/
NN_with_aug.py
171 lines (141 loc) · 6.03 KB
/
NN_with_aug.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
import numpy as np
import tensorflow as tf
import pandas as pd
import cv2
from autoaugment import CIFAR10Policy
from utils import sparse_cost_sensitive_loss,onehot
from PIL import Image
from model import feature_extractor
from utils import sparse_cost_sensitive_loss,onehot
import itertools
def cv_resize_pad(img,desired_size):
osize = img.shape[:2]
ratio = float(desired_size)/max(osize)
nsize = tuple([int(x*ratio) for x in osize])
img = cv2.resize(img, (nsize[1],nsize[0]))
dw = desired_size - nsize[1]
dh = desired_size - nsize[0]
top,bottom = dh//2,dh-(dh//2)
left,right = dw//2,dw-(dw//2)
color = [0,0,0]
return cv2.copyMakeBorder(img,top,bottom,left,right,cv2.BORDER_CONSTANT,value=color)
def load_images(data_dir,resize=None,seq_len=None):
data_desc = pd.read_csv(data_dir+"labelsImgPath.csv",sep=",")
img_list = []
shapes = []
for i in data_desc['filename']:
img_raw = cv2.imread(data_dir+i,cv2.IMREAD_COLOR)
img_list.append(img_raw)
shapes.append((img_raw.shape[0],img_raw.shape[1]))
shapes = np.array(shapes)
if(resize==None):
resize = np.int8((np.mean(shapes[:,0])+np.mean(shapes[:,1]))/2.0)
res_imgs = []
for img in img_list:
nimg = cv_resize_pad(img,resize)
# More augmentation
res_imgs.append(nimg)
images = np.array(res_imgs)
g_images,labels = generate_seqs(images,data_desc)
g_images = tf.keras.preprocessing.sequence.pad_sequences(g_images,maxlen=seq_len,dtype='float32')
return g_images,labels
def generate_seqs(images,data_desc,onehot_lab=True):
idx = []
runn_idx = 0
img_seqs = []
labels = []
label = None
tid = 0
for _,row in data_desc.iterrows():
if(tid != row['trackid']):
if(len(idx)!=0):
idx = list(map(lambda x: x+runn_idx,idx))
img_seqs.append(np.array(images[idx]))
labels.append(label)
runn_idx = runn_idx + len(idx)
tid = row['trackid']
idx = [row['framenr']-2] #TODO
else:
idx.append(row['framenr']-2)
label = row['class']
if(onehot_lab):
labels = onehot(labels,label_dict={'boat':1,'nature':0})
return img_seqs,labels
def make_img_from_tensor(np_ary):
img_list=[]
for i in range(0,20):
img_list.append(Image.fromarray(np_ary[i,:,:],'RGB'))
return img_list
def augment(np_ary):
transformed_list = []
img_list = make_img_from_tensor(np_ary)
policy = CIFAR10Policy()
policy.draw()
for img in img_list:
transformed_list.append(np.array(policy(img)))
return transformed_list
def run(data_dir,batchsize=50, n_epochs=50):
tf.reset_default_graph()
train_dir = data_dir +'train/'
test_dir = data_dir +'test/'
train_imgs,train_labels = load_images(train_dir)
test_imgs,test_labels = load_images(test_dir,resize=train_imgs.shape[2],seq_len=train_imgs.shape[1])
n_samples = train_imgs.shape[0]
#train_dataset = tf.data.Dataset.from_tensor_slices((train_imgs,train_labels))
test_dataset = tf.data.Dataset.from_tensor_slices((test_imgs,test_labels))
train_img = tf.data.Dataset.from_tensor_slices(train_imgs)
train_labels = tf.data.Dataset.from_tensor_slices(train_labels)
train_dataset = tf.data.Dataset.zip((train_img, train_labels))
train_dataset = train_dataset.shuffle(buffer_size=100,reshuffle_each_iteration=True).batch(batchsize).repeat()
test_dataset = test_dataset.shuffle(buffer_size=100,reshuffle_each_iteration=True).batch(batchsize)
iterator = tf.data.Iterator.from_structure(train_dataset.output_types,train_dataset.output_shapes)
x,y = iterator.get_next()
train_iterator = iterator.make_initializer(train_dataset)
test_iterator = iterator.make_initializer(test_dataset)
# apply random augmentations
ft_extr = feature_extractor()
logits = ft_extr.create_3dconv_model(x)
#loss = weighted_ce(next_element[1],model,.1)
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits,labels=y))
prediction = tf.nn.softmax(logits)
#cnf_matrix = tf.math.confusion_matrix(predictions=tf.to_float(tf.argmax(prediction,1)),labels=tf.to_float(tf.argmax(y, 1)),num_classes=2)
equality = tf.equal(tf.to_float(tf.argmax(prediction,1)), tf.to_float(tf.argmax(y, 1)))
accuracy = tf.reduce_mean(tf.to_float(equality))
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer(learning_rate=.001).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(train_iterator)
saver = tf.train.Saver()
for epoch in range(n_epochs):
ep_loss = []
ep_cnf = []
for _ in range(int(n_samples/batchsize)):
_,b_loss = sess.run([optimizer,loss],feed_dict={'is_training:0':True}) #cnf_mat cnf_matrix
ep_loss.append(b_loss)
#ep_cnf.append(cnf_mat)
print(np.mean(ep_loss))
print(np.mean(ep_cnf,axis=0))
if(n_epochs%10==0):
save_path = saver.save(sess,data_dir+str(epoch)+"_checkpoint.ckpt")
print('predicting..')
save_path = saver.save(sess,data_dir+"final_checkpoint.ckpt")
with tf.SessionL() as sess:
sess.run(test_iterator)
result_set = []
try:
while True:
pred = sess.run(prediction,feed_dict={'is_training:0':False})
result_set.append(pred)
except:
pass
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--dir', type=str, default="/fast/AG_Kainmueller/jrumber/Hackathon/data/")
args = parser.parse_args()
data_dir = args.__dict__['dir']
tf.reset_default_graph()
run(data_dir,batchsize=50, n_epochs=200)
print("done")