tflearn 中文汉字识别，训练后模型存为pb给TensorFlow使用。
数据目录在data，data下放了汉字识别图片：
data$ ls
0 1 10 11 12 13 14 15 16 2 3 4 5 6 7 8 9
datag$ ls 0
xxx.png yyy.png ....
代码：
如果将get model里的模型层数加非常深，训练时候很可能不会收敛，精度一直停留下1%以内。
# -*- coding: utf-8 -*-
from __future__ import division, print_function, absolute_import
import os
import numpy as np
import pickle
import tflearn
from PIL import Image
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d, avg_pool_2d
from tflearn.layers.merge_ops import merge
from tflearn.layers.estimator import regression
from tflearn.data_utils import to_categorical, shuffle
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report, confusion_matrix
from tflearn.layers.conv import highway_conv_2d, max_pool_2d
from tflearn.layers.normalization import local_response_normalization, batch_normalization
def resize_image(in_image, new_width, new_height, out_image=None,
resize_mode=Image.ANTIALIAS):
""" Resize an image.
Arguments:
in_image: `PIL.Image`. The image to resize.
new_ `int`. The image new width.
new_height: `int`. The image new height.
out_image: `str`. If specified, save the image to the given path.
resize_mode: `PIL.Image.mode`. The resizing mode.
Returns:
`PIL.Image`. The resize image.
"""
img = in_image.resize((new_width, new_height), resize_mode)
if out_image:
img.save(out_image)
return img
def convert_color(in_image, mode):
""" Convert image color with provided `mode`. """
return in_image.convert(mode)
def pil_to_nparray(pil_image):
""" Convert a PIL.Image to numpy array. """
pil_image.load()
return np.asarray(pil_image, dtype="float32")
def iterbrowse(path):
for home, dirs, files in os.walk(path):
for filename in files:
yield os.path.join(home, filename)
def directory_to_samples(directory, flags):
""" Read a directory, and list all subdirectories files as class sample """
samples = []
targets = []
# label class is from 0 !!!
label = 0
try: # Python 2
classes = sorted(os.walk(directory).next()[1])
except Exception: # Python 3
classes = sorted(os.walk(directory).__next__()[1])
for c in classes:
c_dir = os.path.join(directory, c)
try: # Python 2
walk = os.walk(c_dir).next()
except Exception: # Python 3
walk = os.walk(c_dir).__next__()
for sample in walk[2]:
if any(flag in sample for flag in flags):
samples.append(os.path.join(c_dir, sample))
targets.append(label)
label += 1
return samples, targets
# Get the pixel from the given image
def get_pixel(image, i, j):
# Inside image bounds?
width, height = image.size
if i > width or j > height:
return None
# Get Pixel
pixel = image.getpixel((i, j))
return pixel
# Create a Grayscale version of the image
def convert_to_one_channel(image):
# !!! I assume that the png file is grayscale. And R == G == B !!!! So I check it...
"""
for i in range(len(image)):
for j in range(len(image[i])):
pixel = image[i][j]
# Get R, G, B values (This are int from 0 to 255)
assert len(pixel) == 3
red = pixel[0]
green = pixel[1]
blue = pixel[2]
assert red == green == blue
assert 0 <= red <= 1
"""
# Just extract 1 channel data
return image[:, :, [0]]
def image_dirs_to_samples(directory, resize=None, convert_gray=False,
filetypes=None):
print("Starting to parse images...")
samples, targets = directory_to_samples(directory, flags=filetypes)
for i, s in enumerate(samples):
print("Process %d th file %s" % (i+1, s))
samples[i] = Image.open(s) # Load an image, returns PIL.Image.
if resize:
######################## TODO #######################
samples[i] = resize_image(samples[i], resize[0],
resize[1])
######################### TODO ####################### convert to more data
# if convert_gray:
# samples[i] = convert_color(samples[i], 'L')
samples[i] = pil_to_nparray(samples[i])
samples[i] /= 255. # ormalize a list of sample image data in the range of 0 to 1
samples[i] = convert_to_one_channel(samples[i]) # just want 1 channel data
print("Parsing Done!")
