In [1]:
%matplotlib inline
import os
import sys
import pylab
import random
from random import randint, uniform
from skimage.util import crop
from skimage import transform
import numpy as np
import pandas as pd
import cPickle as pkl
from lasagne import layers
from bs4 import BeautifulSoup as bs
from lasagne import updates
import lasagne as nn
from theano.tensor.nnet import softmax
from scipy.misc import imread, imresize
from nolearn.lasagne import NeuralNet, BatchIterator, visualize
from sklearn.cross_validation import train_test_split
from sklearn.metrics import classification_report, accuracy_score
repo_location = '/workspace/.project/project/'
data_root = os.path.join(os.path.expanduser('~') + repo_location + 'datasets/')
script_root = os.path.join(os.path.expanduser('~') + repo_location + 'scripts/')
model_root = os.path.join(os.path.expanduser('~') + repo_location + 'models/')
In [2]:
# Load dataset
train_soup = bs(open(data_root + 'icdar03/train/char/char.xml').read(), 'lxml-xml')
test_soup = bs(open(data_root + 'icdar03/test/char/char.xml').read(), 'lxml-xml')
X_train = []
y_train = []
X_test = []
y_test = []
for image in train_soup('image'):
try:
img = imread(data_root + 'icdar03/train/char/' + image['file'])
X_train.append(img)
y_train.append(image['tag'])
except:
pass
for image in test_soup('image'):
try:
img = imread(data_root + 'icdar03/test/char/' + image['file'])
X_test.append(img)
y_test.append(image['tag'])
except:
pass
data_train = pd.DataFrame({'image' : X_train, 'label' : y_train})
data_test = pd.DataFrame({'image' : X_test, 'label' : y_test})
# drop extra labels
data_train = data_train.loc[~data_train['label'].isin([':', '-', '.', '\'', '!', '(', '"', ')', '&', '?', u'\xa3', u'\xc9', u'\xd1', u'\xe9', ','])]
data_test = data_test.loc[~data_test['label'].isin([':', '-', '.', '\'', '!', '(', '"', ')', '&', '?', u'\xa3', u'\xc9', u'\xd1', u'\xe9', ','])]
print 'Loaded icdar03'
In [3]:
# Reshape images to 32x32 and convert to grayscale
data_train_x = np.zeros((data_train['image'].count(), 1, 32, 32))
data_train_y = data_train['label'].values
data_test_x = np.zeros((data_test['image'].count(), 1, 32, 32))
data_test_y = data_test['label'].values
for idx, img in enumerate(data_train['image']):
img = imresize(img, (32, 32))
if len(img.shape) == 3:
data_train_x[idx, ...] = img.dot([0.299, 0.587, 0.144])
else:
data_train_x[idx, ...] = img
for idx, img in enumerate(data_test['image']):
img = imresize(img, (32, 32))
if len(img.shape) == 3:
data_test_x[idx, ...] = img.dot([0.299, 0.587, 0.144])
else:
data_test_x[idx, ...] = img
data_train_x = data_train_x.astype('float32')
data_test_x = data_test_x.astype('float32')
print 'icdar03 reshaped and grayscaled'
In [4]:
# Normalize by MuSigma
data_train_x /= data_train_x.std(axis = None)
data_train_x -= data_train_x.mean()
data_test_x /= data_test_x.std(axis = None)
data_test_x -= data_test_x.mean()
In [5]:
print data_train_x.shape, data_train_y.shape, data_test_x.shape, data_test_y.shape
In [6]:
class TransIterator(BatchIterator):
def fast_warp(self, img, tf, output_shape, mode='nearest'):
return transform._warps_cy._warp_fast(img, tf.params, output_shape=output_shape, mode=mode)
def transform(self, Xb, yb):
Xb, yb = super(TransIterator, self).transform(Xb, yb)
Xb_aug = np.empty(shape = (Xb.shape[0], 1, 32, 32), dtype = 'float32')
yb_aug = yb
# random rotations betweein -5 and 5 degrees
dorotate = randint(-5,5)
# random translations
trans_1 = randint(-3,3)
trans_2 = randint(-3,3)
# random zooms
zoom = uniform(0.8, 1.2)
# shearing
shear_deg = uniform(-10, 10)
# set the transform parameters for skimage.transform.warp
# have to shift to center and then shift back after transformation otherwise
# rotations will make image go out of frame
center_shift = np.array((32, 32)) / 2. - 0.5
tform_center = transform.SimilarityTransform(translation=-center_shift)
tform_uncenter = transform.SimilarityTransform(translation=center_shift)
tform_aug = transform.AffineTransform(rotation = np.deg2rad(dorotate),
scale =(1/zoom, 1/zoom),
shear = np.deg2rad(shear_deg),
translation = (trans_1, trans_2))
tform = tform_center + tform_aug + tform_uncenter
for j in range(Xb.shape[0]):
Xb_aug[j][0] = self.fast_warp(Xb[j][0], tform,
output_shape = (32, 32))
return Xb_aug, yb_aug
In [7]:
# setting nn
net = NeuralNet(
layers = [
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer),
('conv2', layers.Conv2DLayer),
('pool3', layers.MaxPool2DLayer),
('dropout4', layers.DropoutLayer),
('conv5', layers.Conv2DLayer),
('conv6', layers.Conv2DLayer),
('pool7', layers.MaxPool2DLayer),
('dropout8', layers.DropoutLayer),
('conv9', layers.Conv2DLayer),
('conv10', layers.Conv2DLayer),
('dropout12', layers.DropoutLayer),
('hidden13', layers.DenseLayer),
('dropout14', layers.DropoutLayer),
('hidden15', layers.DenseLayer),
('dropout16', layers.DropoutLayer),
('output', layers.DenseLayer),
],
input_shape = (None, 1, 32, 32),
conv1_num_filters = 128, conv1_filter_size = (3, 3),
conv2_num_filters = 128, conv2_filter_size = (3, 3),
pool3_pool_size = (2, 2),
dropout4_p = 0,
conv5_num_filters = 256, conv5_filter_size = (3, 3),
conv6_num_filters = 256, conv6_filter_size = (3, 3),
pool7_pool_size = (2, 2),
dropout8_p = 0.2,
conv9_num_filters = 512, conv9_filter_size = (3, 3),
conv10_num_filters = 512, conv10_filter_size = (3, 3),
dropout12_p = 0.2,
hidden13_num_units = 1024,
dropout14_p = 0.5,
hidden15_num_units = 1024,
dropout16_p = 0.5,
output_num_units = 62, output_nonlinearity = softmax,
batch_iterator_train = TransIterator(batch_size = 2500),
batch_iterator_test = BatchIterator(batch_size = 2500),
update = updates.adam,
use_label_encoder = True,
regression = False,
max_epochs = 300,
verbose = 1,
)
In [9]:
# train nn
net.load_params_from(os.path.join(model_root, 'recog_for_icdar.pkl')); # or load a pretrained model!
net.fit(data_train_x, data_train_y);
In [12]:
pred = net.predict(data_test_x)
print accuracy_score(data_test_y, pred)
In [13]:
print classification_report(data_test_y, pred)
In [ ]:
net.save_params_to(os.path.join(model_root, 'recog_for_icdar.pkl'))
In [ ]:
In [ ]: