def __init__(self,
mean,
valid_target_len=float('inf'),
img_width_range=(12, 320),
word_len=30,
channel=3):
img_height = 32
# '[(16,32),(27,32),(35,32),(64,32),(80,32)]'
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)),
(int(math.floor(108 / 4)), int(word_len + 2)),
(int(math.floor(140 / 4)), int(word_len + 2)),
(int(math.floor(256 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)),
int(word_len + 2))]
self.bucket_min_width, self.bucket_max_width = img_width_range #(12,320)
self.image_height = img_height
self.valid_target_len = valid_target_len
self.mean = mean
self.channel = channel
assert len(self.mean) == self.channel
#320个bucket,每個放相應大小爲width圖像
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
def __init__(self,
data_root,
annotation_fn,
evaluate=False,
valid_target_len=float('inf'),
img_width_range=(100, 800),
word_len=60):
"""
:param data_root:
:param annotation_fn:
:param lexicon_fn:
:param img_width_range: only needed for training set
:return:
"""
print("DATA GEN")
img_height = 32
self.data_root = data_root
if os.path.exists(annotation_fn):
self.annotation_path = annotation_fn
else:
self.annotation_path = os.path.join(data_root, annotation_fn)
# fixed width and len of text. It have to be changed for long handwritten line
'''if evaluate:
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)), (int(math.floor(108 / 4)), int(word_len + 2)),
(int(math.floor(140 / 4)), int(word_len + 2)), (int(math.floor(256 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)), int(word_len + 2))]
else:
self.bucket_specs = [(int(64 / 4), 9 + 2), (int(108 / 4), 15 + 2),
(int(140 / 4), 17 + 2), (int(256 / 4), 20 + 2),
(int(math.ceil(img_width_range[1] / 4)), word_len + 2)]'''
self.bucket_specs = [(int(100 / 4), int(word_len / 4) + 2),
(int(200 / 4), int(word_len / 4) + 2),
(int(300 / 4), int(word_len / 2) + 2),
(int(400 / 4), int(word_len / 2) + 2),
(int(500 / 4), word_len + 2),
(int(600 / 4), word_len + 2),
(int(700 / 4), word_len + 2),
(int(800 / 4), word_len + 2)]
self.max_len = word_len + 2
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = img_height
self.valid_target_len = valid_target_len
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
self.char2index = []
with open(os.path.join(self.data_root, 'sample.txt'), "r") as ins:
for line in ins:
self.char2index.append(line.strip())
self.char2index.append(' ')
print(self.char2index)
def __init__(
self,
data_root,
annotation_fn,
evaluate=False,
valid_target_len=float('inf'),
img_width_range=(83, 2083), # iamdb train set
word_len=81):
# img_width_range = (354,1990), # sgdb
# word_len = 74):
# img_width_range = (135,2358), # rimes
# word_len = 100):
"""
:param data_root:
:param annotation_fn:
:param lexicon_fn:
:param img_width_range: only needed for training set
:return:
"""
img_height = 32
self.data_root = data_root
if os.path.exists(annotation_fn):
self.annotation_path = annotation_fn
else:
self.annotation_path = os.path.join(data_root, annotation_fn)
if evaluate:
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)),
(int(math.floor(108 / 4)), int(word_len + 2)),
(int(math.floor(140 / 4)), int(word_len + 2)),
(int(math.floor(256 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)),
int(word_len + 2))]
else:
self.bucket_specs = [
(int(64 / 4), 9 + 2), (int(108 / 4), 15 + 2),
(int(140 / 4), 17 + 2), (int(256 / 4), 20 + 2),
(int(math.ceil(img_width_range[1] / 4)), word_len + 2)
]
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = img_height
self.valid_target_len = valid_target_len
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
def __init__(
self,
data_root,
annotation_fn,
evaluate=False,
valid_target_len=float('inf'),
img_width_range=(170, 1016), # iam training width range
word_len=81): # iam training max
"""
:param data_root:
:param annotation_fn:
:param lexicon_fn:
