dir', type=str, help='', default='/raid5data/dplearn/megaface/MegaFace_Features')
+  parser.add_argument('--facescrub-feature-dir', type=str, help='', default='/raid5data/dplearn/megaface/FaceScrub_Features')
+  #parser.add_argument('--megaface-feature-dir-out', type=str, help='', default='/raid5data/dplearn/megaface/MegaFace_Features_cm')
+  #parser.add_argument('--facescrub-feature-dir-out', type=str, help='', default='/raid5data/dplearn/megaface/FaceScrub_Features_cm')
+  parser.add_argument('--megaface-feature-dir-out', type=str, help='', default='/opt/jiaguo/MegaFace_Features_cm')
+  parser.add_argument('--facescrub-feature-dir-out', type=str, help='', default='/opt/jiaguo/FaceScrub_Features_cm')
+  return parser.parse_args(argv)
+
+if __name__ == '__main__':
+  main(parse_arguments(sys.argv[1:]))
+
+
--- a/src/train_softmax.py
+++ b/src/train_softmax.py
@@ -17,6 +17,7 @@ import argparse
 import mxnet.optimizer as optimizer
 sys.path.append(os.path.join(os.path.dirname(__file__), 'common'))
 import face_image
+from noise_sgd import NoiseSGD
 sys.path.append(os.path.join(os.path.dirname(__file__), 'eval'))
 sys.path.append(os.path.join(os.path.dirname(__file__), 'symbols'))
 import fresnet
@@ -53,7 +54,7 @@ class AccMetric(mx.metric.EvalMetric):

  def update(self, labels, preds):
    self.count+=1
-    if args.loss_type>=2 and args.loss_type<=4 and args.margin_verbose>0:
+    if args.loss_type>=2 and args.loss_type<=5 and args.margin_verbose>0:
      if self.count%args.ctx_num==0:
        mbatch = self.count//args.ctx_num
        if mbatch==1 or mbatch%args.margin_verbose==0:
@@ -123,10 +124,9 @@ def parse_args():
  parser.add_argument('--version-unit', type=int, default=3, help='')
  parser.add_argument('--end-epoch', type=int, default=100000,
      help='training epoch size.')
-  parser.add_argument('--lr', type=float, default=0.1,
-      help='')
-  parser.add_argument('--wd', type=float, default=0.0005,
-      help='')
+  parser.add_argument('--noise-sgd', type=float, default=0.0, help='')
+  parser.add_argument('--lr', type=float, default=0.1, help='')
+  parser.add_argument('--wd', type=float, default=0.0005, help='')
  parser.add_argument('--mom', type=float, default=0.9,
      help='')
  parser.add_argument('--emb-size', type=int, default=512,
@@ -176,6 +176,7 @@ def parse_args():
      help='')
  parser.add_argument('--rand-mirror', type=int, default=1,
      help='')
+  parser.add_argument('--cutoff', type=int, default=0, help='')
  parser.add_argument('--patch', type=str, default='0_0_96_112_0',
      help='')
  parser.add_argument('--lr-steps', type=str, default='', help='')
@@ -300,6 +301,20 @@ def get_symbol(args, arg_params, aux_params):
    cos_t = zy/s
    if args.margin_verbose>0:
      margin_symbols.append(mx.symbol.mean(cos_t))
+    if m>1.0:
+      t = mx.sym.arccos(cos_t)
+      t = t*m
+      body = mx.sym.cos(t)
+      new_zy = body*s
+      if args.margin_verbose>0:
+        new_cos_t = new_zy/s
+        margin_symbols.append(mx.symbol.mean(new_cos_t))
+      diff = new_zy - zy
+      diff = mx.sym.expand_dims(diff, 1)
+      gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
+      body = mx.sym.broadcast_mul(gt_one_hot, diff)
+      fc7 = fc7+body
+
    #threshold = math.cos(args.margin_m)
    #cond_v = cos_t - threshold
    #cond = mx.symbol.Activation(data=cond_v, act_type='relu')
@@ -376,6 +391,21 @@ def get_symbol(args, arg_params, aux_params):
    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = 1.0, off_value = 0.0)
    body = mx.sym.broadcast_mul(gt_one_hot, diff)
    fc7 = fc7+body
+  elif args.loss_type==5:
+    s = args.margin_s
+    m = args.margin_m
+    assert s>0.0
+    assert m>=0.0
+    assert m<(math.pi/2)
+    _weight = mx.symbol.Variable("fc7_weight", shape=(args.num_classes, args.emb_size), lr_mult=1.0)