return samples, targets
def load_data(dirname, resize_pics=(128, 128), shuffle_data=True):
dataset_file = os.path.join(dirname, 'data.pkl')
try:
X, Y, org_labels = pickle.load(open(dataset_file, 'rb'))
except Exception:
# X, Y = image_dirs_to_samples(os.path.join(dirname, 'train/'), resize_pics, False, ['.jpg', '.png'])
X, Y = image_dirs_to_samples(dirname, resize_pics, False,
['.jpg', '.png']) # TODO, memory is too small to load all data
org_labels = Y
Y = to_categorical(Y, np.max(Y) + 1) # First class is '0', Convert class vector (integers from 0 to nb_classes)
if shuffle_data:
X, Y, org_labels = shuffle(X, Y, org_labels)
pickle.dump((X, Y, org_labels), open(dataset_file, 'wb'))
return X, Y, org_labels
class EarlyStoppingCallback(tflearn.callbacks.Callback):
def __init__(self, val_acc_thresh):
# Store a validation accuracy threshold, which we can compare against
# the current validation accuracy at, say, each epoch, each batch step, etc.
self.val_acc_thresh = val_acc_thresh
def on_epoch_end(self, training_state):
"""
This is the final method called in trainer.py in the epoch loop.
We can stop training and leave without losing any information with a simple exception.
"""
# print dir(training_state)
print("Terminating training at the end of epoch", training_state.epoch)
if training_state.val_acc >= self.val_acc_thresh and training_state.acc_value >= self.val_acc_thresh:
raise StopIteration
def on_train_end(self, training_state):
"""
Furthermore, tflearn will then immediately call this method after we terminate training,
(or when training ends regardless). This would be a good time to store any additional
information that tflearn doesn't store already.
"""
print("Successfully left training! Final model accuracy:", training_state.acc_value)
def get_model(width, height, classes=40):
# TODO, modify model
# Real-time data preprocessing
img_prep = tflearn.ImagePreprocessing()
# Real-time data preprocessing
img_prep = tflearn.ImagePreprocessing()
img_prep.add_featurewise_zero_center(per_channel=True)
img_prep.add_featurewise_stdnorm()
network = input_data(shape=[None, width, height, 1], data_preprocessing=img_prep) # if RGB, 224,224,3
network = conv_2d(network, 32, 3, activation='relu')
network = max_pool_2d(network, 2)
network = conv_2d(network, 64, 3, activation='relu')
network = conv_2d(network, 64, 3, activation='relu')
network = max_pool_2d(network, 2)
network = fully_connected(network, 512, activation='relu')
network = dropout(network, 0.5)
network = fully_connected(network, classes, activation='softmax')
network = regression(network, optimizer='adam',
loss='categorical_crossentropy',
learning_rate=0.001)
model = tflearn.DNN(network, tensorboard_verbose=0)
return model
if __name__ == "__main__":
width, height = 32, 32
X, Y, org_labels = load_data(dirname="data", resize_pics=(width, height))
trainX, testX, trainY, testY = train_test_split(X, Y, test_size=0.2, random_state=666)
print("sample data:")
print(trainX[0])
print(trainY[0])
print(testX[-1])
print(testY[-1])
model = get_model(width, height, classes=100)
filename = 'cnn_handwrite-acc0.8.tflearn'
# try to load model and resume training
#try:
# model.load(filename)
# print("Model loaded OK. Resume training!")
#except:
# pass
# Initialize our callback with desired accuracy threshold.
early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.9)
try:
model.fit(trainX, trainY, validation_set=(testX, testY), n_epoch=500, shuffle=True,
snapshot_epoch=True, # Snapshot (save & evaluate) model every epoch.
show_metric=True, batch_size=32, callbacks=early_stopping_cb, run_id='cnn_handwrite')
except StopIteration as e:
print("OK, stop iterate!Good!")
model.save(filename)
# predict all data and calculate confusion_matrix
model.load(filename)
pro_arr =model.predict(X)
predict_labels = np.argmax(pro_arr, axis=1)
print(classification_report(org_labels, predict_labels))
print(confusion_matrix(org_labels, predict_labels))
运行效果：100个汉字2分钟就可以达到95%精度。
---------------------------------
Run id: cnn_handwrite
Log directory: /tmp/tflearn_logs/
---------------------------------
Preprocessing... Calculating mean over all dataset (this may take long)...
