/
scratch_alex.py
213 lines (190 loc) · 8.26 KB
/
scratch_alex.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
from __future__ import division, print_function, absolute_import
import tflearn
from tflearn.data_utils import shuffle, to_categorical
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.normalization import local_response_normalization
from tflearn.layers.estimator import regression
from scipy.misc import imread, imresize, imshow, imsave
from tflearn.data_augmentation import ImageAugmentation
import numpy as np
from matplotlib import pyplot as plt
import os
import random
from sklearn.metrics import confusion_matrix, accuracy_score
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]="2"
def resize_batch(im_batch,size):
im_batch_resized=[]
for im in im_batch:
im_batch_resized.append(imresize(im,(size,size,3)))
return np.array(im_batch_resized)
# class PlottingCallback(tflearn.callbacks.Callback):
# def __init__(self, model, x,
# layers_to_observe=(),
# kernels=10,
# inputs=1):
# self.model = model
# self.x = x
# self.kernels = kernels
# self.inputs = inputs
# self.train_losses = []
# self.observers=[]
# for l in layers_to_observe:
# print(tflearn.DNN(l))
# # self.observers.append(tflearn.DNN(l))
# self.observers = [tflearn.DNN(l) for l in layers_to_observe]
# def on_batch_end(self, training_state, snapshot=False):
# print("The training loss is: ", training_state.global_loss)
# self.train_losses.append(training_state.global_loss)
# # def on_epoch_end(self, training_state):
# # outputs = [o.predict(self.x) for o in self.observers]
# # print([d.shape for d in outputs])#.shape
# # for i in range(self.inputs):
# # plt.figure(frameon=False)
# # plt.subplots_adjust(wspace=0.1, hspace=0.1)
# # ix = 1
# # for o in outputs:
# # for kernel in range(self.kernels):
# # plt.subplot(len(outputs), self.kernels, ix)
# # plt.imshow(o[i, :, :, kernel])
# # plt.axis('off')
# # ix += 1
# # plt.savefig('outputs-for-image:%i-at-epoch:%i.png'
# # % (i, training_state.epoch))
def load_info(dirpath): # image, labels, filename, features
files = os.listdir(dirpath)
im = []
lbl = []
filenames = []
features = []
data_type = dirpath[12:-6]
feature_array = np.load('../features_' + data_type + '_db.npy')
filename_array = np.load('../filenames_' + data_type + '_db.npy')
filename = [str(filename_array[i][43:]) for i in range(filename_array.shape[0])]
for f in files:
im.append(imread(dirpath + f)) # find where this filename is in the array
# we need to get index of dirpath[9:]+f in filename
idx = filename.index(dirpath[12:] + f)
lbl.append(f[0])
features.append(feature_array[idx])
filenames.append(dirpath + f)
return im, lbl, filenames, np.asarray(features)
def test_train_split(im,lbl,files,features,ratio=0.25):
#ratio is num(test samples)/num(train samples)
total_num=len(files)
num_train=(int)((1-ratio)*total_num)
# num_test=total_num-num_train
shuffle_ids=random.sample(range(total_num),total_num)
train_ids=shuffle_ids[:num_train]
test_ids=shuffle_ids[num_train:]
im1=[im[x] for x in train_ids]
im2=[im[x] for x in test_ids]
lbl1=[lbl[x] for x in train_ids]
lbl2=[lbl[x] for x in test_ids]
files1=[files[x] for x in train_ids]
files2=[files[x] for x in test_ids]
features1=[features[x] for x in train_ids]
features2=[features[x] for x in test_ids]
return im1,im2,lbl1,lbl2,files1,files2,features1,features2
def err_calc(im,lbl,model):
probs=model.predict(im)
preds=np.argmax(probs,axis=1)
targets=np.array([int(x) for x in lbl])
err=1-accuracy_score(targets,preds)
return err
def calculate_ERE(f1,t1,model,k):
ere=0
C=k
N=len(t1)
ere_list=[]
#find probs of all points
for i in range(len(t1)):
ere=0
probs=model.predict(f1[i].reshape(1,50,50,3))[0]
for c in range(C):
if c==np.argmax(probs):
i=0
else:
i=1
ere+=probs[c]*i
ere_list.append(ere)
return ere_list
class MonitorCallback(tflearn.callbacks.Callback):
def __init__(self):
self.train_losses = []
def on_batch_end(self, training_state, snapshot=False):
print("The training loss is: ", training_state.global_loss)
self.train_losses.append(training_state.global_loss)
#loading the data
g_src='../new_data/golden_data/' # golden data path, This is the data we initialize on
s_src='../new_data/seed_data/' #seed path, this is the data, the validation is reported on
d_src='../new_data/mturk_data/' # this is rest of the data, split into test and train
g_im, g_lbl, g_files, g_features = load_info(g_src)
s_im, s_lbl, s_files, s_features = load_info(s_src)
d_im, d_lbl, d_files, d_features = load_info(d_src)
tr_im, test_im, tr_lbl, test_lbl, tr_files, test_files, tr_features, test_features = test_train_split(d_im,d_lbl,d_files,d_features)
#Extract a constant training error measure of 50
shuffle_ids=random.sample(range(len(g_lbl)),50)
g_tr_im=[g_im[x] for x in shuffle_ids]
g_tr_lbl=[g_lbl[x] for x in shuffle_ids]
g_tr_files=[g_files[x] for x in shuffle_ids]
g_tr_features=[g_features[x] for x in shuffle_ids]
# Building 'AlexNet'
#modified for image size
img_aug=ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_blur(sigma_max=3.)
