-
Notifications
You must be signed in to change notification settings - Fork 3.6k
/
classifier.py
executable file
·293 lines (252 loc) · 10.5 KB
/
classifier.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
#!/usr/bin/env python2
#
# Example to classify faces.
# Brandon Amos
# 2015/10/11
#
# Copyright 2015-2016 Carnegie Mellon University
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import time
start = time.time()
import argparse
import cv2
import os
import pickle
import sys
from operator import itemgetter
import numpy as np
np.set_printoptions(precision=2)
import pandas as pd
import openface
from sklearn.pipeline import Pipeline
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA
from sklearn.preprocessing import LabelEncoder
from sklearn.svm import SVC
from sklearn.grid_search import GridSearchCV
from sklearn.mixture import GMM
from sklearn.tree import DecisionTreeClassifier
from sklearn.naive_bayes import GaussianNB
fileDir = os.path.dirname(os.path.realpath(__file__))
modelDir = os.path.join(fileDir, '..', 'models')
dlibModelDir = os.path.join(modelDir, 'dlib')
openfaceModelDir = os.path.join(modelDir, 'openface')
def getRep(imgPath, multiple=False):
start = time.time()
bgrImg = cv2.imread(imgPath)
if bgrImg is None:
raise Exception("Unable to load image: {}".format(imgPath))
rgbImg = cv2.cvtColor(bgrImg, cv2.COLOR_BGR2RGB)
if args.verbose:
print(" + Original size: {}".format(rgbImg.shape))
if args.verbose:
print("Loading the image took {} seconds.".format(time.time() - start))
start = time.time()
if multiple:
bbs = align.getAllFaceBoundingBoxes(rgbImg)
else:
bb1 = align.getLargestFaceBoundingBox(rgbImg)
bbs = [bb1]
if len(bbs) == 0 or (not multiple and bb1 is None):
raise Exception("Unable to find a face: {}".format(imgPath))
if args.verbose:
print("Face detection took {} seconds.".format(time.time() - start))
reps = []
for bb in bbs:
start = time.time()
alignedFace = align.align(
args.imgDim,
rgbImg,
bb,
landmarkIndices=openface.AlignDlib.OUTER_EYES_AND_NOSE)
if alignedFace is None:
raise Exception("Unable to align image: {}".format(imgPath))
if args.verbose:
print("Alignment took {} seconds.".format(time.time() - start))
print("This bbox is centered at {}, {}".format(bb.center().x, bb.center().y))
start = time.time()
rep = net.forward(alignedFace)
if args.verbose:
print("Neural network forward pass took {} seconds.".format(
time.time() - start))
reps.append((bb.center().x, rep))
sreps = sorted(reps, key=lambda x: x[0])
return sreps
def train(args):
print("Loading embeddings.")
fname = "{}/labels.csv".format(args.workDir)
labels = pd.read_csv(fname, header=None).as_matrix()[:, 1]
labels = map(itemgetter(1),
map(os.path.split,
map(os.path.dirname, labels))) # Get the directory.
fname = "{}/reps.csv".format(args.workDir)
embeddings = pd.read_csv(fname, header=None).as_matrix()
le = LabelEncoder().fit(labels)
labelsNum = le.transform(labels)
nClasses = len(le.classes_)
print("Training for {} classes.".format(nClasses))
if args.classifier == 'LinearSvm':
clf = SVC(C=1, kernel='linear', probability=True)
elif args.classifier == 'GridSearchSvm':
print("""
Warning: In our experiences, using a grid search over SVM hyper-parameters only
gives marginally better performance than a linear SVM with C=1 and
is not worth the extra computations of performing a grid search.
""")
param_grid = [
{'C': [1, 10, 100, 1000],
'kernel': ['linear']},
{'C': [1, 10, 100, 1000],
'gamma': [0.001, 0.0001],
'kernel': ['rbf']}
]
clf = GridSearchCV(SVC(C=1, probability=True), param_grid, cv=5)
elif args.classifier == 'GMM': # Doesn't work best
clf = GMM(n_components=nClasses)
# ref:
# http://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html#example-classification-plot-classifier-comparison-py
elif args.classifier == 'RadialSvm': # Radial Basis Function kernel
# works better with C = 1 and gamma = 2
clf = SVC(C=1, kernel='rbf', probability=True, gamma=2)
elif args.classifier == 'DecisionTree': # Doesn't work best
clf = DecisionTreeClassifier(max_depth=20)
elif args.classifier == 'GaussianNB':
clf = GaussianNB()
# ref: https://jessesw.com/Deep-Learning/
elif args.classifier == 'DBN':
from nolearn.dbn import DBN
clf = DBN([embeddings.shape[1], 500, labelsNum[-1:][0] + 1], # i/p nodes, hidden nodes, o/p nodes