+    _weight = mx.symbol.L2Normalization(_weight, mode='instance')
+    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
+    cos_a = mx.sym.FullyConnected(data=nembedding, weight = _weight, no_bias = True, num_hidden=args.num_classes, name='fc7')
+    theta_a = mx.sym.arccos(cos_a)
+    gt_one_hot = mx.sym.one_hot(gt_label, depth = args.num_classes, on_value = m, off_value = 0.0)
+    theta_a = theta_a+gt_one_hot
+    fc7 = math.pi/2 - theta_a
+    fc7 = fc7*s
  elif args.loss_type==10: #marginal loss
    nembedding = mx.symbol.L2Normalization(embedding, mode='instance', name='fc1n')
    params = [1.2, 0.3, 1.0]
@@ -700,6 +730,7 @@ def train_net(args):
          shuffle              = True,
          rand_mirror          = args.rand_mirror,
          mean                 = mean,
+          cutoff               = args.cutoff,
          c2c_threshold        = args.c2c_threshold,
          output_c2c           = args.output_c2c,
          c2c_mode             = args.c2c_mode,
@@ -723,6 +754,7 @@ def train_net(args):
            shuffle              = True,
            rand_mirror          = args.rand_mirror,
            mean                 = mean,
+            cutoff               = args.cutoff,
            c2c_threshold        = args.c2c_threshold,
            output_c2c           = args.output_c2c,
            c2c_mode             = args.c2c_mode,
@@ -752,7 +784,11 @@ def train_net(args):
    else:
      initializer = mx.init.Xavier(rnd_type='uniform', factor_type="in", magnitude=2)
    _rescale = 1.0/args.ctx_num
-    opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
+    if args.noise_sgd>0.0:
+      print('use noise sgd')
+      opt = NoiseSGD(scale = args.noise_sgd, learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
+    else:
+      opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=_rescale)
    som = 20
    if args.loss_type==12 or args.loss_type==13:
      som = 2
--- a/src/train_tripletloss.py
+++ b/src/train_tripletloss.py
@@ -1,377 +0,0 @@
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import sys
-import random
-import logging
-import numpy as np
-from data import FaceIter
-import mxnet as mx
-from mxnet import ndarray as nd
-import argparse
-import mxnet.optimizer as optimizer
-import mxcommon.resnet_dcn as resnet_dcn
-import lfw
-import sklearn
-from sklearn.decomposition import PCA
-from center_loss import *
-
-
-logger = logging.getLogger()
-logger.setLevel(logging.INFO)
-
-class AccMetric(mx.metric.EvalMetric):
-  def __init__(self):
-    self.axis = 1
-    super(AccMetric, self).__init__(
-        'acc', axis=self.axis,
-        output_names=None, label_names=None)
-    self.losses = []
-
-  def update(self, labels, preds):
-    #print(len(labels), len(preds))
-    #print(preds[1].asnumpy())
-    loss = preds[2].asnumpy()[0]
-    if len(self.losses)==100:
-      print('triplet loss', sum(self.losses)/len(self.losses))
-      self.losses = []
-    self.losses.append(loss)
-    preds = [preds[1]] #use softmax output
-    for label, pred_label in zip(labels, preds):
-        #print(label.shape, pred_label.shape)
-        if pred_label.shape != label.shape:
-            pred_label = mx.ndarray.argmax(pred_label, axis=self.axis)
-        pred_label = pred_label.asnumpy().astype('int32').flatten()
-        label = label.asnumpy().astype('int32').flatten()
-        #print(label)
-        #print(label, pred_label)
-        assert label.shape==pred_label.shape
-        self.sum_metric += (pred_label.flat == label.flat).sum()
-        self.num_inst += len(pred_label.flat)
-
-def parse_args():
-  parser = argparse.ArgumentParser(description='Train face network')
-  # general
-  parser.add_argument('--prefix', default='../model/face',
-      help='directory to save model.')
