Skip to content

Commit 22673f4

Browse files
rrlymanrrlyman
committed
misc commits
1 parent 2ecd5f4 commit 22673f4

9 files changed

+2196
-0
lines changed

‎n0_network.py

+96
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,96 @@
1+
import tensorflow as tf
2+
import numpy as np
3+
from collections import namedtuple
4+
import datetime
5+
import ocr_utils
6+
7+
class base_network(object):
8+
''' definition of the network
9+
'''
10+
11+
12+
13+
14+
def fit(self, truthed_data, nEpochs=5000):
15+
16+
perfect_count=10
17+
for i in range(nEpochs):
18+
19+
batch = truthed_data.next_batch(100)
20+
# assign feature data to each placeholder
21+
# the batch list is returned in the same order as the features requested
22+
feed = {self._keep_prob: 0.5}
23+
for j in range(truthed_data.num_features):
24+
feed[self._ph[j]] = batch[j]
25+
26+
if i%100 == 0:
27+
28+
feed[self._keep_prob] = 1.0
29+
result = self._sess.run([self._merged, self._accuracy ], feed_dict=feed)
30+
summary_str = result[0]
31+
32+
self._writer.add_summary(summary_str, i)
33+
train_accuracy = self._accuracy.eval(feed)
34+
if train_accuracy <= (1.0 - 1e-5 ):
35+
perfect_count=10;
36+
else:
37+
perfect_count -= 1
38+
if perfect_count==0:
39+
break;
40+
41+
print ("step %d, training accuracy %g"%(i, train_accuracy),flush=True)
42+
self._train_step.run(feed_dict=feed)
43+
44+
45+
46+
47+
48+
def test(self, truthed_data, title = ''):
49+
50+
# assign feature data to each placeholder
51+
error_images = np.empty((0,self._nRows,self._nCols))
52+
53+
test_accuracy=0
54+
m=0
55+
56+
for i in range(int(len(truthed_data.features[0])/100)):
57+
58+
batch = truthed_data.next_batch(100)
59+
# assign feature data to each placeholder
60+
# the batch list is returned in the same order as the features requested
61+
feed = {self._keep_prob: 1.0}
62+
for j in range(truthed_data.num_features):
63+
feed[self._ph[j]] = batch[j]
64+
65+
66+
result = self._sess.run([self._accuracy, self._x_image, self._correct_prediction], feed_dict=feed)
67+
68+
test_accuracy += result[0]
69+
error_images = np.append(error_images, result[1][:,:,:,0][result[2]==False],axis=0)
70+
m += 1
71+
try:
72+
print ("test accuracy {} for : {}".format(test_accuracy/m, title),flush=True)
73+
ocr_utils.montage(error_images,title='TensorFlow {} Error Images'.format(title))
74+
except:
75+
if m==0:
76+
print ("test accuracy 1",flush=True)
77+
else:
78+
print ("test accuracy {}".format(test_accuracy/m),flush=True)
79+
ocr_utils.montage(error_images,title='TensorFlow Error Images')
80+
81+
82+
def predict(self, truthed_features):
83+
feed={self._keep_prob: 1.0}
84+
# assign feature data to each placeholder
85+
error_images = np.empty((truthed_features.num_rows,truthed_features.num_columns))
86+
87+
test_accuracy=0
88+
m=0
89+
90+
for j in range(1,truthed_features.num_features):
91+
feed[self._ph[j]] = truthed_features.features[j]
92+
result = self._sess.run([self._prediction], feed_dict=feed)
93+
94+
return result[0]
95+
96+

