/
main.py
280 lines (250 loc) · 14.9 KB
/
main.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
import os, random, warnings, argparse
warnings.filterwarnings('ignore')
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import torch
import numpy as np
seed_value = 1000
os.environ['PYTHONHASHSEED'] = str(seed_value)
torch.manual_seed(seed_value)
torch.cuda.manual_seed(seed_value)
torch.cuda.manual_seed_all(seed_value)
np.random.seed(seed_value)
random.seed(seed_value)
torch.manual_seed(seed_value)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads(1)
import sys, pickle, time, gc
from poutyne.framework import Model
import multiprocessing
from functools import partial
from multiprocessing.pool import Pool
import util
from sklearn.utils import shuffle
from sklearn import preprocessing
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import confusion_matrix
def first_deriv(x, wavelengths):
# First derivative of measurements with respect to wavelength
x = np.copy(x)
for i, xx in enumerate(x):
dx = np.zeros(xx.shape, np.float)
dx[0:-1] = np.diff(xx)/np.diff(wavelengths)
dx[-1] = (xx[-1] - xx[-2])/(wavelengths[-1] - wavelengths[-2])
x[i] = dx
return x
def prepare_data(x_spectral_train, x_image_train, y_train, x_spectral_test, x_image_test, y_test, wavelengths, deriv=True, data_type='spectral_image', image_preprocess='resnet', objs=None):
if deriv:
# Finite difference (Numerical differentiation)
x_spectral_train = np.concatenate([x_spectral_train, first_deriv(x_spectral_train, wavelengths)], axis=-1)
x_spectral_test = np.concatenate([x_spectral_test, first_deriv(x_spectral_test, wavelengths)], axis=-1)
if data_type == 'spectral':
# Zero mean unit variance
scaler_spectral = preprocessing.StandardScaler()
x_spectral_train = scaler_spectral.fit_transform(x_spectral_train)
x_spectral_test = scaler_spectral.transform(x_spectral_test)
elif data_type == 'image':
if 'none' in image_preprocess:
scaler_image = preprocessing.StandardScaler()
x_image_train = np.reshape(scaler_image.fit_transform(np.reshape(x_image_train, (len(x_image_train), -1))), np.shape(x_image_train))
x_image_test = np.reshape(scaler_image.transform(np.reshape(x_image_test, (len(x_image_test), -1))), np.shape(x_image_test))
else:
scaler_image = preprocessing.StandardScaler()
x_image_train = scaler_image.fit_transform(x_image_train)
x_image_test = scaler_image.transform(x_image_test)
elif data_type == 'spectral_image':
scaler_image = preprocessing.StandardScaler()
x = scaler_image.fit_transform(np.concatenate([x_spectral_train, x_image_train], axis=-1))
x_spectral_train = x[:, :np.shape(x_spectral_train)[-1]]
x_image_train = x[:, np.shape(x_spectral_train)[-1]:]
x = scaler_image.transform(np.concatenate([x_spectral_test, x_image_test], axis=-1))
x_spectral_test = x[:, :np.shape(x_spectral_test)[-1]]
x_image_test = x[:, np.shape(x_spectral_test)[-1]:]
return x_spectral_train, x_image_train, y_train, x_spectral_test, x_image_test, y_test
def nn(input_size, layers, dropout, material_count=None, batchnorm=True):
modules = []
for i in range(len(layers)-1):
modules.append(torch.nn.Linear(input_size if i == 0 else layers[i-1], layers[i]))
if batchnorm:
modules.append(torch.nn.BatchNorm1d(layers[i]))
modules.append(torch.nn.LeakyReLU())
if dropout > 0:
modules.append(torch.nn.Dropout(dropout))
modules.append(torch.nn.Linear(layers[-2], layers[-1]) if len(layers) > 1 else torch.nn.Linear(input_size, layers[-1]))
if batchnorm:
modules.append(torch.nn.BatchNorm1d(layers[-1]))
modules.append(torch.nn.LeakyReLU())
if material_count is not None:
modules.append(torch.nn.Linear(layers[-1], material_count))
return torch.nn.Sequential(*modules)
def learn(o, X_spectral, X_image, y, objs, wavelengths, data_type='spectral_image', image_preprocess='resnet', epochs=50, batch_size=128, material_count=8, layers=[64]*2, dropout=0.0, lr=0.0005, seed=1000, test='looo', verbose=False):
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
torch.set_num_threads(1)
if 'test' in test:
train_idx, test_idx = o