-  parser.add_argument('--load-epoch', type=int, default=0,
-      help='load epoch.')
-  parser.add_argument('--end-epoch', type=int, default=20,
-      help='training epoch size.')
-  parser.add_argument('--retrain', action='store_true', default=False,
-      help='true means continue training.')
-  args = parser.parse_args()
-  return args
-
-def get_symbol(args, arg_params, aux_params):
-  new_args = dict({k:arg_params[k] for k in arg_params if 'fc' not in k})
-  data_shape = (3,args.image_size,args.image_size)
-  image_shape = ",".join([str(x) for x in data_shape])
-  layers = 152
-  _,_,embeddings,_ = resnet_dcn.get_cls_symbol(128, layers, image_shape, use_deformable=False)
-  all_layers = embeddings.get_internals()
-  #print(all_layers)
-  layer_names = ['_plus10','_plus46', '_plus49']
-  #layer_names = ['plus2', '_plus10','_plus46', '_plus49']
-  layers = []
-  for name in layer_names:
-    layers.append( all_layers[name+"_output"] )
-  out_sym = mx.symbol.Group(layers)
-  out_name = out_sym.list_outputs()
-  #arg_name = embeddings.list_arguments()
-  #aux_name = embeddings.list_auxiliary_states()
-  data_shape_dict = {'data' : (args.batch_size,)+data_shape}
-  arg_shape, out_shape, aux_shape = out_sym.infer_shape(**data_shape_dict)
-  #print(out_shape)
-  out_shape_dict = dict(zip(out_name, out_shape))
-  for k,v in out_shape_dict.iteritems():
-    print(k,v)
-
-  layers = []
-  for i in xrange(len(layer_names)):
-    name = layer_names[i]+"_output"
-    _layer = all_layers[name]
-    _kernel = out_shape_dict[name][2]//5
-    if _kernel>1:
-      layer = mx.sym.Pooling(data=_layer, kernel=(_kernel, _kernel), stride=(_kernel,_kernel), pad=(0,0), pool_type='max')
-    else:
-      layer = _layer
-    layer = mx.symbol.Convolution(data=layer, kernel=(3, 3), pad=(1,1), num_filter=128)
-    layers.append(layer)
-  body = mx.symbol.concat(*layers, dim=1)
-  body = mx.symbol.Convolution(data=body, kernel=(1, 1), pad=(0,0), num_filter=128)
-  body = mx.sym.Pooling(data=body, global_pool=True, kernel=(5, 5), pool_type='avg', name='last_pool')
-  embeddings = mx.sym.Flatten(data=body)
-  _, out_shape, _= embeddings.infer_shape(**data_shape_dict)
-  print(out_shape)
-  #print(arg_shape)
-  #sys.exit(0)
-  l2_embeddings = mx.symbol.L2Normalization(embeddings)
-  batch_size = args.batch_size//args.ctx_num
-  anchor = mx.symbol.slice_axis(l2_embeddings, axis=0, begin=0, end=batch_size//3)
-  positive = mx.symbol.slice_axis(l2_embeddings, axis=0, begin=batch_size//3, end=2*batch_size//3)
-  negative = mx.symbol.slice_axis(l2_embeddings, axis=0, begin=2*batch_size//3, end=batch_size)
-  ap = anchor - positive
-  an = anchor - negative
-  ap = ap*ap
-  an = an*an
-  ap = mx.symbol.sum(ap, axis=1, keepdims=1) #(T,1)
-  an = mx.symbol.sum(an, axis=1, keepdims=1) #(T,1)
-  loss_scale = [1.0, 0.0, 0.0]
-  #triplet_loss = mx.symbol.broadcast_maximum(0.0, ap-an+args.margin) #(T,1)
-  triplet_loss = mx.symbol.Activation(data = (ap-an+args.margin), act_type='relu')
-  triplet_loss = mx.symbol.sum(triplet_loss)/(batch_size//3)
-  triplet_loss = mx.symbol.MakeLoss(data = triplet_loss, grad_scale = loss_scale[0])
-  data = mx.symbol.Variable('data')
-  gt_label = mx.symbol.Variable('softmax_label')
-  fc = mx.symbol.FullyConnected(data = embeddings, num_hidden = args.num_classes, name="fc2")