:param img_width_range: only needed for training set
:return:
"""
img_height = 32
self.data_root = data_root
if os.path.exists(annotation_fn):
self.annotation_path = annotation_fn
else:
self.annotation_path = os.path.join(data_root, annotation_fn)
if evaluate:
self.bucket_specs = [
(int(math.floor(img_width_range[0])),
int(95 + 2)), # iam test max
(int(math.ceil(img_width_range[1] / 5)), int(95 + 2)),
(int(math.ceil(img_width_range[1] / 4)), int(95 + 2)),
(int(math.ceil(img_width_range[1] / 3)), int(95 + 2)),
(int(math.ceil(img_width_range[1] / 2)), int(95 + 2))
]
else:
self.bucket_specs = [
(int(math.floor(img_width_range[0])), int(word_len + 2)),
(int(math.ceil(img_width_range[1] / 5)), int(word_len + 2)),
(int(math.ceil(img_width_range[1] / 4)), int(word_len + 2)),
(int(math.ceil(img_width_range[1] / 3)), int(word_len + 2)),
(int(math.ceil(img_width_range[1] / 2)), int(word_len + 2))
]
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = img_height
self.valid_target_len = valid_target_len
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
def __init__(self,
data_root, annotation_fn, train_sample_size=None,
evaluate = False,
valid_target_len = float('inf'),
img_width_range = (12, 320),
word_len = 60):
"""
:param data_root:
:param annotation_fn:
:param lexicon_fn:
:param img_width_range: only needed for training set
:return:
"""
img_height = 32
self.data_root = data_root
if os.path.exists(annotation_fn):
self.annotation_path = annotation_fn
else:
self.annotation_path = os.path.join(data_root, annotation_fn)
if evaluate:
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)), (int(math.floor(108 / 4)), int(word_len + 2)),
(int(math.floor(140 / 4)), int(word_len + 2)), (int(math.floor(256 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)), int(word_len + 2)),
(int(math.ceil(img_width_range[1] / 2)), word_len + 2)]
else:
self.bucket_specs = [(int(64 / 4), 9 + 2), (int(108 / 4), 15 + 2),
(int(140 / 4), 17 + 2), (int(256 / 4), 20 + 2),
(int(math.ceil(img_width_range[1] / 4)), word_len + 2),
(int(math.ceil(img_width_range[1] / 2)), word_len + 2)]
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = img_height
self.valid_target_len = valid_target_len
self.train_sample_size = train_sample_size
self.bucket_data = {i: BucketData()
for i in range(self.bucket_max_width + 1)}
self.OnLineDataGen = DataGenTextOnLine()
def __init__(self,
data_root, annotation_fn,
evaluate = False,
valid_target_len = float('inf'),
img_width_range = (39,936),
word_len = 50):
"""
:param data_root:
:param annotation_fn:
:param lexicon_fn:
:param img_width_range: only needed for training set
:return:
"""
img_height = 48
self.data_root = data_root
if os.path.exists(annotation_fn):
self.annotation_path = annotation_fn # abs path
else:
self.annotation_path = os.path.join(data_root, annotation_fn) # relative path
if evaluate:
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)), (int(math.floor(128 / 4)), int(word_len + 2)),
(int(math.floor(256 / 4)), int(word_len + 2)), (int(math.floor(512 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)), int(word_len + 2))]
else:
self.bucket_specs = [(int(64 / 4), 9 + 2), (int(128 / 4), 15 + 2),
(int(256 / 4), 17 + 2), (int(512 / 4), 25 + 2),
(int(math.ceil(img_width_range[1] / 4)), word_len + 2)]
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = img_height
self.valid_target_len = valid_target_len
self.bucket_data = {i: BucketData()
for i in range(self.bucket_max_width + 1)}
def clear(self):
self.bucket_data = {i: BucketData()
for i in range(self.bucket_max_width + 1)}
def __init__(self):
img_width_range = cfg.img_width_range
word_len = cfg.word_len
self.batch_size = cfg.batch_size
self.visualize = cfg.visualize
gpu_device_id = '/gpu:' + str(cfg.gpu_id)
if cfg.gpu_id == -1:
gpu_device_id = '/cpu:0'
print("Using CPU model!")