img_aug.add_random_rotation(max_angle=25.)
img_aug.add_random_flip_updown()
input_data = input_data(shape=[None, 227,227, 3], data_augmentation=img_aug)
conv0 = conv_2d(input_data, 96, 11, strides=4, activation='relu')
max0 = max_pool_2d(conv0, 3, strides=2)
local0 = local_response_normalization(max0)
conv1 = conv_2d(local0, 256, 5, activation='relu')
max1 = max_pool_2d(conv1, 3, strides=2)
local1 = local_response_normalization(max1)
conv2 = conv_2d(local1, 384, 3, activation='relu')
conv3 = conv_2d(conv2, 384, 3, activation='relu')
conv4 = conv_2d(conv3, 256, 3, activation='relu')
max2 = max_pool_2d(conv4, 3, strides=2)
local2 = local_response_normalization(max2)
full1 = fully_connected(local2, 4096, activation='tanh')
drop1 = dropout(full1, 0.5)
full2 = fully_connected(drop1, 4096, activation='tanh',name='layer')
drop2 = dropout(full2, 0.5)
full3 = fully_connected(drop2, 6, activation='softmax')
network = regression(full3, optimizer='momentum',loss='categorical_crossentropy',learning_rate=0.001)
n_iter=100
suffix='_scratch_'+str(n_iter)
#initial data used is the golden data
t1=to_categorical(g_lbl,6)
f1=np.array(g_im)
f1=resize_batch(f1,227)
model = tflearn.DNN(network,tensorboard_verbose=3)
model.fit(f1,t1, n_epoch=50, validation_set=0.2, shuffle=True,show_metric=True, batch_size=300, snapshot_step=200, snapshot_epoch=False, run_id='training')
# model.load('model_alex_'+suffix+'.tflearn')
#visualize during test
observed = [input_data, conv1, max1,local1, conv2,conv3, conv4,max2,local2]#, full1,drop1,full2,drop2,full3]
layer_names=['input_data', 'conv0','max0','local0','conv1', 'max1','local1', 'conv2','conv3', 'conv4','max2','local2']#, 'full1','drop1','full2','drop2','full3']
observers = [tflearn.DNN(v, session=model.session) for v in observed]
for i in range(f1.shape[0]):
outputs = [m.predict(f1[i].reshape(-1,227,227,3)) for m in observers]
# print([d.shape for d in outputs])
k=0
for d in outputs:
# for i in range(d.shape[3]):
#print(d, i)
# imshow(d[0,:,:,i])
# print(g_files[i],g_files[i][24:])
idx=random.randint(0,d.shape[3]-1)
im=np.array(d[0,:,:,idx])
# print(im.shape,idx)
name='layers_227/'+g_files[i][24:]+'_'+layer_names[k]+'_'+str(idx)+'.png'
imsave(name,im) # image number and class label
k=k+1
# callbacks=PlottingCallback(model, f1[0], (max1, max2),14)
# X_test=f1[0]
# model.fit(f1,t1, n_epoch=5, validation_set=0.2,batch_size=32, callbacks=PlottingCallback(model, f1[0], (max1)))#MonitorCallback())
model.save('model_alex_227_'+suffix+'.tflearn')
#model.load('model_alex_'+suffix+'.tflearn')