序列標註(sequence labelling),輸入序列每一幀預測一個類別。 OCR(Optical Character Recognition 光學字元辨識)。
MIT口語系統研究小組Rob Kassel收集,史丹佛大學人工智慧實驗室Ben Taskar預處理OCR資料集(http://ai.stanford.edu/~btaskar/ocr/ ),包含大量單獨手寫小寫字母,每個樣本對應16X8像素二值影像。字線組合序列,序列對應單字。 6800個,長度不超過14字母的單字。 gzip壓縮,內容用Tab分隔文字檔。 Python csv模組直接讀取。檔案每行一個歸一化字母屬性,ID號碼、標籤、像素值、下一字母ID號碼等。
下一字母ID值排序,依照正確順序讀取每個單字字母。收集字母,直到下一個ID對應欄位未被設定為止。讀取新序列。讀取完目標字母及資料像素,以零影像填滿序列對象,能納入兩個較大目標字母所有像素資料NumPy陣列。
時間步之間共用softmax層。資料和目標數組包含序列,每個目標字母對應一個影像幀。 RNN擴展,每個字母輸出加入softmax分類器。分類器對每幀資料而非整個序列評估預測結果。計算序列長度。一個softmax層加入所有幀:或為所有幀添加幾個不同分類器,或令所有幀共享同一個分類器。共享分類器,權值在訓練中被調整次數更多,訓練單字每個字母。一個全連接層權值矩陣維數batch_size*in_size*out_size。現需要在兩個輸入維度batch_size、sequence_steps更新權值矩陣。令輸入(RNN輸出活性值)扁平為形狀batch_size*sequence_steps*in_size。權值矩陣變成較大的批次資料。結果反扁平化(unflatten)。
代價函數,序列每一幀有預測目標對,在對應維度平均。依據張量長度(序列最大長度)歸一化的tf.reduce_mean無法使用。需依照實際序列長度歸一化,手工呼叫tf.reduce_sum和除法運算平均值。
損失函數,tf.argmax針對軸2非軸1,各幀填充,依據序列實際長度計算平均值。 tf.reduce_mean對批次資料所有單字取均值。
TensorFlow自動導數計算,可使用序列分類相同最佳化運算,只需要代入新代價函數。對所有RNN梯度裁剪,防止訓練發散,避免負面影響。
訓練模型,get_sataset下載手寫體圖像,預處理,小寫字母獨熱編碼向量。隨機打亂資料順序,分偏劃分訓練集、測試集。
單字相鄰字母存在依賴關係(或互資訊),RNN保存相同單字全部輸入資訊到隱含活性值。前幾個字母分類,網路無大量輸入推斷額外信息,雙向RNN(bidirectional RNN)克服缺陷。
兩個RNN觀測輸入序列,一個按照通常順序從左端讀取單詞,另一個按照相反順序從右端讀取單詞。每個時間步驟得到兩個輸出活性值。送上共用softmax層前,拼接。分類器從每個字母獲取完整單字資訊。 tf.modle.rnn.bidirectional_rnn已實作。
實現雙向RNN。劃分預測屬性到兩個函數,只關注較少內容。 _shared_softmax函數,傳入函數張量data推斷輸入尺寸。重複使用其他架構函數,相同扁平化技巧在所有時間步驟共用同一個softmax層。 rnn.dynamic_rnn建立兩個RNN。
序列反轉,比實現新反向傳遞RNN運算容易。 tf.reverse_sequence函數反轉幀資料中sequence_lengths幀。資料流程圖節點有名稱。 scope參數是rnn_dynamic_cell變數scope名稱,預設值RNN。兩個參數不同RNN,需要不同域。
反轉序列送入後向RNN,網路輸出反轉,和前向輸出對齊。沿RNN神經元輸出維度拼接兩個張量,並返回。雙向RNN模型表現更優。
import gzipimport csvimport numpy as npfrom helpers import downloadclass OcrDataset: URL = 'http://ai.stanford.edu/~btaskar/ocr/letter.data.gz'def __init__(self, cache_dir): path = download(type(self).URL, cache_dir) lines = self._read(path) data, target = self._parse(lines) self.data, self.target = self._pad(data, target) @staticmethoddef _read(filepath): with gzip.open(filepath, 'rt') as file_: reader = csv.reader(file_, delimiter='\t') lines = list(reader)return lines @staticmethoddef _parse(lines): lines = sorted(lines, key=lambda x: int(x[0])) data, target = [], [] next_ = Nonefor line in lines:if not next_: data.append([]) target.append([])else:assert next_ == int(line[0]) next_ = int(line[2]) if int(line[2]) > -1 else None pixels = np.array([int(x) for x in line[6:134]]) pixels = pixels.reshape((16, 8)) data[-1].append(pixels) target[-1].append(line[1])return data, target @staticmethoddef _pad(data, target): max_length = max(len(x) for x in target) padding = np.zeros((16, 8)) data = [x + ([padding] * (max_length - len(x))) for x in data] target = [x + ([''] * (max_length - len(x))) for x in target]return np.array(data), np.array(target)import tensorflow as tffrom helpers import lazy_propertyclass SequenceLabellingModel:def __init__(self, data, target, params): self.data = data self.target = target self.params = params self.prediction self.cost self.error self.optimize @lazy_propertydef length(self): used = tf.sign(tf.reduce_max(tf.abs(self.data), reduction_indices=2)) length = tf.reduce_sum(used, reduction_indices=1) length = tf.cast(length, tf.int32)return length @lazy_propertydef prediction(self): output, _ = tf.nn.dynamic_rnn( tf.nn.rnn_cell.GRUCell(self.params.rnn_hidden), self.data, dtype=tf.float32, sequence_length=self.length, )# Softmax layer.max_length = int(self.target.get_shape()[1]) num_classes = int(self.target.get_shape()[2]) weight = tf.Variable(tf.truncated_normal( [self.params.rnn_hidden, num_classes], stddev=0.01)) bias = tf.Variable(tf.constant(0.1, shape=[num_classes]))# Flatten to apply same weights to all time steps.output = tf.reshape(output, [-1, self.params.rnn_hidden]) prediction = tf.nn.softmax(tf.matmul(output, weight) + bias) prediction = tf.reshape(prediction, [-1, max_length, num_classes])return prediction @lazy_propertydef cost(self):# Compute cross entropy for each frame.cross_entropy = self.target * tf.log(self.prediction) cross_entropy = -tf.reduce_sum(cross_entropy, reduction_indices=2) mask = tf.sign(tf.reduce_max(tf.abs(self.target), reduction_indices=2)) cross_entropy *= mask# Average over actual sequence lengths.cross_entropy = tf.reduce_sum(cross_entropy, reduction_indices=1) cross_entropy /= tf.cast(self.length, tf.float32)return tf.reduce_mean(cross_entropy) @lazy_propertydef error(self): mistakes = tf.not_equal( tf.argmax(self.target, 2), tf.argmax(self.prediction, 2)) mistakes = tf.cast(mistakes, tf.float32) mask = tf.sign(tf.reduce_max(tf.abs(self.target), reduction_indices=2)) mistakes *= mask# Average over actual sequence lengths.mistakes = tf.reduce_sum(mistakes, reduction_indices=1) mistakes /= tf.cast(self.length, tf.float32)return tf.reduce_mean(mistakes) @lazy_propertydef optimize(self): gradient = self.params.optimizer.compute_gradients(self.cost)try: limit = self.params.gradient_clipping gradient = [ (tf.clip_by_value(g, -limit, limit), v)if g is not None else (None, v)for g, v in gradient]except AttributeError:print('No gradient clipping parameter specified.') optimize = self.params.optimizer.apply_gradients(gradient)return optimizeimport randomimport tensorflow as tfimport numpy as npfrom helpers import AttrDictfrom OcrDataset import OcrDatasetfrom SequenceLabellingModel import SequenceLabellingModelfrom batched import batched params = AttrDict( rnn_cell=tf.nn.rnn_cell.GRUCell, rnn_hidden=300, optimizer=tf.train.RMSPropOptimizer(0.002), gradient_clipping=5, batch_size=10, epochs=5, epoch_size=50)def get_dataset(): dataset = OcrDataset('./ocr')# Flatten images into vectors.dataset.data = dataset.data.reshape(dataset.data.shape[:2] + (-1,))# One-hot encode targets.target = np.zeros(dataset.target.shape + (26,))for index, letter in np.ndenumerate(dataset.target):if letter: target[index][ord(letter) - ord('a')] = 1dataset.target = target# Shuffle order of examples.order = np.random.permutation(len(dataset.data)) dataset.data = dataset.data[order] dataset.target = dataset.target[order]return dataset# Split into training and test data.dataset = get_dataset() split = int(0.66 * len(dataset.data)) train_data, test_data = dataset.data[:split], dataset.data[split:] train_target, test_target = dataset.target[:split], dataset.target[split:]# Compute graph._, length, image_size = train_data.shape num_classes = train_target.shape[2] data = tf.placeholder(tf.float32, [None, length, image_size]) target = tf.placeholder(tf.float32, [None, length, num_classes]) model = SequenceLabellingModel(data, target, params) batches = batched(train_data, train_target, params.batch_size) sess = tf.Session() sess.run(tf.initialize_all_variables())for index, batch in enumerate(batches): batch_data = batch[0] batch_target = batch[1] epoch = batch[2]if epoch >= params.epochs:breakfeed = {data: batch_data, target: batch_target} error, _ = sess.run([model.error, model.optimize], feed)print('{}: {:3.6f}%'.format(index + 1, 100 * error)) test_feed = {data: test_data, target: test_target} test_error, _ = sess.run([model.error, model.optimize], test_feed)print('Test error: {:3.6f}%'.format(100 * error))import tensorflow as tffrom helpers import lazy_propertyclass BidirectionalSequenceLabellingModel:def __init__(self, data, target, params): self.data = data self.target = target self.params = params self.prediction self.cost self.error self.optimize @lazy_propertydef length(self): used = tf.sign(tf.reduce_max(tf.abs(self.data), reduction_indices=2)) length = tf.reduce_sum(used, reduction_indices=1) length = tf.cast(length, tf.int32)return length @lazy_propertydef prediction(self): output = self._bidirectional_rnn(self.data, self.length) num_classes = int(self.target.get_shape()[2]) prediction = self._shared_softmax(output, num_classes)return predictiondef _bidirectional_rnn(self, data, length): length_64 = tf.cast(length, tf.int64) forward, _ = tf.nn.dynamic_rnn( cell=self.params.rnn_cell(self.params.rnn_hidden), inputs=data, dtype=tf.float32, sequence_length=length, scope='rnn-forward') backward, _ = tf.nn.dynamic_rnn( cell=self.params.rnn_cell(self.params.rnn_hidden), inputs=tf.reverse_sequence(data, length_64, seq_dim=1), dtype=tf.float32, sequence_length=self.length, scope='rnn-backward') backward = tf.reverse_sequence(backward, length_64, seq_dim=1) output = tf.concat(2, [forward, backward])return outputdef _shared_softmax(self, data, out_size): max_length = int(data.get_shape()[1]) in_size = int(data.get_shape()[2]) weight = tf.Variable(tf.truncated_normal( [in_size, out_size], stddev=0.01)) bias = tf.Variable(tf.constant(0.1, shape=[out_size]))# Flatten to apply same weights to all time steps.flat = tf.reshape(data, [-1, in_size]) output = tf.nn.softmax(tf.matmul(flat, weight) + bias) output = tf.reshape(output, [-1, max_length, out_size])return output @lazy_propertydef cost(self):# Compute cross entropy for each frame.cross_entropy = self.target * tf.log(self.prediction) cross_entropy = -tf.reduce_sum(cross_entropy, reduction_indices=2) mask = tf.sign(tf.reduce_max(tf.abs(self.target), reduction_indices=2)) cross_entropy *= mask# Average over actual sequence lengths.cross_entropy = tf.reduce_sum(cross_entropy, reduction_indices=1) cross_entropy /= tf.cast(self.length, tf.float32)return tf.reduce_mean(cross_entropy) @lazy_propertydef error(self): mistakes = tf.not_equal( tf.argmax(self.target, 2), tf.argmax(self.prediction, 2)) mistakes = tf.cast(mistakes, tf.float32) mask = tf.sign(tf.reduce_max(tf.abs(self.target), reduction_indices=2)) mistakes *= mask# Average over actual sequence lengths.mistakes = tf.reduce_sum(mistakes, reduction_indices=1) mistakes /= tf.cast(self.length, tf.float32)return tf.reduce_mean(mistakes) @lazy_propertydef optimize(self): gradient = self.params.optimizer.compute_gradients(self.cost)try: limit = self.params.gradient_clipping gradient = [ (tf.clip_by_value(g, -limit, limit), v)if g is not None else (None, v)for g, v in gradient]except AttributeError:print('No gradient clipping parameter specified.') optimize = self.params.optimizer.apply_gradients(gradient)return optimize
參考資料:
《機器智能的TensorFlow實踐》
歡迎加我微信交流:qingxingfengzi
我的微信公眾號:qingxingfengzigz
我老婆張幸清的微信公眾號:qingqingfeifangz
以上是序列標註、手寫小寫字母OCR資料集、雙向RNN的詳細內容。更多資訊請關注PHP中文網其他相關文章!