X_spectral_train = X_spectral[train_idx]
X_image_train = X_image[train_idx]
y_train = y[train_idx]
X_spectral_test = X_spectral[test_idx]
X_image_test = X_image[test_idx]
y_test = y[test_idx]
objs_train = objs[train_idx]
elif 'looo' in test:
_, obj = o
# Set up leave-one-object-out training and test sets
X_spectral_train = X_spectral[objs != obj]
X_image_train = X_image[objs != obj]
y_train = y[objs != obj]
X_spectral_test = X_spectral[objs == obj]
X_image_test = X_image[objs == obj]
y_test = y[objs == obj]
objs_train = objs[objs != obj]
X_spectral_train, X_image_train, y_train, X_spectral_test, X_image_test, y_test = prepare_data(X_spectral_train, X_image_train, y_train, X_spectral_test, X_image_test, y_test, wavelengths, deriv=True, data_type=data_type, image_preprocess=image_preprocess, objs=objs)
if data_type == 'spectral':
X_spectral_train, y_train = shuffle(X_spectral_train, y_train)
elif data_type == 'image':
X_image_train, y_train = shuffle(X_image_train, y_train)
elif data_type == 'spectral_image':
X_spectral_train, X_image_train, y_train = shuffle(X_spectral_train, X_image_train, y_train)
y_train = torch.tensor(y_train, dtype=torch.long)
y_test = torch.tensor(y_test, dtype=torch.long)
# Image layers
if data_type == 'spectral_image':
spectral_layers = [64, 64, 32, 32]
spectral_dropout = 0.25
spectral_epochs = 50
image_layers = [128, 64, 32]
image_dropout = 0.1
image_epochs = 50
class ConcatPretrainedNetwork(torch.nn.Module):
def __init__(self):
# global spectral_accuracy, image_accuracy
super().__init__()
spectral_net = nn(np.shape(X_spectral_train)[-1], spectral_layers, spectral_dropout, material_count)
opt = torch.optim.Adam(spectral_net.parameters(), lr=lr)
spectral_model = Model(spectral_net, opt, 'cross_entropy', batch_metrics=['accuracy'])
spectral_model.fit(X_spectral_train, y_train, epochs=spectral_epochs, batch_size=batch_size, verbose=False)
image_net = nn(np.shape(X_image_train)[-1], image_layers, image_dropout, material_count)
opt = torch.optim.Adam(image_net.parameters(), lr=lr)
image_model = Model(image_net, opt, 'cross_entropy', batch_metrics=['accuracy'])
image_model.fit(X_image_train, y_train, epochs=image_epochs, batch_size=batch_size, verbose=False)
# Disable dropout, remove last layer and freeze network
self.trained_spectral_model = torch.nn.Sequential(*(list(spectral_net.children())[:-1]))
for p in self.trained_spectral_model.parameters():
p.requires_grad = False
self.trained_image_model = torch.nn.Sequential(*(list(image_net.children())[:-1]))
for p in self.trained_image_model.parameters():
p.requires_grad = False
self.concat_net = nn(spectral_layers[-1] + image_layers[-1], layers, dropout, material_count, batchnorm=False)
def forward(self, x_spectral, x_image):
y1 = self.trained_spectral_model(x_spectral)
y2 = self.trained_image_model(x_image)
concat = torch.cat((y1, y2), -1)
return self.concat_net(concat)
net = ConcatPretrainedNetwork()
X_train = [X_spectral_train, X_image_train]
X_test = [X_spectral_test, X_image_test]
elif data_type == 'image':
net = nn(np.shape(X_image_train)[-1], layers, dropout, material_count)
X_train = X_image_train
X_test = X_image_test
elif data_type == 'spectral':
net = nn(np.shape(X_spectral_train)[-1], layers, dropout, material_count)
X_train = X_spectral_train
X_test = X_spectral_test
opt = torch.optim.Adam(net.parameters(), lr=lr)
model = Model(net, opt, 'cross_entropy', batch_metrics=['accuracy'])
model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=epochs, batch_size=batch_size, verbose=False)
cm = confusion_matrix(y_test, model.predict(X_test).argmax(axis=-1), labels=range(material_count))
# Return accuracy and confusion matrices
if 'backprop' in test:
return {'accuracy': model.evaluate(X_test, y_test)[-1], 'cm': cm, 'obj_cm': np.copy(cm[y_test[0]]), 'model': model, 'net': net, 'X_test': X_test, 'y_test': y_test}
else:
if verbose:
print(obj, model.evaluate(X_test, y_test)[-1])
return {'accuracy': model.evaluate(X_test, y_test)[-1], 'cm': cm, 'obj_cm': np.copy(cm[y_test[0]])}