Mean: [ 0.89235026] (To avoid repetitive computation, add it to argument 'mean' of `add_featurewise_zero_center`)
---------------------------------
Preprocessing... Calculating std over all dataset (this may take long)...
STD: 0.192279 (To avoid repetitive computation, add it to argument 'std' of `add_featurewise_stdnorm`)
---------------------------------
Training samples: 19094
Validation samples: 4774
--
Training Step: 597 | total loss: 0.70288 | time: 40.959ss
| Adam | epoch: 001 | loss: 0.70288 - acc: 0.7922 | val_loss: 0.54380 - val_acc: 0.8460 -- iter: 19094/19094
--
Terminating training at the end of epoch 1
Training Step: 1194 | total loss: 0.48860 | time: 40.245s
| Adam | epoch: 002 | loss: 0.48860 - acc: 0.8783 | val_loss: 0.37020 - val_acc: 0.8923 -- iter: 19094/19094
--
Terminating training at the end of epoch 2
Training Step: 1791 | total loss: 0.35790 | time: 41.315ss
| Adam | epoch: 003 | loss: 0.35790 - acc: 0.9090 | val_loss: 0.34719 - val_acc: 0.9049 -- iter: 19094/19094
--
Terminating training at the end of epoch 3
Successfully left training! Final model accuracy: 0.908959209919
OK, stop iterate!Good!
precision recall f1-score support
0 1.00 0.99 0.99 239
1 0.95 0.96 0.96 237
2 0.91 0.98 0.94 240
3 0.90 0.98 0.94 239
4 0.96 0.98 0.97 239
5 0.94 0.97 0.96 239
6 0.98 0.98 0.98 239
7 0.84 0.99 0.91 240
8 0.99 0.87 0.93 239
9 0.95 0.98 0.96 239
10 0.97 0.94 0.96 240
11 0.95 0.98 0.97 240
12 0.92 0.99 0.95 240
13 0.95 0.96 0.96 239
14 0.94 0.94 0.94 236
15 0.94 0.97 0.96 240
16 0.94 0.98 0.96 240
17 0.97 0.99 0.98 240
18 0.94 0.93 0.94 240
19 1.00 0.95 0.98 239
20 0.96 0.98 0.97 240
21 0.98 0.91 0.95 239
22 0.97 0.95 0.96 239
23 1.00 0.97 0.98 239
24 0.94 0.98 0.96 240
25 0.98 0.98 0.98 237
26 0.91 1.00 0.95 239
27 0.91 0.96 0.93 239
28 0.97 0.88 0.92 239
29 1.00 0.98 0.99 240
30 0.99 0.94 0.96 239
31 0.97 0.97 0.97 237
32 0.94 0.98 0.96 236
33 0.94 0.96 0.95 239
34 0.98 0.99 0.98 239
35 0.99 0.98 0.99 240
36 0.96 0.92 0.94 239
37 1.00 0.93 0.96 240
38 0.96 0.99 0.98 238
39 0.98 0.97 0.97 238
40 0.92 0.90 0.91 240
41 0.96 0.97 0.96 237
42 0.98 0.97 0.97 240
43 0.95 0.96 0.95 239
44 0.97 0.96 0.97 239
45 0.95 0.94 0.95 239