‎n1_2cnv1fc.py

+240
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,240 @@
1+
import tensorflow as tf
2+
import numpy as np
3+
from collections import namedtuple
4+
import datetime
5+
import ocr_utils
6+
from n0_network import base_network as b_network
7+
8+
class network(b_network):
9+
''' definition of the network
10+
'''
11+
def __init__(self, truthed_features, dtype=np.float32):
12+
13+
self._sess = tf.InteractiveSession()
14+
15+
lst = []
16+
extra_features_width = 0 # width of extra features
17+
18+
"""# ==============================================================================
19+
20+
Placeholders
21+
22+
Compute the size of various layers
23+
24+
Create a tensorflow Placeholder for each feature of data returned from the
25+
dataset
26+
27+
"""# ==============================================================================
28+
29+
for i,nm in enumerate(truthed_features.feature_names):
30+
31+
# features[0], is always the target. For instance it may be m_label_one_hot
32+
# the second features[1] is the 'image' that is passed to the convolution layers
33+
# Any additional features bypass the convolution layers and go directly
34+
# into the fully connected layer.
35+
36+
# The width of the extra features is calculated in order to allocate
37+
# the correct widths of weights, # and inputs
38+
# names are assigned to make the look pretty on the tensorboard graph.
39+
40+
if i == 0:
41+
nm = 'y_'+nm
42+
else:
43+
nm = 'x_'+nm
44+
if i>1:
45+
extra_features_width += truthed_features.feature_width[i]
46+
lst.append(tf.placeholder(dtype, shape=[None, truthed_features.feature_width[i]], name=nm))
47+
48+
# ph is a named tuple with key names like 'image', 'm_label', and values that
49+
# are tensors. The display name on the Chrome graph are 'y_m_label', 'x_image,
50+
# x_upper_case etc.
51+
52+
53+
Place_Holders = namedtuple('Place_Holders', truthed_features.feature_names)
54+
self._ph = Place_Holders(*lst) # unpack placeholders into named Tuple
55+
self._keep_prob = tf.placeholder(dtype,name='keep_prob')
56+
self._nRows = truthed_features.num_rows #image height
57+
self._nCols = truthed_features.num_columns #image width
58+
nFc = 1024 # size of fully connected layer
59+
nConv1 = 32 # size of first convolution layer
60+
nConv2 = 64 # size of second convolution layer
61+
nTarget = truthed_features.feature_width[0] # the number of one_hot features in the target, 'm_label'
62+
n_h_pool2_outputs = int(self._nRows/4) * int(self._nCols/4) * nConv2 # second pooling layer
63+
n_h_pool2_outputsx = n_h_pool2_outputs + extra_features_width # fully connected
64+
65+
"""# ==============================================================================
66+
67+
Build a Multilayer Convolutional Network
68+
69+
Weight Initialization
70+
71+
"""# ==============================================================================
72+
73+
def weight_variable(shape, dtype):
74+
initial = tf.truncated_normal(shape, stddev=0.1,dtype=dtype)
75+
return tf.Variable(initial)
76+
77+
def bias_variable(shape, dtype):
78+
initial = tf.constant(0.1, shape=shape, dtype=dtype)
79+
return tf.Variable(initial)
80+
81+
"""# ==============================================================================
82+
83+
Convolution and Pooling
84+
85+
keep our code cleaner, let's also abstract those operations into functions.
86+
87+
"""# ==============================================================================
88+
89+
def conv2d(x, W):
90+
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
91+
92+
def max_pool_2x2(x):
93+
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
94+
strides=[1, 2, 2, 1], padding='SAME')
95+
96+
"""# ==============================================================================
97+
98+
First Convolutional Layer
99+
100+
"""# ==============================================================================
101+
with tf.name_scope("w_conv1") as scope:
102+
W_conv1 = weight_variable([5, 5, 1, nConv1],dtype)
103+
b_conv1 = bias_variable([nConv1],dtype)
104+
105+
with tf.name_scope("reshape_x_image") as scope:
106+
self._x_image = tf.reshape(self._ph.image, [-1,self._nCols,self._nRows,1])
107+
108+
image_summ = tf.image_summary("x_image", self._x_image)
109+
110+
"""# ==============================================================================
111+
112+
We then convolve x_image with the weight tensor, add the bias, apply the ReLU function,
113+
and finally max pool.
114+
115+
"""# ==============================================================================
116+
117+
with tf.name_scope("convolve_1") as scope:
118+