-  softmax = mx.symbol.SoftmaxOutput(data=fc, label = gt_label, name='softmax', grad_scale = loss_scale[1])
-  if loss_scale[2]>0.0:
-    _center_loss = mx.symbol.Custom(data = l2_embeddings, label = gt_label, name='center_loss', op_type='centerloss'
-        , num_class= args.num_classes, alpha = 0.5, scale=loss_scale[2], batchsize=batch_size)
-    out = mx.symbol.Group([mx.symbol.BlockGrad(l2_embeddings), softmax, triplet_loss, _center_loss])
-  else:
-    out = mx.symbol.Group([mx.symbol.BlockGrad(l2_embeddings), softmax, triplet_loss])
-  #out = triplet_loss
-  #out = softmax
-  return (out, new_args, aux_params)
-
-def train_net(args):
-    ctx = []
-    cvd = os.environ['CUDA_VISIBLE_DEVICES'].strip()
-    if len(cvd)>0:
-      for i in xrange(len(cvd.split(','))):
-        ctx.append(mx.gpu(i))
-    if len(ctx)==0:
-      ctx = [mx.cpu()]
-      print('use cpu')
-    else:
-      print('gpu num:', len(ctx))
-    prefix = args.prefix
-    end_epoch = args.end_epoch
-    pretrained = '../model/resnet-152'
-    load_epoch = args.load_epoch
-    args.image_size = 160
-    per_batch_size = 60
-    args.ctx_num = len(ctx)
-    args.batch_size = per_batch_size*args.ctx_num
-    #args.all_batch_size = args.batch_size*args.ctx_num
-    args.bag_size = 3600
-    args.margin = 0.2
-    args.num_classes = 10575 #webface
-
-
-    data_shape = (3,args.image_size,args.image_size)
-
-    begin_epoch = 0
-    base_lr = 0.05
-    base_wd = 0.0002
-    base_mom = 0.0
-    lr_decay = 0.98
-    if not args.retrain:
-      #load and initialize params
-      print(pretrained)
-      _, arg_params, aux_params = mx.model.load_checkpoint(pretrained, load_epoch)
-      sym, arg_params, aux_params = get_symbol(args, arg_params, aux_params)
-      #arg_params, aux_params = load_param(pretrained, epoch, convert=True)
-      data_shape_dict = {'data': (args.batch_size, 3, args.image_size, args.image_size), 'softmax_label': (args.batch_size,)}
-      resnet_dcn.init_weights(sym, data_shape_dict, arg_params, aux_params)
-    else:
-      pretrained = args.prefix
-      sym, arg_params, aux_params = mx.model.load_checkpoint(pretrained, load_epoch)
-      begin_epoch = load_epoch
-      end_epoch = begin_epoch+10
-      base_wd = 0.00005
-      lr_decay = 0.5
-      base_lr = 0.015
-    # infer max shape
-
-    model = mx.mod.Module(
-        context       = ctx,
-        symbol        = sym,
-        #label_names   = [],
-        #fixed_param_prefix = fixed_param_prefix,
-    )
-
-    train_dataiter = FaceIter(
-        path_imglist         = "/raid5data/dplearn/faceinsight_align_webface.lst",
-        data_shape           = data_shape,
-        mod                  = model,
-        ctx_num              = args.ctx_num,
-        batch_size           = args.batch_size,
-        bag_size             = args.bag_size,
-        images_per_person    = 5,
-    )
-
-    #_dice = DiceMetric()
-    _acc = AccMetric()
-    eval_metrics = [mx.metric.create(_acc)]
-
-    # rpn_eval_metric, rpn_cls_metric, rpn_bbox_metric, eval_metric, cls_metric, bbox_metric
-    #for child_metric in [fcn_loss_metric]:
-    #    eval_metrics.add(child_metric)
-
-    # callback
-    #batch_end_callback = callback.Speedometer(input_batch_size, frequent=args.frequent)