46 0.93 0.96 0.94 232
47 0.98 0.91 0.94 237
48 0.95 0.97 0.96 239
49 0.97 0.80 0.88 226
50 0.90 0.95 0.92 240
51 0.98 0.99 0.99 236
52 0.96 0.90 0.93 240
53 0.99 0.96 0.97 235
54 0.97 0.93 0.95 240
55 0.99 0.98 0.99 240
56 0.97 0.92 0.95 239
57 0.97 0.97 0.97 239
58 1.00 0.98 0.99 238
59 0.92 0.98 0.95 240
60 0.99 0.90 0.94 240
61 1.00 0.99 0.99 238
62 0.92 0.95 0.94 239
63 0.92 0.98 0.95 238
64 0.98 0.92 0.95 240
65 0.99 0.92 0.95 239
66 0.98 0.99 0.99 240
67 0.95 0.95 0.95 240
68 0.96 0.98 0.97 239
69 0.97 0.97 0.97 239
70 0.98 0.94 0.96 240
71 0.91 0.96 0.93 239
72 0.98 0.97 0.97 239
73 0.99 0.89 0.94 240
74 0.97 0.99 0.98 237
75 0.89 0.97 0.92 240
76 0.97 0.96 0.97 241
77 0.89 0.91 0.90 240
78 1.00 0.89 0.94 239
79 0.90 0.98 0.94 239
80 0.89 0.96 0.92 240
81 0.96 0.71 0.81 225
82 0.95 1.00 0.97 238
83 0.67 0.96 0.79 239
84 0.97 0.85 0.91 240
85 0.95 0.98 0.96 239
86 0.99 0.93 0.96 240
87 0.98 0.91 0.94 239
88 0.97 0.97 0.97 240
89 0.97 0.94 0.95 239
90 0.97 0.96 0.96 236
91 0.91 0.97 0.94 239
92 0.98 0.95 0.96 240
93 0.98 0.97 0.98 239
94 0.98 0.95 0.97 240
95 0.98 0.99 0.99 239
96 0.95 0.97 0.96 240
97 0.98 0.97 0.98 239
98 0.95 0.98 0.97 237
99 0.97 0.96 0.97 239
avg / total 0.96 0.95 0.95 23868
[[237 0 0 ..., 0 0 0]
[ 0 228 0 ..., 0 0 0]
[ 0 0 235 ..., 0 0 0]
...,
[ 0 0 0 ..., 233 0 0]
[ 0 0 0 ..., 0 233 0]
[ 0 0 0 ..., 0 0 230]]
更多模型见：http://www.cnblogs.com/bonelee/p/8978060.html
将上述模型保存并给TensorFlow使用，仅仅在保存模型前加del tf.get_collection_ref(tf.GraphKeys.TRAIN_OPS)[:]，仅仅保留inference时候的OP（如果需要retrain注意），如下：
model = get_model(width, height, classes=100)
filename = 'cnn_handwrite-acc0.8.tflearn'
# try to load model and resume training
#try:
# model.load(filename)
# print("Model loaded OK. Resume training!")
#except:
# pass
# Initialize our callback with desired accuracy threshold.
early_stopping_cb = EarlyStoppingCallback(val_acc_thresh=0.8)
try:
model.fit(trainX, trainY, validation_set=(testX, testY), n_epoch=500, shuffle=True,
snapshot_epoch=True, # Snapshot (save & evaluate) model every epoch.
show_metric=True, batch_size=32, callbacks=early_stopping_cb, run_id='cnn_handwrite')
except StopIteration as e:
print("OK, stop iterate!Good!")