learn_rates=0.3,
# Smaller steps mean a possibly more accurate result, but the
# training will take longer
learn_rate_decays=0.9,
# a factor the initial learning rate will be multiplied by
# after each iteration of the training
epochs=300, # no of iternation
# dropouts = 0.25, # Express the percentage of nodes that
# will be randomly dropped as a decimal.
verbose=1)
if args.ldaDim > 0:
clf_final = clf
clf = Pipeline([('lda', LDA(n_components=args.ldaDim)),
('clf', clf_final)])
clf.fit(embeddings, labelsNum)
fName = "{}/classifier.pkl".format(args.workDir)
print("Saving classifier to '{}'".format(fName))
with open(fName, 'w') as f:
pickle.dump((le, clf), f)
def infer(args, multiple=False):
with open(args.classifierModel, 'rb') as f:
if sys.version_info[0] < 3:
(le, clf) = pickle.load(f)
else:
(le, clf) = pickle.load(f, encoding='latin1')
for img in args.imgs:
print("\n=== {} ===".format(img))
reps = getRep(img, multiple)
if len(reps) > 1:
print("List of faces in image from left to right")
for r in reps:
rep = r[1].reshape(1, -1)
bbx = r[0]
start = time.time()
predictions = clf.predict_proba(rep).ravel()
maxI = np.argmax(predictions)
person = le.inverse_transform(maxI)
confidence = predictions[maxI]
if args.verbose:
print("Prediction took {} seconds.".format(time.time() - start))
if multiple:
print("Predict {} @ x={} with {:.2f} confidence.".format(person.decode('utf-8'), bbx,
confidence))
else:
print("Predict {} with {:.2f} confidence.".format(person.decode('utf-8'), confidence))
if isinstance(clf, GMM):
dist = np.linalg.norm(rep - clf.means_[maxI])
print(" + Distance from the mean: {}".format(dist))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument(
'--dlibFacePredictor',
type=str,
help="Path to dlib's face predictor.",
default=os.path.join(
dlibModelDir,
"shape_predictor_68_face_landmarks.dat"))
parser.add_argument(
'--networkModel',
type=str,
help="Path to Torch network model.",
default=os.path.join(
openfaceModelDir,
'nn4.small2.v1.t7'))
parser.add_argument('--imgDim', type=int,
help="Default image dimension.", default=96)
parser.add_argument('--cuda', action='store_true')
parser.add_argument('--verbose', action='store_true')
subparsers = parser.add_subparsers(dest='mode', help="Mode")
trainParser = subparsers.add_parser('train',
help="Train a new classifier.")
trainParser.add_argument('--ldaDim', type=int, default=-1)
trainParser.add_argument(
'--classifier',
type=str,
choices=[
'LinearSvm',
'GridSearchSvm',
'GMM',
'RadialSvm',
'DecisionTree',
'GaussianNB',
'DBN'],
help='The type of classifier to use.',
default='LinearSvm')
trainParser.add_argument(
'workDir',
type=str,
help="The input work directory containing 'reps.csv' and 'labels.csv'. Obtained from aligning a directory with 'align-dlib' and getting the representations with 'batch-represent'.")
inferParser = subparsers.add_parser(
'infer', help='Predict who an image contains from a trained classifier.')
inferParser.add_argument(
'classifierModel',
type=str,
help='The Python pickle representing the classifier. This is NOT the Torch network model, which can be set with --networkModel.')
inferParser.add_argument('imgs', type=str, nargs='+',
help="Input image.")
inferParser.add_argument('--multi', help="Infer multiple faces in image",
action="store_true")
args = parser.parse_args()
if args.verbose:
print("Argument parsing and import libraries took {} seconds.".format(
time.time() - start))
if args.mode == 'infer' and args.classifierModel.endswith(".t7"):
raise Exception("""
Torch network model passed as the classification model,
which should be a Python pickle (.pkl)
See the documentation for the distinction between the Torch
network and classification models:
http://cmusatyalab.github.io/openface/demo-3-classifier/
http://cmusatyalab.github.io/openface/training-new-models/
Use `--networkModel` to set a non-standard Torch network model.""")
start = time.time()
align = openface.AlignDlib(args.dlibFacePredictor)
net = openface.TorchNeuralNet(args.networkModel, imgDim=args.imgDim,
cuda=args.cuda)
if args.verbose:
print("Loading the dlib and OpenFace models took {} seconds.".format(
time.time() - start))
start = time.time()
if args.mode == 'train':
train(args)
elif args.mode == 'infer':
infer(args, args.multi)