-    #epoch_end_callback = mx.callback.module_checkpoint(mod, prefix, period=1, save_optimizer_states=True)
-
-    # decide learning rate
-    #lr_step = '10,20,30'
-    #train_size = 4848
-    #nrof_batch_in_epoch = int(train_size/input_batch_size)
-    #print('nrof_batch_in_epoch:', nrof_batch_in_epoch)
-    #lr_factor = 0.1
-    #lr_epoch = [float(epoch) for epoch in lr_step.split(',')]
-    #lr_epoch_diff = [epoch - begin_epoch for epoch in lr_epoch if epoch > begin_epoch]
-    #lr = base_lr * (lr_factor ** (len(lr_epoch) - len(lr_epoch_diff)))
-    #lr_iters = [int(epoch * train_size / batch_size) for epoch in lr_epoch_diff]
-    #print 'lr', lr, 'lr_epoch_diff', lr_epoch_diff, 'lr_iters', lr_iters
-
-    #lr_scheduler = MultiFactorScheduler(lr_iters, lr_factor)
-
-    # optimizer
-    #optimizer_params = {'momentum': 0.9,
-    #                    'wd': 0.0005,
-    #                    'learning_rate': base_lr,
-    #                    'rescale_grad': 1.0,
-    #                    'clip_gradient': None}
-    initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="in", magnitude=2)
-    #opt = optimizer.SGD(learning_rate=base_lr, momentum=0.9, wd=base_wd, rescale_grad=(1.0/args.batch_size))
-    opt = optimizer.SGD(learning_rate=base_lr, momentum=base_mom, wd=base_wd, rescale_grad=1.0)
-    #opt = optimizer.AdaGrad(learning_rate=base_lr, wd=base_wd, rescale_grad=1.0)
-    _cb = mx.callback.Speedometer(args.batch_size, 10)
-
-    lfw_dir = '/raid5data/dplearn/lfw_mtcnn'
-    lfw_pairs = lfw.read_pairs(os.path.join(lfw_dir, 'pairs.txt'))
-    lfw_paths, issame_list = lfw.get_paths(lfw_dir, lfw_pairs, 'png')
-    imgs = []
-    lfw_data_list = []
-    for flip in [0,1]:
-      lfw_data = nd.empty((len(lfw_paths), 3, args.image_size, args.image_size))
-      i = 0
-      for path in lfw_paths:
-        with open(path, 'rb') as fin:
-          _bin = fin.read()
-          img = mx.image.imdecode(_bin)
-          img = nd.transpose(img, axes=(2, 0, 1))
-          if flip==1:
-            img = img.asnumpy()
-            for c in xrange(img.shape[0]):
-              img[c,:,:] = np.fliplr(img[c,:,:])
-            img = nd.array( img )
-          #print(img.shape)
-          lfw_data[i][:] = img
-          i+=1
-          if i%1000==0:
-            print('loading lfw', i)
-      print(lfw_data.shape)
-      lfw_data_list.append(lfw_data)
-
-    def lfw_test(nbatch):
-      print('testing lfw..')
-      embeddings_list = []
-      for i in xrange( len(lfw_data_list) ):
-        lfw_data = lfw_data_list[i]
-        embeddings = None
-        ba = 0
-        while ba<lfw_data.shape[0]:
-          bb = min(ba+args.batch_size, lfw_data.shape[0])
-          _data = nd.slice_axis(lfw_data, axis=0, begin=ba, end=bb)
-          _label = nd.ones( (bb-ba,) )
-          db = mx.io.DataBatch(data=(_data,), label=(_label,))
-          model.forward(db, is_train=False)
-          net_out = model.get_outputs()
-          _embeddings = net_out[0].asnumpy()
-          if embeddings is None:
-            embeddings = np.zeros( (lfw_data.shape[0], _embeddings.shape[1]) )
-          embeddings[ba:bb,:] = _embeddings
-          ba = bb
-        embeddings_list.append(embeddings)
-
-      acc_list = []
-      embeddings = embeddings_list[0]
-      _, _, accuracy, val, val_std, far = lfw.evaluate(embeddings, issame_list, nrof_folds=10)