model.save(filename)
del tf.get_collection_ref(tf.GraphKeys.TRAIN_OPS)[:]
"""
# print op name
with tf.Session() as sess:
init_op = tf.initialize_all_variables()
sess.run(init_op)
for v in sess.graph.get_operations():
print(v.name)
"""
filename = 'cnn_handwrite-acc0.8.infer.tflearn'
model.save(filename)
参考：http://www.cnblogs.com/bonelee/p/8445261.html 里的脚本，修改：
output_node_names = "FullyConnected/Softmax"
通常为：
output_node_names = "FullyConnected/Softmax"
或者
output_node_names = "FullyConnected_1/Softmax"
output_node_names = "FullyConnected_2/Softmax"
就看你使用的全连接层数，上面分别是1，2，3层。
最后，tensorflow里的使用：
def inference(image):
print('inference')
temp_image = Image.open(image).convert('L')
temp_image = temp_image.resize((FLAGS.image_size, FLAGS.image_size), Image.ANTIALIAS)
temp_image = np.asarray(temp_image) / 255.0
temp_image = temp_image.reshape([-1, 32, 32, 1])
from tensorflow.python.platform import gfile
with tf.Graph().as_default():
output_graph_def = tf.GraphDef()
with open("frozen_model.pb", "rb") as f:
output_graph_def.ParseFromString(f.read())
tensors = tf.import_graph_def(output_graph_def, name="")
#print tensors
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
op = sess.graph.get_operations()
"""
for m in op:
print(m.values())
"""
op = sess.graph.get_tensor_by_name("FullyConnected_1/Softmax:0")
input_tensor = sess.graph.get_tensor_by_name('InputData/X:0')
probs = sess.run(op,feed_dict = {input_tensor:temp_image})
print probs
result = []
for word in probs:
result.append(np.argsort(-word)[:3])
return result
def main(_):
image_path = './data/test/00098/104405.png'
#image_path = '../data/00010/17724.png'
final_predict_val = inference(image_path)
logger.info('the result info label {0} predict index {1}'.format(98, final_predict_val))
一般，输入TensorFlow input name默认为InputData/X，但只是op，如果要tensor的话，加上数字0，也就是：InputData/X:0
同理，FullyConnected_1/Softmax:0。
最后预测效果：
[[ 8.42533936e-08 1.60850794e-11 2.60133332e-10 2.42555542e-14
4.96124599e-08 4.45251297e-15 3.98175590e-11 1.64476592e-11
7.03968351e-13 5.42319011e-12 8.55469237e-11 4.91866422e-13
1.77282828e-07 4.05237593e-10 3.13049003e-10 1.34780919e-11
2.05803235e-06 2.87827305e-07 1.47789994e-12 2.53391891e-11
3.77086790e-13 2.02639586e-10 9.03167027e-13 3.96698889e-11
1.30850096e-11 5.71980917e-12 3.03487374e-11 2.04132298e-14
6.25303683e-13 1.46122332e-07 2.17450633e-07 1.69623715e-09
6.80857757e-12 2.52643609e-13 6.56771096e-11 8.55152287e-16
1.34496514e-09 1.22644633e-06 1.12011307e-07 7.93476283e-05
8.24334611e-12 4.77531155e-14 9.39397757e-13 2.38438267e-14
2.11416329e-10 5.54395712e-08 2.30046147e-12 2.63584043e-10
4.70621564e-16 5.14432724e-12 6.42602327e-09 1.62485829e-13
7.39078274e-08 3.19146315e-12 5.25887156e-09 1.35877786e-13
1.39127886e-13 2.11998293e-13 9.09501097e-09 9.46486750e-07
2.47498733e-09 2.74523763e-12 1.02716433e-14 1.02069058e-17
3.09356682e-16 1.51022904e-15 9.34333665e-13 2.62195051e-14
3.38079781e-16 7.43019903e-13 1.92409091e-13 3.86611994e-13
2.61276265e-12 1.07969211e-09 1.30814548e-09 2.44038188e-14
9.79275905e-13 1.41007803e-10 6.15137758e-12 2.08893070e-10
1.34751668e-14 2.76824767e-15 7.84100464e-16 7.70873335e-15
5.45704757e-12 3.69386271e-18 2.06012223e-13 1.62567273e-14
1.54544960e-03 2.05292008e-06 1.31726174e-09 7.04993663e-09
4.11338266e-03 3.19344110e-07 3.96519717e-05 2.26919351e-12
2.39114349e-12 2.35558744e-07 9.94213998e-01 1.10125060e-11]]
the result info label 98 predict index [array([98, 92, 88])]