# NOTE: Leave-one-object-out cross-validation.
def looo_cv(X_spectral, X_image, y, objs, wavelengths, data_type, image_preprocess, layers, dropout, epochs, batch_size, lr, seeds, material_count, jobs, test='looo', verbose=False):
objs_set = []
for o in objs:
if o not in objs_set:
objs_set.append(o)
accuracies = []
confusion_mat = None
object_confusion_matrix = []
for s in seeds:
pool = Pool(processes=jobs)
results = pool.imap(partial(learn, X_spectral=X_spectral, X_image=X_image, y=y, objs=objs, wavelengths=wavelengths, data_type=data_type, image_preprocess=image_preprocess, epochs=epochs, batch_size=batch_size, material_count=material_count, layers=layers, dropout=dropout, lr=lr, seed=s, test=test, verbose=verbose), list(enumerate(objs_set)))
pool.close()
pool.join()
seed_accuracy = []
for result in list(results):
accuracies.append(result['accuracy'])
seed_accuracy.append(result['accuracy'])
if confusion_mat is None:
confusion_mat = result['cm']
else:
confusion_mat += result['cm']
object_confusion_matrix.append(result['obj_cm'])
print('Accuracy with seed', s, ':', np.mean(seed_accuracy))
print('Results for:', data_type, '5 materials' if five_mats else '8 materials', image_preprocess, '- Accuracy:', np.mean(accuracies))
# print(confusion_mat.astype('float') / confusion_mat.sum(axis=1)[:, np.newaxis], '\n')
sys.stdout.flush()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Training and Evaluation for SpectroVision')
parser.add_argument('-s', '--seed', default=-1, help='Seed used for random number generators. (default -1) for averaging of 10 seeds', type=int)
parser.add_argument('-v', '--verbose', help='Verbose', action='store_true')
args = parser.parse_args()
verbose = args.verbose
jobs = multiprocessing.cpu_count()
layers = {'spectral': [64, 64, 32, 32], 'image': [128, 64, 32], 'spectral_image': [32]}
dropout = {'spectral': 0.25, 'image': 0.1, 'spectral_image': 0.0}
epochs = {'spectral': 50, 'image': 50, 'spectral_image': 10}
lr = 0.0005
batch_size = 128
seeds = range(8000, 8010) if args.seed == -1 else [args.seed]
parent_directory = os.path.join(os.path.dirname(os.path.realpath(__file__)), 'dataset')
# Table I, Leave-one-object-out
print('Table I: leave-one-object-out')
for five_mats in [True, False]:
for data_type in ['spectral', 'image', 'spectral_image']:
X_spectral, X_image, y, objs, wavelengths = util.load_data(parent_directory, data_type, 'densenet201_240_320', (240, 320), five_mats)
# print(np.shape(X_spectral), np.shape(X_image))
print('Computing results for:', data_type, '5 materials' if five_mats else '8 materials')
looo_cv(X_spectral, X_image, y, objs, wavelengths, data_type, 'densenet201_240_320', layers[data_type], dropout[data_type], epochs[data_type], batch_size, lr, seeds, 5 if five_mats else 8, jobs, 'looo', verbose)
# Table II, Test set
print('Table II: test set')
for five_mats in [True, False]:
for data_type in ['spectral', 'image', 'spectral_image']:
X_spectral, X_image, y, objs, wavelengths = util.load_data(parent_directory, data_type, 'densenet201_240_320', (240, 320), five_mats)
X_spectral_test, X_image_test, y_test, objs_test, wavelengths_test = util.load_data(parent_directory, data_type, 'densenet201_240_320', (240, 320), five_mats, test_set=True)
accuracies = []
confusion_mat = None
for seed in seeds:
results = learn([list(range(len(y))), list(range(len(y), len(y)+len(y_test)))], np.concatenate([X_spectral, X_spectral_test], axis=0), np.concatenate([X_image, X_image_test], axis=0), np.concatenate([y, y_test], axis=0), np.concatenate([objs, objs_test], axis=0), wavelengths, data_type, 'densenet201_240_320', epochs[data_type], batch_size, 5 if five_mats else 8, layers[data_type], dropout[data_type], lr, seed, 'test', verbose)
accuracies.append(results['accuracy'])
if confusion_mat is None:
confusion_mat = results['cm']
else:
confusion_mat += results['cm']
print(data_type, '5 materials' if five_mats else '8 materials', '- Accuracy:', np.mean(accuracies))
# print(confusion_mat.astype('float') / confusion_mat.sum(axis=1)[:, np.newaxis], '\n')
sys.stdout.flush()
# Table IV, Resizing and cropping images
print('Table IV: resizing and cropping images')
for image_shape, crop in [((240, 320), False), ((240, 320), True), ((480, 640), False), ((480, 640), True), ((960, 1280), False)]:
five_mats = False
data_type = 'image'
image_preprocess = 'densenet201_%d_%d%s' % (image_shape[0], image_shape[1], '_crop' if crop else '')
X_spectral, X_image, y, objs, wavelengths = util.load_data(parent_directory, data_type, image_preprocess, (240, 320), five_mats)
looo_cv(X_spectral, X_image, y, objs, wavelengths, data_type, image_preprocess, layers[data_type], dropout[data_type], epochs[data_type], batch_size, lr, seeds, 5 if five_mats else 8, jobs, 'looo', verbose)
# Table V, ImageNet models
print('Table V: ImageNet models')
for model in ['vgg19', 'resnet50', 'resnet101', 'resnet152', 'densenet201', 'resnext101', 'efficientnet-b5']:
five_mats = False
data_type = 'image'
image_preprocess = ('%s_240_320' % model) if 'efficient' not in model else ('%s_456_608' % model)
X_spectral, X_image, y, objs, wavelengths = util.load_data(parent_directory, data_type, image_preprocess, (240, 320), five_mats)
looo_cv(X_spectral, X_image, y, objs, wavelengths, data_type, image_preprocess, layers[data_type], dropout[data_type], epochs[data_type], batch_size, lr, seeds, 5 if five_mats else 8, jobs, 'looo', verbose)