-      acc_list.append(np.mean(accuracy))
-      print('[%d]Accuracy: %1.3f+-%1.3f' % (nbatch, np.mean(accuracy), np.std(accuracy)))
-      print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-      embeddings = np.concatenate(embeddings_list, axis=1)
-      embeddings = sklearn.preprocessing.normalize(embeddings)
-      print(embeddings.shape)
-      _, _, accuracy, val, val_std, far = lfw.evaluate(embeddings, issame_list, nrof_folds=10)
-      acc_list.append(np.mean(accuracy))
-      print('[%d]Accuracy-Flip: %1.3f+-%1.3f' % (nbatch, np.mean(accuracy), np.std(accuracy)))
-      print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-      pca = PCA(n_components=128)
-      embeddings = pca.fit_transform(embeddings)
-      embeddings = sklearn.preprocessing.normalize(embeddings)
-      print(embeddings.shape)
-      _, _, accuracy, val, val_std, far = lfw.evaluate(embeddings, issame_list, nrof_folds=10)
-      acc_list.append(np.mean(accuracy))
-      print('[%d]Accuracy-PCA: %1.3f+-%1.3f' % (nbatch, np.mean(accuracy), np.std(accuracy)))
-      print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))
-      return max(*acc_list)
-
-
-    #global_step = 0
-    highest_acc = [0.0]
-    last_save_acc = [0.0]
-    def _batch_callback(param):
-      #global global_step
-      mbatch = param.nbatch+1
-      if mbatch % 4000 == 0:
-        opt.lr *= lr_decay
-
-      #print(param.nbatch, opt.lr)
-      _cb(param)
-      if param.nbatch%100==0:
-        print('lr-batch-epoch:',opt.lr,param.nbatch,param.epoch)
-      if param.nbatch>=0 and param.nbatch%400==0:
-        acc = lfw_test(param.nbatch)
-        if acc>highest_acc[0]:
-          highest_acc[0] = acc
-        if acc>0.9 and acc-last_save_acc[0]>=0.01:
-          print('saving', mbatch, acc, last_save_acc[0])
-          _arg, _aux = model.get_params()
-          mx.model.save_checkpoint(args.prefix, mbatch, model.symbol, _arg, _aux)
-          last_save_acc[0] = acc
-        print('[%d]highest Accu: %1.3f'%(param.nbatch, highest_acc[0]))
-
-
-
-
-      sys.stdout.flush()
-      sys.stderr.flush()
-
-    epoch_cb = mx.callback.do_checkpoint(prefix, 1)
-    #epoch_cb = None
-
-
-
-    def _epoch_callback(epoch, sym, arg_params, aux_params):
-      print('epoch-end', epoch)
-
-    model.fit(train_dataiter,
-        begin_epoch        = begin_epoch,
-        num_epoch          = end_epoch,
-        #eval_data          = val_dataiter,
-        eval_metric        = eval_metrics,
-        kvstore            = 'device',
-        optimizer          = opt,
-        #optimizer_params   = optimizer_params,
-        initializer        = initializer,
-        arg_params         = arg_params,
-        aux_params         = aux_params,
-        allow_missing      = True,
-        batch_end_callback = _batch_callback,
-        epoch_end_callback = epoch_cb )
-
-def main():
-    #time.sleep(3600*6.5)
-    args = parse_args()
-    print('Called with argument:', args)
-    train_net(args)
-
-if __name__ == '__main__':
-    main()
--- a/src/val_lfw.py
+++ b/src/val_lfw.py
@@ -1,92 +0,0 @@
-from __future__ import division
-
-import numpy as np
-import os
-import sys
-sys.path.append(os.path.join(os.path.dirname(__file__), 'eval'))
-import lfw
-import sklearn
-from sklearn.decomposition import PCA
-
