def folds_to_sets(f_csv=dir_db + 'journal_data/Pairs/folds_5splits/',
dir_out=dir_db + "journal_data/Pairs/sets/"):
""" Method used to merge 5 fold splits into 3 sets for RFIW (train, val, and test)"""
f_in = glob.glob(f_csv + '*-folds.csv')
for file in f_in:
# each list of pairs <FOLD, LABEL, PAIR_1, PAIR_2>
f_name = io.file_base(file)
print("\nProcessing {}\n".format(f_name))
df_pairs = pd.read_csv(file)
# merge to form train set
df_train = df_pairs[(df_pairs['fold'] == 1) | (df_pairs['fold'] == 5)]
df_train.to_csv(dir_out + "train/" +
f_name.replace("-folds", "-train") + ".csv")
# merge to form val set
df_val = df_pairs[(df_pairs['fold'] == 2) | (df_pairs['fold'] == 4)]
df_val.to_csv(dir_out + "val/" + f_name.replace("-folds", "-val") +
".csv")
# merge to form test set
df_test = df_pairs[(df_pairs['fold'] == 3)]
df_test.to_csv(dir_out + "test/" + f_name.replace("-folds", "-test") +
".csv")
# print stats
print("{} Training;\t {} Val;\t{} Test".format(
df_train['fold'].count(), df_val['fold'].count(),
df_test['fold'].count()))
def prepare_fids(dir_fid=dir_home() +
"/Dropbox/Families_In_The_Wild/Database/Ann/FW_FIDs/",
dirs_out=dir_home() +
"/Dropbox/Families_In_The_Wild/Database/FIDs/",
do_save=False):
"""
Parses FID CSV files and places in DB. Additionally, checks are made for inconsistency in labels.
:param dir_fid:
:param dirs_out:
:return:
"""
# load fids (CSV files)
fid_files = glob.glob(dir_fid + "F????.csv")
fid_dicts = {io.file_base(f): pd.read_csv(f) for f in fid_files}
# dfs_fids, fids = [(pd.read_csv(f), io.file_base(f)) for f in fid_files]
dfs_fams = []
for fid in fid_dicts:
# for each fid (i.e., iterate keys of dictionary)
# print(fid)
df_rel_mat = fid_dicts[fid]
col_gender = df_rel_mat.Gender
# get MIDs
ids = [i for i, c in enumerate(col_gender) if '-1' not in c]
tmp = list(range(0, len(ids)))
same_size, same_contents = helpers.compare_mid_lists(tmp, ids)
if not (same_size and same_contents):
logger.error(
"MIDs and row indices of relationship table differ in {}.".
format(fid))
mids = list(np.array(ids) + 1)
cols = df_rel_mat.columns[mids]
# df_rel_mat.loc[ids][cols]
rel_mat = np.array(df_rel_mat.loc[ids][cols])
success, messages = helpers.check_rel_matrix(rel_mat, fid=fid)
if not success:
logger.error(
"Relationship matrix failed inspection {}.".format(fid))
[logger.error("\t{}".format(m)) for m in messages]
continue
genders = [g for g in list(col_gender[ids])]
names = list(df_rel_mat.Name[ids])
df_fam = pd.DataFrame({'MID': mids})
df_fam = df_fam.join(df_rel_mat.loc[ids][cols])
df_fam = df_fam.join(pd.DataFrame({'Gender': genders, 'Name': names}))
if do_save:
df_fam.to_csv(dirs_out + "/" + fid + "/mid.csv", index=False)
dfs_fams.append(df_fam)
return dfs_fams
# instantiate CNN face detector
f_model = args.model_path
cnn_face_detector = dlib.cnn_face_detection_model_v1(f_model)
dir_fids = glob.glob(dir_images + "F????/")
dir_fids.sort()
f_pids = glob.glob(dir_images + "F????/P*.jpg")
f_pids.sort()
f_prefix = [f.replace(dir_images, "").replace("/", "_").replace(".jpg", "_") for f in f_pids]
# f_pids =list(np.array(f_pids)[ids])
fids = [myio.file_base(myio.filepath(p)) for p in f_pids]
pids = [myio.file_base(p) for p in f_pids]
# f_prefix =list( np.array(f_prefix)[ids])
npids = len(pids)
print("Processing {} images".format(npids))