with tf.device(gpu_device_id):
self.img_data = tf.placeholder(tf.float32,
shape=(None, 1, 32, None),
name='img_data')
self.zero_paddings = tf.placeholder(tf.float32,
shape=(None, None, 512),
name='zero_paddings')
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)),
(int(math.floor(108 / 4)), int(word_len + 2)),
(int(math.floor(140 / 4)), int(word_len + 2)),
(int(math.floor(256 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)),
int(word_len + 2))]
buckets = self.buckets = self.bucket_specs
self.decoder_inputs = []
self.encoder_masks = []
self.target_weights = []
with tf.device(gpu_device_id):
for i in xrange(int(buckets[-1][0] + 1)):
self.encoder_masks.append(
tf.placeholder(tf.float32,
shape=[None, 1],
name="encoder_mask{0}".format(i)))
for i in xrange(buckets[-1][1] + 1):
self.decoder_inputs.append(
tf.placeholder(tf.int32,
shape=[None],
name="decoder{0}".format(i)))
self.target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = cfg.img_height
self.valid_target_len = cfg.valid_target_len
self.forward_only = True
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
with tf.device(gpu_device_id):
cnn_model = CNN(self.img_data, True) #(not self.forward_only))
self.conv_output = cnn_model.tf_output()
self.concat_conv_output = tf.concat(
axis=1, values=[self.conv_output, self.zero_paddings])
self.perm_conv_output = tf.transpose(self.concat_conv_output,
perm=[1, 0, 2])
with tf.device(gpu_device_id):
self.attention_decoder_model = Seq2SeqModel(
encoder_masks=self.encoder_masks,
encoder_inputs_tensor=self.perm_conv_output,
decoder_inputs=self.decoder_inputs,
target_weights=self.target_weights,
target_vocab_size=cfg.target_vocab_size,
buckets=self.buckets,
target_embedding_size=cfg.target_embedding_size,
attn_num_layers=cfg.attn_num_layers,
attn_num_hidden=cfg.attn_num_hidden,
forward_only=self.forward_only,
use_gru=cfg.use_gru)
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=1.0)
self.sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
self.saver_all = tf.train.Saver(tf.global_variables())
self.saver_all.restore(self.sess, cfg.ocr_model_path)
def __init__(
self,
data_root,
annotation_fn,
evaluate=False,
valid_target_len=float('inf'),
img_width_range=(83, 2083), # iamdb train set
word_len=81):
# img_width_range = (354,1990), # sgdb
# word_len = 74):
# img_width_range = (306,911), # pardb
# word_len = 70):
# img_width_range = (135,2358), # rimes
# word_len = 100):
# img_width_range = (175,1801), # gwdb
# word_len = 87):
"""
:param data_root:
:param annotation_fn:
:param lexicon_fn:
:param img_width_range: only needed for training set
:return:
"""
img_height = 64
self.data_root = data_root
if os.path.exists(annotation_fn):
self.annotation_path = annotation_fn
else:
self.annotation_path = os.path.join(data_root, annotation_fn)
if evaluate:
self.bucket_specs = [(int(math.ceil(img_width_range[0])),
int(math.ceil(img_width_range[1] / 8))),
(int(math.ceil(img_width_range[1] / 8)),
int(math.ceil(img_width_range[1] / 6))),
(int(math.ceil(img_width_range[1] / 6)),
int(math.ceil(img_width_range[1] / 4))),
(int(math.ceil(img_width_range[1] / 4)),
int(math.ceil(img_width_range[1] / 3))),
(int(math.ceil(img_width_range[1] / 3)),
int(math.ceil(img_width_range[1] / 2)))]
else:
self.bucket_specs = [(int(math.ceil(img_width_range[0])),
int(math.ceil(img_width_range[1] / 8))),
(int(math.ceil(img_width_range[1] / 8)),
int(math.ceil(img_width_range[1] / 6))),
(int(math.ceil(img_width_range[1] / 6)),
int(math.ceil(img_width_range[1] / 4))),
(int(math.ceil(img_width_range[1] / 4)),
int(math.ceil(img_width_range[1] / 3))),
(int(math.ceil(img_width_range[1] / 3)),
int(math.ceil(img_width_range[1] / 2)))]
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = img_height
self.valid_target_len = valid_target_len
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