h_conv1 = tf.nn.relu(conv2d(self._x_image, W_conv1) + b_conv1)
119+
120+
with tf.name_scope("pool_1") as scope:
121+
h_pool1 = max_pool_2x2(h_conv1)
122+
123+
"""# ==============================================================================
124+
125+
Second Convolutional Layer
126+
127+
In order to build a deep network, we stack several layers of this type. The second
128+
layer will have 64 features for each 5x5 patch.
129+
130+
"""# ==============================================================================
131+
132+
with tf.name_scope("convolve_2") as scope:
133+
W_conv2 = weight_variable([5, 5, nConv1, nConv2],dtype)
134+
b_conv2 = bias_variable([64],dtype)
135+
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
136+
137+
with tf.name_scope("pool_2") as scope:
138+
h_pool2 = max_pool_2x2(h_conv2)
139+
140+
"""# ==============================================================================
141+
142+
Densely Connected Layer
143+
144+
Now that the image size has been reduced to 7x7, we add a fully-connected layer
145+
with neurons to allow processing on the entire image. We reshape the tensor
146+
from the pooling layer into a batch of vectors, multiply by a weight matrix, add
147+
a bias, and apply a ReLU.
148+
149+
"""# ==============================================================================
150+
151+
with tf.name_scope("W_fc1_b") as scope:
152+
W_fc1 = weight_variable([n_h_pool2_outputsx, nFc],dtype)
153+
b_fc1 = bias_variable([nFc],dtype)
154+
155+
h_pool2_flat = tf.reshape(h_pool2, [-1, n_h_pool2_outputs])
156+
157+
# append the features, the 2nd on, that go directly to the fully connected layer
158+
for i in range(2,truthed_features.num_features ):
159+
h_pool2_flat = tf.concat(1, [h_pool2_flat, self._ph[i]])
160+
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
161+
162+
"""# ==============================================================================
163+
164+
Dropout
165+
166+
"""# ==============================================================================
167+
168+
169+
with tf.name_scope("drop") as scope:
170+
h_fc1_drop = tf.nn.dropout(h_fc1, self._keep_prob)
171+
172+
"""# ==============================================================================
173+
174+
Readout Layer
175+
176+
"""# ==============================================================================
177+
with tf.name_scope("softmax") as scope:
178+
W_fc2 = weight_variable([nFc, nTarget],dtype)
179+
b_fc2 = bias_variable([nTarget],dtype)
180+
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
181+
182+
183+
with tf.name_scope("xent") as scope:
184+
185+
# 1e-8 added to eliminate the crash of training when taking log of 0
186+
cross_entropy = -tf.reduce_sum(self._ph[0]*tf.log(y_conv+ 1e-8 ))
187+
ce_summ = tf.scalar_summary("cross entropy", cross_entropy)
188+
189+
with tf.name_scope("train") as scope:
190+
self._train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
191+
192+
with tf.name_scope("test") as scope:
193+
self._correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(self._ph[0],1))
194+
self._prediction = tf.argmax(y_conv,1)
195+
196+
self._accuracy = tf.reduce_mean(tf.cast(self._correct_prediction, dtype))
197+
accuracy_summary = tf.scalar_summary("accuracy", self._accuracy)
198+
"""# ==============================================================================
199+
200+
Start TensorFlow Interactive Session
201+
202+
"""# ==============================================================================
203+
204+
self._sess.run(tf.initialize_all_variables())
205+
self._merged = tf.merge_all_summaries()
206+
tm = ""
207+
tp = datetime.datetime.now().timetuple()
208+
for i in range(4):
209+
tm += str(tp[i])+'-'
210+
tm += str(tp[4])
211+
self._writer = tf.train.SummaryWriter("/tmp/ds_logs/"+ tm, self._sess.graph)
212+
213+
def computeSize(s,tens):
214+
sumC = 1
215+
tShape = tens.get_shape()
216+
nDims = len(tShape)
217+
for i in range(nDims):
218+
sumC *= tShape[i].value
219+
print ('\t{}\t{}'.format(s,sumC),flush=True)
220+
return sumC
221+
222+
print ('network size:',flush=True)
223+
total = computeSize("W_fc1",W_fc1)+ \
224+
computeSize ("b_fc1",b_fc1) + \
225+
computeSize ("W_conv1",W_conv1) + \
226+
computeSize ("b_conv1",b_conv1) + \
227+
computeSize ("W_conv2",W_conv2) + \
228+
computeSize ("b_conv2",b_conv2) + \
229+
computeSize ("W_fc2",W_fc2) + \
230+
computeSize ("b_fc2",b_fc2)
231+
print('\ttotal\t{}'.format(total),flush=True )
232+
233+
234+
def reset_graph(self):
235+
tf.reset_default_graph() # only necessary when iterating through fonts
236+
self._sess.close()
237+
238+
239+
240+

0 commit comments

Comments
 (0)