-lfw_dir = '/raid5data/dplearn/lfw_mtcnn2'
-lfw_pairs = lfw.read_pairs(os.path.join(lfw_dir, 'pairs.txt'))
-lfw_paths, issame_list = lfw.get_paths(lfw_dir, lfw_pairs, 'jpg')
-
-model_dir = '../model'
-
-models = ['sphereface_p0_20-lfw-0006.npy']
-models = ['sphereface-64-p0_0_96_95_0-lfw-0001.npy']
-models = ['sphereface_p0_20-lfw-0006.npy', 'sphereface-64-p0_0_96_95_0-lfw-0001.npy']
-models = ['sphereface-20-p0_0_96_112_0-lfw-0022.npy','sphereface-20-p0_0_96_95_0-lfw-0021.npy', 'sphereface-20-p0_40_96_112_0-lfw-0022.npy']
-models = ['sphereface-20-p0_0_96_112_0-lfw-0022.npy','sphereface-20-p0_0_96_95_0-lfw-0021.npy', 'sphereface-36-p0_0_96_112_0-lfw-0022.npy']
-models = ['sphereface-20-p0_0_96_112_0-lfw-0022.npy','sphereface-20-p0_0_96_95_0-lfw-0021.npy']
-models = ['sphereface-20-p0_0_96_112_0-lfw-0022.npy','sphereface-20-p0_0_96_95_0-lfw-0021.npy', 'sphereface-36-p0_0_96_112_0-lfw-0022.npy']
-models = ['sphereface-20-p0_0_96_112_0-lfw-0022.npy','sphereface-20-p0_0_96_95_0-lfw-0021.npy', 'sphereface-36-p0_0_96_95_0-lfw-0021.npy']
-models = [
-          #'sphereface-20-p0_0_96_112_0-lfw-0022.npy',
-          #'sphereface-20-p0_0_96_95_0-lfw-0021.npy', 
-          #'sphereface-20-p0_0_80_95_0-lfw-0021.npy', 
-          #'sphereface-36-p0_0_96_95_0-lfw-0021.npy',
-          'sphereface-s60-p0_0_96_112_0-lfw-0031.npy',
-          'sphereface-s60-p0_0_96_95_0-lfw-0021.npy',
-          'sphereface2-s60-p0_0_96_112_0-lfw-0021.npy',
-          'sphereface3-s60-p0_0_96_95_0-lfw-0023.npy',
-          #'sphereface-s60-p0_0_80_95_0-lfw-0025.npy',
-          #'sphereface-s60-p16_0_96_112_0-lfw-0023.npy',
-          #'spherefacec-s60-p0_0_96_112_0-lfw-0021.npy',
-          ]
-models = [
-          '../model31/sphere-m51-p0_0_96_112_0-lfw-0083.npy',
-          '../model/softmax-m53-p0_0_96_112_0-lfw-0026.npy',
-          #'../model32/sphere-m30-p0_0_96_112_0-lfw-0092.npy',
-          ]
-#models = models[0:3]
-concat = True
-pca = False
-weights = None
-#weights = [0.5, 1.0, 0.5]
-
-F = None
-ii = 0
-for m in models:
-  model = m
-  #model = os.path.join(model_dir, m)
-  X = np.load(model)
-  X1 = X[0:(X.shape[0]//2),:]
-  X2 = X[(X.shape[0]//2):,:]
-  print(X.shape, X1.shape, X2.shape)
-  #X1 = sklearn.preprocessing.normalize(X1)
-  #X2 = sklearn.preprocessing.normalize(X2)
-  XX = X1+X2
-  XX = sklearn.preprocessing.normalize(XX)
-  if weights is not None:
-    weight = weights[ii]
-    XX *= weight
-  if F is None:
-    F = XX
-  else:
-    if concat:
-      F = np.concatenate((F,XX), axis=1)
-    else:
-      F += XX
-  ii+=1
-  #if concat:
-  #  F = np.concatenate((F,X2), axis=1)
-  #else:
-  #  F += X2
-
-print(F.shape)
-npca = 0
-if concat and pca:
-  #F = sklearn.preprocessing.normalize(F)
-  npca = 180
-  #pca = PCA(n_components=512)
-  #F = pca.fit_transform(F)
-  for npca in xrange(512,513,1):
-    _, _, accuracy, val, val_std, far = lfw.evaluate(F, issame_list, nrof_folds=10, pca=npca)
-    print('[%d]Accuracy: %1.5f+-%1.5f' % (npca, np.mean(accuracy), np.std(accuracy)))
-else:
-  F = sklearn.preprocessing.normalize(F)
-  _, _, accuracy, val, val_std, far = lfw.evaluate(F, issame_list, nrof_folds=10, pca=npca)
-  print('[%d]Accuracy: %1.5f+-%1.5f' % (0, np.mean(accuracy), np.std(accuracy)))
-  print('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val, val_std, far))