# Second argument indicates that we should upsample the image 1 time (i.e., bigger image to detect of more faces).
dets = [cnn_face_detector(io.imread(f), 1) for f in f_pids]
df = pd.DataFrame(columns=['FID', 'PID', 'face_id', 'filename', 'left', 'top', 'right', 'bottom', 'confidence'])
print("Number of faces detected: {}".format(len(dets)))
counter = 0
for faces, prefix in zip(dets, f_prefix):
# build dataframe of face detections and corresponding metadata
for i, d in enumerate(faces):
do_init=True)
dirs_fid, fids = fiwdb.load_fids(args.input)
ifiles = glob.glob(args.input + "*/MID*/*.jpg")
ofiles = [
dout + str(f).replace(args.input, "").replace(".jpg", ".csv")
for f in ifiles
]
# layers = args.layers
for ifile in ifiles:
ofile = dout + str(ifile).replace(args.input, "").replace(
".jpg", ".csv")
if os.path.isfile(ofile):
continue
logger.info("Extracting features: {}\n".format(ofile))
fname = io.file_base(ifile)
image = caffe_tools.load_prepare_image_vgg(ifile)
my_net.net.blobs['data'].data[...] = image
output = my_net.net.forward()
io.mkdir(io.filepath(ofile))
feat = my_net.net.blobs[args.layer].data[0]
np.savetxt(ofile, feat.flatten(), delimiter=',') # X is an array
#
# # net = net.net
# # # dir_data = '/data/FIDs2/'
# # dir_data = '/data/FIW_Extended/FIDs/'
# # dir_out = '/data/FIW_Extended/feats/vgg_face/'
#
# # fam_dirs = glob.glob(dir_data + "F????/")
nimages = len(im_files)
for i in range(nimages):
in_file = im_files[i]
f_file = feat_files[i]
if os.path.exists(f_file):
continue
print(f_file)
# layers = layers
# ofile = dout + str(ifile).replace(input, "").replace(".jpg", ".csv")
# if not io.mkdir(io.parent_dir(f_file)):
# continue
# continue
# if os.path.isfile(in_file):
logger.info("Extracting feature for: {}\n".format(in_file))
fname = io.file_base(in_file)
image = caffe_tools.load_prepare_resnet_centerloss(in_file, (112, 96))
# import pdb
# pdb.set_trace()
# my_net.net.blobs['data'].data[...] = image
output = my_net.net.forward()
# io.mkdir(io.filepath(in_file))
feat = my_net.net.blobs[l].data[0]
logger.info("Writing feature to disk: {}\n".format(f_file))
np.savetxt(f_file, feat.flatten(), delimiter=',') # X is an array
#
sub_dirs = ['father-dau'] #, 'father-son', 'mother-dau', 'mother-son']
dir_root = '/media/jrobby/Seagate Backup Plus Drive1/DATA/Kinship/KinFaceW-II/'
dir_features = dir_root + '/features/fine-tuned/'
dir_results = dir_features + 'results_spca/'
io.mkdir(dir_results)
dir_perms = dir_root + 'perm/'
dir_lists = dir_root + 'meta_data/'
do_pca = True
k = 200
# load experimental settings for 5-fold verification
f_lists = glob.glob(dir_lists + "*.csv")
pair_types = [io.file_base(f) for f in f_lists]
dir_feats = [dir_features + p + "/" for p in sub_dirs]
fold_list = [1, 2, 3, 4, 5]
for ids in fold_list:
folds, labels, pairs1, pairs2 = kinwild.read_pair_list(f_lists[ids])
d_out = dir_results + pair_types[ids] + "/"
io.mkdir(d_out)
for feature in features:
print("processing features from layer", feature)
feats1 = kinwild.load_all_features(dir_feats[ids] + feature + "/",
pairs1)
feats2 = kinwild.load_all_features(dir_feats[ids] + feature + "/",
layers = ['conv5_2', 'conv5_3', 'pool5', 'fc6', 'fc7']
layers = ['res5a']
# lid = 4
# sub_dirs = ['father-dau', 'father-son', 'mother-dau', 'mother-son']
dir_root = io.sys_home() + '/Dropbox/Families_In_The_Wild/Database/journal_data/'
dir_features = '/media/jrobby/Seagate Backup Plus Drive1/FIW_dataset/FIW_Extended/features/vgg_face/resnet/'
dir_results = io.sys_home() + '/Dropbox/Families_In_The_Wild/Database/journal_results/verification/res_net/'
io.mkdir(dir_results)
dir_pairs = dir_root + "Pairs/folds_5splits/"
# load experimental settings for 5-fold verification
f_lists = glob.glob(dir_pairs + "*.csv")
pair_types = [io.file_base(f).replace('-folds', '') for f in f_lists]
# dir_feats = [dir_features + p + "/" for p in pair_types]
import os
for i in range(0, 11):
df_list = pd.read_csv(f_lists[i])
pair_type = io.file_base(f_lists[i]).replace('-folds', '')
labels = np.array(df_list.label)
folds = np.array(df_list.fold)
for layer in layers:
fold_list = list(set(folds))
dir_out = dir_results + pair_type + "/" + layer + "/"
if os.path.isdir(dir_out) is True:
do_continue = True
for fold in fold_list:
if not os.path.isdir(dir_out + str(fold) + "/"):